log only on rank 0

Co-authored-by: Tristan Stahnke <tristan@melchior.lan>
Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>

.circleci/config.yml

@@ -1,100 +0,0 @@
-version: 2.1
-
-orbs:
-  apple: ml-explore/pr-approval@0.1.0
- 
-jobs:
-  linux_build_and_test:
-    docker:
-      - image: cimg/python:3.9
-
-    steps:
-      - checkout
-      - run:
-          name: Run style checks
-          command: |
-            pip install pre-commit
-            pre-commit run --all
-            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
-
-  mlx_lm_build_and_test:
-    macos:
-      xcode: "15.2.0"
-    resource_class: m2pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            brew install python@3.9
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install sentencepiece
-            pip install unittest-xml-reporting
-            pip install -e ".[test]"
-      - run:
-          name: Run Python tests
-          command: |
-            source env/bin/activate
-            python -m xmlrunner discover -v tests -o test-results/
-      - store_test_results:
-          path: test-results
-
-  build_release:
-    macos:
-      xcode: "15.2.0"
-    resource_class: m2pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            brew install python@3.9
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install build
-            pip install twine
-      - run:
-          name: Build and upload
-          command: |
-            source env/bin/activate
-            python -m build
-            twine upload dist/*
-      - store_artifacts:
-          path: dist/
-
-workflows:
-  build_and_test:
-    when:
-      matches:
-        pattern: "^(?!pull/)[-\\w]+$"
-        value: << pipeline.git.branch >>
-    jobs:
-      - mlx_lm_build_and_test
-      - linux_build_and_test
-
-  build_pypi_release:
-    jobs:
-      - build_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-
-  prb:
-    when:
-      matches:
-        pattern: "^pull/\\d+(/head)?$"
-        value: << pipeline.git.branch >>
-    jobs:
-      - hold:
-          type: approval
-      - apple/authenticate:
-          context: pr-approval
-      - mlx_lm_build_and_test:
-          requires: [ hold ]
-      - linux_build_and_test:
-          requires: [ hold ]

.github/actions/setup-macos/action.yml

@@ -0,0 +1,16 @@
+name: 'Setup macOS Environment'
+description: 'Install dependencies for macOS'
+
+inputs:
+  python-version:
+    description: 'Python version to use'
+    required: false
+    default: '3.10'
+
+runs:
+  using: "composite"
+  steps:
+    - uses: conda-incubator/setup-miniconda@v3
+      with:
+        miniconda-version: "latest"
+        python-version: ${{ inputs.python-version }}

.github/workflows/pull_request.yml

@@ -0,0 +1,44 @@
+name: Build and Test
+
+on:
+  push:
+    branches: ["main"]
+  pull_request:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.ref != 'refs/head/main' }}
+
+jobs:
+  check_lint:
+    if: github.repository == 'ml-explore/mlx-lm'
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.10"
+      - uses: pre-commit/action@v3.0.1
+
+  mac_build_and_test:
+    if: github.repository == 'ml-explore/mlx-lm'
+    runs-on: [self-hosted, macos]
+    needs: check_lint
+    steps:
+      - uses: actions/checkout@v5
+      - uses: ./.github/actions/setup-macos
+      - name: Install test dependencies
+        shell: bash -l {0}
+        run: |
+          pip install unittest-xml-reporting
+          pip install -e ".[test]"
+      - name: Run tests
+        shell: bash -l {0}
+        run: |
+          curl -o test_data.zip -L https://github.com/ml-explore/mlx-lm/releases/download/test_data/test_data.zip
+          unzip test_data.zip
+          METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 HF_HOME="." python -m xmlrunner discover -v tests -o test-results/
+          mlx.launch -n 2 tests/model_parallel_tests.py

.github/workflows/release.yml

@@ -0,0 +1,41 @@
+name: PyPI Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+
+  build_release:
+    if: github.repository == 'ml-explore/mlx-lm'
+    runs-on: ubuntu-22.04
+    permissions:
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/p/mlx-lm
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.10"
+      - name: Build package
+        shell: sh
+        run: |
+          pip install build
+          python -m build
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: mlx-lm
+          path: dist/*
+      - name: Publish package distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/

ACKNOWLEDGMENTS.md

@@ -8,5 +8,26 @@ with a short description of your contribution(s) below. For example:
 MLX LM was developed with contributions from the following individuals:
 
 - Shunta Saito: Added support for PLaMo models.
-- Gökdeniz Gülmez: Added support for the following architectures: OpenBMB's `MiniCPM` and `MiniCPM3`, Kyutai's `Helium`, State-Space's`Mamba v1`, Z.ai & THUKEG's `GLM4`, Rednote `dots.llm1`, Baisu's `Ernie4.5 MoE`, inclusionAI's `Ling-lite, Ling-Coder-lite, Ling-plus`, and Allenai's `OLMoE`; Added support for the following training algorithms: `Full Weight Fine-Tuning`, and the `Muon` optimizer; Added support for the following other features: `Multiple Optimizers to choose for training`, and `reporting training metrics to WandB (Weights & Biases)`.
-- Prince Canuma: Helped add support for the following model architectures: HuggingFace's `Starcoder2`, Cohere's `Cohere (1 and 2)`, Alibaba Qwen's `Qwen (2, 3 and MoE)`, Microsoft's `Phi (3 and 3.5 MoE)`, `BitNet1.58`, Meta's `Llama (3 and 4)`, Google DeepMind's `Gemma 3`, and InterLM's `InternLM 2.5`.
+- Gökdeniz Gülmez: Added support for the following architectures: 
+OpenBMB's `MiniCPM` and `MiniCPM3`, Kyutai's `Helium`, State-Space's `Mamba v1` and 
+`Mamba v2`, Z.ai & THUKEG's `GLM`, `GLM4`, `GLM5 (GLM MoE DSA)`, Rednote `dots.llm1`, Baidu's `Ernie4.5 MoE`, 
+inclusionAI's `Bailing MoE e.g. Ling-family`, `Bailing MoE Linear e.g. Ling-Linear-family`, 
+Klear team - Kuaishou Technology's `Klear`, AI21 Lab's `Jamba` IBM's `Granite MoE`, 
+Meituan's `LongCat`, Nvidia's `Nemotron H`, Swiss-AI's `Apertus`, Nikity's `Lille130m`, 
+Alibaba Qwen's `Qwen3Next`, Tele-AI's `TeleChat3`, and Allenai's `OLMoE` and `Olmo 3`; 
+Helped add support for the following model architectures: 
+Alibaba Qwen's `Qwen3 & Qwen3MoE)`; Added support for the following training algorithms: 
+`Full Weight Fine-Tuning`, and the `Muon` optimizer; 
+Added support for the following other features: 
+`Multiple Optimizers to choose for training`, and `reporting training metrics to WandB (Weights & Biases)`.
+- Prince Canuma: Helped add support for the following model architectures:
+  HuggingFace's `Starcoder2`, Cohere's `Cohere (1 and 2)`, Alibaba Qwen's `Qwen (2, 3 and MoE)`, 
+  Microsoft's `Phi (3 and 3.5 MoE)`, `BitNet1.58`, Meta's `Llama (3 and 4)`, MinimaxAI's `MiniMax`, 
+  MoonshotAI's `Kimi-Linear`, LiquidAI's `LFM2` and `LFM2 MoE`, 
+  Google DeepMind's `Gemma 3`, TII's `Falcon H1` and InterLM's `InternLM 2.5`.
+- Ivan Fioravanti: Added support for the following architectures:
+ ServiceNow-AI's `Apriel 1.5`, Tencent's `Hunyuan Dense V1` and `Hunyuan MoE V1`.
+- Tarjei Mandt: Added support for the following architectures: `Step 3.5 Flash`, 
+MoonshotAI's `Kimi K2.5`, Upstage's `Solar Open`, LG AI Research's `K-Exaone MoE`, 
+Meituan's `LongCat Flash Lite` Helped add support for the following model architectures: 
+Z.ai & THUKEG's `GLM5 (GLM MoE DSA)`

README.md

@@ -12,7 +12,6 @@ Some key features include:
   fine-tuning](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/LORA.md)
   with support for quantized models.
 * Distributed inference and fine-tuning with `mx.distributed`
-* Supports experiment tracking using SwanLab and W&B.
 
 The easiest way to get started is to install the `mlx-lm` package:
 
@@ -53,6 +52,12 @@ options for a command, e.g.:
 mlx_lm.generate -h
 ```
 
+The default model for generation and chat is
+`mlx-community/Llama-3.2-3B-Instruct-4bit`.  You can specify any MLX-compatible
+model with the `--model` flag. Thousands are available in the
+[MLX Community](https://huggingface.co/mlx-community) Hugging Face
+organization.
+
 ### Python API
 
 You can use `mlx-lm` as a module:
@@ -66,7 +71,7 @@ prompt = "Write a story about Einstein"
 
 messages = [{"role": "user", "content": prompt}]
 prompt = tokenizer.apply_chat_template(
-    messages, add_generation_prompt=True
+    messages, add_generation_prompt=True,
 )
 
 text = generate(model, tokenizer, prompt=prompt, verbose=True)
@@ -80,7 +85,9 @@ To see a description of all the arguments you can do:
 
 Check out the [generation
 example](https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/examples/generate_response.py)
-to see how to use the API in more detail.
+to see how to use the API in more detail. Check out the [batch generation
+example](https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/examples/batch_generate_response.py)
+to see how to efficiently generate continuations for a batch of prompts.
 
 The `mlx-lm` package also comes with functionality to quantize and optionally
 upload models to the Hugging Face Hub.
@@ -123,7 +130,7 @@ prompt = "Write a story about Einstein"
 
 messages = [{"role": "user", "content": prompt}]
 prompt = tokenizer.apply_chat_template(
-    messages, add_generation_prompt=True
+    messages, add_generation_prompt=True,
 )
 
 for response in stream_generate(model, tokenizer, prompt, max_tokens=512):
@@ -163,7 +170,7 @@ mlx_lm.generate --help
 To quantize a model from the command line run:
 
 ```
-mlx_lm.convert --hf-path mistralai/Mistral-7B-Instruct-v0.3 -q
+mlx_lm.convert --model mistralai/Mistral-7B-Instruct-v0.3 -q
 ```
 
 For more options run:
@@ -178,7 +185,7 @@ You can upload new models to Hugging Face by specifying `--upload-repo` to
 
 ```
 mlx_lm.convert \
-    --hf-path mistralai/Mistral-7B-Instruct-v0.3 \
+    --model mistralai/Mistral-7B-Instruct-v0.3 \
     -q \
     --upload-repo mlx-community/my-4bit-mistral
 ```
@@ -229,45 +236,19 @@ for more usage details.
 
 ### Supported Models
 
-`mlx-lm` supports thousands of Hugging Face format LLMs. If the model you want to
-run is not supported, file an
-[issue](https://github.com/ml-explore/mlx-lm/issues/new) or better yet,
-submit a pull request.
+`mlx-lm` supports thousands of LLMs available on the Hugging Face Hub. If the
+model you want to run is not supported, file an
+[issue](https://github.com/ml-explore/mlx-lm/issues/new) or better yet, submit
+a pull request. Many supported models are available in various quantization
+formats in the [MLX Community](https://huggingface.co/mlx-community) Hugging
+Face organization.
 
-Here are a few examples of Hugging Face models that work with this example:
+For some models the tokenizer may require you to enable the `trust_remote_code`
+option. You can do this by passing `--trust-remote-code` in the command line.
+If you don't specify the flag explicitly, you will be prompted to trust remote
+code in the terminal when running the model. 
 
-- [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
-- [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf)
-- [deepseek-ai/deepseek-coder-6.7b-instruct](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct)
-- [01-ai/Yi-6B-Chat](https://huggingface.co/01-ai/Yi-6B-Chat)
-- [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
-- [mistralai/Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1)
-- [Qwen/Qwen-7B](https://huggingface.co/Qwen/Qwen-7B)
-- [pfnet/plamo-13b](https://huggingface.co/pfnet/plamo-13b)
-- [pfnet/plamo-13b-instruct](https://huggingface.co/pfnet/plamo-13b-instruct)
-- [stabilityai/stablelm-2-zephyr-1_6b](https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b)
-- [internlm/internlm2-7b](https://huggingface.co/internlm/internlm2-7b)
-- [tiiuae/falcon-mamba-7b-instruct](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct)
-
-Most
-[Mistral](https://huggingface.co/models?library=transformers,safetensors&other=mistral&sort=trending),
-[Llama](https://huggingface.co/models?library=transformers,safetensors&other=llama&sort=trending),
-[Phi-2](https://huggingface.co/models?library=transformers,safetensors&other=phi&sort=trending),
-and
-[Mixtral](https://huggingface.co/models?library=transformers,safetensors&other=mixtral&sort=trending)
-style models should work out of the box.
-
-For some models (such as `Qwen` and `plamo`) the tokenizer requires you to
-enable the `trust_remote_code` option. You can do this by passing
-`--trust-remote-code` in the command line. If you don't specify the flag
-explicitly, you will be prompted to trust remote code in the terminal when
-running the model. 
-
-For `Qwen` models you must also specify the `eos_token`. You can do this by
-passing `--eos-token "<|endoftext|>"` in the command
-line. 
-
-These options can also be set in the Python API. For example:
+Tokenizer options can also be set in the Python API. For example:
 
 ```python
 model, tokenizer = load(

benchmarks/server_benchmark.py

@@ -0,0 +1,348 @@
+"""
+Spin up the local server:
+
+    mlx_lm.server
+
+Then run the benchmark:
+
+    python server_benchmark.py --concurrency 4
+"""
+
+import argparse
+import asyncio
+import json
+import math
+import time
+from collections import defaultdict
+from itertools import cycle
+from typing import Any, Dict, List, Optional, Tuple
+
+import aiohttp
+from tqdm import tqdm
+
+# Default prompts if no file is provided
+DEFAULT_PROMPTS = [
+    "Explain quantum computing in simple terms.",
+    "What are the main differences between Python and JavaScript?",
+    "Describe the process of photosynthesis in plants.",
+    "How does a neural network learn from data?",
+    "What is the significance of the Turing test in AI?",
+    "Explain the concept of blockchain technology.",
+    "What causes seasons on Earth?",
+    "How do vaccines work in the human body?",
+    "Describe the water cycle and its importance.",
+    "What is the theory of relativity proposed by Einstein?",
+    "How do electric cars help reduce carbon emissions?",
+    "What are the key features of a market economy?",
+    "Explain how DNA replication works in cells.",
+    "What is machine learning and its real-world applications?",
+    "Describe the structure and function of the human heart.",
+]
+
+
+def tokens_per_second(tokens):
+    start = math.floor(tokens[0])
+    stop = math.ceil(tokens[-1])
+    n_bins = int(stop - start) * 10
+    bins = [0] * n_bins
+    for t in tokens:
+        bins[int(n_bins * (t - start) / (stop - start))] += 1
+
+    result = []
+
+    ms = 0
+    cnt = 0
+    for i, b in enumerate(bins):
+        ms += b
+        if cnt == 10:
+            ms -= bins[i - 10]
+        else:
+            cnt += 1
+
+        result.append(10 * ms / cnt)
+
+    times = [start]
+    while times[-1] < stop:
+        times.append(times[-1] + 0.1)
+
+    return times, result
+
+
+def plot_generation(times, tokens_per_sec, start=None, interval=1.0, width=50):
+    c = "█"
+    start = start or times[0]
+    stop = times[-1]
+
+    bar_times = [start]
+    while bar_times[-1] < stop:
+        bar_times.append(bar_times[-1] + interval)
+
+    bar_values = [[] for _ in bar_times]
+    bar_idx = 0
+
+    for t, v in zip(times, tokens_per_sec):
+        while t > bar_times[bar_idx] + interval:
+            bar_idx += 1
+        bar_values[bar_idx].append(v)
+
+    bar_values = [sum(v) / len(v) if v else 0 for v in bar_values]
+    m = max(bar_values)
+
+    for t, v in zip(bar_times, bar_values):
+        t = t - start
+        b = c * int(v * width / m)
+        print(f"{t:3.2f} {b} ({v})")
+
+
+def percentile(data, percent):
+    if not data:
+        return 0
+    data = sorted(data)
+    k = (len(data) - 1) * percent / 100
+    f = math.floor(k)
+    c = math.ceil(k)
+    return (
+        data[int(f)]
+        if f == c
+        else data[int(f)] + (data[int(c)] - data[int(f)]) * (k - f)
+    )
+
+
+def median(data):
+    return percentile(data, 50)
+
+
+async def make_request(
+    session: aiohttp.ClientSession,
+    url: str,
+    api_key: str,
+    model: str,
+    prompt: str,
+    max_tokens: int,
+) -> Tuple[bool, float, list]:
+    """
+    Make a single streaming API request and return
+
+        - whether the request succeeded
+        - the request start time
+        - the time of every generated token
+    """
+    payload = {
+        "model": model,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+        "stream": True,
+    }
+    headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
+
+    start_time = time.perf_counter()
+    tokens = []
+
+    try:
+        async with session.post(url, json=payload, headers=headers) as response:
+            if response.status != 200:
+                error_body = await response.text()
+                print(f"Error {response.status}: {error_body}")
+                return (False, 0, [])
+
+            # Process streaming response
+            async for chunk in response.content:
+                if chunk:
+                    chunk_str = chunk.decode("utf-8").strip()
+                    if chunk_str.startswith("data:"):
+                        data_str = chunk_str[5:].strip()
+                        if data_str == "[DONE]":
+                            break
+
+                        try:
+                            data = json.loads(data_str)
+                            if choices := data.get("choices", False):
+                                if choices[0].get("finish_reason") != "length":
+                                    tokens.append(time.perf_counter())
+                        except json.JSONDecodeError:
+                            continue
+
+            return (bool(tokens), start_time, tokens)
+
+    except Exception as e:
+        print(f"Request failed: {str(e)}")
+        return (False, 0, [])
+
+
+async def run_benchmark(
+    url: str,
+    api_key: str,
+    model: str,
+    max_tokens: int,
+    concurrency: int,
+    total_requests: int,
+    prompts: List[str],
+) -> Dict[str, Any]:
+    prompt_cycle = cycle(prompts)
+    semaphore = asyncio.Semaphore(concurrency)
+    results = []
+    request_times = []
+    bar = tqdm(total=total_requests)
+
+    async def worker():
+        async with semaphore:
+            prompt = next(prompt_cycle)
+            result = await make_request(
+                session, url, api_key, model, prompt, max_tokens
+            )
+            bar.update(1)
+            return result
+
+    async with aiohttp.ClientSession() as session:
+        tasks = []
+        for _ in range(total_requests):
+            task = asyncio.create_task(worker())
+            tasks.append(task)
+            await asyncio.sleep(0.01)  # Stagger requests slightly
+
+        for task in tasks:
+            result = await task
+            results.append(result)
+        bar.close()
+
+    successful_requests = [r for r in results if r[0]]
+    total_tokens = sum(len(r[2]) for r in successful_requests)
+
+    # Gather all the tokens generated with their corresponding timestamps
+    all_tokens = []
+    for r in successful_requests:
+        all_tokens.extend(r[2])
+    all_tokens.sort()
+    full_generation = tokens_per_second(all_tokens)
+    start = min(r[1] for r in successful_requests)
+
+    # Aggregate metrics
+    metrics = {
+        "total_requests": total_requests,
+        "successful_requests": len(successful_requests),
+        "failed_requests": total_requests - len(successful_requests),
+        "total_tokens": total_tokens,
+        "total_time": all_tokens[-1] - start,
+        "aggregate_tokens_per_sec": median(full_generation[1]),
+        "per_request": [],
+        "start": start,
+        "full_generation": full_generation,
+    }
+
+    # Per-request metrics
+    for i, (_, start, tokens) in enumerate(successful_requests):
+        metrics["per_request"].append(
+            {
+                "request_id": i + 1,
+                "time_to_first_token": tokens[0] - start,
+                "total_time": tokens[-1] - start,
+                "tokens_received": len(tokens),
+                "tokens_per_sec": median(tokens_per_second(tokens)[1]),
+            }
+        )
+
+    # Calculate percentiles
+    ttft_values = [m["time_to_first_token"] for m in metrics["per_request"]]
+    tps_values = [m["tokens_per_sec"] for m in metrics["per_request"]]
+
+    metrics["aggregate_metrics"] = {
+        "time_to_first_token": {
+            "min": min(ttft_values) if ttft_values else 0,
+            "max": max(ttft_values) if ttft_values else 0,
+            "avg": sum(ttft_values) / len(ttft_values) if ttft_values else 0,
+            "p95": percentile(ttft_values, 95) if ttft_values else 0,
+        },
+        "tokens_per_sec": {
+            "min": min(tps_values) if tps_values else 0,
+            "max": max(tps_values) if tps_values else 0,
+            "avg": sum(tps_values) / len(tps_values) if tps_values else 0,
+            "p95": percentile(tps_values, 95) if tps_values else 0,
+        },
+    }
+
+    return metrics
+
+
+def main():
+    parser = argparse.ArgumentParser(description="LLM API Benchmark Tool")
+    parser.add_argument(
+        "--url",
+        default="http://localhost:8080/v1/chat/completions",
+        help="Chat completions API endpoint URL",
+    )
+    parser.add_argument("--api-key", default="none", help="API key")
+    parser.add_argument("--model", default="default_model", help="Model name")
+    parser.add_argument(
+        "--max-tokens", type=int, default=100, help="Max tokens to generate"
+    )
+    parser.add_argument(
+        "--concurrency", type=int, default=1, help="Number of concurrent requests"
+    )
+    parser.add_argument(
+        "--total-requests", type=int, default=10, help="Total requests to make"
+    )
+    parser.add_argument("--prompt-file", help="File containing prompts (one per line)")
+    parser.add_argument("--output", help="Output file for results (JSON format)")
+
+    args = parser.parse_args()
+
+    # Load prompts
+    if args.prompt_file:
+        with open(args.prompt_file, "r") as f:
+            prompts = [line.strip() for line in f if line.strip()]
+    else:
+        prompts = DEFAULT_PROMPTS
+
+    print(
+        f"Starting benchmark with {args.concurrency} concurrency and {args.total_requests} total requests..."
+    )
+    start_time = time.perf_counter()
+
+    # Run benchmark
+    results = asyncio.run(
+        run_benchmark(
+            url=args.url,
+            api_key=args.api_key,
+            model=args.model,
+            max_tokens=args.max_tokens,
+            concurrency=args.concurrency,
+            total_requests=args.total_requests,
+            prompts=prompts,
+        )
+    )
+
+    duration = time.perf_counter() - start_time
+    print(f"\nBenchmark completed in {duration:.2f} seconds")
+    print(
+        f"Successful requests: {results['successful_requests']}/{args.total_requests}"
+    )
+    print(f"Total tokens generated: {results['total_tokens']}")
+    print(f"Aggregate tokens/sec: {results['aggregate_tokens_per_sec']:.2f}")
+
+    # Print summary
+    if results["successful_requests"] > 0:
+        ttft = results["aggregate_metrics"]["time_to_first_token"]
+        tps = results["aggregate_metrics"]["tokens_per_sec"]
+
+        print("\nTime to First Token (seconds):")
+        print(
+            f"  Min: {ttft['min']:.4f} | Max: {ttft['max']:.4f} | Avg: {ttft['avg']:.4f} | P95: {ttft['p95']:.4f}"
+        )
+
+        print("\nTokens per Second (per request):")
+        print(
+            f"  Min: {tps['min']:.2f} | Max: {tps['max']:.2f} | Avg: {tps['avg']:.2f} | P95: {tps['p95']:.2f}"
+        )
+
+        print()
+        plot_generation(*results["full_generation"], results["start"])
+
+    # Save results
+    if args.output:
+        with open(args.output, "w") as f:
+            json.dump(results, f, indent=2)
+        print(f"\nResults saved to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

mlx_lm/BENCHMARKS.md

@@ -0,0 +1,63 @@
+# Benchmarks
+
+## Commands 
+
+The command for evaluating on MMLU Pro:
+
+```
+mlx_lm.evaluate --model model/repo --task mmlu_pro
+```
+ 
+The command for efficiency benchmarks:
+
+```
+mlx_lm.benchmark --model model/repo -p 2048 -g 128
+```
+
+To get the package versions run:
+
+```
+python -m mlx --version && python -m mlx_lm --version
+```
+
+## Models
+
+<details>
+
+ <summary> Qwen/Qwen3-4B-Instruct-2507 </summary>
+
+Precision | MMLU Pro | Prompt (2048) tok/sec | Generation (128) tok/sec | Memory GB | Repo
+--------- | -------- | ------------------- | ------------------------ | --------- | ----
+bf16      | 64.05    | 1780.63             | 52.47                    | 9.02    | Qwen/Qwen3-4B-Instruct-2507
+q8 | 63.85 | 1606.573| 86.907 | 5.254 | mlx-community/Qwen3-4B-Instruct-2507-8bit
+q6 | 63.53 | 1576.73 | 104.68 | 4.25 | mlx-community/Qwen3-4B-Instruct-2507-6bit
+q5 g32 | 63.16 | 1570.80 | 110.29 | 4.00 | mlx-community/Qwen3-4B-Instruct-2507-5bit-g32
+q5 | 62.38 | 1584.33 | 116.39 | 3.86 | mlx-community/Qwen3-4B-Instruct-2507-5bit
+q4 g32 | 61.46 | 1610.03 | 126.00 | 3.603 | mlx-community/Qwen3-4B-Instruct-2507-4bit-g32
+q4 | 60.72 | 1622.27 | 134.52 | 3.35 | mlx-community/Qwen3-4B-Instruct-2507-4bit
+
+- Performance benchmark on 64GB M4 Max
+- mlx 0.29.2.dev20251008+85a8824a8
+- mlx-lm 0.28.2
+- macOS 26.1
+ 
+</details>
+
+<details>
+<summary> Qwen/Qwen3-30B-A3B-Instruct-2507 </summary>
+
+Precision | MMLU Pro | Prompt (2048) tok/sec | Generation (128) tok/sec | Memory GB | Repo
+--------- | -------- | ------------------- | ------------------------ | --------- | ----
+bf16 | 72.62 | :skull: | :skull: | :skull: | Qwen/Qwen3-30B-A3B-Instruct-2507
+q8 | 72.46 | 1719.47 | 83.16 | 33.46 | mlx-community/Qwen3-30B-A3B-Instruct-2507-8bit 
+q6 | 72.41 | 1667.45 | 94.14 | 25.82 | mlx-community/Qwen3-30B-A3B-Instruct-2507-6bit
+q5 | 71.97 | 1664.24 | 101.00 |22.01 | mlx-community/Qwen3-30B-A3B-Instruct-2507-5bit
+q4 | 70.71 | 1753.90 | 113.33 |18.20 | mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit
+
+ 
+- Performance benchmarks on 64GB M4 Max
+- mlx 0.29.2.dev20251008+85a8824a8
+- mlx-lm 0.28.2
+- macOS 26.1
+
+</details>

mlx_lm/LEARNED_QUANTS.md

@@ -129,7 +129,7 @@ mlx_lm.awq --help
 Use `mlx_lm.gptq` to run GPTQ on a given model. For example:
 
 ```bash
-mlx_lm.awq --model Qwen/Qwen3-0.6B
+mlx_lm.gptq --model Qwen/Qwen3-0.6B
 ```
 
 The script can take anywhere from a few minutes to several hours depending on

mlx_lm/LORA.md

@@ -66,9 +66,10 @@ mlx_lm.lora \
 To fine-tune the full model weights, add the `--fine-tune-type full` flag.
 Currently supported fine-tuning types are `lora` (default), `dora`, and `full`.
 
-The `--data` argument must specify a path to a `train.jsonl`, `valid.jsonl`
-when using `--train` and a path to a `test.jsonl` when using `--test`. For more
-details on the data format see the section on [Data](#Data).
+The `--data` argument must specify a path to a `train.jsonl` when using
+`--train` and a path to a `test.jsonl` when using `--test`. A `valid.jsonl` is
+optional; if provided, validation loss will be reported during training. For
+more details on the data format see the section on [Data](#Data).
 
 For example, to fine-tune a Mistral 7B you can use `--model
 mistralai/Mistral-7B-v0.1`.
@@ -184,9 +185,10 @@ Face.
 
 ### Local Datasets
 
-For fine-tuning (`--train`), the data loader expects a `train.jsonl` and a
-`valid.jsonl` to be in the data directory. For evaluation (`--test`), the data
-loader expects a `test.jsonl` in the data directory. 
+For fine-tuning (`--train`), the data loader expects a `train.jsonl` to be in
+the data directory. A `valid.jsonl` is optional; if present, validation loss
+will be reported periodically during training. For evaluation (`--test`), the
+data loader expects a `test.jsonl` in the data directory. 
 
 Currently, `*.jsonl` files support `chat`, `tools`, `completions`, and `text`
 data formats. Here are examples of these formats:
@@ -371,7 +373,10 @@ of memory. Here are some tips to reduce memory use should you need to do so:
 
 2. Try using a smaller batch size with `--batch-size`. The default is `4` so
    setting this to `2` or `1` will reduce memory consumption. This may slow
-   things down a little, but will also reduce the memory use.
+   things down a little, but will also reduce the memory use. You can increase
+   the effective batch size without increasing the memory use by accumulating
+   gradients using `--grad-accumulation-steps <N>` which will accumulate the
+   gradient of `<N>` batches before updating the parameters.
 
 3. Reduce the number of layers to fine-tune with `--num-layers`. The default
    is `16`, so you can try `8` or `4`. This reduces the amount of memory

mlx_lm/SERVER.md

@@ -72,12 +72,24 @@ curl localhost:8080/v1/chat/completions \
 - `min_p`: (Optional) A float specifying the min-p sampling parameter.
   Defaults to `0.0` (disabled).
 
-- `repetition_penalty`: (Optional) Applies a penalty to repeated tokens.
-  Defaults to `1.0`.
+- `repetition_penalty`: (Optional) Applies a multiplicative penalty to repeated
+  tokens. Defaults to `0.0` (disabled).
 
 - `repetition_context_size`: (Optional) The size of the context window for
   applying repetition penalty. Defaults to `20`.
 
+- `presence_penalty`: (Optional) Applies an additive penalty to tokens
+  that appeared before. Defaults to `0.0` (disabled).
+
+- `presence_context_size`: (Optional) The size of the context window for
+  applying presence penalty. Defaults to `20`.
+
+- `frequency_penalty`: (Optional) Applies an additive penalty proportional to
+  how many times a token appeared previously. Defaults to `0.0` (disabled).
+
+- `frequency_context_size`: (Optional) The size of the context window for
+  applying frequency penalty. Defaults to `20`.
+
 - `logit_bias`: (Optional) A dictionary mapping token IDs to their bias
   values. Defaults to `None`.
 

mlx_lm/__init__.py

@@ -7,5 +7,14 @@ from ._version import __version__
 os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
 
 from .convert import convert
-from .generate import generate, stream_generate
+from .generate import batch_generate, generate, stream_generate
 from .utils import load
+
+__all__ = [
+    "__version__",
+    "convert",
+    "batch_generate",
+    "generate",
+    "stream_generate",
+    "load",
+]

mlx_lm/__main__.py

@@ -1,31 +1,6 @@
 # Copyright © 2025 Apple Inc.
 
-import importlib
-import sys
-
 if __name__ == "__main__":
-    subcommands = {
-        "quant.awq",
-        "quant.dwq",
-        "quant.dynamic_quant",
-        "quant.gptq",
-        "benchmark",
-        "cache_prompt",
-        "chat",
-        "convert",
-        "evaluate",
-        "fuse",
-        "generate",
-        "lora",
-        "perplexity",
-        "server",
-        "manage",
-        "upload",
-    }
-    if len(sys.argv) < 2:
-        raise ValueError(f"CLI requires a subcommand in {subcommands}")
-    subcommand = sys.argv.pop(1)
-    if subcommand not in subcommands:
-        raise ValueError(f"CLI requires a subcommand in {subcommands}")
-    submodule = importlib.import_module(f"mlx_lm.{subcommand}")
-    submodule.main()
+    from . import cli
+
+    cli.main()

mlx_lm/_version.py

@@ -1,3 +1,3 @@
 # Copyright © 2023-2025 Apple Inc.
 
-__version__ = "0.27.0"
+__version__ = "0.31.3"

mlx_lm/benchmark.py

@@ -1,16 +1,13 @@
 # Copyright © 2025 Apple Inc.
 
 import argparse
+import time
 
 import mlx.core as mx
 
-from mlx_lm import stream_generate
+from mlx_lm import batch_generate, load, stream_generate
 from mlx_lm.generate import DEFAULT_MODEL
-from mlx_lm.tokenizer_utils import load_tokenizer
-from mlx_lm.utils import (
-    fetch_from_hub,
-    get_model_path,
-)
+from mlx_lm.utils import pipeline_load, sharded_load
 
 
 def setup_arg_parser():
@@ -25,11 +22,6 @@ def setup_arg_parser():
         ),
         default=None,
     )
-    parser.add_argument(
-        "--trust-remote-code",
-        action="store_true",
-        help="Enable trusting remote code for tokenizer",
-    )
     parser.add_argument(
         "--prompt-tokens",
         "-p",
@@ -44,6 +36,13 @@ def setup_arg_parser():
         help="Length of completion",
         type=int,
     )
+    parser.add_argument(
+        "--batch-size",
+        "-b",
+        default=1,
+        help="Batch size",
+        type=int,
+    )
     parser.add_argument(
         "--num-trials",
         "-n",
@@ -51,6 +50,29 @@ def setup_arg_parser():
         help="Number of timing trials",
         type=int,
     )
+    parser.add_argument(
+        "--pipeline",
+        action="store_true",
+        help="Use pipelining instead of tensor parallelism",
+    )
+    parser.add_argument(
+        "--quantize-activations",
+        "-qa",
+        action="store_true",
+        help="Quantize activations using the same quantization config as the corresponding layer.",
+    )
+    parser.add_argument(
+        "--prefill-step-size",
+        type=int,
+        default=2048,
+        help="Step size for prefill processing (default: 2048)",
+    )
+    parser.add_argument(
+        "--delay",
+        type=int,
+        default=0,
+        help="Delay between each test in seconds (default: 0)",
+    )
     return parser
 
 
@@ -59,39 +81,81 @@ def main():
     args = parser.parse_args()
     mx.random.seed(0)
 
+    group = mx.distributed.init()
+    rank = group.rank()
+    pipeline_group = group if args.pipeline else None
+    tensor_group = group if not args.pipeline else None
+
+    def rprint(*args, **kwargs):
+        if rank == 0:
+            print(*args, **kwargs)
+
     model_path = args.model or DEFAULT_MODEL
 
-    model_path, _ = get_model_path(model_path, revision=None)
-    model, config, _ = fetch_from_hub(model_path, trust_remote_code=True)
-    tokenizer = load_tokenizer(
-        model_path,
-        eos_token_ids=[],  # Empty to avoid early stopping
-        tokenizer_config_extra={"trust_remote_code": True},
-    )
+    if group.size() > 1:
+        model, tokenizer, config = sharded_load(
+            model_path, pipeline_group, tensor_group, return_config=True
+        )
+    else:
+        model, tokenizer, config = load(
+            model_path,
+            return_config=True,
+            tokenizer_config={"trust_remote_code": True},
+            model_config={"quantize_activations": args.quantize_activations},
+        )
+
+    # Empty to avoid early stopping
+    tokenizer._eos_token_ids = {}
 
     prompt_tokens = args.prompt_tokens
     generation_tokens = args.generation_tokens
-    prompt = mx.random.randint(0, config["vocab_size"], (prompt_tokens,))
+    batch_size = args.batch_size
+    vocab_size = config.get("vocab_size") or config["text_config"]["vocab_size"]
+    prompts = mx.random.randint(0, vocab_size, (batch_size, prompt_tokens)).tolist()
+    prompt = prompts[0]
 
-    def _bench():
+    def single_bench():
         for response in stream_generate(
-            model, tokenizer, prompt, max_tokens=generation_tokens
+            model,
+            tokenizer,
+            prompt,
+            max_tokens=generation_tokens,
+            prefill_step_size=args.prefill_step_size,
         ):
             pass
         return response
 
-    print("Running warmup..")
+    def batch_bench():
+        return batch_generate(
+            model,
+            tokenizer,
+            prompts,
+            max_tokens=generation_tokens,
+            prefill_step_size=args.prefill_step_size,
+        ).stats
+
+    if batch_size == 1:
+        _bench = single_bench
+    else:
+        _bench = batch_bench
+
+    rprint("Running warmup..")
     _bench()
 
     report_keys = ["prompt_tps", "generation_tps", "peak_memory"]
-    print(f"Timing with {prompt_tokens=} and {generation_tokens=}.")
+    rprint(f"Timing with {prompt_tokens=}, {generation_tokens=}, {batch_size=}.")
     responses = []
     for i in range(args.num_trials):
+        if args.delay > 0:
+            time.sleep(args.delay)
+        tic = time.perf_counter()
         response = _bench()
+        toc = time.perf_counter()
         responses.append(response)
         results = [(k, getattr(response, k)) for k in report_keys]
         results = [f"{k}={v:.3f}" for k, v in results]
-        print(f"Trial {i+1}:  " + ", ".join(results))
+        results.append(f"total_time={toc - tic:.3f}")
+        rprint(f"Trial {i+1}:  " + ", ".join(results))
 
     def avg(k):
         vals = (getattr(response, k) for response in responses)
@@ -99,7 +163,7 @@ def main():
 
     results = [(k, avg(k)) for k in report_keys]
     results = [f"{k}={v:.3f}" for k, v in results]
-    print(f"Averages: " + ", ".join(results))
+    rprint(f"Averages: " + ", ".join(results))
 
 
 if __name__ == "__main__":

mlx_lm/cache_prompt.py

@@ -41,16 +41,6 @@ def setup_arg_parser():
         default=None,
         help="End of sequence token for tokenizer",
     )
-    parser.add_argument(
-        "--ignore-chat-template",
-        action="store_true",
-        help="Use the raw prompt without the tokenizer's chat template.",
-    )
-    parser.add_argument(
-        "--use-default-chat-template",
-        action="store_true",
-        help="Use the default chat template",
-    )
     parser.add_argument(
         "--max-kv-size",
         type=int,
@@ -107,14 +97,12 @@ def main():
 
     args.prompt = sys.stdin.read() if args.prompt == "-" else args.prompt
 
-    if args.use_default_chat_template:
-        if tokenizer.chat_template is None:
-            tokenizer.chat_template = tokenizer.default_chat_template
-
-    if not args.ignore_chat_template and tokenizer.chat_template is not None:
+    if tokenizer.has_chat_template:
         messages = [{"role": "user", "content": args.prompt}]
         prompt = tokenizer.apply_chat_template(
-            messages, add_generation_prompt=False, continue_final_message=True
+            messages,
+            add_generation_prompt=False,
+            continue_final_message=True,
         )
 
     else:
@@ -153,7 +141,6 @@ def main():
     print("Saving...")
     metadata = {}
     metadata["model"] = args.model
-    metadata["chat_template"] = json.dumps(tokenizer.chat_template)
     metadata["tokenizer_config"] = json.dumps(tokenizer_config)
     save_prompt_cache(args.prompt_cache_file, cache, metadata)
 

mlx_lm/chat.py

@@ -1,23 +1,43 @@
 # Copyright © 2023-2024 Apple Inc.
 
 import argparse
+import sys
 
 import mlx.core as mx
 
+from .cli_ui import (
+    make_console,
+    make_corridor_prompt,
+    print_chat_help,
+    print_header_panel,
+)
 from .generate import stream_generate
 from .models.cache import make_prompt_cache
 from .sample_utils import make_sampler
-from .utils import load
+from .utils import load, sharded_load
 
 DEFAULT_TEMP = 0.0
 DEFAULT_TOP_P = 1.0
 DEFAULT_XTC_PROBABILITY = 0.0
 DEFAULT_XTC_THRESHOLD = 0.0
-DEFAULT_SEED = None
+DEFAULT_SEED = 0
 DEFAULT_MAX_TOKENS = 256
 DEFAULT_MODEL = "mlx-community/Llama-3.2-3B-Instruct-4bit"
 
 
+def _print_chat_header(args, console):
+    rows = [("model", str(args.model))]
+    if args.adapter_path:
+        rows.append(("adapter", str(args.adapter_path)))
+    rows.append(("max tokens", f"{args.max_tokens:,}"))
+    if args.system_prompt:
+        sp = args.system_prompt
+        if len(sp) > 60:
+            sp = sp[:57] + "..."
+        rows.append(("system", sp))
+    print_header_panel(console, "mlx_lm.chat", rows)
+
+
 def setup_arg_parser():
     """Set up and return the argument parser."""
     parser = argparse.ArgumentParser(description="Chat with an LLM")
@@ -79,6 +99,11 @@ def setup_arg_parser():
         default=None,
         help="System prompt to be used for the chat template",
     )
+    parser.add_argument(
+        "--pipeline",
+        action="store_true",
+        help="Use pipelining instead of tensor parallelism",
+    )
     return parser
 
 
@@ -86,41 +111,74 @@ def main():
     parser = setup_arg_parser()
     args = parser.parse_args()
 
-    if args.seed is not None:
-        mx.random.seed(args.seed)
+    group = mx.distributed.init()
+    rank = group.rank()
+    pipeline_group = group if args.pipeline else None
+    tensor_group = group if not args.pipeline else None
 
-    model, tokenizer = load(
-        args.model,
-        adapter_path=args.adapter_path,
-        tokenizer_config={
-            "trust_remote_code": True if args.trust_remote_code else None
-        },
-    )
+    console = make_console()
 
-    def print_help():
-        print("The command list:")
-        print("- 'q' to exit")
-        print("- 'r' to reset the chat")
-        print("- 'h' to display these commands")
+    def rprint(*args, **kwargs):
+        if rank == 0:
+            print(*args, **kwargs)
+
+    mx.random.seed(args.seed)
+
+    if group.size() > 1:
+        if args.adapter_path:
+            parser.error("Adapters not supported in distributed mode")
+        model, tokenizer = sharded_load(args.model, pipeline_group, tensor_group)
+    else:
+        model, tokenizer = load(
+            args.model,
+            adapter_path=args.adapter_path,
+            tokenizer_config={
+                "trust_remote_code": True if args.trust_remote_code else None
+            },
+        )
+
+    if rank == 0:
+        _print_chat_header(args, console)
+        print_chat_help(console)
+
+    if rank == 0:
+        prompt_console = make_console(force_terminal=True, color_system="truecolor")
+        _draw_corridor_prompt = make_corridor_prompt(prompt_console)
+    else:
+        _draw_corridor_prompt = lambda: ""
 
-    print(f"[INFO] Starting chat session with {args.model}.")
-    print_help()
     prompt_cache = make_prompt_cache(model, args.max_kv_size)
     while True:
-        query = input(">> ")
+        prompt = _draw_corridor_prompt()
+        query = input(prompt)
+        if rank == 0:
+            # Cursor is now on the bottom-rule row; advance past it.
+            sys.stdout.write("\n")
+            sys.stdout.flush()
         if query == "q":
+            if rank == 0:
+                console.print("[ui.muted]bye[/ui.muted]")
             break
         if query == "r":
             prompt_cache = make_prompt_cache(model, args.max_kv_size)
+            if rank == 0:
+                console.print(
+                    "  [ui.good]reset[/ui.good] [ui.muted]context cleared[/ui.muted]"
+                )
             continue
         if query == "h":
-            print_help()
+            if rank == 0:
+                print_chat_help(console)
             continue
         messages = []
         if args.system_prompt is not None:
             messages.append({"role": "system", "content": args.system_prompt})
         messages.append({"role": "user", "content": query})
-        prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+        prompt = tokenizer.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+        )
+        last_response = None
         for response in stream_generate(
             model,
             tokenizer,
@@ -137,8 +195,16 @@ def main():
             ),
             prompt_cache=prompt_cache,
         ):
-            print(response.text, flush=True, end="")
-        print()
+            rprint(response.text, flush=True, end="")
+            last_response = response
+        rprint()
+        if rank == 0 and last_response is not None:
+            console.print(
+                f"  [ui.muted]{last_response.generation_tokens} tokens · "
+                f"{last_response.generation_tps:.1f} tok/s · "
+                f"prompt {last_response.prompt_tps:.1f} tok/s · "
+                f"peak {last_response.peak_memory:.2f} GB[/ui.muted]"
+            )
 
 
 if __name__ == "__main__":

mlx_lm/chat_templates/deepseek_v32.py

@@ -0,0 +1,345 @@
+# Copyright © 2025 Apple Inc.
+
+import copy
+import json
+import re
+from inspect import isfunction
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from transformers.utils.chat_template_utils import get_json_schema
+
+TOOLS_SYSTEM_TEMPLATE = """## Tools
+
+You have access to a set of tools you can use to answer the user's question.
+You can invoke functions by writing a "<{dsml_token}function_calls>" block like the following as part of your reply to the user:
+<{dsml_token}function_calls>
+<{dsml_token}invoke name="$FUNCTION_NAME">
+<{dsml_token}parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</{dsml_token}parameter>
+...
+</{dsml_token}invoke>
+<{dsml_token}invoke name="$FUNCTION_NAME2">
+...
+</{dsml_token}invoke>
+</{dsml_token}function_calls>
+
+String and scalar parameters should be specified as is without any escaping or quotes, while lists and objects should use JSON format. The "string" attribute should be set to "true" for string type parameters and "false" for other types (numbers, booleans, arrays, objects).
+
+If the thinking_mode is enabled, then after function results you should strongly consider outputting a thinking block. Here is an example:
+
+<{dsml_token}function_calls>
+...
+</{dsml_token}function_calls>
+
+<function_results>
+...
+</function_results>
+
+{thinking_start_token}...thinking about results{thinking_end_token}
+
+Here are the functions available in JSONSchema format:
+<functions>
+{tool_schemas}
+</functions>
+"""
+
+bos_token: str = "<｜begin▁of▁sentence｜>"
+eos_token: str = "<｜end▁of▁sentence｜>"
+thinking_start_token: str = "<think>"
+thinking_end_token: str = "</think>"
+dsml_token: str = "｜DSML｜"
+system_msg_template: str = "{content}"
+user_msg_template: str = "<｜User｜>{content}<｜Assistant｜>"
+assistant_msg_template: str = "{reasoning}{content}{tool_calls}<｜end▁of▁sentence｜>"
+thinking_template = "{reasoning_content}"
+
+response_format_template: str = (
+    "## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n{schema}"
+)
+tool_call_template: str = (
+    '<{dsml_token}invoke name="{name}">\n{arguments}\n</{dsml_token}invoke>'
+)
+tool_calls_template = (
+    "<{dsml_token}function_calls>\n{tool_calls}\n</{dsml_token}function_calls>"
+)
+
+tool_output_template: str = "\n<result>{content}</result>"
+
+
+def to_json(value: Any) -> str:
+    try:
+        return json.dumps(value, ensure_ascii=False)
+    except:
+        return json.dumps(value, ensure_ascii=True)
+
+
+def tools_from_openai_format(tools):
+    def normalize_tool(tool):
+        if isfunction(tool):
+            return get_json_schema(tool)
+        return tool["function"]
+
+    return [normalize_tool(tool) for tool in tools]
+
+
+def tool_calls_from_openai_format(tool_calls):
+    return [
+        {
+            "name": tool_call["function"]["name"],
+            "arguments": tool_call["function"]["arguments"],
+        }
+        for tool_call in tool_calls
+    ]
+
+
+def encode_arguments_to_dsml(tool_call: Dict[str, str]) -> str:
+    p_dsml_template = """<{dsml_token}parameter name="{key}" string="{is_str}">{value}</{dsml_token}parameter>"""
+    P_dsml_strs = []
+
+    arguments = json.loads(tool_call["arguments"])
+
+    for k, v in arguments.items():
+        p_dsml_str = p_dsml_template.format(
+            dsml_token=dsml_token,
+            key=k,
+            is_str="true" if isinstance(v, str) else "false",
+            value=v if isinstance(v, str) else to_json(v),
+        )
+
+        P_dsml_strs.append(p_dsml_str)
+
+    return "\n".join(P_dsml_strs)
+
+
+def decode_dsml_to_arguments(
+    tool_name: str, tool_args: Dict[str, Tuple[str, str]]
+) -> Dict[str, str]:
+    def _decode_value(key: str, value: str, string: str):
+        if string == "true":
+            value = to_json(value)
+        return f"{to_json(key)}: {value}"
+
+    tool_args_json = (
+        "{"
+        + ", ".join(
+            [_decode_value(k, v, string=is_str) for k, (v, is_str) in tool_args.items()]
+        )
+        + "}"
+    )
+    return dict(name=tool_name, arguments=tool_args_json)
+
+
+def render_tools(tools: List[Dict[str, Union[str, Dict[str, Any]]]]) -> str:
+    tools_json = [to_json(t) for t in tools]
+
+    return TOOLS_SYSTEM_TEMPLATE.format(
+        tool_schemas="\n".join(tools_json),
+        dsml_token=dsml_token,
+        thinking_start_token=thinking_start_token,
+        thinking_end_token=thinking_end_token,
+    )
+
+
+def find_last_user_index(messages: List[Dict[str, Any]]) -> int:
+    last_user_index = -1
+    for idx in range(len(messages) - 1, -1, -1):
+        if messages[idx].get("role") in ["user", "developer"]:
+            last_user_index = idx
+            break
+    return last_user_index
+
+
+def render_message(
+    index: int,
+    messages: List[Dict[str, Any]],
+    thinking_mode: str,
+    tools: Any = None,
+) -> str:
+    assert 0 <= index < len(messages)
+    assert thinking_mode in [
+        "chat",
+        "thinking",
+    ], f"Invalid thinking_mode `{thinking_mode}`"
+
+    prompt = ""
+    msg = messages[index]
+    last_user_idx = find_last_user_index(messages)
+
+    role = msg.get("role")
+    content = msg.get("content")
+    tools = tools or msg.get("tools")
+    response_format = msg.get("response_format")
+    tool_calls = msg.get("tool_calls")
+    reasoning_content = msg.get("reasoning_content")
+
+    if tool_calls:
+        tool_calls = tool_calls_from_openai_format(tool_calls)
+
+    if role == "system":
+        prompt += system_msg_template.format(content=content or "")
+        if tools:
+            prompt += "\n\n" + render_tools(tools_from_openai_format(tools))
+
+        if response_format:
+            prompt += "\n\n" + response_format_template.format(
+                schema=to_json(response_format)
+            )
+
+    elif role == "developer":
+        assert content, f"Invalid message for role `{role}`: {msg}"
+        content_developer = ""
+        if tools:
+            content_developer += "\n\n" + render_tools(tools_from_openai_format(tools))
+
+        if response_format:
+            content_developer += "\n\n" + response_format_template.format(
+                schema=to_json(response_format)
+            )
+
+        content_developer += "\n\n# The user's message is: {}".format(content)
+
+        prompt += user_msg_template.format(content=content_developer)
+        if index == last_user_idx and thinking_mode == "thinking":
+            prompt += thinking_start_token
+        else:
+            prompt += thinking_end_token
+
+    elif role == "user":
+        prompt += user_msg_template.format(content=content)
+
+        if index == last_user_idx and thinking_mode == "thinking":
+            prompt += thinking_start_token
+        else:
+            prompt += thinking_end_token
+
+    elif role == "tool":
+        prev_assistant_idx = index - 1
+        assistant_msg = messages[prev_assistant_idx]
+        while prev_assistant_idx >= 0 and assistant_msg.get("role") == "tool":
+            prev_assistant_idx -= 1
+            assistant_msg = messages[prev_assistant_idx]
+
+        assert (
+            index == 0
+            or prev_assistant_idx >= 0
+            and assistant_msg.get("role") == "assistant"
+        ), f"Invalid messages at {index}:\n{assistant_msg}"
+
+        tool_call_order = index - prev_assistant_idx
+        assistant_tool_calls = assistant_msg.get("tool_calls")
+        assert (
+            assistant_tool_calls and len(assistant_tool_calls) >= tool_call_order
+        ), "No tool calls but found tool output"
+
+        if tool_call_order == 1:
+            prompt += "\n\n<function_results>"
+
+        prompt += tool_output_template.format(content=content)
+
+        if tool_call_order == len(assistant_tool_calls):
+            prompt += "\n</function_results>"
+
+            if index >= last_user_idx and thinking_mode == "thinking":
+                prompt += "\n\n" + thinking_start_token
+            else:
+                prompt += "\n\n" + thinking_end_token
+
+    elif role == "assistant":
+        prev_assistant_idx = index
+        thinking_part = ""
+
+        tool_calls_content = ""
+        if tool_calls:
+            tool_calls = [
+                tool_call_template.format(
+                    dsml_token=dsml_token,
+                    name=tool_call.get("name"),
+                    arguments=encode_arguments_to_dsml(tool_call),
+                )
+                for tool_call in tool_calls
+            ]
+            tool_calls_content += "\n\n" + tool_calls_template.format(
+                dsml_token=dsml_token, tool_calls="\n".join(tool_calls)
+            )
+
+        summary_content = content or ""
+
+        if thinking_mode == "thinking" and index > last_user_idx:
+            assert (
+                reasoning_content or tool_calls
+            ), f"ThinkingMode: {thinking_mode}, invalid message without reasoning_content/tool_calls `{msg}` after last user message"
+            thinking_part = (
+                thinking_template.format(reasoning_content=reasoning_content or "")
+                + thinking_end_token
+            )
+
+        prompt += assistant_msg_template.format(
+            reasoning=thinking_part,
+            content=summary_content,
+            tool_calls=tool_calls_content,
+        )
+    else:
+        raise NotImplementedError(f"Unknown role: {role}")
+
+    return prompt
+
+
+def drop_thinking_messages(
+    messages: List[Dict[str, Any]], last_user_idx: Optional[int] = None
+) -> List[Dict[str, Any]]:
+    messages_wo_thinking: List[Dict[str, Any]] = []
+    last_user_idx = (
+        find_last_user_index(messages) if last_user_idx is None else last_user_idx
+    )
+    for idx, msg in enumerate(messages):
+        role = msg.get("role")
+        if role in ["user", "system", "tool"] or idx >= last_user_idx:
+            messages_wo_thinking.append(msg)
+            continue
+
+        elif role == "assistant":
+            msg_wo_thinking = copy.copy(msg)
+            msg_wo_thinking.pop("reasoning_content", None)
+            messages_wo_thinking.append(msg_wo_thinking)
+
+    return messages_wo_thinking
+
+
+def encode_messages(
+    messages: List[Dict[str, Any]],
+    thinking_mode: str = "thinking",
+    context: Optional[List[Dict[str, Any]]] = None,
+    drop_thinking: bool = True,
+    add_default_bos_token: bool = True,
+    tools: Any = None,
+) -> str:
+    context = context if context else []
+    full_messages = context + messages
+    prompt = bos_token if add_default_bos_token and len(context) == 0 else ""
+
+    if thinking_mode == "thinking" and drop_thinking:
+        full_messages = drop_thinking_messages(full_messages)
+
+    for idx in range(len(messages)):
+        prompt += render_message(
+            idx + len(context),
+            full_messages,
+            thinking_mode=thinking_mode,
+            tools=tools,
+        )
+
+    return prompt
+
+
+def apply_chat_template(
+    messages, continue_final_message=False, add_generation_prompt=False, **kwargs
+):
+    out = encode_messages(messages, **kwargs)
+    if continue_final_message and add_generation_prompt:
+        raise ValueError(
+            "Only one of continue_final_message or add_generation_prompt can be True"
+        )
+    if not add_generation_prompt and messages[-1]["role"] == "user":
+        out = out.removesuffix("<｜Assistant｜><think>")
+    if continue_final_message and messages[-1]["role"] == "assistant":
+        out = out.removesuffix(eos_token)
+    return out

mlx_lm/cli.py

@@ -0,0 +1,53 @@
+# Copyright © 2025 Apple Inc.
+
+import importlib
+import sys
+
+
+def main():
+    subcommands = (
+        "benchmark",
+        "cache_prompt",
+        "chat",
+        "convert",
+        "evaluate",
+        "fuse",
+        "generate",
+        "lora",
+        "manage",
+        "perplexity",
+        "awq",
+        "dwq",
+        "dynamic_quant",
+        "gptq",
+        "server",
+        "upload",
+        "share",
+    )
+    subpackages = {
+        "awq": "quant",
+        "dwq": "quant",
+        "dynamic_quant": "quant",
+        "gptq": "quant",
+    }
+    if len(sys.argv) < 2:
+        raise ValueError(f"CLI requires a subcommand in {subcommands}")
+    subcommand = sys.argv.pop(1)
+    if subcommand in subcommands:
+        if subpackage := subpackages.get(subcommand):
+            subcommand = f"{subpackage}.{subcommand}"
+        submodule = importlib.import_module(f"mlx_lm.{subcommand}")
+        submodule.main()
+    elif subcommand == "--version":
+        from mlx_lm import __version__
+
+        print(__version__)
+    elif subcommand in ("-h", "--help"):
+        print(f"The supported subcommands are {subcommands}")
+        print()
+        print(
+            "For help on an individual subcommand, pass --help "
+            "to the subcommand. For example: mlx_lm.generate --help"
+        )
+    else:
+        raise ValueError(f"CLI requires a subcommand in {subcommands}")

mlx_lm/cli_ui.py

@@ -0,0 +1,172 @@
+# Copyright © 2024 Apple Inc.
+
+"""Shared UI helpers for the mlx_lm command-line tools.
+
+Centralizes the rich-based panel/progress/prompt rendering used by the chat
+and training entry points. The theme is hardcoded for a light terminal
+background.
+"""
+
+import os
+import re
+import shutil
+import sys
+
+from rich.box import ROUNDED
+from rich.console import Console
+from rich.panel import Panel
+from rich.progress import (
+    Progress,
+    ProgressColumn,
+    TextColumn,
+)
+from rich.text import Text
+from rich.theme import Theme
+
+
+def _terminal_width(default: int = 120) -> int:
+    """Best-effort terminal width.
+
+    Under launchers like ``mlx.launch`` the worker's stdout is a pipe, so
+    Rich's auto-detection falls back to 80 columns. Honor an explicit
+    ``MLX_LM_WIDTH`` override, then ``COLUMNS``, then a real TTY query, and
+    finally a generous default that's nicer than 80 on modern terminals.
+    """
+    for var in ("MLX_LM_WIDTH", "COLUMNS"):
+        value = os.environ.get(var)
+        if value and value.isdigit():
+            return int(value)
+    width = shutil.get_terminal_size(fallback=(0, 0)).columns
+    return width if width > 0 else default
+
+
+def _make_theme() -> Theme:
+    return Theme(
+        {
+            "ui.strong": "bold #000000",
+            "ui.label": "#2a2a2a",
+            "ui.muted": "grey42",
+            "ui.heading": "bold #1a1a1a",
+            "ui.dim": "grey62",
+            "ui.accent": "bold purple",
+            "ui.border": "blue",
+            "ui.good": "bold green",
+            "ui.warn": "yellow",
+            "progress.percentage": "bold blue",
+        }
+    )
+
+
+def make_console(**kwargs) -> Console:
+    """Return a rich Console pre-loaded with the mlx_lm theme."""
+    kwargs.setdefault("highlight", False)
+    kwargs.setdefault("color_system", "truecolor")
+    kwargs.setdefault("width", _terminal_width())
+    return Console(theme=_make_theme(), **kwargs)
+
+
+def print_header_panel(
+    console: Console, title: str, rows: list[tuple[str, str]]
+) -> None:
+    """Render the boxed header used by the chat and training entry points."""
+    label_w = max(len(k) for k, _ in rows)
+    body = "\n".join(
+        f"  [ui.label]{k.ljust(label_w)}[/ui.label]   [ui.strong]{v}[/ui.strong]"
+        for k, v in rows
+    )
+    console.print(
+        Panel(
+            body,
+            title=f"[ui.accent]{title}[/ui.accent]",
+            title_align="left",
+            border_style="ui.border",
+            box=ROUNDED,
+            padding=(0, 2),
+        )
+    )
+
+
+def print_chat_help(console: Console) -> None:
+    console.print(
+        "  [ui.label]commands[/ui.label]    "
+        "[ui.strong]q[/ui.strong] [ui.muted]exit[/ui.muted]   "
+        "[ui.strong]r[/ui.strong] [ui.muted]reset[/ui.muted]   "
+        "[ui.strong]h[/ui.strong] [ui.muted]help[/ui.muted]"
+    )
+
+
+def make_corridor_prompt(console: Console):
+    """Return a callable that draws the chat input corridor.
+
+    The returned callable draws the top/bottom rules around the input line,
+    repositions the cursor onto the middle line, and returns the styled
+    "›" prompt string. Pass that string to ``input()`` so readline treats
+    the marker as part of the prompt — otherwise backspace will erase it.
+    """
+
+    _ANSI_RE = re.compile(r"\x1b\[[0-9;]*[A-Za-z]")
+
+    def _readline_safe(text: str) -> str:
+        # Wrap escape sequences in \x01..\x02 so readline doesn't count
+        # them when computing the prompt's visible width.
+        return _ANSI_RE.sub(lambda m: f"\x01{m.group(0)}\x02", text)
+
+    def _draw() -> str:
+        width = console.width
+        dashes = "─" * max(width - 1, 10)
+        with console.capture() as cap:
+            console.print(f"[ui.muted]{dashes}[/ui.muted]")
+            console.print()
+            console.print(f"[ui.muted]{dashes}[/ui.muted]")
+        sys.stdout.write(cap.get())
+        # Move the cursor up two rows back onto the blank middle line.
+        sys.stdout.write("\033[2A\r")
+        sys.stdout.flush()
+        with console.capture() as cap2:
+            console.print("[ui.accent]›[/ui.accent] ", end="")
+        return _readline_safe(cap2.get())
+
+    return _draw
+
+
+class SquareBar(ProgressColumn):
+    """Progress bar rendered with █/░ blocks plus eighth-block sub-precision."""
+
+    _EIGHTHS = "▏▎▍▌▋▊▉"  # 1/8 .. 7/8
+
+    def __init__(self, bar_width: int = 40, complete_style: str = "blue"):
+        super().__init__()
+        self.bar_width = bar_width
+        self.complete_style = complete_style
+
+    def render(self, task):
+        if not task.total:
+            return Text("░" * self.bar_width, style="ui.dim")
+        pct = min(max(task.completed / task.total, 0.0), 1.0)
+        total_eighths = int(pct * self.bar_width * 8)
+        full = total_eighths // 8
+        rem = total_eighths % 8
+        text = Text()
+        text.append("█" * full, style=self.complete_style)
+        used = full
+        if rem > 0 and full < self.bar_width:
+            text.append(self._EIGHTHS[rem - 1], style=self.complete_style)
+            used += 1
+        text.append("░" * (self.bar_width - used), style="ui.dim")
+        return text
+
+
+def make_train_progress(console: Console, *, disable: bool = False) -> Progress:
+    return Progress(
+        TextColumn("[bold blue]train[/bold blue]"),
+        SquareBar(bar_width=30, complete_style="blue"),
+        TextColumn("[progress.percentage]{task.percentage:>3.0f}%"),
+        TextColumn("[ui.muted]·[/ui.muted]"),
+        TextColumn(
+            "[bold blue]{task.completed:>5,}[/bold blue]"
+            "[ui.muted]/{task.total:,}[/ui.muted]"
+        ),
+        console=console,
+        transient=False,
+        disable=disable,
+    )

mlx_lm/convert.py

@@ -10,8 +10,7 @@ from mlx.utils import tree_map_with_path
 
 from .utils import (
     dequantize_model,
-    fetch_from_hub,
-    get_model_path,
+    load,
     quantize_model,
     save,
     upload_to_hub,
@@ -19,10 +18,10 @@ from .utils import (
 
 
 def mixed_quant_predicate_builder(
-    recipe: str, model: nn.Module
+    recipe: str, model: nn.Module, group_size: int = 64
 ) -> Callable[[str, nn.Module, dict], Union[bool, dict]]:
+    mode = "affine"
     high_bits = 6
-    group_size = 64
 
     if recipe == "mixed_2_6":
         low_bits = 2
@@ -34,7 +33,7 @@ def mixed_quant_predicate_builder(
     elif recipe == "mixed_4_6":
         low_bits = 4
     else:
-        raise ValueError("Invalid quant recipe {recipe}")
+        raise ValueError(f"Invalid quant recipe {recipe}")
 
     down_keys = [k for k, _ in model.named_modules() if "down_proj" in k]
     if len(down_keys) == 0:
@@ -64,14 +63,16 @@ def mixed_quant_predicate_builder(
             or index >= 7 * num_layers // 8
             or (index - num_layers // 8) % 3 == 2
         )
-        if "v_proj" in path and use_more_bits:
-            return {"group_size": group_size, "bits": high_bits}
+        if (
+            "v_proj" in path or "v_a_proj" in path or "v_b_proj" in path
+        ) and use_more_bits:
+            return {"group_size": group_size, "bits": high_bits, "mode": mode}
         if "down_proj" in path and use_more_bits:
-            return {"group_size": group_size, "bits": high_bits}
+            return {"group_size": group_size, "bits": high_bits, "mode": mode}
         if "lm_head" in path:
-            return {"group_size": group_size, "bits": high_bits}
+            return {"group_size": group_size, "bits": high_bits, "mode": mode}
 
-        return {"group_size": group_size, "bits": low_bits}
+        return {"group_size": group_size, "bits": low_bits, "mode": mode}
 
     return mixed_quant_predicate
 
@@ -85,8 +86,8 @@ def convert(
     hf_path: str,
     mlx_path: str = "mlx_model",
     quantize: bool = False,
-    q_group_size: int = 64,
-    q_bits: int = 4,
+    q_group_size: Optional[int] = None,
+    q_bits: Optional[int] = None,
     q_mode: str = "affine",
     dtype: Optional[str] = None,
     upload_repo: str = None,
@@ -108,16 +109,27 @@ def convert(
         )
 
     print("[INFO] Loading")
-    model_path, hf_path = get_model_path(hf_path, revision=revision)
-    model, config, tokenizer = fetch_from_hub(
-        model_path, lazy=True, trust_remote_code=trust_remote_code
+    model, tokenizer, config = load(
+        hf_path,
+        revision=revision,
+        return_config=True,
+        tokenizer_config={"trust_remote_code": trust_remote_code},
+        lazy=True,
     )
 
     if isinstance(quant_predicate, str):
-        quant_predicate = mixed_quant_predicate_builder(quant_predicate, model)
+        if q_mode != "affine":
+            raise ValueError(f"Quant predicates only support 'affine' quantization.")
+        quant_predicate = mixed_quant_predicate_builder(
+            quant_predicate,
+            model,
+            q_group_size,
+        )
 
     if dtype is None:
         dtype = config.get("torch_dtype", None)
+    if dtype is None and (text_config := config.get("text_config", None)):
+        dtype = text_config.get("dtype", None)
     if dtype in MODEL_CONVERSION_DTYPES:
         print("[INFO] Using dtype:", dtype)
         dtype = getattr(mx, dtype)
@@ -153,11 +165,10 @@ def convert(
 
     save(
         mlx_path,
-        model_path,
+        hf_path,
         model,
         tokenizer,
         config,
-        hf_repo=hf_path,
     )
 
     if upload_repo is not None:
@@ -175,7 +186,12 @@ def configure_parser() -> argparse.ArgumentParser:
         description="Convert Hugging Face model to MLX format"
     )
 
-    parser.add_argument("--hf-path", type=str, help="Path to the Hugging Face model.")
+    parser.add_argument(
+        "--hf-path",
+        "--model",
+        type=str,
+        help="Path to the model. This can be a local path or a Hugging Face Hub model identifier.",
+    )
     parser.add_argument(
         "--mlx-path", type=str, default="mlx_model", help="Path to save the MLX model."
     )
@@ -183,17 +199,23 @@ def configure_parser() -> argparse.ArgumentParser:
         "-q", "--quantize", help="Generate a quantized model.", action="store_true"
     )
     parser.add_argument(
-        "--q-group-size", help="Group size for quantization.", type=int, default=64
+        "--q-group-size",
+        help="Group size for quantization.",
+        type=int,
+        default=None,
     )
     parser.add_argument(
-        "--q-bits", help="Bits per weight for quantization.", type=int, default=4
+        "--q-bits",
+        help="Bits per weight for quantization.",
+        type=int,
+        default=None,
     )
     parser.add_argument(
         "--q-mode",
         help="The quantization mode.",
         type=str,
         default="affine",
-        choices=["affine", "mxfp4"],
+        choices=["affine", "mxfp4", "nvfp4", "mxfp8"],
     )
     parser.add_argument(
         "--quant-predicate",

mlx_lm/evaluate.py

@@ -12,7 +12,7 @@ import logging
 import os
 from importlib.metadata import version
 from pathlib import Path
-from typing import Any, Optional
+from typing import Any, Callable, Optional
 
 import lm_eval
 import mlx.core as mx
@@ -20,13 +20,14 @@ import mlx.nn as nn
 import numpy as np
 from lm_eval.api.model import LM
 from lm_eval.api.registry import register_model
-from lm_eval.models import huggingface
 from tqdm import tqdm
 
-from .generate import stream_generate
-from .models.base import create_causal_mask
+from .generate import batch_generate
 from .models.cache import make_prompt_cache
-from .utils import common_prefix_len, load
+from .sample_utils import make_sampler
+from .utils import load
+
+DEFAULT_MAX_TOKENS = 8192
 
 
 def _rstrip_until(s, untils):
@@ -37,6 +38,13 @@ def _rstrip_until(s, untils):
     return s[: min(f)]
 
 
+def _lstrip(s, pattern):
+    """Truncate the prefix of the string after the first occurrence of pattern."""
+    if (idx := s.find(pattern)) != -1:
+        return s[idx + len(pattern) :]
+    return s
+
+
 def _pad_inputs(inputs):
     lengths = np.array([len(x) for x in inputs])
     maxlen = lengths.max()
@@ -63,26 +71,28 @@ def chat_template_fn(**extra_kwargs):
 @register_model("mlxlm")
 class MLXLM(LM):
 
-    tokenizer_name = huggingface.HFLM.tokenizer_name
     apply_chat_template = chat_template_fn()
 
     def __init__(
         self,
         path_or_hf_repo: str,
         max_tokens: Optional[int] = None,
+        batch_size: int = 8,
         use_chat_template: Optional[bool] = None,
         trust_remote_code: bool = False,
+        sampler: Optional[Callable[[mx.array], mx.array]] = None,
     ) -> None:
         super().__init__()
         tokenizer_config = {"trust_remote_code": True if trust_remote_code else None}
         self._model, self.tokenizer = load(
             path_or_hf_repo, tokenizer_config=tokenizer_config
         )
-        self._max_tokens = max_tokens or self.tokenizer.model_max_length
-        self._batch_size = 8
+        self._max_tokens = max_tokens
+        self._batch_size = batch_size
         self.use_chat_template = use_chat_template
         if use_chat_template is None:
             self.use_chat_template = self.tokenizer.chat_template is not None
+        self._sampler = sampler
 
     def _process_prompt(self, prompt, step_size: int = 2048):
         prompt = mx.array(prompt)[None]
@@ -99,30 +109,28 @@ class MLXLM(LM):
         inputs, targets = inputs[..., :-1], inputs[..., 1:]
 
         cache = cache or make_prompt_cache(self._model)
-        lengths += cache[0].offset
-
+        offset = 0
         scores, is_greedy = [], []
         for i in range(0, inputs.shape[1], step_size):
             inp = inputs[:, i : i + step_size]
             T = inp.shape[1]
 
-            offset = cache[0].offset
-            mask = create_causal_mask(T, offset, lengths=lengths)
-
-            logits = self._model(inp, cache=cache, mask=mask)
+            logits = self._model(inp, cache=cache)
             log_probs = nn.log_softmax(logits.astype(mx.float32))
 
             score = mx.take_along_axis(
                 log_probs, targets[:, i : i + step_size, mx.newaxis], axis=-1
             )[..., 0]
+
             ig = targets[:, i : i + step_size] == mx.argmax(logits, axis=-1)
-            ig = mx.where(mx.arange(T) + offset < lengths[:, None], ig, False)
+            ig = mx.where(mx.arange(offset, T + offset) < lengths[:, None], ig, False)
 
             mx.eval(score, ig)
             mx.clear_cache()
 
             is_greedy.append(ig)
             scores.append(score)
+            offset += T
 
         scores = mx.concatenate(scores, axis=1)
         is_greedy = mx.concatenate(is_greedy, axis=1)
@@ -137,6 +145,10 @@ class MLXLM(LM):
             for t in texts
         ]
 
+    @property
+    def tokenizer_name(self) -> str:
+        return self.tokenizer.name_or_path.replace("/", "__")
+
     def loglikelihood(self, requests) -> list[tuple[float, bool]]:
         """Compute log-likelihood of generating a continuation from a context.
         Downstream tasks should attempt to use loglikelihood instead of other
@@ -185,7 +197,8 @@ class MLXLM(LM):
             max_completed_l = max(len(s) for s in full_sequences)
 
             # compute truncation length
-            truncation = max(0, max_completed_l - self._max_tokens - 1)
+            max_tokens = self._max_tokens or DEFAULT_MAX_TOKENS
+            truncation = max(0, max_completed_l - max_tokens - 1)
             orig_prefix_l = len(prefix)
             prefix_l = max(len(prefix) - truncation, 0)
             prefix = prefix[len(prefix) - prefix_l :]
@@ -307,32 +320,77 @@ class MLXLM(LM):
             continuation: str
                 The generated continuation.
         """
+        group = mx.distributed.init()
+
+        # split data accross ranks
+        total_requests = len(requests)
+        requests = requests[group.rank() :: group.size()]
+
         logging.info("Generating continuation for %d sequences." % len(requests))
         contexts, options = zip(*[req.args for req in requests])
-        # contrary to the doc the second element of the tuple contains
+        # The second element of the tuple contains:
         # {'do_sample': False, 'until': ['\n\n'], 'temperature': 0}
-        completions = []
 
-        for context, opt in tqdm(zip(contexts, options), total=len(contexts)):
-            until = opt["until"]
-            context = self.tokenizer.encode(
+        # Tokenize all contexts
+        contexts = [
+            self.tokenizer.encode(
                 context, add_special_tokens=not self.use_chat_template
             )
-            max_tokens = min(
-                opt.get("max_gen_tokens", self._max_tokens),
-                self.tokenizer.model_max_length - len(context),
-            )
-            text = ""
-            for response in stream_generate(
-                self._model, self.tokenizer, prompt=context, max_tokens=max_tokens
-            ):
-                text += response.text
-                if any(u in text for u in until):
-                    text = _rstrip_until(text, until)
-                    completions.append(text)
-                    break
-            else:
-                completions.append(text)
+            for context in contexts
+        ]
+
+        # TODO consider multi-token, per-prompt stop conditions
+        max_tokens = [
+            self._max_tokens or opt.get("max_gen_tokens", DEFAULT_MAX_TOKENS)
+            for opt in options
+        ]
+
+        completions = batch_generate(
+            model=self._model,
+            tokenizer=self.tokenizer,
+            prompts=contexts,
+            max_tokens=max_tokens,
+            verbose=True,
+            sampler=self._sampler,
+        ).texts
+
+        for e, (text, opt) in enumerate(zip(completions, options)):
+            completions[e] = _rstrip_until(text, opt["until"])
+            if self.tokenizer.has_thinking:
+                completions[e] = _lstrip(text, self.tokenizer.think_end)
+
+        # Gather the completions
+        if group.size() > 1:
+            with mx.stream(mx.cpu):
+                pad_to = (total_requests + group.size() - 1) // group.size()
+                pad = pad_to - len(completions)
+                completions = [list(c.encode("utf-8")) for c in completions]
+                max_len = mx.array(max(len(c) for c in completions))
+                max_len = mx.distributed.all_max(max_len).item()
+                lengths = mx.array([len(c) for c in completions] + [0] * pad)
+                completions = mx.array(
+                    [c + [0] * (max_len - len(c)) for c in completions]
+                    + [[0] * max_len] * pad,
+                    mx.uint8,
+                )
+                completions = (
+                    mx.distributed.all_gather(completions[None])
+                    .swapaxes(0, 1)
+                    .flatten(0, 1)
+                    .tolist()
+                )
+                lengths = (
+                    mx.distributed.all_gather(lengths[None])
+                    .swapaxes(0, 1)
+                    .flatten(0, 1)
+                    .tolist()
+                )
+                completions = completions[:total_requests]
+                lengths = lengths[:total_requests]
+                completions = [
+                    bytearray(c[:l]).decode() for c, l in zip(completions, lengths)
+                ]
+
         return completions
 
 
@@ -350,7 +408,9 @@ def main():
     parser.add_argument(
         "--max-tokens",
         type=int,
-        help="Maximum nunber of tokens to generate. Defaults to the model's max context length.",
+        help="Maximum number of tokens to generate. When set, this value takes"
+        " precedence over task specific defaults.",
+        default=None,
     )
     parser.add_argument(
         "--limit",
@@ -392,7 +452,9 @@ def main():
         action="store_true",
         help="Enable trusting remote code for tokenizer",
     )
-
+    parser.add_argument("--temp", type=float, default=0.0, help="Sampling temperature")
+    parser.add_argument("--top-p", type=float, default=1.0, help="Sampling top-p")
+    parser.add_argument("--top-k", type=int, default=0, help="Sampling top-k")
     args = parser.parse_args()
 
     output_dir = Path(args.output_dir)
@@ -409,11 +471,18 @@ def main():
     if world.size() > 1 and world.rank() == 0:
         print(f"Evaluating with {world.size()} nodes")
 
+    sampler = make_sampler(
+        temp=args.temp,
+        top_p=args.top_p,
+        top_k=args.top_k,
+    )
     lm = MLXLM(
         args.model,
         max_tokens=args.max_tokens,
+        batch_size=args.batch_size,
         use_chat_template=args.apply_chat_template,
         trust_remote_code=args.trust_remote_code,
+        sampler=sampler,
     )
     MLXLM.apply_chat_template = chat_template_fn(**args.chat_template_args)
 

mlx_lm/examples/batch_generate_response.py

@@ -0,0 +1,51 @@
+# Copyright © 2025 Apple Inc.
+
+from mlx_lm import batch_generate, load
+
+# Specify the checkpoint
+checkpoint = "mlx-community/Llama-3.2-3B-Instruct-4bit"
+
+# Load the corresponding model and tokenizer
+model, tokenizer = load(path_or_hf_repo=checkpoint)
+
+# A batch of prompts
+prompts = [
+    "Write a story about Einstein.",
+    "Why is the sky blue?",
+    "What time is it?",
+    "How tall is Mt Everest?",
+]
+
+# Apply the chat template and encode to tokens
+prompts = [
+    tokenizer.apply_chat_template(
+        [{"role": "user", "content": p}],
+        add_generation_prompt=True,
+    )
+    for p in prompts
+]
+
+# Set `verbose=True` to see generation statistics
+result = batch_generate(
+    model, tokenizer, prompts, verbose=False, return_prompt_caches=True, max_tokens=2048
+)
+print(result.texts[-1])
+
+prompts = [
+    "Could you summarize that?",
+    "And what about the sea?",
+    "Try again?",
+    "And Mt Olympus?",
+]
+prompts = [
+    tokenizer.apply_chat_template(
+        [{"role": "user", "content": p}],
+        add_generation_prompt=True,
+    )
+    for p in prompts
+]
+
+result = batch_generate(
+    model, tokenizer, prompts, verbose=False, prompt_caches=result.caches
+)
+print(result.texts[-1])

mlx_lm/examples/chat.py

@@ -15,7 +15,10 @@ prompt_cache = make_prompt_cache(model)
 # User turn
 prompt = "Hi my name is <Name>."
 messages = [{"role": "user", "content": prompt}]
-prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+prompt = tokenizer.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+)
 
 # Assistant response
 response = generate(
@@ -29,7 +32,10 @@ response = generate(
 # User turn
 prompt = "What's my name?"
 messages = [{"role": "user", "content": prompt}]
-prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+prompt = tokenizer.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+)
 
 # Assistant response
 response = generate(

mlx_lm/examples/generate_response.py

@@ -14,7 +14,8 @@ conversation = [{"role": "user", "content": prompt}]
 
 # Transform the prompt into the chat template
 prompt = tokenizer.apply_chat_template(
-    conversation=conversation, add_generation_prompt=True
+    conversation=conversation,
+    add_generation_prompt=True,
 )
 
 # Specify the maximum number of tokens

mlx_lm/examples/lora_config.yaml

@@ -1,5 +1,5 @@
 # The path to the local model directory or Hugging Face repo.
-model: "mlx-community/Llama-3.2-1B-Instruct"
+model: "mlx-community/Llama-3.2-1B-Instruct-bf16"
 
 # Whether or not to train (boolean)
 train: true
@@ -47,6 +47,9 @@ steps_per_report: 10
 # Number of training steps between validations.
 steps_per_eval: 200
 
+# Number of micro-steps to accumulate before each optimizer update.
+grad_accumulation_steps: 1
+
 # Load path to resume training with the given adapter weights.
 resume_adapter_file: null
 
@@ -90,4 +93,3 @@ lora_parameters:
 #  valid_split: "train[-100:]"
 #  prompt_feature: "text"
 #  completion_feature: "summary"
-

mlx_lm/examples/openai_reasoning_content.py

@@ -0,0 +1,40 @@
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="not-needed",
+    base_url="http://localhost:8080/v1",
+)
+
+model = "mlx-community/Qwen3-4B-Thinking-2507-4bit"
+
+messages = [{"role": "user", "content": "9.11 and 9.8, which is greater?"}]
+
+# Non-streaming example
+
+response = client.chat.completions.create(
+    model=model, messages=messages, max_tokens=2048
+)
+
+reasoning = response.choices[0].message.reasoning
+content = response.choices[0].message.content
+
+print("=== reasoning ===\n")
+print(f"\033[37m{reasoning}\033[0m")
+print("=== content ===\n")
+print(content)
+
+# Streaming example
+
+stream = client.chat.completions.create(
+    model=model,
+    messages=messages,
+    stream=True,
+    max_tokens=2048,
+)
+
+for chunk in stream:
+    if (reasoning := chunk.choices[0].delta.reasoning) is not None:
+        print(f"\033[37m{reasoning}\033[0m", end="")
+    if (content := chunk.choices[0].delta.content) is not None:
+        print(f"{content}", end="")
+print()

mlx_lm/examples/openai_tool_use.py

@@ -8,11 +8,13 @@ To run, first start the server:
 
 Then run this script.
 """
+import json
+
 from openai import OpenAI
 
 client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-needed")
 
-model = "mlx-community/qwen3-4b-4bit-DWQ"
+model = "mlx-community/Qwen3-4B-Instruct-2507-4bit"
 messages = [{"role": "user", "content": "What's the weather in Boston?"}]
 
 tools = [

mlx_lm/examples/pipeline_generate.py

@@ -1,135 +0,0 @@
-# Copyright © 2024 Apple Inc.
-
-"""
-Run with:
-
-```
-mlx.launch \
- --hostfile /path/to/hosts.json \
- /path/to/pipeline_generate.py \
- --prompt "hello world"
-```
-
-Make sure you can run MLX over MPI on two hosts. For more information see the
-documentation:
-
-https://ml-explore.github.io/mlx/build/html/usage/distributed.html).
-"""
-
-import argparse
-import json
-import resource
-from pathlib import Path
-
-import mlx.core as mx
-from huggingface_hub import snapshot_download
-from mlx.utils import tree_flatten
-
-from mlx_lm import load, stream_generate
-from mlx_lm.utils import load_model, load_tokenizer
-
-# Needed for 8 bit model
-resource.setrlimit(resource.RLIMIT_NOFILE, (2048, 4096))
-
-
-def download(repo: str, allow_patterns: list[str]) -> Path:
-    return Path(
-        snapshot_download(
-            repo,
-            allow_patterns=allow_patterns,
-        )
-    )
-
-
-def shard_and_load(repo):
-    # Get model path with everything but weight safetensors
-    model_path = download(
-        args.model,
-        allow_patterns=["*.json", "*.py", "tokenizer.model", "*.tiktoken", "*.txt"],
-    )
-
-    # Lazy load and shard model to figure out
-    # which weights we need
-    model, config = load_model(model_path, lazy=True, strict=False)
-
-    group = mx.distributed.init()
-    rank = group.rank()
-    model.model.pipeline(group)
-
-    # Figure out which files we need for the local shard
-    with open(model_path / "model.safetensors.index.json", "r") as fid:
-        weight_index = json.load(fid)["weight_map"]
-
-    local_files = set()
-    for k, _ in tree_flatten(model.parameters()):
-        local_files.add(weight_index[k])
-
-    # Download weights for local shard
-    download(args.model, allow_patterns=local_files)
-
-    # Load and shard the model, and load the weights
-    tokenizer = load_tokenizer(
-        model_path,
-        {"trust_remote_code": True},
-        eos_token_ids=config.get("eos_token_id", None),
-    )
-    model, _ = load_model(model_path, lazy=True, strict=False)
-    model.model.pipeline(group)
-    mx.eval(model.parameters())
-
-    # Synchronize processes before generation to avoid timeout if downloading
-    # model for the first time.
-    mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu))
-    return model, tokenizer
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="LLM pipelined inference example")
-    parser.add_argument(
-        "--model",
-        default="mlx-community/DeepSeek-R1-3bit",
-        help="HF repo or path to local model.",
-    )
-    parser.add_argument(
-        "--prompt",
-        "-p",
-        default="Write a quicksort in C++.",
-        help="Message to be processed by the model ('-' reads from stdin)",
-    )
-    parser.add_argument(
-        "--max-tokens",
-        "-m",
-        type=int,
-        default=256,
-        help="Maximum number of tokens to generate",
-    )
-    args = parser.parse_args()
-
-    group = mx.distributed.init()
-    rank = group.rank()
-
-    def rprint(*args, **kwargs):
-        if rank == 0:
-            print(*args, **kwargs)
-
-    model, tokenizer = shard_and_load(args.model)
-
-    messages = [{"role": "user", "content": args.prompt}]
-    prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
-
-    for response in stream_generate(
-        model, tokenizer, prompt, max_tokens=args.max_tokens
-    ):
-        rprint(response.text, end="", flush=True)
-
-    rprint()
-    rprint("=" * 10)
-    rprint(
-        f"Prompt: {response.prompt_tokens} tokens, "
-        f"{response.prompt_tps:.3f} tokens-per-sec"
-    )
-    rprint(
-        f"Generation: {response.generation_tokens} tokens, "
-        f"{response.generation_tps:.3f} tokens-per-sec"
-    )
-    rprint(f"Peak memory: {response.peak_memory:.3f} GB")

mlx_lm/examples/sharded_generate.py

@@ -0,0 +1,86 @@
+# Copyright © 2025 Apple Inc.
+
+"""
+Run with:
+
+```
+mlx.launch \
+    --backend jaccl \
+    --env MLX_METAL_FAST_SYNCH=1 \
+    --hostfile /path/to/hosts.json \
+    /path/to/sharded_generate.py \
+    --prompt 'Hello world'
+```
+
+For more information on running distributed programs with MLX see the documentation:
+
+https://ml-explore.github.io/mlx/build/html/usage/distributed.html .
+"""
+
+import argparse
+
+import mlx.core as mx
+
+from mlx_lm import stream_generate
+from mlx_lm.utils import sharded_load
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="LLM distributed inference example")
+    parser.add_argument(
+        "--model",
+        default="mlx-community/Llama-3.3-70B-Instruct-4bit",
+        help="HF repo or path to local model.",
+    )
+    parser.add_argument(
+        "--prompt",
+        "-p",
+        default="Write a quicksort in C++.",
+        help="Message to be processed by the model ('-' reads from stdin)",
+    )
+    parser.add_argument(
+        "--max-tokens",
+        "-m",
+        type=int,
+        default=256,
+        help="Maximum number of tokens to generate",
+    )
+    parser.add_argument(
+        "--pipeline",
+        action="store_true",
+        help="Use pipelining instead of tensor parallelism",
+    )
+    args = parser.parse_args()
+
+    group = mx.distributed.init()
+    rank = group.rank()
+    pipeline_group = group if args.pipeline else None
+    tensor_group = group if not args.pipeline else None
+
+    def rprint(*args, **kwargs):
+        if rank == 0:
+            print(*args, **kwargs)
+
+    model, tokenizer = sharded_load(args.model, pipeline_group, tensor_group)
+
+    messages = [{"role": "user", "content": args.prompt}]
+    prompt = tokenizer.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+    )
+
+    for response in stream_generate(
+        model, tokenizer, prompt, max_tokens=args.max_tokens
+    ):
+        rprint(response.text, end="", flush=True)
+
+    rprint()
+    rprint("=" * 10)
+    rprint(
+        f"Prompt: {response.prompt_tokens} tokens, "
+        f"{response.prompt_tps:.3f} tokens-per-sec"
+    )
+    rprint(
+        f"Generation: {response.generation_tokens} tokens, "
+        f"{response.generation_tps:.3f} tokens-per-sec"
+    )
+    rprint(f"Peak memory: {response.peak_memory:.3f} GB")

mlx_lm/examples/tool_use.py

@@ -6,7 +6,7 @@ from mlx_lm import generate, load
 from mlx_lm.models.cache import make_prompt_cache
 
 # Specify the checkpoint
-checkpoint = "mlx-community/Qwen2.5-32B-Instruct-4bit"
+checkpoint = "mlx-community/Qwen3-4B-Instruct-2507-4bit"
 
 # Load the corresponding model and tokenizer
 model, tokenizer = load(path_or_hf_repo=checkpoint)
@@ -31,7 +31,9 @@ prompt = "Multiply 12234585 and 48838483920."
 messages = [{"role": "user", "content": prompt}]
 
 prompt = tokenizer.apply_chat_template(
-    messages, add_generation_prompt=True, tools=list(tools.values())
+    messages,
+    add_generation_prompt=True,
+    tools=list(tools.values()),
 )
 
 prompt_cache = make_prompt_cache(model)
@@ -47,12 +49,11 @@ response = generate(
 )
 
 # Parse the tool call:
-# (Note, the tool call format is model specific)
-tool_open = "<tool_call>"
-tool_close = "</tool_call>"
-start_tool = response.find(tool_open) + len(tool_open)
-end_tool = response.find(tool_close)
-tool_call = json.loads(response[start_tool:end_tool].strip())
+# - The tool call format is model specific.
+# - The tokenizer's tool parser expects tool call text to be already extracted.
+start_tool = response.find(tokenizer.tool_call_start) + len(tokenizer.tool_call_start)
+end_tool = response.find(tokenizer.tool_call_end)
+tool_call = tokenizer.tool_parser(response[start_tool:end_tool].strip())
 tool_result = tools[tool_call["name"]](**tool_call["arguments"])
 
 # Put the tool result in the prompt

mlx_lm/fuse.py

@@ -4,10 +4,9 @@ from pathlib import Path
 from mlx.utils import tree_flatten, tree_unflatten
 
 from .gguf import convert_to_gguf
-from .tuner.utils import dequantize, load_adapters
 from .utils import (
-    fetch_from_hub,
-    get_model_path,
+    dequantize_model,
+    load,
     save,
     upload_to_hub,
 )
@@ -40,8 +39,8 @@ def parse_arguments() -> argparse.Namespace:
         default=None,
     )
     parser.add_argument(
-        "--de-quantize",
-        help="Generate a de-quantized model.",
+        "--dequantize",
+        help="Generate a dequantized model.",
         action="store_true",
     )
     parser.add_argument(
@@ -62,14 +61,12 @@ def main() -> None:
     print("Loading pretrained model")
     args = parse_arguments()
 
-    model_path, hf_path = get_model_path(args.model)
-    model, config, tokenizer = fetch_from_hub(model_path)
-
-    model.freeze()
-    model = load_adapters(model, args.adapter_path)
+    model, tokenizer, config = load(
+        args.model, adapter_path=args.adapter_path, return_config=True
+    )
 
     fused_linears = [
-        (n, m.fuse(de_quantize=args.de_quantize))
+        (n, m.fuse(dequantize=args.dequantize))
         for n, m in model.named_modules()
         if hasattr(m, "fuse")
     ]
@@ -77,19 +74,19 @@ def main() -> None:
     if fused_linears:
         model.update_modules(tree_unflatten(fused_linears))
 
-    if args.de_quantize:
-        print("De-quantizing model")
-        model = dequantize(model)
+    if args.dequantize:
+        print("Dequantizing model")
+        model = dequantize_model(model)
         config.pop("quantization", None)
+        config.pop("quantization_config", None)
 
     save_path = Path(args.save_path)
     save(
         save_path,
-        model_path,
+        args.model,
         model,
         tokenizer,
         config,
-        hf_repo=hf_path,
         donate_model=False,
     )
 
@@ -100,13 +97,9 @@ def main() -> None:
                 f"Model type {model_type} not supported for GGUF conversion."
             )
         weights = dict(tree_flatten(model.parameters()))
-        convert_to_gguf(model_path, weights, config, str(save_path / args.gguf_path))
+        convert_to_gguf(save_path, weights, config, str(save_path / args.gguf_path))
 
     if args.upload_repo is not None:
-        if hf_path is None:
-            raise ValueError(
-                "Must provide original Hugging Face repo to upload local model."
-            )
         upload_to_hub(args.save_path, args.upload_repo)
 
 

mlx_lm/lora.py

@@ -12,6 +12,7 @@ import mlx.optimizers as optim
 import numpy as np
 import yaml
 
+from .cli_ui import make_console, print_header_panel
 from .tuner.callbacks import get_reporting_callbacks
 from .tuner.datasets import CacheDataset, load_dataset
 from .tuner.trainer import TrainingArgs, TrainingCallback, evaluate, train
@@ -21,7 +22,13 @@ from .tuner.utils import (
     load_adapters,
     print_trainable_parameters,
 )
-from .utils import load, save_config
+from .utils import _parse_size, load, save_config
+
+
+def printf(*args, **kwargs):
+    if mx.distributed.init().rank() == 0:
+        print(*args, **kwargs)
+
 
 yaml_loader = yaml.SafeLoader
 yaml_loader.add_implicit_resolver(
@@ -40,7 +47,7 @@ yaml_loader.add_implicit_resolver(
 )
 
 CONFIG_DEFAULTS = {
-    "model": "mlx_model",
+    "model": "Qwen/Qwen3-0.6b",
     "train": False,
     "fine_tune_type": "lora",
     "optimizer": "adam",
@@ -51,7 +58,7 @@ CONFIG_DEFAULTS = {
         "sgd": {},
         "adafactor": {},
     },
-    "data": "data/",
+    "data": "mlx-community/WikiSQL",
     "seed": 0,
     "num_layers": 16,
     "batch_size": 4,
@@ -68,10 +75,11 @@ CONFIG_DEFAULTS = {
     "max_seq_length": 2048,
     "config": None,
     "grad_checkpoint": False,
+    "grad_accumulation_steps": 1,
+    "clear_cache_threshold": 0,
     "lr_schedule": None,
     "lora_parameters": {"rank": 8, "dropout": 0.0, "scale": 20.0},
     "mask_prompt": False,
-    "wandb": None,  # will be deprecated in a future release
     "report_to": None,
     "project_name": None,
 }
@@ -142,6 +150,11 @@ def build_parser():
         type=int,
         help="Number of training steps between validations.",
     )
+    parser.add_argument(
+        "--grad-accumulation-steps",
+        type=int,
+        help="Number of steps to accumulate before each optimizer update.",
+    )
     parser.add_argument(
         "--resume-adapter-file",
         type=str,
@@ -185,14 +198,11 @@ def build_parser():
         help="Use gradient checkpointing to reduce memory use.",
         default=None,
     )
-    parser.add_argument(  # will be deprecated in a future release
-        "--wandb",
-        type=str,
-        default=None,
-        help=(
-            "The 'wandb' argument is deprecated and will be removed in a future release. "
-            "Use 'report_to: wandb' and 'project_name' in the configuration instead."
-        ),
+    parser.add_argument(
+        "--clear-cache-threshold",
+        type=_parse_size,
+        default=0,
+        help="Clear the allocator cache between steps if it grows too large.",
     )
     parser.add_argument(
         "--report-to",
@@ -243,7 +253,7 @@ def train_model(
 
     # Resume from weights if provided
     if args.resume_adapter_file is not None:
-        print(f"Loading fine-tuned weights from {args.resume_adapter_file}")
+        printf(f"Loading fine-tuned weights from {args.resume_adapter_file}")
         model.load_weights(args.resume_adapter_file, strict=False)
 
     print_trainable_parameters(model)
@@ -265,6 +275,7 @@ def train_model(
         adapter_file=adapter_file,
         max_seq_length=args.max_seq_length,
         grad_checkpoint=args.grad_checkpoint,
+        grad_accumulation_steps=args.grad_accumulation_steps,
     )
 
     # Initialize the selected optimizer
@@ -287,6 +298,8 @@ def train_model(
 
     opt = opt_class(learning_rate=lr, **optimizer_config)
 
+    _print_run_header(args)
+
     # Train model
     train(
         model=model,
@@ -298,30 +311,54 @@ def train_model(
     )
 
 
+def _print_run_header(args):
+    rank = mx.distributed.init().rank()
+    if rank != 0:
+        return
+
+    type_label = args.fine_tune_type
+    if args.fine_tune_type in ("lora", "dora"):
+        rank = args.lora_parameters.get("rank", "?")
+        type_label = f"{args.fine_tune_type} · {args.num_layers} layers · rank {rank}"
+    elif args.fine_tune_type == "full":
+        type_label = f"full · {args.num_layers} layers"
+
+    lr = (
+        args.learning_rate
+        if isinstance(args.learning_rate, (int, float))
+        else "schedule"
+    )
+    lr_str = f"{lr:.1e}" if isinstance(lr, (int, float)) else lr
+
+    rows = [
+        ("model", str(args.model)),
+        ("type", type_label),
+        ("dataset", str(args.data)),
+        ("optimizer", f"{args.optimizer} · lr {lr_str}"),
+        ("batch · iters", f"{args.batch_size} · {args.iters:,}"),
+        ("max seq", f"{args.max_seq_length:,}"),
+    ]
+    print_header_panel(make_console(), "mlx_lm.lora", rows)
+
+
 def evaluate_model(args, model: nn.Module, test_set):
+    rank = mx.distributed.init().rank()
     test_loss = evaluate(
         model=model,
         dataset=CacheDataset(test_set),
         batch_size=args.batch_size,
         num_batches=args.test_batches,
         max_seq_length=args.max_seq_length,
+        progress=(rank == 0),
     )
 
     test_ppl = math.exp(test_loss)
 
-    print(f"Test loss {test_loss:.3f}, Test ppl {test_ppl:.3f}.")
+    printf(f"Test loss {test_loss:.3f}, Test ppl {test_ppl:.3f}.")
 
 
 def run(args, training_callback: TrainingCallback = None):
     np.random.seed(args.seed)
-    if args.wandb is not None:
-        warnings.warn(
-            "The 'wandb' argument is deprecated and will be removed in a future release. "
-            "Use 'report_to: wandb' and 'project_name' in the configuration instead.",
-            DeprecationWarning,
-        )
-        args.report_to = "wandb"
-        args.project_name = args.wandb
     training_callback = get_reporting_callbacks(
         args.report_to,
         project_name=args.project_name,
@@ -329,10 +366,10 @@ def run(args, training_callback: TrainingCallback = None):
         config=vars(args),
     )
 
-    print("Loading pretrained model")
+    printf("Loading pretrained model")
     model, tokenizer = load(args.model, tokenizer_config={"trust_remote_code": True})
 
-    print("Loading datasets")
+    printf("Loading datasets")
     train_set, valid_set, test_set = load_dataset(args, tokenizer)
 
     if args.test and not args.train:
@@ -341,13 +378,13 @@ def run(args, training_callback: TrainingCallback = None):
             load_adapters(model, args.adapter_path)
 
     elif args.train:
-        print("Training")
+        printf("Training")
         train_model(args, model, train_set, valid_set, training_callback)
     else:
         raise ValueError("Must provide at least one of --train or --test")
 
     if args.test:
-        print("Testing")
+        printf("Testing")
         evaluate_model(args, model, test_set)
 
 
@@ -355,10 +392,11 @@ def main():
     os.environ["TOKENIZERS_PARALLELISM"] = "true"
     parser = build_parser()
     args = parser.parse_args()
+
     config = args.config
     args = vars(args)
     if config:
-        print("Loading configuration file", config)
+        printf("Loading configuration file", config)
         with open(config, "r") as file:
             config = yaml.load(file, yaml_loader)
         # Prefer parameters from command-line arguments

mlx_lm/models/Klear.py

@@ -0,0 +1,263 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    attention_bias: bool
+    mlp_only_layers: List[int]
+    num_experts: int
+    num_experts_per_tok: int
+    decoder_sparse_step: int
+    n_shared_experts: int
+    moe_intermediate_size: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    rope_theta: float
+    max_position_embeddings: int
+    norm_topk_prob: bool
+
+
+class KlearAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+
+        self.head_dim = args.hidden_size // args.num_attention_heads
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size,
+            self.num_attention_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.o_proj = nn.Linear(
+            self.num_attention_heads * self.head_dim,
+            args.hidden_size,
+            bias=args.attention_bias,
+        )
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.rope = nn.RoPE(
+            self.head_dim,
+            traditional=False,
+            base=args.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = self.q_norm(
+            queries.reshape(B, L, self.num_attention_heads, -1)
+        ).transpose(0, 2, 1, 3)
+        keys = self.k_norm(keys.reshape(B, L, self.num_key_value_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class KlearMLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class KlearSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.norm_topk_prob = args.norm_topk_prob
+        self.num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+
+        self.gate = nn.Linear(args.hidden_size, args.num_experts, bias=False)
+        self.experts = SwitchGLU(
+            args.hidden_size, args.moe_intermediate_size, args.num_experts
+        )
+        self.shared_experts = KlearMLP(
+            args.hidden_size,
+            hidden_dim=args.moe_intermediate_size * args.n_shared_experts,
+        )
+        self.coefficient = nn.Linear(args.hidden_size, 2)
+        self.expert_bias = mx.zeros((self.num_experts,), dtype=mx.float32)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        routing_weights = mx.sigmoid(self.gate(x).astype(mx.float32))
+        biased_weights = routing_weights + self.expert_bias.reshape((1, 1, -1))
+        k = self.top_k
+        inds = mx.argpartition(-biased_weights, kth=k - 1, axis=-1)[..., :k]
+        scores = mx.take_along_axis(routing_weights, inds, axis=-1)
+        if self.norm_topk_prob:
+            scores = scores / mx.sum(scores, axis=-1, keepdims=True)
+        scores = scores.astype(x.dtype)
+        expert_out = self.experts(x, inds)
+        y_experts = (expert_out * scores[..., None]).sum(axis=-2)
+        coef = mx.softmax(self.coefficient(x), axis=-1, precise=True)
+        shared = self.shared_experts(x)
+        y = y_experts * coef[..., :1] + shared * coef[..., 1:]
+        return y
+
+
+class KlearDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = KlearAttention(args)
+
+        if (layer_idx not in args.mlp_only_layers) and (
+            args.num_experts > 0 and (layer_idx + 1) % args.decoder_sparse_step == 0
+        ):
+            self.mlp = KlearSparseMoeBlock(args)
+        else:
+            self.mlp = KlearMLP(args.hidden_size, args.intermediate_size)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class KlearModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            KlearDecoderLayer(args=args, layer_idx=i)
+            for i in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = KlearModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+    def sanitize(self, weights):
+        if "model.layers.0.mlp.experts.0.gate_proj.weight" not in weights:
+            return weights
+
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}.mlp.experts"
+            for name in ["gate_proj", "up_proj", "down_proj"]:
+                stacked = [
+                    weights.pop(f"{prefix}.{e}.{name}.weight")
+                    for e in range(self.args.num_experts)
+                ]
+                weights[f"{prefix}.{name}.weight"] = mx.stack(stacked)
+
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("mlp.gate"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "expert_bias" not in k
+
+        return predicate

mlx_lm/models/activations.py

@@ -0,0 +1,43 @@
+# Copyright © 2023-2026 Apple Inc.
+
+from functools import partial
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+@partial(mx.compile, shapeless=True)
+def swiglu(gate, x):
+    return nn.silu(gate) * x
+
+
+@partial(mx.compile, shapeless=True)
+def xielu(x, alpha_p, alpha_n, beta, eps):
+    alpha_p = nn.softplus(alpha_p)
+    alpha_n = beta + nn.softplus(alpha_n)
+    return mx.where(
+        x > 0,
+        alpha_p * mx.square(x) + beta * x,
+        (mx.expm1(mx.minimum(x, eps)) - x) * alpha_n + beta * x,
+    )
+
+
+class XieLU(nn.Module):
+    def __init__(
+        self,
+        alpha_p_init=0.8,
+        alpha_n_init=0.8,
+        beta=0.5,
+        eps=-1e-6,
+    ):
+        super().__init__()
+        alpha_p_tensor = mx.array(alpha_p_init)
+        alpha_n_tensor = mx.array(alpha_n_init - beta)
+        self.alpha_p = mx.log(mx.exp(alpha_p_tensor) - 1)
+        self.alpha_n = mx.log(mx.exp(alpha_n_tensor) - 1)
+
+        self.beta = mx.array(beta)
+        self.eps = mx.array(eps)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return xielu(x, self.alpha_p, self.alpha_n, self.beta, self.eps)

mlx_lm/models/afm7.py

@@ -9,6 +9,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import ConcatenateKVCache, KVCache
 from .rope_utils import initialize_rope
@@ -50,7 +51,7 @@ class FusedLoRALinear(nn.Module):
         ]
         self.lora_b = [mx.zeros((r, od)) for od in output_dims]
 
-    def fuse(self, de_quantize: bool = False):
+    def fuse(self, dequantize: bool = False):
         linear = self.linear
         weight = linear.weight
         is_quantized = isinstance(linear, FusedQuantizedLinear)
@@ -79,7 +80,7 @@ class FusedLoRALinear(nn.Module):
         delta = mx.concatenate(deltas, axis=0)
         fused_linear.weight = weight + delta
 
-        if is_quantized and not de_quantize:
+        if is_quantized and not dequantize:
             fused_linear = fused_linear.to_quantized(linear.group_size, linear.bits)
 
         return fused_linear
@@ -262,11 +263,6 @@ class KVReuseAttention(nn.Module):
         return self.out_proj(output)
 
 
-@partial(mx.compile, shapeless=True)
-def _swiglu(g, x):
-    return nn.silu(g) * x
-
-
 class MLP(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
@@ -281,7 +277,7 @@ class MLP(nn.Module):
     def __call__(self, x) -> mx.array:
         g = self.gate_proj(x)
         x = self.up_proj(x)
-        return self.down_proj(_swiglu(g, x))
+        return self.down_proj(swiglu(g, x))
 
 
 class TransformerBlock(nn.Module):
@@ -350,18 +346,16 @@ class AFMModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embedding(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
             cache[-1] = ConcatenateKVCache()
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -382,10 +376,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.embedding.as_linear(out)
         return out
 

mlx_lm/models/afmoe.py

@@ -0,0 +1,405 @@
+# Copyright © 2024 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    layer_types: List[str]
+    vocab_size: int = 200192
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    moe_intermediate_size: int = 1024
+    num_hidden_layers: int = 32
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 4
+    head_dim: int = 64
+    max_position_embeddings: int = 131072
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 10000
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+    # MoE config
+    num_experts: int = 128
+    num_experts_per_tok: int = 8
+    num_shared_experts: int = 1
+    num_dense_layers: int = 2
+    route_norm: bool = True
+    route_scale: float = 2.826
+    score_func: str = "sigmoid"
+    n_group: int = 1
+    topk_group: int = 1
+    sliding_window: int = 2048
+    mup_enabled: bool = True
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, is_local_attention: bool = False):
+        super().__init__()
+
+        self.hidden_size = args.hidden_size
+        self.n_heads = args.num_attention_heads
+        self.n_kv_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim
+        self.is_local_attention = is_local_attention
+
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.n_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.n_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.gate_proj = nn.Linear(
+            self.hidden_size, self.n_heads * self.head_dim, bias=False
+        )
+
+        if is_local_attention:
+            self.rope = initialize_rope(
+                self.head_dim,
+                args.rope_theta,
+                False,  # traditional
+                args.rope_scaling,
+                args.max_position_embeddings,
+            )
+        else:
+            self.rope = None
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        keys = keys.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+
+        queries = self.q_norm(queries)
+        keys = self.k_norm(keys)
+
+        if self.is_local_attention and self.rope is not None:
+            if cache is not None:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+            else:
+                queries = self.rope(queries)
+                keys = self.rope(keys)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        gate = mx.sigmoid(self.gate_proj(x))
+        output = output * gate
+
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs, intermediate_size: Optional[int] = None):
+        super().__init__()
+
+        dim = args.hidden_size
+        hidden_dim = (
+            intermediate_size
+            if intermediate_size is not None
+            else args.intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class MoERouter(nn.Module):
+    """Router module that wraps the gate for proper weight naming."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.gate = nn.Linear(args.hidden_size, args.num_experts, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.gate(x)
+
+
+class AfmoeMoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.num_experts = args.num_experts
+        self.num_experts_per_tok = args.num_experts_per_tok
+        self.route_norm = args.route_norm
+        self.route_scale = args.route_scale
+        self.score_func = args.score_func
+        self.n_group = args.n_group
+        self.topk_group = args.topk_group
+
+        self.router = MoERouter(args)
+
+        self.expert_bias = mx.zeros((args.num_experts,))
+
+        self.experts = SwitchGLU(
+            args.hidden_size,
+            args.moe_intermediate_size,
+            args.num_experts,
+        )
+
+        if args.num_shared_experts > 0:
+            shared_intermediate_size = (
+                args.moe_intermediate_size * args.num_shared_experts
+            )
+            self.shared_experts = MLP(args, intermediate_size=shared_intermediate_size)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        gates = self.router(x)
+
+        if self.score_func == "sigmoid":
+            scores = mx.sigmoid(gates.astype(mx.float32))
+        else:
+            scores = mx.softmax(gates.astype(mx.float32), axis=-1)
+
+        # Add expert bias for selection
+        selection_scores = scores + self.expert_bias
+
+        # Group-based expert selection if n_group > 1
+        if self.n_group > 1:
+            selection_scores = mx.unflatten(
+                selection_scores, axis=-1, shape=(self.n_group, -1)
+            )
+            group_scores = mx.topk(selection_scores, 2, axis=-1).sum(
+                axis=-1, keepdims=True
+            )
+            k = self.n_group - self.topk_group
+            group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+            selection_scores = mx.put_along_axis(
+                selection_scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+            )
+            selection_scores = mx.flatten(selection_scores, -2, -1)
+
+        # Select top-k experts
+        k = self.num_experts_per_tok
+        inds = mx.argpartition(-selection_scores, kth=k - 1, axis=-1)[..., :k]
+
+        selected_scores = mx.take_along_axis(scores, inds, axis=-1)
+
+        if self.route_norm and self.num_experts_per_tok > 1:
+            denominator = selected_scores.sum(axis=-1, keepdims=True)
+            selected_scores = selected_scores / denominator
+
+        selected_scores = selected_scores * self.route_scale
+
+        y = self.experts(x, inds)
+        y = (y * selected_scores[..., None]).sum(axis=-2).astype(y.dtype)
+
+        if self.args.num_shared_experts > 0:
+            y = y + self.shared_experts(x)
+
+        return y
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int, use_sliding: bool = False):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.use_sliding = use_sliding
+        self.layer_idx = layer_idx
+
+        self.self_attn = Attention(args, is_local_attention=use_sliding)
+
+        if layer_idx < args.num_dense_layers:
+            self.mlp = MLP(args)
+        else:
+            self.mlp = AfmoeMoE(args)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.pre_mlp_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_mlp_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        r = self.post_attention_layernorm(r)
+        h = x + r
+
+        r = self.mlp(self.pre_mlp_layernorm(h))
+        r = self.post_mlp_layernorm(r)
+        return h + r
+
+
+class AfmoeModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.layer_types = args.layer_types
+        self.sliding_window = args.sliding_window
+        self.mup_enabled = args.mup_enabled
+        self.hidden_size = args.hidden_size
+
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            DecoderLayer(
+                args=args, layer_idx=idx, use_sliding=layer_type == "sliding_attention"
+            )
+            for idx, layer_type in enumerate(self.layer_types)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+        self.fa_idx = self.layer_types.index("full_attention")
+        self.swa_idx = None
+        for idx, layer in enumerate(self.layers):
+            if layer.use_sliding:
+                self.swa_idx = idx
+                break
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if self.mup_enabled:
+            h = h * math.sqrt(self.hidden_size)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        fa_mask = create_attention_mask(h, cache[self.fa_idx])
+        swa_mask = None
+        if self.swa_idx is not None:
+            swa_mask = create_attention_mask(
+                h, cache[self.swa_idx], window_size=self.sliding_window
+            )
+
+        for layer, c in zip(self.layers, cache):
+            mask = swa_mask if layer.use_sliding else fa_mask
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = AfmoeModel(args)
+
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        # Remove unused precomputed rotary freqs
+        weights = {k: v for k, v in weights.items() if "rotary_emb.inv_freq" not in k}
+
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        # Stack experts weights for SwitchGLU
+        for l in range(self.args.num_hidden_layers):
+            if l < self.args.num_dense_layers:
+                continue
+            prefix = f"model.layers.{l}"
+            for n in ["up_proj", "down_proj", "gate_proj"]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{n}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{n}.{k}")
+                            for e in range(self.args.num_experts)
+                        ]
+                        weights[f"{prefix}.mlp.experts.{n}.{k}"] = mx.stack(to_join)
+
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        return [
+            (
+                RotatingKVCache(max_size=self.model.sliding_window)
+                if layer.use_sliding
+                else KVCache()
+            )
+            for layer in self.layers
+        ]
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "expert_bias" not in k
+
+        return predicate
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if "router.gate" in path:
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate

mlx_lm/models/apertus.py

@@ -0,0 +1,195 @@
+# Copyright © 2023-2025 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import XieLU
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    mlp_bias: bool
+    num_attention_heads: int
+    attention_bias: bool
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    max_position_embeddings: int
+    rope_theta: float
+    post_norm: bool
+    qk_norm: bool
+    tie_word_embeddings: bool
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+
+class ApertusMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.up_proj = nn.Linear(
+            args.hidden_size, args.intermediate_size, bias=args.mlp_bias
+        )
+        self.down_proj = nn.Linear(
+            args.intermediate_size, args.hidden_size, bias=args.mlp_bias
+        )
+        self.act_fn = XieLU()
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(self.act_fn(self.up_proj(x)))
+
+
+class ApertusAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+
+        self.head_dim = args.hidden_size // args.num_attention_heads
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size, args.num_attention_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size, args.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size, args.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            args.num_attention_heads * self.head_dim, args.hidden_size, bias=False
+        )
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            args.rope_traditional,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+        queries = self.q_norm(
+            queries.reshape(B, L, self.num_attention_heads, -1)
+        ).transpose(0, 2, 1, 3)
+        keys = self.k_norm(keys.reshape(B, L, self.num_key_value_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class ApertusDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = ApertusAttention(args)
+        self.mlp = ApertusMLP(args)
+
+        self.attention_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.feedforward_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = x + self.self_attn(self.attention_layernorm(x), mask, cache)
+        out = h + self.mlp(self.feedforward_layernorm(h))
+        return out
+
+
+class ApertusModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            ApertusDecoderLayer(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask=mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = ApertusModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        for k, v in weights.items():
+            if k.endswith("alpha_p") or k.endswith("alpha_n"):
+                weights[k] = v.squeeze()
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers

mlx_lm/models/baichuan_m1.py

@@ -6,8 +6,9 @@ from typing import Any, List, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
-from .cache import CacheList, KVCache, MambaCache, RotatingKVCache
+from .cache import ArraysCache, CacheList, KVCache, RotatingKVCache
 
 
 @dataclass
@@ -96,7 +97,10 @@ class Attention(nn.Module):
         k = k.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
         v = v.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
 
-        if cache is not None:
+        if cache is None:
+            cache = (None, None)
+
+        if cache[0] is not None:
             offset = cache[1].offset
             last_k, last_v = cache[0][0], cache[0][1]
         else:
@@ -110,7 +114,7 @@ class Attention(nn.Module):
         q = self.rope(q, offset=offset)
         k = self.rope(k, offset=offset)
 
-        if cache is not None:
+        if cache[0] is not None:
             k, v = cache[1].update_and_fetch(k, v)
             if L > 0:
                 cache[0][0] = k_init[:, :, -1:, :]
@@ -137,7 +141,7 @@ class MLP(nn.Module):
         )
 
     def __call__(self, x: mx.array) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class DecoderLayer(nn.Module):
@@ -167,17 +171,40 @@ class BaichuanModel(nn.Module):
         self.layers = [DecoderLayer(config, i) for i in range(config.num_hidden_layers)]
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-    def __call__(
-        self, inputs: mx.array, mask: mx.array = None, cache: Any = None
-    ) -> mx.array:
+        self.sliding_window = config.sliding_window
+        self.first_swa_idx = None
+        if config.sliding_window_layers:
+            self.first_swa_idx = config.sliding_window_layers[0]
+
+        self.first_global_idx = None
+        self.swa_layers = set(config.sliding_window_layers)
+        for i in range(config.num_hidden_layers):
+            if i in self.swa_layers:
+                continue
+            self.first_global_idx = i
+            break
+
+    def __call__(self, inputs: mx.array, cache: Any = None) -> mx.array:
         x = self.embed_tokens(inputs)
-        if mask is None:
-            if cache is not None:
-                c = [cache[0][1]]
-            mask = create_attention_mask(x, c)
+
         if cache is None:
-            cache = [None] * len(self.layers)
-        for layer, c in zip(self.layers, cache):
+            cache = [(None, None)] * len(self.layers)
+
+        if self.first_global_idx is None:
+            c_global = None
+        else:
+            c_global = cache[self.first_global_idx][1]
+
+        if self.first_swa_idx is None:
+            c_swa = None
+        else:
+            c_swa = cache[self.first_swa_idx][1]
+
+        global_mask = create_attention_mask(x, c_global)
+        swa_mask = create_attention_mask(x, c_swa, window_size=self.sliding_window)
+
+        for l, (layer, c) in enumerate(zip(self.layers, cache)):
+            mask = swa_mask if l in self.swa_layers else global_mask
             x = layer(x, mask, c)
         return self.norm(x)
 
@@ -196,7 +223,7 @@ class Model(nn.Module):
         caches = []
         for i, layer in enumerate(self.model.layers):
             is_swa = i in self.config.sliding_window_layers
-            conv_cache = MambaCache()
+            conv_cache = ArraysCache(size=2)
             if is_swa:
                 kv_cache = RotatingKVCache(max_size=self.config.sliding_window)
             else:
@@ -215,10 +242,8 @@ class Model(nn.Module):
             weights["lm_head.weight"] = w
         return weights
 
-    def __call__(
-        self, inputs: mx.array, mask: mx.array = None, cache: Any = None
-    ) -> mx.array:
-        outputs = self.model(inputs, mask, cache)
+    def __call__(self, inputs: mx.array, cache: Any = None) -> mx.array:
+        outputs = self.model(inputs, cache)
         return self.lm_head(outputs)
 
     @property

mlx_lm/models/bailing_moe.py

@@ -1,11 +1,13 @@
 # Copyright © 2025 Apple Inc.
 
 from dataclasses import dataclass
+from functools import partial
 from typing import Any, Dict, Optional, Union
 
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 from .switch_layers import SwitchGLU
@@ -34,6 +36,25 @@ class ModelArgs(BaseModelArgs):
     use_qkv_bias: bool = False
     norm_head: bool = False
     norm_softmax: bool = False
+    use_qk_norm: bool = False
+    tie_word_embeddings: bool = False
+    partial_rotary_factor: float = 1.0
+    rotary_dim: Optional[int] = None
+    moe_router_enable_expert_bias: bool = False
+    moe_router_enable_routed_scaling: bool = True
+    routed_scaling_factor: float = 1.0
+    score_function: str = "softmax"
+    n_group: int = 1
+    topk_group: int = 4
+    moe_shared_expert_intermediate_size: Optional[int] = None
+    moe_router_enable_shared_expert: bool = True
+
+
+@partial(mx.compile, shapeless=True)
+def aggregate_expert_outputs(expert_outputs, scores):
+    return (
+        (expert_outputs * scores[..., None]).sum(axis=-2).astype(expert_outputs.dtype)
+    )
 
 
 class BailingMoeMLP(nn.Module):
@@ -56,12 +77,13 @@ class BailingMoeMLP(nn.Module):
         )
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class BailingMoeAttention(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
+        self.use_qk_norm = args.use_qk_norm
         self.num_attention_heads = args.num_attention_heads
         self.num_key_value_heads = args.num_key_value_heads
         self.head_dim = args.hidden_size // self.num_attention_heads
@@ -78,8 +100,14 @@ class BailingMoeAttention(nn.Module):
             bias=args.use_bias,
         )
 
+        if args.use_qk_norm:
+            self.key_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+            self.query_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        if (rope_dim := args.rotary_dim) is None:
+            rope_dim = int(self.head_dim * args.partial_rotary_factor)
         self.rope = initialize_rope(
-            self.head_dim,
+            rope_dim,
             args.rope_theta,
             traditional=False,
             scaling_config=args.rope_scaling,
@@ -110,6 +138,10 @@ class BailingMoeAttention(nn.Module):
             0, 2, 1, 3
         )
 
+        if self.use_qk_norm:
+            queries = self.query_layernorm(queries)
+            keys = self.key_layernorm(keys)
+
         if cache is not None:
             queries = self.rope(queries, offset=cache.offset)
             keys = self.rope(keys, offset=cache.offset)
@@ -126,32 +158,77 @@ class BailingMoeAttention(nn.Module):
         return self.dense(output)
 
 
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+    score_function,
+):
+
+    in_type = gates.dtype
+    if score_function == "sigmoid":
+        scores = mx.sigmoid(gates.astype(mx.float32))
+    else:
+        scores = mx.softmax(gates.astype(mx.float32), axis=-1)
+    orig_scores = scores
+    if e_score_correction_bias is not None:
+        scores = scores + e_score_correction_bias
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0, scores.dtype), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(scores, kth=-k, axis=-1)[..., -k:]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True) + 1e-20
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores.astype(in_type)
+
+
 class BailingMoeGate(nn.Module):
-    def __init__(self, config):
+    def __init__(self, args: ModelArgs):
         super().__init__()
-        self.config = config
-        self.top_k = config.num_experts_per_tok
-        self.num_experts = config.num_experts
-        self.norm_topk_prob = config.norm_topk_prob
-        self.gating_dim = config.hidden_size
+        self.norm_topk_prob = args.norm_topk_prob
 
-        self.gate_proj = nn.Linear(self.gating_dim, self.num_experts, bias=False)
+        self.top_k = args.num_experts_per_tok
+        self.n_group = args.n_group
+        self.topk_group = args.topk_group
+        self.routed_scaling_factor = args.routed_scaling_factor
+        self.enable_routed_scaling = args.moe_router_enable_routed_scaling
 
-    def __call__(self, hidden_states):
-        B, L, D = hidden_states.shape
-        x = hidden_states.reshape(-1, D)
+        self.gate_proj = nn.Linear(args.hidden_size, args.num_experts, bias=False)
+        self.expert_bias = (
+            mx.zeros((args.num_experts,))
+            if args.moe_router_enable_expert_bias
+            else None
+        )
+        self.score_function = args.score_function
 
-        logits = self.gate_proj(x)
-        scores = mx.softmax(logits, axis=-1, precise=True)
-
-        topk_idx = mx.argpartition(scores, kth=-self.top_k, axis=-1)[..., -self.top_k :]
-        topk_scores = mx.take_along_axis(scores, topk_idx, axis=-1)
-
-        if self.top_k > 1 and self.norm_topk_prob:
-            denom = mx.sum(topk_scores, axis=-1, keepdims=True)
-            topk_scores = topk_scores / mx.maximum(denom, 1e-9)
-
-        return topk_idx, topk_scores
+    def __call__(self, x):
+        return group_expert_select(
+            self.gate_proj(x),
+            self.expert_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+            self.score_function,
+        )
 
 
 class BailingMoeSparseMoeBlock(nn.Module):
@@ -159,42 +236,32 @@ class BailingMoeSparseMoeBlock(nn.Module):
         super().__init__()
         self.args = args
         self.num_experts_per_tok = args.num_experts_per_tok
-
         self.switch_mlp = SwitchGLU(
             args.hidden_size,
             args.moe_intermediate_size,
             args.num_experts,
             bias=args.use_bias,
         )
-
-        self.gate = BailingMoeGate(config=args)
-
-        if args.num_shared_experts > 0:
-            self.shared_experts = BailingMoeMLP(
+        self.gate = BailingMoeGate(args)
+        shared_dim = (
+            args.moe_shared_expert_intermediate_size or args.moe_intermediate_size
+        )
+        self.shared_experts = (
+            BailingMoeMLP(
                 args=args,
-                intermediate_size=args.moe_intermediate_size * args.num_shared_experts,
+                intermediate_size=shared_dim * args.num_shared_experts,
             )
-        else:
-            self.shared_experts = None
-
-    def __call__(self, hidden_states):
-        batch_size, seq_len, hidden_dim = hidden_states.shape
+            if args.num_shared_experts > 0 and args.moe_router_enable_shared_expert
+            else None
+        )
 
+    def __call__(self, x):
+        topk_idx, topk_weight = self.gate(x)
+        out = self.switch_mlp(x, topk_idx)
+        out = aggregate_expert_outputs(out, topk_weight)
         if self.shared_experts is not None:
-            identity = hidden_states
-
-        x = hidden_states.reshape(-1, hidden_dim)
-
-        expert_indices, expert_weights = self.gate(hidden_states)
-        expert_outputs = self.switch_mlp(x, expert_indices)
-
-        weighted_output = mx.sum(expert_outputs * expert_weights[..., None], axis=-2)
-        output = weighted_output.reshape(batch_size, seq_len, hidden_dim)
-
-        if self.shared_experts is not None:
-            output = output + self.shared_experts(hidden_states)
-
-        return output
+            out = out + self.shared_experts(x)
+        return out
 
 
 class BailingMoeDecoderLayer(nn.Module):
@@ -239,17 +306,15 @@ class BailingMoeModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ):
         h = self.word_embeddings(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -263,18 +328,25 @@ class Model(nn.Module):
         self.norm_head = args.norm_head
         self.model_type = args.model_type
         self.model = BailingMoeModel(args)
-        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
 
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        h = self.model(inputs, mask, cache)
-        return self.lm_head(h)
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.word_embeddings.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
 
     def sanitize(self, weights):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
         if self.norm_head:
             w = weights["lm_head.weight"]
             dtype = w.dtype
@@ -311,12 +383,19 @@ class Model(nn.Module):
     @property
     def quant_predicate(self):
         def predicate(path, _):
-            if path.endswith("mlp.gate"):
+            if path.endswith("mlp.gate.gate_proj"):
                 return {"group_size": 64, "bits": 8}
             return True
 
         return predicate
 
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "expert_bias" not in k
+
+        return predicate
+
     @property
     def layers(self):
         return self.model.layers

mlx_lm/models/bailing_moe_linear.py

@@ -0,0 +1,595 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, KVCache
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    max_position_embeddings: int
+    moe_intermediate_size: int
+    num_experts: int
+    num_shared_experts: int
+    norm_topk_prob: bool
+    num_attention_heads: int
+    num_experts_per_tok: int
+    num_hidden_layers: int
+    num_key_value_heads: int
+    rms_norm_eps: float
+    rope_theta: float
+    vocab_size: int
+    first_k_dense_replace: int
+    layer_group_size: int
+    group_norm_size: int
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    rope_traditional: bool = False
+    use_bias: bool = False
+    use_qkv_bias: bool = False
+    norm_head: bool = False
+    norm_softmax: bool = False
+    use_qk_norm: bool = False
+    tie_word_embeddings: bool = False
+    partial_rotary_factor: float = 1.0
+    moe_router_enable_expert_bias: bool = False
+    moe_router_enable_routed_scaling: bool = True
+    routed_scaling_factor: float = 1.0
+    score_function: str = "softmax"
+    n_group: int = 1
+    topk_group: int = 4
+    use_rmsnorm: bool = True
+    moe_shared_expert_intermediate_size: Optional[int] = None
+    moe_router_enable_shared_expert: bool = True
+    head_dim: Optional[int] = None
+
+
+def recurrent_gla(
+    q: mx.array,
+    k: mx.array,
+    v: mx.array,
+    g: mx.array,
+    scale: float,
+    h: Optional[mx.array] = None,
+) -> mx.array:
+    """
+    Recurrence per (b, h):
+        h_t = h_{t-1} * exp(g_t)
+        h_t = h_t + k_t^T @ v_t
+        y_t = (q_t @ h_t) * scale
+    Returns y with shape [B, H, T, Dv].
+    """
+    B, Hq, L, K = q.shape
+    Hv = k.shape[1]
+    V = v.shape[-1]
+
+    outputs = []
+    exp_g = mx.exp(g)[:, None, None].astype(q.dtype)
+    q = q * scale
+    for t in range(L):
+        q_t = q[:, :, t : t + 1]
+        k_t = k[:, :, t : t + 1]
+        v_t = v[:, :, t : t + 1]
+        h_up = k_t.transpose(0, 1, 3, 2) @ v_t
+        if h is not None:
+            h = h * exp_g + h_up
+        else:
+            h = h_up
+        o_t = q_t @ h
+        outputs.append(o_t)
+
+    return mx.concatenate(outputs, axis=2), h
+
+
+class GroupRMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-5, groups: int = 1):
+        super().__init__()
+        self.weight = mx.ones((dims,))
+        self.groups = groups
+        self.eps = eps
+
+    def __call__(self, x: mx.array) -> mx.array:
+        shape = x.shape
+        x = mx.unflatten(x, axis=-1, shape=(self.groups, -1))
+        x = mx.fast.rms_norm(x, weight=None, eps=self.eps)
+        return self.weight * mx.flatten(x, -2)
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs, intermediate_size: Optional[int] = None):
+        super().__init__()
+        self.intermediate_size = (
+            intermediate_size
+            if intermediate_size is not None
+            else args.intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(
+            args.hidden_size, self.intermediate_size, bias=args.use_bias
+        )
+        self.down_proj = nn.Linear(
+            self.intermediate_size, args.hidden_size, bias=args.use_bias
+        )
+        self.up_proj = nn.Linear(
+            args.hidden_size, self.intermediate_size, bias=args.use_bias
+        )
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.use_qk_norm = args.use_qk_norm
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim or args.hidden_size // self.num_attention_heads
+        self.scale = self.head_dim**-0.5
+
+        self.query_key_value = nn.Linear(
+            args.hidden_size,
+            (self.num_attention_heads + 2 * self.num_key_value_heads) * self.head_dim,
+            bias=args.use_qkv_bias,
+        )
+        self.dense = nn.Linear(
+            self.num_attention_heads * self.head_dim,
+            args.hidden_size,
+            bias=args.use_bias,
+        )
+
+        if args.use_qk_norm:
+            self.key_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+            self.query_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.rope = initialize_rope(
+            int(self.head_dim * args.partial_rotary_factor),
+            args.rope_theta,
+            traditional=args.rope_traditional,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        qkv = self.query_key_value(x)
+
+        q_size = self.num_attention_heads * self.head_dim
+        kv_size = self.num_key_value_heads * self.head_dim
+        q, k, v = mx.split(qkv, [q_size, q_size + kv_size], axis=-1)
+
+        queries = q.reshape(B, L, self.num_attention_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        keys = k.reshape(B, L, self.num_key_value_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        values = v.reshape(B, L, self.num_key_value_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+
+        if self.use_qk_norm:
+            queries = self.query_layernorm(queries)
+            keys = self.key_layernorm(keys)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.dense(output)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.use_qk_norm = args.use_qk_norm
+        self.num_hidden_layers = args.num_hidden_layers
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_attention_heads
+        self.head_dim = args.hidden_size // self.num_attention_heads
+        self.scale = self.head_dim**-0.5
+        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
+        assert self.num_key_value_groups == 1, "Grouped linear not yet supported."
+
+        self.query_key_value = nn.Linear(
+            args.hidden_size,
+            (self.num_attention_heads + 2 * self.num_key_value_heads) * self.head_dim,
+            bias=args.use_qkv_bias,
+        )
+
+        self.dense = nn.Linear(
+            self.num_attention_heads * self.head_dim,
+            args.hidden_size,
+            bias=args.use_bias,
+        )
+
+        self.g_proj = nn.Linear(
+            args.hidden_size, args.num_attention_heads * self.head_dim, bias=False
+        )
+        self.g_norm = GroupRMSNorm(
+            args.num_attention_heads * self.head_dim,
+            eps=args.rms_norm_eps,
+            groups=args.group_norm_size,
+        )
+
+        if args.use_qk_norm:
+            self.key_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+            self.query_layernorm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.rope = initialize_rope(
+            int(self.head_dim * args.partial_rotary_factor),
+            args.rope_theta,
+            traditional=args.rope_traditional,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+        self._slope = self._get_slopes()
+
+    def _get_slopes(self) -> mx.array:
+        n = self.num_attention_heads
+
+        def power_of_2_slopes(n):
+            return [2 ** (-(2 ** -(math.log2(n) - 3)) * (i + 1)) for i in range(n)]
+
+        if math.log2(n).is_integer():
+            slopes = power_of_2_slopes(n)
+        else:
+            p = 2 ** math.floor(math.log2(n))
+            slopes = power_of_2_slopes(p) + power_of_2_slopes(2 * p)[::2][: n - p]
+
+        slopes = mx.array(slopes, dtype=mx.float32)
+        denom = max(1, self.num_hidden_layers - 1)
+        layer_pos = max(0, self.layer_idx - 1)
+        layer_factor = 1 - (layer_pos / denom) + 1e-5
+        return -slopes * layer_factor
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        offset: int = 0,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        qkv = self.query_key_value(x)
+        qkv_mix = qkv.reshape(
+            B,
+            L,
+            (self.num_attention_heads + 2 * self.num_key_value_heads),
+            self.head_dim,
+        )
+        q, k, v = mx.split(
+            qkv_mix,
+            [
+                self.num_attention_heads,
+                self.num_attention_heads + self.num_key_value_heads,
+            ],
+            axis=2,
+        )
+
+        queries = q.transpose(0, 2, 1, 3)
+        keys = k.transpose(0, 2, 1, 3)
+        values = v.transpose(0, 2, 1, 3)
+
+        if self.use_qk_norm:
+            queries = self.query_layernorm(queries)
+            keys = self.key_layernorm(keys)
+
+        queries = self.rope(queries, offset=offset)
+        keys = self.rope(keys, offset=offset)
+
+        if cache is None:
+            cache = [None]
+        output, cache[0] = recurrent_gla(
+            q=queries,
+            k=keys,
+            v=values,
+            g=self._slope,
+            scale=self.scale,
+            h=cache[0],
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        output = self.g_norm(output) * mx.sigmoid(self.g_proj(x))
+        return self.dense(output)
+
+
+def group_expert_select(
+    gates: mx.array,
+    e_score_correction_bias: mx.array,
+    top_k: int,
+    n_group: int,
+    topk_group: int,
+    routed_scaling_factor: float,
+    norm_topk_prob: bool,
+    score_function: str,
+) -> Tuple[mx.array, mx.array]:
+    in_type = gates.dtype
+    if score_function == "sigmoid":
+        scores = mx.sigmoid(gates.astype(mx.float32))
+    else:
+        scores = mx.softmax(gates.astype(mx.float32), axis=-1)
+    orig_scores = scores
+    if e_score_correction_bias is not None:
+        scores = scores + e_score_correction_bias
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores.astype(in_type)
+
+
+class Gate(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.norm_topk_prob = args.norm_topk_prob
+
+        self.top_k = args.num_experts_per_tok
+        self.n_group = args.n_group
+        self.topk_group = args.topk_group
+        self.routed_scaling_factor = args.routed_scaling_factor
+        self.enable_routed_scaling = args.moe_router_enable_routed_scaling
+
+        self.gate_proj = nn.Linear(args.hidden_size, args.num_experts, bias=False)
+        self.expert_bias = (
+            mx.zeros((args.num_experts,))
+            if args.moe_router_enable_expert_bias
+            else None
+        )
+        self.score_function = args.score_function
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return group_expert_select(
+            self.gate_proj(x),
+            self.expert_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+            self.score_function,
+        )
+
+
+class SparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.num_experts_per_tok = args.num_experts_per_tok
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size,
+            args.moe_intermediate_size,
+            args.num_experts,
+            bias=args.use_bias,
+        )
+        self.gate = Gate(args)
+        shared_dim = (
+            args.moe_shared_expert_intermediate_size or args.moe_intermediate_size
+        )
+        self.shared_experts = (
+            MLP(
+                args=args,
+                intermediate_size=shared_dim * args.num_shared_experts,
+            )
+            if args.num_shared_experts > 0 and args.moe_router_enable_shared_expert
+            else None
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        topk_idx, topk_weight = self.gate(x)
+        out = self.switch_mlp(x, topk_idx)
+        out = (out * topk_weight[..., None]).sum(axis=-2)
+        if self.shared_experts is not None:
+            out = out + self.shared_experts(x)
+        return out
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.is_global = (
+            (layer_idx + 1) % args.layer_group_size == 0
+            or layer_idx
+            >= args.num_hidden_layers // args.layer_group_size * args.layer_group_size
+        )
+
+        if self.is_global:
+            self.attention = Attention(args)
+        else:
+            self.attention = LinearAttention(args, layer_idx=layer_idx)
+
+        self.mlp = (
+            SparseMoeBlock(args)
+            if (
+                args.num_experts is not None and layer_idx >= args.first_k_dense_replace
+            )
+            else MLP(args)
+        )
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        offset: int = 0,
+    ) -> mx.array:
+        if self.is_global:
+            r = self.attention(self.input_layernorm(x), mask, cache)
+        else:
+            r = self.attention(self.input_layernorm(x), mask, cache, offset=offset)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            DecoderLayer(args, layer_idx=i) for i in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.gla_idx = 0
+        self.attn_idx = args.layer_group_size - 1
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.word_embeddings(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        offset = 0
+        attn_mask = create_attention_mask(h, cache[self.attn_idx])
+        gla_mask = create_ssm_mask(h, cache[self.gla_idx])
+        if cache[self.attn_idx] is not None:
+            offset = cache[self.attn_idx].offset
+
+        for layer, c in zip(self.layers, cache):
+            mask = attn_mask if layer.is_global else gla_mask
+            h = layer(h, mask, c, offset=offset)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.norm_head = args.norm_head
+        self.model_type = args.model_type
+        self.model = LanguageModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.word_embeddings.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        if self.norm_head:
+            w = weights["lm_head.weight"]
+            dtype = w.dtype
+            weight_norm = (
+                mx.linalg.norm(w.astype(mx.float32), axis=0, keepdims=True) + 1e-7
+            )
+            weights["lm_head.weight"] = (w / weight_norm).astype(dtype)
+
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            # Handle MoE layers
+            if l >= self.args.first_k_dense_replace:
+                for m in ["gate_proj", "down_proj", "up_proj"]:
+                    for k in ["weight", "scales", "biases"]:
+                        if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                            to_join = [
+                                weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                                for e in range(self.args.num_experts)
+                            ]
+                            weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(
+                                to_join
+                            )
+
+                if f"{prefix}.mlp.gate.weight" in weights:
+                    gate_weight = weights.pop(f"{prefix}.mlp.gate.weight")
+                    weights[f"{prefix}.mlp.gate.gate_proj.weight"] = gate_weight
+
+                if f"{prefix}.mlp.gate.bias" in weights:
+                    gate_bias = weights.pop(f"{prefix}.mlp.gate.bias")
+                    weights[f"{prefix}.mlp.gate.gate_proj.bias"] = gate_bias
+
+        return weights
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("mlp.gate.gate_proj"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "expert_bias" not in k
+
+        return predicate
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        caches = []
+        for l in self.layers:
+            if l.is_global:
+                caches.append(KVCache())
+            else:
+                caches.append(ArraysCache(size=1))
+        return caches

mlx_lm/models/base.py

@@ -7,8 +7,6 @@ from typing import Any, Optional
 import mlx.core as mx
 from mlx.utils import tree_map
 
-from .cache import QuantizedKVCache
-
 
 @dataclass
 class BaseModelArgs:
@@ -27,7 +25,8 @@ def create_causal_mask(
     N: int,
     offset: int = 0,
     window_size: Optional[int] = None,
-    lengths: Optional[mx.array] = None,
+    right_padding: Optional[mx.array] = None,
+    left_padding: Optional[mx.array] = None,
 ):
     rinds = mx.arange(offset + N)
     linds = mx.arange(offset, offset + N) if offset else rinds
@@ -36,33 +35,30 @@ def create_causal_mask(
     mask = linds >= rinds
     if window_size is not None:
         mask = mask & (linds < rinds + window_size)
-    if lengths is not None:
-        lengths = lengths[:, None, None, None]
-        mask = mask & (rinds < lengths)
+    if right_padding is not None:
+        mask = mask & (rinds < mx.expand_dims((offset + N) - right_padding, (1, 2, 3)))
+    if left_padding is not None:
+        mask = mask & (mx.expand_dims(left_padding, (1, 2, 3)) <= rinds)
     return mask
 
 
 def create_attention_mask(
-    h: mx.array, cache: Optional[Any] = None, return_array: bool = False
+    h, cache=None, window_size: Optional[int] = None, return_array: bool = False
 ):
-    T = h.shape[1]
-    if T > 1:
-        offset = 0
-        window_size = None
-        if cache is not None and cache[0] is not None:
-            c = cache[0]
-            offset = c.offset
-            if hasattr(c, "max_size"):
-                window_size = c.max_size
-                offset = min(window_size, offset)
-                return_array = return_array or offset + T > window_size
-        if return_array:
-            return create_causal_mask(T, offset, window_size=window_size)
-        else:
-            return "causal"
-    else:
-        mask = None
-    return mask
+    N = h.shape[1]
+    if cache and hasattr(cache, "make_mask"):
+        return cache.make_mask(N, return_array=return_array, window_size=window_size)
+    if N == 1:
+        return None
+    if return_array or (window_size and N > window_size):
+        return create_causal_mask(N, window_size=window_size)
+    return "causal"
+
+
+def create_ssm_mask(h, cache=None):
+    if cache and hasattr(cache, "make_mask"):
+        return cache.make_mask(h.shape[1])
+    return None
 
 
 def quantized_scaled_dot_product_attention(
@@ -116,8 +112,11 @@ def scaled_dot_product_attention(
     cache,
     scale: float,
     mask: Optional[mx.array],
+    sinks: Optional[mx.array] = None,
 ) -> mx.array:
-    if isinstance(cache, QuantizedKVCache):
+    if hasattr(cache, "bits"):
+        if sinks is not None:
+            raise ValueError("Quantized SDPA does not support attention sinks.")
         return quantized_scaled_dot_product_attention(
             queries,
             keys,
@@ -129,5 +128,10 @@ def scaled_dot_product_attention(
         )
     else:
         return mx.fast.scaled_dot_product_attention(
-            queries, keys, values, scale=scale, mask=mask
+            queries,
+            keys,
+            values,
+            scale=scale,
+            mask=mask,
+            sinks=sinks,
         )

mlx_lm/models/bitnet.py

@@ -93,11 +93,6 @@ class Attention(nn.Module):
         return output
 
 
-@partial(mx.compile, shapeless=True)
-def relu2(x):
-    return mx.square(nn.relu(x))
-
-
 class MLP(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
@@ -116,7 +111,7 @@ class MLP(nn.Module):
         self.ffn_sub_norm = nn.RMSNorm(args.intermediate_size, eps=args.rms_norm_eps)
 
     def __call__(self, x) -> mx.array:
-        x = relu2(self.gate_proj(x)) * self.up_proj(x)
+        x = nn.relu2(self.gate_proj(x)) * self.up_proj(x)
         x = self.ffn_sub_norm(x)
         x = self.down_proj(x)
         return x
@@ -163,17 +158,15 @@ class LlamaModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -192,10 +185,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/cohere.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -109,7 +110,7 @@ class MLP(nn.Module):
         self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
 
     def __call__(self, x):
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -155,17 +156,15 @@ class CohereModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -182,10 +181,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.embed_tokens.as_linear(out)
         out = out * self.model.args.logit_scale
         return out

mlx_lm/models/cohere2.py

@@ -6,6 +6,7 @@ from typing import Optional, Tuple
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import KVCache, RotatingKVCache
 
@@ -83,11 +84,6 @@ class Attention(nn.Module):
         if cache is not None:
             keys, values = cache.update_and_fetch(keys, values)
 
-        if self.use_sliding_window and isinstance(mask, mx.array):
-            key_len = keys.shape[-2]
-            if mask.shape[-1] != key_len:
-                mask = mask[..., -key_len:]
-
         # TODO: maybe remove cast once fused mask is supported since attention
         # may be in higher precision
         sdpa_type = mx.float32 if queries.dtype == mx.float16 else queries.dtype
@@ -111,7 +107,7 @@ class MLP(nn.Module):
         self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
 
     def __call__(self, x):
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -148,6 +144,7 @@ class CohereModel(nn.Module):
         self.vocab_size = args.vocab_size
         self.num_hidden_layers = args.num_hidden_layers
         assert self.vocab_size > 0
+        self.window_size = args.sliding_window
         self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
         self.layers = [
             TransformerBlock(args=args, layer_idx=i)
@@ -160,7 +157,6 @@ class CohereModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
@@ -168,10 +164,9 @@ class CohereModel(nn.Module):
         if cache is None:
             cache = [None] * len(self.layers)
 
-        if mask is None:
-            j = self.args.sliding_window_pattern
-            full_mask = create_attention_mask(h, cache[j - 1 : j])
-            sliding_window_mask = create_attention_mask(h, cache)
+        j = self.args.sliding_window_pattern
+        full_mask = create_attention_mask(h, cache[j - 1])
+        swa_mask = create_attention_mask(h, cache[0], window_size=self.window_size)
 
         for i, (layer, c) in enumerate(zip(self.layers, cache)):
             is_global = (
@@ -179,13 +174,9 @@ class CohereModel(nn.Module):
                 == self.args.sliding_window_pattern - 1
             )
 
-            local_mask = mask
-            if mask is None and is_global:
-                local_mask = full_mask
-            elif mask is None:
-                local_mask = sliding_window_mask
+            mask = full_mask if is_global else swa_mask
 
-            h = layer(h, local_mask, c)
+            h = layer(h, mask, c)
 
         return self.norm(h)
 
@@ -200,10 +191,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.embed_tokens.as_linear(out)
         out = out * self.model.args.logit_scale
         return out

mlx_lm/models/dbrx.py

@@ -7,6 +7,7 @@ import mlx.core as mx
 import mlx.nn as nn
 import numpy as np
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -107,7 +108,7 @@ class MLP(nn.Module):
         self.w2 = nn.Linear(ffn_dim, d_model, bias=False)
 
     def __call__(self, x: mx.array) -> mx.array:
-        current_hidden_states = nn.silu(self.w1(x)) * self.v1(x)
+        current_hidden_states = swiglu(self.w1(x), self.v1(x))
         current_hidden_states = self.w2(current_hidden_states)
         return current_hidden_states
 
@@ -196,17 +197,15 @@ class DBRX(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.wte(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.blocks)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.blocks, cache):
             h = layer(h, mask, c)
 
@@ -224,10 +223,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.transformer(inputs, mask, cache)
+        out = self.transformer(inputs, cache)
         return self.lm_head(out)
 
     @property

mlx_lm/models/deepseek.py

@@ -4,6 +4,7 @@ from typing import Any, Dict, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .switch_layers import SwitchGLU
 
@@ -120,7 +121,7 @@ class DeepseekMLP(nn.Module):
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
 
     def __call__(self, x: mx.array) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class MoEGate(nn.Module):
@@ -210,15 +211,14 @@ class DeepseekModel(nn.Module):
         self,
         x: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ) -> mx.array:
         h = self.embed_tokens(x)
-        if mask is None:
-            mask = create_attention_mask(h, cache)
 
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -237,9 +237,8 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, cache, mask)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
     def sanitize(self, weights):

mlx_lm/models/deepseek_v2.py

@@ -6,8 +6,11 @@ from typing import Any, Dict, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .pipeline import PipelineMixin
 from .switch_layers import SwitchGLU
 
 
@@ -258,7 +261,7 @@ class DeepseekV2MLP(nn.Module):
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
 
     def __call__(self, x):
-        down_proj = self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
         return down_proj
 
 
@@ -314,13 +317,21 @@ class DeepseekV2MoE(nn.Module):
                 config=config, intermediate_size=intermediate_size
             )
 
+        self.sharding_group = None
+
     def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
         inds, scores = self.gate(x)
         y = self.switch_mlp(x, inds)
         y = (y * scores[..., None]).sum(axis=-2)
         if self.config.n_shared_experts is not None:
             y = y + self.shared_experts(x)
 
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
         return y
 
 
@@ -355,7 +366,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         return out
 
 
-class DeepseekV2Model(nn.Module):
+class DeepseekV2Model(PipelineMixin, nn.Module):
     def __init__(self, config: ModelArgs):
         super().__init__()
         self.vocab_size = config.vocab_size
@@ -364,61 +375,38 @@ class DeepseekV2Model(nn.Module):
             DeepseekV2DecoderLayer(config, idx)
             for idx in range(config.num_hidden_layers)
         ]
-        self.start_idx = 0
-        self.end_idx = len(self.layers)
-        self.num_layers = self.end_idx
-
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-        self.pipeline_rank = 0
-        self.pipeline_size = 1
-
-    def pipeline(self, group):
-        # Split layers in reverse so rank=0 gets the last layers and
-        # rank=pipeline_size-1 gets the first
-        self.pipeline_rank = group.rank()
-        self.pipeline_size = group.size()
-        layers_per_rank = len(self.layers) // self.pipeline_size
-        extra = len(self.layers) - layers_per_rank * self.pipeline_size
-        if self.pipeline_rank < extra:
-            layers_per_rank += 1
-
-        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
-        self.end_idx = self.start_idx + layers_per_rank
-        self.num_layers = layers_per_rank
-        self.layers = self.layers[: self.end_idx]
-        self.layers[: self.start_idx] = [None] * self.start_idx
-        self.num_layers = len(self.layers) - self.start_idx
-
     def __call__(
         self,
         x: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ) -> mx.array:
         h = self.embed_tokens(x)
 
         pipeline_rank = self.pipeline_rank
         pipeline_size = self.pipeline_size
-        if mask is None:
-            mask = create_attention_mask(h, cache)
 
         if cache is None:
-            cache = [None] * self.num_layers
+            cache = [None] * len(self.pipeline_layers)
+        mask = create_attention_mask(h, cache[0])
 
         # Receive from the previous process in the pipeline
         if pipeline_rank < pipeline_size - 1:
             h = mx.distributed.recv_like(h, (pipeline_rank + 1))
 
-        for i in range(self.num_layers):
-            h = self.layers[self.start_idx + i](h, mask, cache[i])
+        for l, c in zip(self.pipeline_layers, cache):
+            h = l(h, mask, cache=c)
 
         # Send to the next process in the pipeline
         if pipeline_rank != 0:
             h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+            if cache[-1] is not None:
+                cache[-1].keys = mx.depends(cache[-1].keys, h)
 
         # Broadcast h while keeping it in the graph
-        h = mx.distributed.all_gather(h)[: h.shape[0]]
+        if pipeline_size > 1:
+            h = mx.distributed.all_gather(h)[: h.shape[0]]
 
         return self.norm(h)
 
@@ -435,9 +423,8 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, cache, mask)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
     def sanitize(self, weights):
@@ -453,6 +440,62 @@ class Model(nn.Module):
                         weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
         return weights
 
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        for layer in self.model.layers:
+            # Shard the self attention
+            if layer.self_attn.q_lora_rank is None:
+                layer.self_attn.q_proj = shard_linear(
+                    layer.self_attn.q_proj, "all-to-sharded", group=group
+                )
+            else:
+                layer.self_attn.q_b_proj = shard_linear(
+                    layer.self_attn.q_b_proj, "all-to-sharded", group=group
+                )
+            layer.self_attn.kv_b_proj = shard_linear(
+                layer.self_attn.kv_b_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.num_heads //= N
+
+            # Shard the MLP
+            if isinstance(layer.mlp, DeepseekV2MLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                layer.mlp.sharding_group = group
+                shard_inplace(
+                    layer.mlp.shared_experts.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+
     @property
     def layers(self):
-        return self.model.layers[self.model.start_idx : self.model.end_idx]
+        return self.model.pipeline_layers

mlx_lm/models/deepseek_v3.py

@@ -7,8 +7,13 @@ from typing import Any, Dict, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .mla import MultiLinear
+from .pipeline import PipelineMixin
+from .rope_utils import initialize_rope
 from .switch_layers import SwitchGLU
 
 
@@ -33,9 +38,9 @@ class ModelArgs(BaseModelArgs):
     topk_method: str = "noaux_tc"
     scoring_func: str = "sigmoid"
     norm_topk_prob: bool = True
-    n_group: Optional[int] = None
-    topk_group: Optional[int] = None
-    num_experts_per_tok: Optional[int] = None
+    n_group: int = 1
+    topk_group: int = 1
+    num_experts_per_tok: int = 1
     moe_layer_freq: int = 1
     first_k_dense_replace: int = 0
     max_position_embeddings: int = 2048
@@ -45,85 +50,6 @@ class ModelArgs(BaseModelArgs):
     attention_bias: bool = False
 
 
-def yarn_find_correction_dim(
-    num_rotations, dim, base=10000, max_position_embeddings=2048
-):
-    return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (
-        2 * math.log(base)
-    )
-
-
-def yarn_find_correction_range(
-    low_rot, high_rot, dim, base=10000, max_position_embeddings=2048
-):
-    low = math.floor(
-        yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings)
-    )
-    high = math.ceil(
-        yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings)
-    )
-    return max(low, 0), min(high, dim - 1)
-
-
-def yarn_get_mscale(scale=1, mscale=1):
-    if scale <= 1:
-        return 1.0
-    return 0.1 * mscale * math.log(scale) + 1.0
-
-
-def yarn_linear_ramp_mask(min_val, max_val, dim):
-    if min_val == max_val:
-        max_val += 0.001  # Prevent singularity
-
-    linear_func = (mx.arange(dim, dtype=mx.float32) - min_val) / (max_val - min_val)
-    return mx.clip(linear_func, 0, 1)
-
-
-class DeepseekV3YarnRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        dim,
-        max_position_embeddings=2048,
-        base=10000,
-        scaling_factor=1.0,
-        original_max_position_embeddings=4096,
-        beta_fast=32,
-        beta_slow=1,
-        mscale=1,
-        mscale_all_dim=0,
-    ):
-        super().__init__()
-        self.mscale = yarn_get_mscale(scaling_factor, mscale) / yarn_get_mscale(
-            scaling_factor, mscale_all_dim
-        )
-        freq_extra = base ** (mx.arange(0, dim, 2, dtype=mx.float32) / dim)
-        freq_inter = scaling_factor * freq_extra
-        low, high = yarn_find_correction_range(
-            beta_fast,
-            beta_slow,
-            dim,
-            base,
-            original_max_position_embeddings,
-        )
-        freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2)
-        self._freqs = (freq_inter * freq_extra) / (
-            freq_inter * freq_mask + freq_extra * (1 - freq_mask)
-        )
-
-    def __call__(self, x, offset=0):
-        if self.mscale != 1.0:
-            x = self.mscale * x
-        return mx.fast.rope(
-            x,
-            x.shape[-1],
-            traditional=True,
-            base=None,
-            scale=1.0,
-            offset=offset,
-            freqs=self._freqs,
-        )
-
-
 class DeepseekV3Attention(nn.Module):
     def __init__(self, config: ModelArgs):
         super().__init__()
@@ -160,11 +86,11 @@ class DeepseekV3Attention(nn.Module):
             bias=config.attention_bias,
         )
         self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=1e-6)
-        self.kv_b_proj = nn.Linear(
-            self.kv_lora_rank,
-            self.num_heads
-            * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
-            bias=False,
+        self.embed_q = MultiLinear(
+            self.qk_nope_head_dim, self.kv_lora_rank, self.num_heads
+        )
+        self.unembed_out = MultiLinear(
+            self.kv_lora_rank, self.v_head_dim, self.num_heads
         )
 
         self.o_proj = nn.Linear(
@@ -175,35 +101,19 @@ class DeepseekV3Attention(nn.Module):
 
         if self.config.rope_scaling is not None:
             mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
-            scaling_factor = self.config.rope_scaling["factor"]
             if mscale_all_dim:
-                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
-                self.scale = self.scale * mscale * mscale
+                scaling_factor = self.config.rope_scaling["factor"]
+                if scaling_factor > 1:
+                    s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0
+                    self.scale = self.scale * s * s
 
-            rope_kwargs = {
-                key: self.config.rope_scaling[key]
-                for key in [
-                    "original_max_position_embeddings",
-                    "beta_fast",
-                    "beta_slow",
-                    "mscale",
-                    "mscale_all_dim",
-                ]
-                if key in self.config.rope_scaling
-            }
-            self.rope = DeepseekV3YarnRotaryEmbedding(
-                dim=self.qk_rope_head_dim,
-                max_position_embeddings=self.max_position_embeddings,
-                scaling_factor=scaling_factor,
-                base=self.rope_theta,
-                **rope_kwargs,
-            )
-        else:
-            self.rope = nn.RoPE(
-                dims=self.qk_rope_head_dim,
-                base=self.rope_theta,
-                traditional=True,
-            )
+        self.rope = initialize_rope(
+            dims=self.qk_rope_head_dim,
+            base=self.rope_theta,
+            traditional=True,
+            max_position_embeddings=self.max_position_embeddings,
+            scaling_config=self.config.rope_scaling,
+        )
 
     def __call__(
         self,
@@ -223,29 +133,38 @@ class DeepseekV3Attention(nn.Module):
         compressed_kv = self.kv_a_proj_with_mqa(x)
         compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
         k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
-        kv = self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
-        kv = kv.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+        kv_latent = self.kv_a_layernorm(compressed_kv)
 
-        k_nope, values = mx.split(kv, [self.qk_nope_head_dim], axis=-1)
+        offset = cache.offset if cache is not None else 0
+        q_pe = self.rope(q_pe, offset)
+        k_pe = self.rope(k_pe, offset)
+
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
 
         if cache is not None:
-            q_pe = self.rope(q_pe, cache.offset)
-            k_pe = self.rope(k_pe, cache.offset)
-            k_pe = mx.repeat(k_pe, self.num_heads, axis=1)
-            keys, values = cache.update_and_fetch(
-                mx.concatenate([k_nope, k_pe], axis=-1), values
-            )
-        else:
-            q_pe = self.rope(q_pe)
-            k_pe = self.rope(k_pe)
-            k_pe = mx.repeat(k_pe, self.num_heads, axis=1)
-            keys = mx.concatenate([k_nope, k_pe], axis=-1)
+            kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe)
 
-        queries = mx.concatenate([q_nope, q_pe], axis=-1)
+        pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2)
+        if mask is not None:
+            pe_scores = mx.where(
+                mask,
+                pe_scores,
+                mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype),
+            )
+
+        if L == 1:
+            q_nope = self.embed_q(q_nope)
+            k = v = kv_latent
+        else:
+            k = self.embed_q(kv_latent, transpose=False)
+            v = self.unembed_out(kv_latent)
 
         output = scaled_dot_product_attention(
-            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+            q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores
         )
+        if L == 1:
+            output = self.unembed_out(output)
+
         output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
         return self.o_proj(output)
 
@@ -266,7 +185,7 @@ class DeepseekV3MLP(nn.Module):
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
 
     def __call__(self, x):
-        down_proj = self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
         return down_proj
 
 
@@ -281,18 +200,18 @@ def group_expert_select(
     norm_topk_prob,
 ):
 
-    k = top_k
     scores = mx.sigmoid(gates.astype(mx.float32))
     orig_scores = scores
     scores = scores + e_score_correction_bias
-    scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
-    group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
-    k = n_group - topk_group
-    group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
-    scores = mx.put_along_axis(
-        scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
-    )
-    scores = mx.flatten(scores, -2, -1)
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
 
     k = top_k
     inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
@@ -349,13 +268,21 @@ class DeepseekV3MoE(nn.Module):
                 config=config, intermediate_size=intermediate_size
             )
 
+        self.sharding_group = None
+
     def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
         inds, scores = self.gate(x)
         y = self.switch_mlp(x, inds)
         y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
         if self.config.n_shared_experts is not None:
             y = y + self.shared_experts(x)
 
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
         return y
 
 
@@ -389,7 +316,7 @@ class DeepseekV3DecoderLayer(nn.Module):
         return h + r
 
 
-class DeepseekV3Model(nn.Module):
+class DeepseekV3Model(PipelineMixin, nn.Module):
     def __init__(self, config: ModelArgs):
         super().__init__()
         self.vocab_size = config.vocab_size
@@ -398,59 +325,38 @@ class DeepseekV3Model(nn.Module):
             DeepseekV3DecoderLayer(config, idx)
             for idx in range(config.num_hidden_layers)
         ]
-        self.start_idx = 0
-        self.end_idx = len(self.layers)
-        self.num_layers = self.end_idx
-
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.pipeline_rank = 0
-        self.pipeline_size = 1
-
-    def pipeline(self, group):
-        # Split layers in reverse so rank=0 gets the last layers and
-        # rank=pipeline_size-1 gets the first
-        self.pipeline_rank = group.rank()
-        self.pipeline_size = group.size()
-        layers_per_rank = len(self.layers) // self.pipeline_size
-        extra = len(self.layers) - layers_per_rank * self.pipeline_size
-        if self.pipeline_rank < extra:
-            layers_per_rank += 1
-        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
-        self.end_idx = self.start_idx + layers_per_rank
-        self.layers = self.layers[: self.end_idx]
-        self.layers[: self.start_idx] = [None] * self.start_idx
-        self.num_layers = len(self.layers) - self.start_idx
 
     def __call__(
         self,
         x: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ) -> mx.array:
         h = self.embed_tokens(x)
 
         pipeline_rank = self.pipeline_rank
         pipeline_size = self.pipeline_size
-        if mask is None:
-            mask = create_attention_mask(h, cache)
 
         if cache is None:
-            cache = [None] * self.num_layers
+            cache = [None] * len(self.pipeline_layers)
+        mask = create_attention_mask(h, cache[0], return_array=True)
 
         # Receive from the previous process in the pipeline
-
         if pipeline_rank < pipeline_size - 1:
             h = mx.distributed.recv_like(h, (pipeline_rank + 1))
 
-        for i in range(self.num_layers):
-            h = self.layers[self.start_idx + i](h, mask, cache[i])
+        for l, c in zip(self.pipeline_layers, cache):
+            h = l(h, mask, cache=c)
 
         # Send to the next process in the pipeline
         if pipeline_rank != 0:
             h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+            if cache[-1] is not None:
+                cache[-1].keys = mx.depends(cache[-1].keys, h)
 
         # Broadcast h while keeping it in the graph
-        h = mx.distributed.all_gather(h)[: h.shape[0]]
+        if pipeline_size > 1:
+            h = mx.distributed.all_gather(h)[: h.shape[0]]
 
         return self.norm(h)
 
@@ -467,14 +373,14 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, cache, mask)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
     def sanitize(self, weights):
         def dequant(weight, scale_inv):
-            dtype = weight.dtype
+            dtype = mx.bfloat16
+            weight = mx.from_fp8(weight, dtype=mx.bfloat16)
             bs = 128  # block size
             m, n = weight.shape
             pad_bottom = (-m) % bs
@@ -488,7 +394,22 @@ class Model(nn.Module):
             )
             return weight[:m, :n].astype(dtype)
 
-        # Dequantize
+        # Remap for int4
+        new_weights = {}
+        for k, v in weights.items():
+            if k.endswith("weight_shape"):
+                base = k.replace("weight_shape", "")
+                new_weights[base + "weight"] = weights[base + "weight_packed"].view(
+                    mx.uint32
+                )
+                s = weights[base + "weight_scale"]
+                new_weights[base + "scales"] = s
+                new_weights[base + "biases"] = -8 * s
+            elif not (k.endswith("weight_scale") or k.endswith("weight_packed")):
+                new_weights[k] = v
+        weights = new_weights
+
+        # Dequantize fp8
         new_weights = {}
         for k, v in weights.items():
             if "weight_scale_inv" in k:
@@ -512,6 +433,42 @@ class Model(nn.Module):
                             for e in range(self.args.n_routed_experts)
                         ]
                         weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+            prefix = f"model.layers.{l}.self_attn"
+            if f"{prefix}.kv_b_proj.weight" in weights:
+                layer = self.model.layers[l].self_attn.embed_q
+                quantized = f"{prefix}.kv_b_proj.scales" in weights
+                v = weights.pop(f"{prefix}.kv_b_proj.weight")
+                head_dim = self.args.qk_nope_head_dim + self.args.v_head_dim
+
+                if quantized:
+                    dims = self.args.kv_lora_rank
+                    scales = weights.pop(f"{prefix}.kv_b_proj.scales")
+                    biases = weights.pop(f"{prefix}.kv_b_proj.biases")
+                    # Try to infer bits and group size
+                    bits = (v.shape[-1] * 32) // dims
+                    group_size = dims // scales.shape[-1]
+                    v = mx.dequantize(
+                        v, scales, biases, bits=bits, group_size=group_size
+                    )
+                num_heads = self.args.num_attention_heads
+                v = v.reshape(num_heads, head_dim, -1)
+                wk = mx.contiguous(
+                    v[:, : self.args.qk_nope_head_dim, :].swapaxes(-1, -2)
+                )
+                wv = mx.contiguous(v[:, self.args.qk_nope_head_dim :, :])
+                if quantized:
+                    wk, wk_scales, wk_biases = mx.quantize(
+                        wk, bits=bits, group_size=group_size
+                    )
+                    wv, wv_scales, wv_biases = mx.quantize(
+                        wv, bits=bits, group_size=group_size
+                    )
+                    weights[f"{prefix}.embed_q.scales"] = wk_scales
+                    weights[f"{prefix}.unembed_out.scales"] = wv_scales
+                    weights[f"{prefix}.embed_q.biases"] = wk_biases
+                    weights[f"{prefix}.unembed_out.biases"] = wv_biases
+                weights[f"{prefix}.embed_q.weight"] = wk
+                weights[f"{prefix}.unembed_out.weight"] = wv
 
         # Remove multi-token prediction layer and any unused precomputed rotary freqs
         return {
@@ -520,9 +477,73 @@ class Model(nn.Module):
             if not k.startswith("model.layers.61") and "rotary_emb.inv_freq" not in k
         }
 
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        rank = group.rank()
+        for layer in self.model.layers:
+            # Shard the self attention
+            if layer.self_attn.q_lora_rank is None:
+                layer.self_attn.q_proj = shard_linear(
+                    layer.self_attn.q_proj, "all-to-sharded", group=group
+                )
+            else:
+                layer.self_attn.q_b_proj = shard_linear(
+                    layer.self_attn.q_b_proj, "all-to-sharded", group=group
+                )
+            layer.self_attn.num_heads //= N
+            num_heads = layer.self_attn.num_heads
+            sh = rank * num_heads
+            eh = sh + num_heads
+
+            def shard_heads(w):
+                return w[sh:eh]
+
+            layer.self_attn.embed_q.apply(shard_heads)
+            layer.self_attn.unembed_out.apply(shard_heads)
+
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+
+            # Shard the MLP
+            if isinstance(layer.mlp, DeepseekV3MLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                layer.mlp.sharding_group = group
+                shard_inplace(
+                    layer.mlp.shared_experts.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+
     @property
     def layers(self):
-        return self.model.layers[self.model.start_idx : self.model.end_idx]
+        return self.model.pipeline_layers
 
     @property
     def cast_predicate(self):

mlx_lm/models/deepseek_v32.py

@@ -0,0 +1,654 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import CacheList, KVCache
+from .mla import MultiLinear
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "deepseek_v32"
+    vocab_size: int = 102400
+    hidden_size: int = 4096
+    index_head_dim: int = 128
+    index_n_heads: int = 64
+    index_topk: int = 2048
+    intermediate_size: int = 11008
+    moe_intermediate_size: int = 1407
+    num_hidden_layers: int = 30
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 32
+    n_shared_experts: Optional[int] = None
+    n_routed_experts: Optional[int] = None
+    routed_scaling_factor: float = 1.0
+    kv_lora_rank: int = 512
+    q_lora_rank: int = 1536
+    qk_rope_head_dim: int = 64
+    v_head_dim: int = 128
+    qk_nope_head_dim: int = 128
+    topk_method: str = "noaux_tc"
+    scoring_func: str = "sigmoid"
+    norm_topk_prob: bool = True
+    n_group: int = 1
+    topk_group: int = 1
+    num_experts_per_tok: int = 1
+    moe_layer_freq: int = 1
+    first_k_dense_replace: int = 0
+    max_position_embeddings: int = 2048
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 10000.0
+    rope_scaling: Dict = None
+    attention_bias: bool = False
+
+
+class Indexer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.dim = args.hidden_size
+        self.n_heads = args.index_n_heads
+        self.head_dim = args.index_head_dim
+        self.rope_head_dim = args.qk_rope_head_dim
+        self.index_topk = args.index_topk
+        self.q_lora_rank = args.q_lora_rank
+        self.wq_b = nn.Linear(
+            self.q_lora_rank, self.n_heads * self.head_dim, bias=False
+        )
+        self.wk = nn.Linear(self.dim, self.head_dim, bias=False)
+        self.k_norm = nn.LayerNorm(self.head_dim)
+        self.weights_proj = nn.Linear(self.dim, self.n_heads, bias=False)
+        self.softmax_scale = self.head_dim**-0.5
+        self.rope = initialize_rope(
+            dims=args.qk_rope_head_dim,
+            base=args.rope_theta,
+            traditional=True,
+            max_position_embeddings=args.max_position_embeddings,
+            scaling_config=args.rope_scaling,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        qr: mx.array,
+        mask: Optional[mx.array],
+        cache: Optional[Any] = None,
+    ):
+        # Computes top_k indices for attention
+        b, s, _ = x.shape
+        q = self.wq_b(qr)
+        q = q.reshape(b, s, self.n_heads, self.head_dim).swapaxes(1, 2)
+        k = self.wk(x)
+        k = self.k_norm(k)
+        k = mx.reshape(k, (b, 1, s, self.head_dim))
+
+        offset = cache.offset if cache is not None else 0
+
+        q = self.rope(q, offset=offset)
+        k = self.rope(k, offset=offset)
+
+        if cache is not None:
+            k, _ = cache.update_and_fetch(k, mx.zeros([b, 1, s, 0]))
+        if k.shape[2] <= self.index_topk:
+            return None
+        scores = q @ k.swapaxes(-1, -2)
+        scores = mx.maximum(scores, 0)
+        weights = self.weights_proj(x) * (self.n_heads**-0.5 * self.softmax_scale)
+        weights = weights.swapaxes(-1, -2)[..., None]
+        scores = scores * weights
+        scores = scores.sum(axis=1, keepdims=True)
+        if mask is not None:
+            scores = mx.where(mask, scores, -float("inf"))
+        return mx.argpartition(scores, kth=-self.index_topk, axis=-1)[
+            ..., -self.index_topk :
+        ]
+
+
+class DeepseekV32Attention(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
+
+        self.scale = self.q_head_dim**-0.5
+
+        self.q_a_proj = nn.Linear(
+            self.hidden_size, self.q_lora_rank, bias=config.attention_bias
+        )
+        self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank, eps=1e-6)
+        self.q_b_proj = nn.Linear(
+            self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
+        )
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=1e-6)
+        self.embed_q = MultiLinear(
+            self.qk_nope_head_dim, self.kv_lora_rank, self.num_heads
+        )
+        self.unembed_out = MultiLinear(
+            self.kv_lora_rank, self.v_head_dim, self.num_heads
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        if self.config.rope_scaling is not None:
+            mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
+            if mscale_all_dim:
+                scaling_factor = self.config.rope_scaling["factor"]
+                if scaling_factor > 1:
+                    s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0
+                    self.scale = self.scale * s * s
+
+        self.indexer = Indexer(config)
+        self.rope = initialize_rope(
+            dims=self.qk_rope_head_dim,
+            base=self.rope_theta,
+            traditional=True,
+            max_position_embeddings=self.max_position_embeddings,
+            scaling_config=self.config.rope_scaling,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        qr = self.q_a_layernorm(self.q_a_proj(x))
+        q = self.q_b_proj(qr)
+
+        q = q.reshape(B, L, self.num_heads, self.q_head_dim).transpose(0, 2, 1, 3)
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+        kv_latent = self.kv_a_layernorm(compressed_kv)
+
+        offset = cache[0].offset if cache is not None else 0
+        q_pe = self.rope(q_pe, offset)
+        k_pe = self.rope(k_pe, offset)
+
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
+
+        if cache is not None:
+            kv_latent, k_pe = cache[0].update_and_fetch(kv_latent, k_pe)
+        else:
+            cache = [None] * 2
+
+        topk_indices = self.indexer(x, qr, mask, cache=cache[1])
+        if topk_indices is not None:
+            if L == 1:
+                idx = topk_indices[:, :, 0, :, None]
+                kv_latent = mx.take_along_axis(
+                    kv_latent,
+                    mx.broadcast_to(idx, idx.shape[:-1] + (kv_latent.shape[-1],)),
+                    axis=2,
+                )
+                k_pe = mx.take_along_axis(
+                    k_pe,
+                    mx.broadcast_to(idx, idx.shape[:-1] + (k_pe.shape[-1],)),
+                    axis=2,
+                )
+                if mask is not None:
+                    mask = mx.take_along_axis(mask, topk_indices, axis=-1)
+            else:
+                shape = list(topk_indices.shape)
+                shape[-1] = kv_latent.shape[2]
+                sparse_mask = mx.zeros(shape, dtype=mx.bool_)
+                sparse_mask = mx.put_along_axis(
+                    sparse_mask, topk_indices, mx.array(True), axis=-1
+                )
+                if mask is not None:
+                    sparse_mask = sparse_mask & mask
+                mask = sparse_mask
+        # Ensure the indexer cache is evaluated even if the topk_indices are unused
+        # to keep the graph from getting too large
+        if cache is not None and cache[0] is not None:
+            cache[0].keys = mx.depends(cache[0].keys, (cache[1].keys, cache[1].values))
+
+        pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2)
+        if mask is not None:
+            pe_scores = mx.where(
+                mask,
+                pe_scores,
+                mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype),
+            )
+
+        if L == 1:
+            q_nope = self.embed_q(q_nope)
+            k = v = kv_latent
+        else:
+            k = self.embed_q(kv_latent, transpose=False)
+            v = self.unembed_out(kv_latent)
+
+        output = scaled_dot_product_attention(
+            q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores
+        )
+        if L == 1:
+            output = self.unembed_out(output)
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class DeepseekV32MLP(nn.Module):
+    def __init__(
+        self, config: ModelArgs, hidden_size: int = None, intermediate_size: int = None
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x):
+        down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+        return down_proj
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
+class MoEGate(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, config.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert config.topk_method == "noaux_tc", "Unsupported topk method."
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class DeepseekV32MoE(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.switch_mlp = SwitchGLU(
+            config.hidden_size,
+            config.moe_intermediate_size,
+            config.n_routed_experts,
+        )
+
+        self.gate = MoEGate(config)
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = DeepseekV32MLP(
+                config=config, intermediate_size=intermediate_size
+            )
+
+        self.sharding_group = None
+
+    def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
+        inds, scores = self.gate(x)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.config.n_shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y
+
+
+class DeepseekV32DecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = DeepseekV32Attention(config)
+        self.mlp = (
+            DeepseekV32MoE(config)
+            if (
+                config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+                and layer_idx % config.moe_layer_freq == 0
+            )
+            else DeepseekV32MLP(config)
+        )
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class DeepseekV32Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            DeepseekV32DecoderLayer(config, idx)
+            for idx in range(config.num_hidden_layers)
+        ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.pipeline_rank = 0
+        self.pipeline_size = 1
+
+    def pipeline(self, group):
+        # Split layers in reverse so rank=0 gets the last layers and
+        # rank=pipeline_size-1 gets the first
+        self.pipeline_rank = group.rank()
+        self.pipeline_size = group.size()
+        layers_per_rank = len(self.layers) // self.pipeline_size
+        extra = len(self.layers) - layers_per_rank * self.pipeline_size
+        if self.pipeline_rank < extra:
+            layers_per_rank += 1
+        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
+        self.end_idx = self.start_idx + layers_per_rank
+        self.layers = self.layers[: self.end_idx]
+        self.layers[: self.start_idx] = [None] * self.start_idx
+        self.num_layers = len(self.layers) - self.start_idx
+
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(x)
+
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
+
+        if cache is None:
+            cache = [None] * self.num_layers
+        mask = create_attention_mask(
+            h, cache[0][0] if cache[0] else None, return_array=True
+        )
+
+        # Receive from the previous process in the pipeline
+
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1))
+
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i])
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+            if cache[-1] is not None:
+                cache[-1][0].keys = mx.depends(cache[-1][0].keys, h)
+
+        # Broadcast h while keeping it in the graph
+        if pipeline_size > 1:
+            h = mx.distributed.all_gather(h)[: h.shape[0]]
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.model = DeepseekV32Model(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+    def sanitize(self, weights):
+        # Remove multi-token prediction layers
+        mpt_layer = self.args.num_hidden_layers
+        new_weights = {}
+        for k, v in weights.items():
+            parts = k.split(".")
+            if len(parts) >= 3 and parts[1] == "layers" and int(parts[2]) >= mpt_layer:
+                continue
+            new_weights[k] = v
+        weights = new_weights
+
+        def dequant(weight, scale_inv):
+            dtype = mx.bfloat16
+            weight = mx.from_fp8(weight, dtype=mx.bfloat16)
+            bs = 128  # block size
+            m, n = weight.shape
+            pad_bottom = (-m) % bs
+            pad_side = (-n) % bs
+            weight = mx.pad(weight, ((0, pad_bottom), (0, pad_side)))
+            weight = weight.reshape(
+                ((m + pad_bottom) // bs, bs, (n + pad_side) // bs, bs)
+            )
+            weight = (weight * scale_inv[:, None, :, None]).reshape(
+                m + pad_bottom, n + pad_side
+            )
+            return weight[:m, :n].astype(dtype)
+
+        # Dequantize
+        new_weights = {}
+        for k, v in weights.items():
+            if "weight_scale_inv" in k:
+                scale_inv = v
+                wk = k.replace("_scale_inv", "")
+                weight = weights[wk]
+                weight = dequant(weight, scale_inv)
+                new_weights[wk] = weight
+            elif k not in new_weights:
+                new_weights[k] = v
+        weights = new_weights
+
+        # Stack experts
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+            prefix = f"model.layers.{l}.self_attn"
+            if f"{prefix}.kv_b_proj.weight" in weights:
+                layer = self.model.layers[l].self_attn.embed_q
+                quantized = f"{prefix}.kv_b_proj.scales" in weights
+                v = weights.pop(f"{prefix}.kv_b_proj.weight")
+                head_dim = self.args.qk_nope_head_dim + self.args.v_head_dim
+
+                if quantized:
+                    dims = self.args.kv_lora_rank
+                    scales = weights.pop(f"{prefix}.kv_b_proj.scales")
+                    biases = weights.pop(f"{prefix}.kv_b_proj.biases")
+                    # Try to infer bits and group size
+                    bits = (v.shape[-1] * 32) // dims
+                    group_size = dims // scales.shape[-1]
+                    v = mx.dequantize(
+                        v, scales, biases, bits=bits, group_size=group_size
+                    )
+                num_heads = self.args.num_attention_heads
+                v = v.reshape(num_heads, head_dim, -1)
+                wk = mx.contiguous(
+                    v[:, : self.args.qk_nope_head_dim, :].swapaxes(-1, -2)
+                )
+                wv = mx.contiguous(v[:, self.args.qk_nope_head_dim :, :])
+                if quantized:
+                    wk, wk_scales, wk_biases = mx.quantize(
+                        wk, bits=bits, group_size=group_size
+                    )
+                    wv, wv_scales, wv_biases = mx.quantize(
+                        wv, bits=bits, group_size=group_size
+                    )
+                    weights[f"{prefix}.embed_q.scales"] = wk_scales
+                    weights[f"{prefix}.unembed_out.scales"] = wv_scales
+                    weights[f"{prefix}.embed_q.biases"] = wk_biases
+                    weights[f"{prefix}.unembed_out.biases"] = wv_biases
+                weights[f"{prefix}.embed_q.weight"] = wk
+                weights[f"{prefix}.unembed_out.weight"] = wv
+
+        return weights
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        rank = group.rank()
+        for layer in self.model.layers:
+            layer.self_attn.q_b_proj = shard_linear(
+                layer.self_attn.q_b_proj, "all-to-sharded", group=group
+            )
+
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.num_heads //= N
+            num_heads = layer.self_attn.num_heads
+            sh = rank * num_heads
+            eh = sh + num_heads
+
+            def shard_heads(w):
+                return w[sh:eh]
+
+            layer.self_attn.embed_q.apply(shard_heads)
+            layer.self_attn.unembed_out.apply(shard_heads)
+
+            # Shard the MLP
+            if isinstance(layer.mlp, DeepseekV32MLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                layer.mlp.sharding_group = group = group
+                shard_inplace(
+                    layer.mlp.shared_experts.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+
+    @property
+    def layers(self):
+        return self.model.layers[self.model.start_idx : self.model.end_idx]
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
+
+    def make_cache(self):
+        return [CacheList(KVCache(), KVCache()) for _ in self.layers]

mlx_lm/models/dots1.py

@@ -7,6 +7,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 from .switch_layers import SwitchGLU
@@ -22,10 +23,9 @@ class ModelArgs(BaseModelArgs):
     rms_norm_eps: float
     vocab_size: int
     max_position_embeddings: Optional[int]
-    num_key_value_heads: Optional[int]
+    num_key_value_heads: int
     first_k_dense_replace: int
     moe_intermediate_size: int
-    moe_layer_freq: int
     n_routed_experts: int
     n_shared_experts: int
     norm_topk_prob: bool
@@ -48,7 +48,6 @@ class Dots1Attention(nn.Module):
 
         dim = args.hidden_size
         self.n_heads = n_heads = args.num_attention_heads
-        assert args.num_key_value_heads is not None
         self.n_kv_heads = n_kv_heads = args.num_key_value_heads
 
         head_dim = args.head_dim or args.hidden_size // n_heads
@@ -182,7 +181,7 @@ class Dots1MLP(nn.Module):
         )
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class Dots1MoE(nn.Module):
@@ -254,17 +253,15 @@ class Dots1Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ) -> mx.array:
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -283,10 +280,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/ernie4_5.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 
@@ -87,7 +88,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=use_bias)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class DecoderLayer(nn.Module):
@@ -123,17 +124,15 @@ class Ernie45Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -152,10 +151,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/ernie4_5_moe.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 from .switch_layers import SwitchGLU
@@ -98,7 +99,7 @@ class Ernie4_5_MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=use_bias)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class Ernie4_5_MoeMLP(nn.Module):
@@ -219,17 +220,15 @@ class Ernie45Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -248,10 +247,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/exaone.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 
@@ -91,7 +92,7 @@ class MLP(nn.Module):
         self.c_proj = nn.Linear(hidden_dim, dim, bias=args.mlp_bias)
 
     def __call__(self, x: mx.array) -> mx.array:
-        return self.c_proj(nn.silu(self.c_fc_0(x)) * self.c_fc_1(x))
+        return self.c_proj(swiglu(self.c_fc_0(x), self.c_fc_1(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -123,16 +124,15 @@ class ExaoneModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.wte(inputs)
-        if mask is None:
-            mask = create_attention_mask(h, cache)
 
         if cache is None:
             cache = [None] * len(self.h)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.h, cache):
             h = layer(h, mask, cache=c)
 
@@ -151,10 +151,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.transformer(inputs, mask, cache)
+        out = self.transformer(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.transformer.wte.as_linear(out)
         else:

mlx_lm/models/exaone4.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import KVCache, RotatingKVCache
 from .rope_utils import initialize_rope
@@ -102,7 +103,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -149,23 +150,35 @@ class ExaoneModel(nn.Module):
             )
             for i in range(args.num_hidden_layers)
         ]
+        if pattern:
+            self.swa_idx = pattern.index("L")
+            self.full_idx = pattern.index("G")
+        else:
+            self.swa_idx = None
+            self.full_idx = 0
+
+        self.window_size = args.sliding_window
         self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
 
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
+        global_mask = create_attention_mask(h, cache[self.full_idx])
+        if self.swa_idx is not None:
+            swa_mask = create_attention_mask(
+                h, cache[self.swa_idx], window_size=self.window_size
+            )
+        else:
+            swa_mask = None
 
         for layer, c in zip(self.layers, cache):
+            mask = swa_mask if layer.self_attn.is_local else global_mask
             h = layer(h, mask, c)
 
         return self.norm(h)
@@ -183,10 +196,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/exaone_moe.py

@@ -0,0 +1,439 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    moe_intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: int
+    num_experts: int
+    num_experts_per_tok: int
+    num_shared_experts: int
+    rms_norm_eps: float
+    max_position_embeddings: int
+    sliding_window: int
+    layer_types: List[str]
+    is_moe_layer: List[bool]
+    n_group: int = 1
+    topk_group: int = 1
+    routed_scaling_factor: float = 2.5
+    norm_topk_prob: bool = True
+    scoring_func: str = "sigmoid"
+    topk_method: str = "noaux_tc"
+    rope_theta: float = 1000000.0
+    rope_scaling: Optional[dict] = None
+    rope_parameters: Optional[dict] = None
+    tie_word_embeddings: bool = False
+
+    def __post_init__(self):
+        if self.rope_parameters is not None and "rope_theta" in self.rope_parameters:
+            self.rope_theta = self.rope_parameters["rope_theta"]
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / (denominator + 1e-20)
+
+    scores = scores * routed_scaling_factor
+    return inds, scores
+
+
+class MoEGate(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+        self.n_routed_experts = args.num_experts
+        self.routed_scaling_factor = args.routed_scaling_factor
+        self.n_group = args.n_group
+        self.topk_group = args.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, args.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert args.topk_method == "noaux_tc", "Unsupported topk method."
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs, intermediate_size: Optional[int] = None):
+        super().__init__()
+        hidden_size = args.hidden_size
+        intermediate_size = intermediate_size or args.intermediate_size
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class MoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size,
+            args.moe_intermediate_size,
+            args.num_experts,
+        )
+
+        self.gate = MoEGate(args)
+
+        self.shared_experts = (
+            MLP(
+                args,
+                intermediate_size=args.moe_intermediate_size * args.num_shared_experts,
+            )
+            if args.num_shared_experts is not None and args.num_shared_experts > 0
+            else None
+        )
+
+        self.sharding_group = None
+
+    def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
+        inds, scores = self.gate(x)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+
+        self.hidden_size = args.hidden_size
+        self.n_heads = args.num_attention_heads
+        self.n_kv_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.n_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.n_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+
+        self.is_sliding_window = args.layer_types[layer_idx] == "sliding_attention"
+        self.apply_rope_all_layers = "sliding_attention" not in args.layer_types
+        self.use_rope = self.is_sliding_window or self.apply_rope_all_layers
+
+        if self.use_rope:
+            self.rope = initialize_rope(
+                self.head_dim,
+                base=args.rope_theta,
+                traditional=False,
+                scaling_config=args.rope_scaling,
+                max_position_embeddings=args.max_position_embeddings,
+            )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = self.q_norm(queries.reshape(B, L, self.n_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        keys = self.k_norm(keys.reshape(B, L, self.n_kv_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            if self.use_rope:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        elif self.use_rope:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+
+        self.self_attn = Attention(args, layer_idx)
+        self.mlp = MoE(args) if args.is_moe_layer[layer_idx] else MLP(args)
+        self.is_sliding_window = self.self_attn.is_sliding_window
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class ExaoneMoEModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [DecoderLayer(args, idx) for idx in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+        self.swa_idx = None
+        self.ga_idx = None
+        for i, layer in enumerate(self.layers):
+            if layer.is_sliding_window and self.swa_idx is None:
+                self.swa_idx = i
+            if not layer.is_sliding_window and self.ga_idx is None:
+                self.ga_idx = i
+
+        self.window_size = args.sliding_window
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        global_mask = create_attention_mask(
+            h, cache[self.ga_idx] if self.ga_idx is not None else cache[0]
+        )
+        swa_mask = create_attention_mask(
+            h,
+            cache[self.swa_idx] if self.swa_idx is not None else cache[0],
+            window_size=self.window_size,
+        )
+
+        for layer, c in zip(self.layers, cache):
+            mask = swa_mask if layer.is_sliding_window else global_mask
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = ExaoneMoEModel(args)
+
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        new_weights = {k: v for k, v in weights.items() if not k.startswith("mtp.")}
+        weights = new_weights
+
+        for l in range(self.args.num_hidden_layers):
+            if not self.args.is_moe_layer[l]:
+                continue
+
+            prefix = f"model.layers.{l}"
+
+            bias_key = f"{prefix}.mlp.e_score_correction_bias"
+            if bias_key in weights:
+                weights[f"{prefix}.mlp.gate.e_score_correction_bias"] = weights.pop(
+                    bias_key
+                )
+
+            for m in ["gate_proj", "down_proj", "up_proj"]:
+                for k in ["weight", "scales", "biases"]:
+                    first_key = f"{prefix}.mlp.experts.0.{m}.{k}"
+                    last_key = (
+                        f"{prefix}.mlp.experts.{self.args.num_experts - 1}.{m}.{k}"
+                    )
+                    if first_key in weights and last_key in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.num_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
+
+    def make_cache(self):
+        caches = []
+        for layer in self.layers:
+            if layer.is_sliding_window:
+                caches.append(
+                    RotatingKVCache(max_size=self.args.sliding_window, keep=0)
+                )
+            else:
+                caches.append(KVCache())
+        return caches
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+
+        for layer in self.model.layers:
+            layer.self_attn.q_proj = shard_linear(
+                layer.self_attn.q_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.k_proj = shard_linear(
+                layer.self_attn.k_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.v_proj = shard_linear(
+                layer.self_attn.v_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.n_heads //= N
+            layer.self_attn.n_kv_heads //= N
+
+            if isinstance(layer.mlp, MLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+            else:
+                layer.mlp.sharding_group = group
+                if layer.mlp.shared_experts is not None:
+                    shard_inplace(
+                        layer.mlp.shared_experts.gate_proj,
+                        "all-to-sharded",
+                        group=group,
+                    )
+                    shard_inplace(
+                        layer.mlp.shared_experts.down_proj,
+                        "sharded-to-all",
+                        group=group,
+                    )
+                    shard_inplace(
+                        layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group
+                    )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )

mlx_lm/models/falcon_h1.py

@@ -0,0 +1,504 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass, field
+from typing import List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, CacheList, KVCache
+from .rope_utils import initialize_rope
+from .ssm import ssm_update
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    attention_bias: bool = False
+    attention_in_multiplier: float = 1.0
+    attention_out_multiplier: float = 0.9375
+    embedding_multiplier: float = 5.656854249492381
+    head_dim: int = 64
+    hidden_size: int = 1024
+    initializer_range: float = 0.02
+    intermediate_size: int = 2048
+    key_multiplier: float = 0.390625
+    lm_head_multiplier: float = 0.0390625
+    mamba_chunk_size: int = 128
+    mamba_conv_bias: bool = True
+    mamba_d_conv: int = 4
+    mamba_d_head: int = 64
+    mamba_d_ssm: int = 1536
+    mamba_d_state: int = 128
+    mamba_expand: int = 2
+    mamba_n_groups: int = 1
+    mamba_n_heads: int = 24
+    mamba_norm_before_gate: bool = False
+    mamba_proj_bias: bool = False
+    mamba_rms_norm: bool = False
+    mamba_use_mlp: bool = True
+    max_position_embeddings: int = 131072
+    mlp_bias: bool = False
+    mlp_expansion_factor: int = 8
+    mlp_multipliers: List[float] = field(
+        default_factory=lambda: [0.8838834764831844, 0.5859375]
+    )
+    model_type: str = "falcon_h1"
+    num_attention_heads: int = 8
+    num_hidden_layers: int = 36
+    num_key_value_heads: int = 2
+    projectors_bias: bool = False
+    rms_norm_eps: float = 1e-05
+    rope_traditional: bool = False
+    rope_scaling: Optional[float] = None
+    rope_theta: float = 100000000000.0
+    ssm_in_multiplier: float = 1.25
+    ssm_multipliers: List[float] = field(
+        default_factory=lambda: [
+            0.3535533905932738,
+            0.25,
+            0.3535533905932738,
+            0.5,
+            0.3535533905932738,
+        ]
+    )
+    ssm_out_multiplier: float = 0.23570226039551587
+    vocab_size: int = 32784
+    tie_word_embeddings: bool = True
+
+
+class FalconH1RMSNormGated(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6, n_groups=1, norm_before_gate=True):
+        super().__init__()
+        self.weight = mx.ones((hidden_size,))
+        self.variance_epsilon = eps
+        self.n_groups = n_groups
+        self.norm_before_gate = norm_before_gate
+
+    def __call__(self, hidden_states, gate=None):
+        if not self.norm_before_gate and gate is not None:
+            hidden_states = swiglu(gate, hidden_states)
+
+        hidden_states = mx.fast.rms_norm(
+            hidden_states, self.weight, self.variance_epsilon
+        )
+
+        if self.norm_before_gate and gate is not None:
+            hidden_states = swiglu(gate, hidden_states)
+        return hidden_states
+
+
+def compute_mup_vector(args):
+    intermediate_size = args.mamba_d_ssm
+    groups_time_state_size = args.mamba_n_groups * args.mamba_d_state
+    num_heads = args.mamba_n_heads
+    sizes = [
+        intermediate_size,
+        intermediate_size,
+        groups_time_state_size,
+        groups_time_state_size,
+        num_heads,
+    ]
+    return mx.concatenate(
+        [
+            mx.broadcast_to(mx.array(m), (s,))
+            for s, m in zip(sizes, args.ssm_multipliers)
+        ]
+    )
+
+
+class FalconH1Attention(nn.Module):
+
+    def __init__(self, args):
+        super().__init__()
+
+        self.hidden_size = args.hidden_size
+        self.num_heads = args.num_attention_heads
+        self.num_kv_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.num_heads * self.head_dim, bias=args.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size,
+            self.num_kv_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size,
+            self.num_kv_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=args.attention_bias
+        )
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            args.rope_traditional,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(self, x, mask=None, cache=None):
+        B, L, _ = x.shape
+
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.num_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.num_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, mask=mask, scale=self.scale, cache=cache
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.o_proj(output)
+
+
+class FalconH1Mixer(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.num_heads = args.mamba_n_heads
+        self.hidden_size = args.hidden_size
+        self.ssm_state_size = args.mamba_d_state
+        self.conv_kernel_size = args.mamba_d_conv
+        self.intermediate_size = args.mamba_d_ssm
+        self.use_conv_bias = args.mamba_conv_bias
+
+        self.layer_norm_epsilon = args.rms_norm_eps
+        self.groups_time_state_size = args.mamba_n_groups * self.ssm_state_size
+
+        self.n_groups = args.mamba_n_groups
+        self.head_dim = args.mamba_d_head
+        self.chunk_size = args.mamba_chunk_size
+
+        self.time_step_limit = (0.0, float("inf"))
+        self.time_step_min = 0.001
+        self.time_step_max = 0.1
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=self.use_conv_bias,
+            kernel_size=self.conv_kernel_size,
+            groups=self.conv_dim,
+        )
+
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size,
+            projection_size,
+            bias=args.mamba_proj_bias,
+        )
+
+        self.dt_bias = mx.ones(self.num_heads)
+
+        A = mx.arange(1, self.num_heads + 1)
+        self.A_log = mx.log(A)
+
+        self.mamba_rms_norm = args.mamba_rms_norm
+        if self.mamba_rms_norm:
+            self.norm = FalconH1RMSNormGated(
+                self.intermediate_size,
+                eps=self.layer_norm_epsilon,
+                n_groups=self.n_groups,
+                norm_before_gate=args.mamba_norm_before_gate,
+            )
+
+        self.D = mx.ones(self.num_heads)
+
+        self.out_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=args.projectors_bias
+        )
+
+    def _conv(
+        self,
+        conv_input: mx.array,
+        cache: Optional[ArraysCache],
+        mask: Optional[mx.array],
+    ) -> mx.array:
+        if mask is not None:
+            conv_input = mx.where(mask[..., None], conv_input, 0)
+
+        if cache is not None:
+            if cache[0] is None:
+                conv_state = mx.zeros(
+                    (conv_input.shape[0], self.conv_kernel_size - 1, self.conv_dim),
+                    dtype=conv_input.dtype,
+                )
+            else:
+                conv_state = cache[0]
+            padded_input = mx.concatenate([conv_state, conv_input], axis=1)
+            n_keep = self.conv_kernel_size - 1
+            if cache.lengths is not None:
+                t = padded_input.shape[1]
+                ends = mx.clip(cache.lengths, 0, t - n_keep)
+                positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+                cache[0] = mx.take_along_axis(padded_input, positions, axis=1)
+            else:
+                cache[0] = padded_input[:, -n_keep:, :]
+        else:
+            padded_input = mx.pad(
+                conv_input, [(0, 0), (self.conv_kernel_size - 1, 0), (0, 0)]
+            )
+
+        conv_output = self.conv1d(padded_input)
+        return nn.silu(conv_output)
+
+    def _ssm(
+        self,
+        hidden_states: mx.array,
+        B: mx.array,
+        C: mx.array,
+        dt: mx.array,
+        cache: Optional[ArraysCache],
+        mask: Optional[mx.array],
+    ) -> mx.array:
+        batch_size, seq_len, _ = hidden_states.shape
+        hidden_states = hidden_states.reshape(
+            batch_size, seq_len, self.num_heads, self.head_dim
+        )
+        B = B.reshape(batch_size, seq_len, self.n_groups, self.ssm_state_size)
+        C = C.reshape(batch_size, seq_len, self.n_groups, self.ssm_state_size)
+        if cache:
+            state = cache[1]
+            lengths = cache.lengths
+        else:
+            state, lengths = None, None
+        y, state = ssm_update(
+            hidden_states,
+            self.A_log,
+            B,
+            C,
+            self.D,
+            dt,
+            self.dt_bias,
+            state,
+            self.time_step_limit,
+            mask,
+            lengths,
+        )
+        if cache:
+            cache[1] = state
+        return y.reshape(batch_size, seq_len, self.intermediate_size)
+
+    def __call__(self, input_states, cache=None, mask: Optional[mx.array] = None):
+        projected_states = self.in_proj(input_states)
+
+        gate, conv_input, dt = mx.split(
+            projected_states,
+            [self.intermediate_size, self.intermediate_size + self.conv_dim],
+            axis=-1,
+        )
+
+        conv_output = self._conv(conv_input, cache, mask)
+
+        hidden_states, B, C = mx.split(
+            conv_output,
+            [
+                self.intermediate_size,
+                self.intermediate_size + self.n_groups * self.ssm_state_size,
+            ],
+            axis=-1,
+        )
+
+        y = self._ssm(hidden_states, B, C, dt, cache, mask=mask)
+        if cache:
+            cache.advance(y.shape[1])
+
+        if self.mamba_rms_norm:
+            y = self.norm(y, gate)
+        else:
+            y = swiglu(gate, y)
+
+        return self.out_proj(y)
+
+
+class FalconH1MLP(nn.Module):
+
+    def __init__(self, args):
+        super().__init__()
+
+        hidden_size = args.hidden_size
+        intermediate_size = args.intermediate_size
+
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=args.mlp_bias)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=args.mlp_bias)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=args.mlp_bias)
+
+    def __call__(self, x):
+        y = swiglu(self.gate_proj(x), self.up_proj(x))
+        y = self.down_proj(y)
+        return y
+
+
+class FalconH1DecoderLayer(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.feed_forward = FalconH1MLP(args)
+
+        head_dim = args.head_dim
+        self.channels_attn = (
+            args.num_attention_heads * head_dim
+            + 2 * args.num_key_value_heads * head_dim
+        )
+
+        self.mamba = FalconH1Mixer(args=args)
+
+        self.self_attn = FalconH1Attention(args)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.pre_ff_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        h: mx.array,
+        cache,
+        attn_mask: Optional[mx.array],
+        mamba_mask: Optional[mx.array],
+    ) -> mx.array:
+
+        residual = h
+        h = self.input_layernorm(h)
+
+        mamba_h = self.mamba(input_states=h, cache=cache[0], mask=mamba_mask)
+
+        attn_h = self.self_attn(
+            h,
+            mask=attn_mask,
+            cache=cache[1],
+        )
+
+        h = residual + mamba_h + attn_h
+
+        residual = h
+        h = self.pre_ff_layernorm(h)
+        h = self.feed_forward(h)
+        return residual + h
+
+
+class FalconH1Model(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.hidden_size = args.hidden_size
+
+        self.embed_tokens = nn.Embedding(self.vocab_size, self.hidden_size)
+
+        self._mup_vector = compute_mup_vector(args)
+        self.layers = [
+            FalconH1DecoderLayer(args) for _ in range(args.num_hidden_layers)
+        ]
+        self.final_layernorm = nn.RMSNorm(self.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(self, inputs, cache=None):
+
+        h = self.embed_tokens(inputs)
+
+        h = h
+
+        if cache is None:
+            cache = [(None, None) * len(self.layers)]
+
+        mamba_mask = create_ssm_mask(h, cache[0][0])
+        attn_mask = create_attention_mask(h, cache[0][1])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(
+                h,
+                cache=c,
+                attn_mask=attn_mask,
+                mamba_mask=mamba_mask,
+            )
+
+        return self.final_layernorm(h)
+
+
+class Model(nn.Module):
+
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = FalconH1Model(args=args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(self, inputs, cache=None):
+        hidden_states = self.model(inputs, cache=cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(hidden_states)
+            return out * (self.args.lm_head_multiplier / self.args.embedding_multiplier)
+        else:
+            return self.lm_head(hidden_states)
+
+    def sanitize(self, weights):
+        # Check if needs sanitization
+        c1d = weights["model.layers.0.mamba.conv1d.weight"]
+        if c1d.shape[-1] <= c1d.shape[1]:
+            return weights
+
+        sanitized_weights = {}
+        args = self.args
+
+        for name, param in weights.items():
+            # Fold-in multipliers
+            if name.endswith("embed_tokens.weight"):
+                param *= args.embedding_multiplier
+            elif name.endswith("lm_head.weight"):
+                param *= args.lm_head_multiplier
+            elif name.endswith("q_proj.weight") or name.endswith("k_proj.weight"):
+                param *= args.attention_in_multiplier
+            elif name.endswith("key_proj.weight"):
+                param *= args.attention_in_multiplier * args.key_multiplier
+            elif name.endswith("o_proj.weight"):
+                param *= args.attention_out_multiplier
+            elif name.endswith("out_proj.weight"):
+                param *= args.ssm_out_multiplier
+            elif name.endswith("gate_proj.weight"):
+                param *= args.mlp_multipliers[0]
+            elif name.endswith("down_proj.weight"):
+                param *= args.mlp_multipliers[1]
+            elif name.endswith("in_proj.weight"):
+                param *= (
+                    args.ssm_in_multiplier
+                    * self.model._mup_vector.astype(param.dtype)[:, None]
+                )
+            elif "conv1d.weight" in name:
+                param = param.transpose(0, 2, 1)
+            sanitized_weights[name] = param
+        return sanitized_weights
+
+    def make_cache(self):
+        return [
+            CacheList(ArraysCache(size=2), KVCache())
+            for _ in range(self.args.num_hidden_layers)
+        ]
+
+    @property
+    def layers(self):
+        return self.model.layers

mlx_lm/models/gated_delta.py

@@ -0,0 +1,283 @@
+from functools import partial
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+@partial(mx.compile, shapeless=True)
+def compute_g(A_log, a, dt_bias):
+    return mx.exp(-mx.exp(A_log.astype(mx.float32)) * nn.softplus(a + dt_bias))
+
+
+def _make_gated_delta_kernel(has_mask=False, vectorized=False):
+    if not mx.metal.is_available():
+        return None
+    mask_source = "mask[b_idx * T + t]" if has_mask else "true"
+
+    # Configure g indexing based on whether gating is vectorized
+    if vectorized:
+        g_comment = "// g: [B, T, Hv, Dk]"
+        g_setup = "auto g_ = g + (b_idx * T * Hv + hv_idx) * Dk;"
+        g_access = "g_[s_idx]"
+        g_advance = "g_ += Hv * Dk;"
+    else:
+        g_comment = "// g: [B, T, Hv]"
+        g_setup = "auto g_ = g + b_idx * T * Hv;"
+        g_access = "g_[hv_idx]"
+        g_advance = "g_ += Hv;"
+
+    source = f"""
+        auto n = thread_position_in_grid.z;
+        auto b_idx = n / Hv;
+        auto hv_idx = n % Hv;
+        auto hk_idx = hv_idx / (Hv / Hk);
+        constexpr int n_per_t = Dk / 32;
+
+        // q, k: [B, T, Hk, Dk]
+        auto q_ = q + b_idx * T * Hk * Dk + hk_idx * Dk;
+        auto k_ = k + b_idx * T * Hk * Dk + hk_idx * Dk;
+
+        // v, y: [B, T, Hv, Dv]
+        auto v_ = v + b_idx * T * Hv * Dv + hv_idx * Dv;
+        y += b_idx * T * Hv * Dv + hv_idx * Dv;
+
+        auto dk_idx = thread_position_in_threadgroup.x;
+        auto dv_idx = thread_position_in_grid.y;
+
+        // state_in, state_out: [B, Hv, Dv, Dk]
+        auto i_state = state_in + (n * Dv + dv_idx) * Dk;
+        auto o_state = state_out + (n * Dv + dv_idx) * Dk;
+
+        float state[n_per_t];
+        for (int i = 0; i < n_per_t; ++i) {{
+          auto s_idx = n_per_t * dk_idx + i;
+          state[i] = static_cast<float>(i_state[s_idx]);
+        }}
+
+        {g_comment}
+        {g_setup}
+        auto beta_ = beta + b_idx * T * Hv;
+
+        for (int t = 0; t < T; ++t) {{
+          if ({mask_source}) {{
+            float kv_mem = 0.0f;
+            for (int i = 0; i < n_per_t; ++i) {{
+              auto s_idx = n_per_t * dk_idx + i;
+              state[i] = state[i] * {g_access};
+              kv_mem += state[i] * k_[s_idx];
+            }}
+            kv_mem = simd_sum(kv_mem);
+
+            auto delta = (v_[dv_idx] - kv_mem) * beta_[hv_idx];
+
+            float out = 0.0f;
+            for (int i = 0; i < n_per_t; ++i) {{
+              auto s_idx = n_per_t * dk_idx + i;
+              state[i] = state[i] + k_[s_idx] * delta;
+              out += state[i] * q_[s_idx];
+            }}
+            out = simd_sum(out);
+            if (thread_index_in_simdgroup == 0) {{
+              y[dv_idx] = static_cast<InT>(out);
+            }}
+          }} else {{
+            y[dv_idx] = static_cast<InT>(0);
+          }}
+          // Increment data pointers to next time step
+          q_ += Hk * Dk;
+          k_ += Hk * Dk;
+          v_ += Hv * Dv;
+          y += Hv * Dv;
+          {g_advance}
+          beta_ += Hv;
+        }}
+        for (int i = 0; i < n_per_t; ++i) {{
+          auto s_idx = n_per_t * dk_idx + i;
+          o_state[s_idx] = static_cast<StT>(state[i]);
+        }}
+    """
+    inputs = ["q", "k", "v", "g", "beta", "state_in", "T"]
+    if has_mask:
+        inputs.append("mask")
+
+    suffix = ""
+    if vectorized:
+        suffix += "_vec"
+    if has_mask:
+        suffix += "_mask"
+
+    return mx.fast.metal_kernel(
+        name=f"gated_delta_step{suffix}",
+        input_names=inputs,
+        output_names=["y", "state_out"],
+        source=source,
+    )
+
+
+_gated_delta_kernel = _make_gated_delta_kernel(has_mask=False, vectorized=False)
+_gated_delta_kernel_masked = _make_gated_delta_kernel(has_mask=True, vectorized=False)
+_gated_delta_kernel_vec = _make_gated_delta_kernel(has_mask=False, vectorized=True)
+_gated_delta_kernel_vec_masked = _make_gated_delta_kernel(
+    has_mask=True, vectorized=True
+)
+
+
+@mx.compile
+def _gated_delta_step_ops(
+    q: mx.array,
+    k: mx.array,
+    v: mx.array,
+    g: mx.array,
+    beta: mx.array,
+    state: mx.array,
+    mask: Optional[mx.array] = None,
+) -> Tuple[mx.array, mx.array]:
+    """
+    Ops-based reference implementation for a single recurrent step.
+
+    Shapes:
+      - q, k: [B, H, Dk]
+      - v: [B, H, Dv]
+      - g: [B, H] or [B, H, Dk]
+      - beta: [B, H]
+      - state: [B, H, Dv, Dk]
+    Returns:
+      - y: [B, H, Dv]
+      - new_state: [B, H, Dv, Dk]
+    """
+
+    # Decay
+    old_state = state
+    if g.ndim == 2:
+        decay = g[..., None, None]
+    elif g.ndim == 3:
+        decay = g[..., None, :]
+    else:
+        raise ValueError(f"Unsupported gating shape {g.shape}")
+    state = state * decay
+    kv_mem = (state * k[..., None, :]).sum(axis=-1)  # [B, H, Dv]
+    delta = (v - kv_mem) * beta[..., None]  # [B, H, Dv]
+    state = state + k[..., None, :] * delta[..., None]
+    # Output projection along key dim with q
+    y = (state * q[..., None, :]).sum(axis=-1)  # [B, H, Dv]
+
+    if mask is not None:
+        mask = mx.expand_dims(mask, axis=(1, 2, 3))
+        state = mx.where(mask, state, old_state)
+    return y.astype(q.dtype), state
+
+
+def gated_delta_kernel(
+    q: mx.array,
+    k: mx.array,
+    v: mx.array,
+    g: mx.array,
+    beta: mx.array,
+    state: mx.array,
+    mask: Optional[mx.array] = None,
+) -> Tuple[mx.array, mx.array]:
+    B, T, Hk, Dk = k.shape
+    Hv, Dv = v.shape[2:]
+    input_type = q.dtype
+    state_type = state.dtype
+    if g.ndim == 4:
+        kernel = _gated_delta_kernel_vec
+        inputs = [q, k, v, g, beta, state, T]
+        if mask is not None:
+            kernel = _gated_delta_kernel_vec_masked
+            inputs.append(mask)
+    else:
+        kernel = _gated_delta_kernel
+        inputs = [q, k, v, g, beta, state, T]
+        if mask is not None:
+            kernel = _gated_delta_kernel_masked
+            inputs.append(mask)
+
+    return kernel(
+        inputs=inputs,
+        template=[
+            ("InT", input_type),
+            ("StT", state_type),
+            ("Dk", Dk),
+            ("Dv", Dv),
+            ("Hk", Hk),
+            ("Hv", Hv),
+        ],
+        grid=(32, Dv, B * Hv),
+        threadgroup=(32, 4, 1),
+        output_shapes=[(B, T, Hv, Dv), state.shape],
+        output_dtypes=[input_type, state_type],
+    )
+
+
+def gated_delta_ops(
+    q: mx.array,
+    k: mx.array,
+    v: mx.array,
+    g: mx.array,
+    beta: mx.array,
+    state: Optional[mx.array] = None,
+    mask: Optional[mx.array] = None,
+) -> Tuple[mx.array, mx.array]:
+    """
+    Ops-based reference implementation for prompt prefill (sequential loop).
+    Supports both scalar and vectorized gating.
+
+    Shapes:
+      - q, k: [B, T, Hk, Dk]
+      - v: [B, T, Hv, Dv]
+      - g: [B, T, Hv] (scalar) or [B, T, Hv, Dk] (vectorized)
+      - beta: [B, T, Hv]
+      - state: [B, Hv, Dv, Dk]
+    Returns:
+      - y: [B, T, Hv, Dv]
+      - state: [B, Hv, Dv, Dk]
+    """
+    B, T, Hk, Dk = q.shape
+    Hv, Dv = v.shape[-2:]
+    if state is None:
+        state = mx.zeros((B, Hv, Dv, Dk), dtype=mx.float32)
+
+    if (repeat_factor := Hv // Hk) > 1:
+        q = mx.repeat(q, repeat_factor, -2)
+        k = mx.repeat(k, repeat_factor, -2)
+
+    ys = []
+    for t in range(T):
+        y, state = _gated_delta_step_ops(
+            q[:, t],
+            k[:, t],
+            v[:, t],
+            g[:, t],
+            beta[:, t],
+            state,
+            None if mask is None else mask[:, t],
+        )
+        ys.append(y)
+    y = mx.stack(ys, axis=1)
+    return y, state
+
+
+def gated_delta_update(
+    q: mx.array,
+    k: mx.array,
+    v: mx.array,
+    a: mx.array,
+    b: mx.array,
+    A_log: mx.array,
+    dt_bias: mx.array,
+    state: Optional[mx.array] = None,
+    mask: Optional[mx.array] = None,
+    use_kernel: bool = True,
+) -> Tuple[mx.array, mx.array]:
+    beta = mx.sigmoid(b)
+    g = compute_g(A_log, a, dt_bias)
+    if state is None:
+        B, _, Hk, Dk = q.shape
+        Hv, Dv = v.shape[-2:]
+        state = mx.zeros((B, Hv, Dv, Dk), dtype=mx.float32)
+
+    if not use_kernel or mx.default_device() != mx.gpu or not mx.metal.is_available():
+        return gated_delta_ops(q, k, v, g, beta, state, mask)
+    return gated_delta_kernel(q, k, v, g, beta, state, mask)

mlx_lm/models/gemma.py

@@ -138,18 +138,16 @@ class GemmaModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
         h = h * (self.args.hidden_size**0.5)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -166,10 +164,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.embed_tokens.as_linear(out)
         return out
 

mlx_lm/models/gemma2.py

@@ -165,18 +165,16 @@ class GemmaModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
         h = h * (self.args.hidden_size**0.5)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache, return_array=True)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0], return_array=True)
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -194,10 +192,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.embed_tokens.as_linear(out)
         out = mx.tanh(out / self.final_logit_softcapping)
         out = out * self.final_logit_softcapping

mlx_lm/models/gemma3.py

@@ -40,11 +40,10 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache=None,
-        mask: Optional[mx.array] = None,
         input_embeddings: Optional[mx.array] = None,
     ):
         return self.language_model(
-            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+            inputs, cache=cache, input_embeddings=input_embeddings
         )
 
     def sanitize(self, weights):

mlx_lm/models/gemma3_text.py

@@ -2,13 +2,14 @@
 
 from dataclasses import dataclass
 from functools import partial
-from typing import Any, Optional
+from typing import Any, Dict, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
 
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import KVCache, RotatingKVCache
+from .rope_utils import initialize_rope
 
 
 @dataclass
@@ -22,12 +23,13 @@ class ModelArgs(BaseModelArgs):
     rms_norm_eps: float = 1.0e-6
     vocab_size: int = 262144
     num_key_value_heads: int = 1
-    rope_global_base_freq: float = 1_000_000.0
+    rope_theta: float = 1_000_000.0
     rope_local_base_freq: float = 10_000.0
-    rope_traditional: bool = False
     query_pre_attn_scalar: float = 256
     sliding_window: int = 512
     sliding_window_pattern: int = 6
+    max_position_embeddings: int = 32768
+    rope_scaling: Dict = None
 
 
 class Attention(nn.Module):
@@ -52,15 +54,20 @@ class Attention(nn.Module):
         self.k_norm = RMSNorm(dims=head_dim, eps=args.rms_norm_eps)
         self.is_sliding = (layer_idx + 1) % args.sliding_window_pattern != 0
 
-        self.rope = nn.RoPE(
-            head_dim,
-            traditional=args.rope_traditional,
-            base=(
-                args.rope_local_base_freq
-                if self.is_sliding
-                else args.rope_global_base_freq
-            ),
-        )
+        if self.is_sliding:
+            self.rope = initialize_rope(
+                dims=head_dim,
+                base=args.rope_local_base_freq,
+                traditional=False,
+            )
+        else:
+            self.rope = initialize_rope(
+                dims=head_dim,
+                base=args.rope_theta,
+                traditional=False,
+                max_position_embeddings=args.max_position_embeddings,
+                scaling_config=args.rope_scaling,
+            )
 
     def __call__(
         self,
@@ -87,8 +94,6 @@ class Attention(nn.Module):
             keys = self.rope(keys)
 
         # Sliding window
-        if isinstance(mask, mx.array) and mask.shape[-1] != keys.shape[-2]:
-            mask = mask[..., -keys.shape[-2] :]
         output = scaled_dot_product_attention(
             queries, keys, values, cache=cache, scale=self.scale, mask=mask
         )
@@ -160,6 +165,8 @@ class Gemma3Model(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
         self.args = args
+        self.window_size = args.sliding_window
+        self.sliding_window_pattern = args.sliding_window_pattern
         self.vocab_size = args.vocab_size
         self.num_hidden_layers = args.num_hidden_layers
         assert self.vocab_size > 0
@@ -173,7 +180,6 @@ class Gemma3Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
@@ -186,24 +192,22 @@ class Gemma3Model(nn.Module):
         if cache is None:
             cache = [None] * len(self.layers)
 
-        if mask is None:
-            j = self.args.sliding_window_pattern
-            full_mask = create_attention_mask(h, cache[j - 1 : j])
-            sliding_window_mask = create_attention_mask(h, cache)
+        global_mask = create_attention_mask(h, cache[self.sliding_window_pattern - 1])
 
+        if self.sliding_window_pattern > 1:
+            sliding_window_mask = create_attention_mask(
+                h,
+                cache[0],
+                window_size=self.window_size,
+            )
+        else:
+            sliding_window_mask = None
         for i, (layer, c) in enumerate(zip(self.layers, cache)):
             is_global = (
-                i % self.args.sliding_window_pattern
-                == self.args.sliding_window_pattern - 1
+                i % self.sliding_window_pattern == self.sliding_window_pattern - 1
             )
-
-            local_mask = mask
-            if mask is None and is_global:
-                local_mask = full_mask
-            elif mask is None:
-                local_mask = sliding_window_mask
-
-            h = layer(h, local_mask, c)
+            mask = global_mask if is_global else sliding_window_mask
+            h = layer(h, mask, c)
 
         return self.norm(h)
 
@@ -221,10 +225,9 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache=None,
-        mask: Optional[mx.array] = None,
         input_embeddings: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, mask, cache, input_embeddings)
+        out = self.model(inputs, cache, input_embeddings)
         if self.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:
@@ -250,7 +253,5 @@ class Model(nn.Module):
             ):
                 caches.append(KVCache())
             else:
-                caches.append(
-                    RotatingKVCache(max_size=self.args.sliding_window, keep=0)
-                )
+                caches.append(RotatingKVCache(max_size=self.args.sliding_window))
         return caches

mlx_lm/models/gemma3n.py

@@ -151,9 +151,6 @@ class Gemma3nAttention(nn.Module):
         queries = queries.transpose(0, 2, 1, 3)
         queries = self.rope(queries, offset=offset)
 
-        if isinstance(mask, mx.array) and mask.shape[-1] != keys.shape[-2]:
-            mask = mask[:, : keys.shape[-2]]
-
         output = scaled_dot_product_attention(
             queries, keys, values, cache=cache, scale=self.scale, mask=mask
         )
@@ -273,12 +270,11 @@ class Gemma3nAltUp(nn.Module):
             )
 
         all_coefs = self.correction_coefs(modalities) + 1.0
-
         active_x = predictions[self.config.altup_active_idx]
         innovation = activated - active_x
 
-        all_coefs = all_coefs.transpose(2, 1, 0)
-        corrected = innovation[None] * all_coefs[:, None]
+        all_coefs = all_coefs.moveaxis(2, 0)
+        corrected = innovation[None] * all_coefs[..., None]
         corrected += predictions
 
         return corrected.astype(activated.dtype)
@@ -310,7 +306,6 @@ class Gemma3nDecoderLayer(nn.Module):
             eps=config.rms_norm_eps,
         )
         self.is_sliding = self.self_attn.is_sliding
-        self.sliding_window = config.sliding_window
 
         self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
 
@@ -355,7 +350,6 @@ class Gemma3nDecoderLayer(nn.Module):
         attn_ffw = self.mlp(attn_norm)
         attn_ffw_norm = self.post_feedforward_layernorm(attn_ffw)
         attn_ffw_laurel_gated = attn_laurel + attn_ffw_norm
-
         corrected_predictions = self.altup.correct(predictions, attn_ffw_laurel_gated)
 
         first_prediction = corrected_predictions[self.config.altup_active_idx]
@@ -437,10 +431,11 @@ class LanguageModel(nn.Module):
             eps=config.rms_norm_eps,
         )
 
-        self.first_sliding_idx = self.config.layer_types.index("sliding_attention")
-        self.first_full_idx = self.config.layer_types.index("full_attention")
+        self.first_sliding_idx = config.layer_types.index("sliding_attention")
+        self.first_full_idx = config.layer_types.index("full_attention")
+        self.sliding_window = config.sliding_window
 
-        concrete_layers = self.config.layer_types[: self.first_kv_shared_layer_idx]
+        concrete_layers = config.layer_types[: self.first_kv_shared_layer_idx]
         shared_full_idx = (
             len(concrete_layers) - 1 - concrete_layers[::-1].index("full_attention")
         )
@@ -463,7 +458,6 @@ class LanguageModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array = None,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: mx.array = None,
     ):
@@ -476,17 +470,17 @@ class LanguageModel(nn.Module):
         per_layer_inputs = self.project_per_layer_inputs(h, per_layer_inputs)
 
         if cache is None:
-            cache = self.make_cache()
+            cache = [None] * len(self.layers)
 
-        if mask is None:
-            full_mask = create_attention_mask(
-                h,
-                cache[self.first_full_idx :],
-            )
-            sliding_window_mask = create_attention_mask(
-                h,
-                cache[self.first_sliding_idx :],
-            )
+        global_mask = create_attention_mask(
+            h,
+            cache[self.first_full_idx],
+        )
+        sliding_window_mask = create_attention_mask(
+            h,
+            cache[self.first_sliding_idx],
+            window_size=self.sliding_window,
+        )
         h0 = h
 
         # Expand hidden_states to support per-layer inputs
@@ -497,21 +491,19 @@ class LanguageModel(nn.Module):
         h = mx.stack(h_list, axis=0)
         mags = mx.mean(h[1:] ** 2, axis=-1, keepdims=True) ** 0.5
         h[1:] = h[1:] * (target_magnitude / mx.maximum(mags, mx.finfo(h0.dtype).min))
-
         for i, layer in enumerate(self.layers):
             per_layer_input = per_layer_inputs[:, :, i, :]
 
             is_global = self.config.layer_types[i] == "full_attention"
 
-            local_mask = mask
-            if mask is None and is_global:
-                local_mask = full_mask
-            elif mask is None:
-                local_mask = sliding_window_mask
+            if is_global:
+                mask = global_mask
+            else:
+                mask = sliding_window_mask
 
             h = layer(
                 h,
-                local_mask,
+                mask,
                 cache[self.layer_idx_to_cache_idx[i]],
                 per_layer_input,
             )
@@ -582,11 +574,10 @@ class Gemma3n(nn.Module):
         self,
         inputs: mx.array,
         cache=None,
-        mask: Optional[mx.array] = None,
         input_embeddings: Optional[mx.array] = None,
     ):
         return self.language_model(
-            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+            inputs, cache=cache, input_embeddings=input_embeddings
         )
 
     def make_cache(self):
@@ -598,17 +589,15 @@ class Model(nn.Module):
         super().__init__()
         self.args = args
         self.model = Gemma3n(args)
+        self.model_type = args.model_type
 
     def __call__(
         self,
         inputs: mx.array,
         cache=None,
-        mask: Optional[mx.array] = None,
         input_embeddings: Optional[mx.array] = None,
     ):
-        return self.model(
-            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
-        )
+        return self.model(inputs, cache=cache, input_embeddings=input_embeddings)
 
     def sanitize(self, weights):
         weights = tree_unflatten(list(weights.items()))

mlx_lm/models/gemma4.py

@@ -0,0 +1,92 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from . import gemma4_text
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "gemma4"
+    text_config: dict = None
+    vocab_size: int = 262144
+
+    def __post_init__(self):
+        if self.text_config is None:
+            self.text_config = {}
+        self.text_config["vocab_size"] = self.vocab_size
+        self.text_config["num_attention_heads"] = self.text_config.get(
+            "num_attention_heads", 8
+        )
+        self.text_config["num_key_value_heads"] = self.text_config.get(
+            "num_key_value_heads", 1
+        )
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = gemma4_text.Model(
+            gemma4_text.ModelArgs.from_dict(args.text_config)
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+        per_layer_inputs: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs,
+            cache=cache,
+            input_embeddings=input_embeddings,
+            per_layer_inputs=per_layer_inputs,
+        )
+
+    def sanitize(self, weights):
+        new_weights = {}
+        for k, v in weights.items():
+            starts_w_model = k.startswith("model.")
+
+            k = k.removeprefix("model.")
+            if k.startswith(
+                (
+                    "vision_tower",
+                    "multi_modal_projector",
+                    "audio_tower",
+                    "embed_audio",
+                    "embed_vision",
+                )
+            ):
+                continue
+
+            if not starts_w_model:
+                new_weights[k] = v
+                continue
+
+            if k.startswith("language_model"):
+                k = k.replace("language_model.", "language_model.model.")
+
+            new_weights[k] = v
+
+        return self.language_model.sanitize(new_weights)
+
+    @property
+    def layers(self):
+        return self.language_model.layers
+
+    @property
+    def quant_predicate(self):
+        return self.language_model.quant_predicate
+
+    def make_cache(self):
+        return self.language_model.make_cache()

mlx_lm/models/gemma4_text.py

@@ -0,0 +1,688 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache, _BaseCache
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "gemma4_text"
+    hidden_size: int = 1536
+    num_hidden_layers: int = 35
+    intermediate_size: int = 6144
+    num_attention_heads: int = 8
+    head_dim: int = 256
+    global_head_dim: int = 512
+    global_partial_rotary_factor: float = 0.25
+    rms_norm_eps: float = 1e-6
+    vocab_size: int = 262144
+    vocab_size_per_layer_input: int = 262144
+    num_key_value_heads: int = 1
+    num_global_key_value_heads: Optional[int] = None
+    num_kv_shared_layers: int = 20
+    pad_token_id: int = 0
+    hidden_size_per_layer_input: int = 256
+    rope_traditional: bool = False
+    partial_rotary_factor: float = 1.0
+    rope_parameters: Optional[Dict] = None
+    sliding_window: int = 512
+    sliding_window_pattern: int = 5
+    max_position_embeddings: int = 131072
+    attention_k_eq_v: bool = False
+    final_logit_softcapping: float = 30.0
+    use_double_wide_mlp: bool = True
+    enable_moe_block: bool = False
+    num_experts: Optional[int] = None
+    top_k_experts: Optional[int] = None
+    moe_intermediate_size: Optional[int] = None
+    layer_types: Optional[List[str]] = None
+    tie_word_embeddings: bool = True
+
+    def __post_init__(self):
+        if self.rope_parameters is None:
+            self.rope_parameters = {
+                "full_attention": {
+                    "partial_rotary_factor": 0.25,
+                    "rope_theta": 1000000.0,
+                    "rope_type": "proportional",
+                },
+                "sliding_attention": {
+                    "partial_rotary_factor": 1.0,
+                    "rope_theta": 10000.0,
+                    "rope_type": "default",
+                },
+            }
+        if self.layer_types is None:
+            pattern = ["sliding_attention"] * (self.sliding_window_pattern - 1) + [
+                "full_attention"
+            ]
+            self.layer_types = (pattern * (self.num_hidden_layers // len(pattern) + 1))[
+                : self.num_hidden_layers
+            ]
+
+
+class RMSNormNoScale(nn.Module):
+    """RMSNorm without learnable scale."""
+
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return mx.fast.rms_norm(x, None, self.eps)
+
+
+@partial(mx.compile, shapeless=True)
+def logit_softcap(softcap, x):
+    return mx.tanh(x / softcap) * softcap
+
+
+@partial(mx.compile, shapeless=True)
+def _complete_square(x2, y2, xy):
+    return x2 + mx.expand_dims(y2, -1) - 2 * xy
+
+
+@partial(mx.compile, shapeless=True)
+def geglu(gate, x):
+    return nn.gelu_approx(gate) * x
+
+
+class MLP(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int = 0):
+        super().__init__()
+        first_kv_shared_layer_idx = (
+            config.num_hidden_layers - config.num_kv_shared_layers
+        )
+        is_kv_shared_layer = layer_idx >= first_kv_shared_layer_idx > 0
+        use_double_wide = config.use_double_wide_mlp and is_kv_shared_layer
+        intermediate_size = config.intermediate_size * (2 if use_double_wide else 1)
+
+        self.gate_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, config.hidden_size, bias=False)
+        self.up_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(geglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class Router(nn.Module):
+    """Expert router: norm -> scale -> project -> top-k -> renormalize."""
+
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.eps = config.rms_norm_eps
+        self.proj = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.scale = mx.ones((config.hidden_size,))
+        self.per_expert_scale = mx.ones((config.num_experts,))
+        self._root_size = config.hidden_size**-0.5
+
+    def __call__(self, x: mx.array):
+        x = mx.fast.rms_norm(x, self.scale * self._root_size, self.eps)
+
+        expert_scores = self.proj(x)
+
+        top_k_indices = mx.argpartition(
+            expert_scores, kth=-self.config.top_k_experts, axis=-1
+        )
+        top_k_indices = top_k_indices[..., -self.config.top_k_experts :]
+
+        top_k_weights = mx.take_along_axis(expert_scores, top_k_indices, axis=-1)
+        top_k_weights = mx.softmax(top_k_weights, axis=-1)
+        top_k_weights = top_k_weights * self.per_expert_scale[top_k_indices]
+
+        return top_k_indices, top_k_weights
+
+
+class GeGLU(nn.Module):
+    """GELU-gated linear unit activation for SwitchGLU."""
+
+    def __call__(self, x, gate):
+        return geglu(gate, x)
+
+
+class Experts(nn.Module):
+    """Sparse MoE using SwitchGLU with gather_mm."""
+
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+
+        self.switch_glu = SwitchGLU(
+            input_dims=config.hidden_size,
+            hidden_dims=config.moe_intermediate_size,
+            num_experts=config.num_experts,
+            activation=GeGLU(),
+            bias=False,
+        )
+
+    def __call__(
+        self, x: mx.array, top_k_indices: mx.array, top_k_weights: mx.array
+    ) -> mx.array:
+        w = mx.expand_dims(top_k_weights, -1)
+        y = self.switch_glu(x, top_k_indices)
+
+        return (w * y).sum(-2)
+
+
+class Attention(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.layer_type = config.layer_types[layer_idx]
+        self.is_sliding = self.layer_type == "sliding_attention"
+        self.has_kv = layer_idx < config.num_hidden_layers - config.num_kv_shared_layers
+
+        self.head_dim = (
+            config.global_head_dim
+            if self.layer_type == "full_attention"
+            and hasattr(config, "global_head_dim")
+            and config.global_head_dim
+            else config.head_dim
+        )
+
+        dim = config.hidden_size
+        self.n_heads = config.num_attention_heads
+
+        # K-eq-V for full attention layers (26B/31B models)
+        self.use_k_eq_v = config.attention_k_eq_v and not self.is_sliding
+        if self.use_k_eq_v and config.num_global_key_value_heads is not None:
+            self.n_kv_heads = config.num_global_key_value_heads
+        else:
+            self.n_kv_heads = config.num_key_value_heads
+
+        self.scale = 1.0
+
+        self.q_proj = nn.Linear(dim, self.n_heads * self.head_dim, bias=False)
+        if self.has_kv:
+            self.k_proj = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+            if not self.use_k_eq_v:
+                self.v_proj = nn.Linear(
+                    dim, self.n_kv_heads * self.head_dim, bias=False
+                )
+        self.o_proj = nn.Linear(self.n_heads * self.head_dim, dim, bias=False)
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        if self.has_kv:
+            self.k_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+            self.v_norm = RMSNormNoScale(self.head_dim, eps=config.rms_norm_eps)
+
+        # RoPE (with partial rotation support)
+        layer_key = "sliding_attention" if self.is_sliding else "full_attention"
+        rope_params = config.rope_parameters.get(layer_key, {})
+        rope_theta = rope_params.get("rope_theta", 10000.0)
+        self.rope = initialize_rope(
+            dims=self.head_dim,
+            traditional=config.rope_traditional,
+            base=rope_theta,
+            scaling_config=rope_params,
+            max_position_embeddings=config.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        shared_kv: Optional[tuple] = None,
+        offset: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        queries = self.q_proj(x).reshape(B, L, self.n_heads, self.head_dim)
+        queries = self.q_norm(queries)
+
+        if shared_kv is not None:
+            keys, values = shared_kv
+        elif not self.has_kv:
+            raise ValueError(
+                f"Layer {self.layer_idx} is a KV-shared layer but received no shared_kv"
+            )
+        else:
+            keys = self.k_proj(x).reshape(B, L, self.n_kv_heads, self.head_dim)
+            values = keys
+            if not self.use_k_eq_v:
+                values = self.v_proj(x).reshape(B, L, self.n_kv_heads, self.head_dim)
+
+            offset = mx.array(cache.offset) if cache is not None else 0
+
+            keys = self.k_norm(keys)
+            keys = keys.transpose(0, 2, 1, 3)
+            keys = self.rope(keys, offset=offset)
+
+            values = self.v_norm(values)
+            values = values.transpose(0, 2, 1, 3)
+
+        queries = queries.transpose(0, 2, 1, 3)
+        queries = self.rope(queries, offset=offset)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.o_proj(output), (keys, values), offset
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.layer_type = config.layer_types[layer_idx]
+        self.self_attn = Attention(config, layer_idx)
+        self.mlp = MLP(config, layer_idx)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.pre_feedforward_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_feedforward_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+        # MoE (26B model)
+        self.enable_moe = config.enable_moe_block
+        if self.enable_moe:
+            self.router = Router(config)
+            self.experts = Experts(config)
+            self.post_feedforward_layernorm_1 = nn.RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+            self.post_feedforward_layernorm_2 = nn.RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+            self.pre_feedforward_layernorm_2 = nn.RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+
+        # Per-layer input gating (2B/4B models)
+        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
+        if self.hidden_size_per_layer_input:
+            self.per_layer_input_gate = nn.Linear(
+                config.hidden_size, self.hidden_size_per_layer_input, bias=False
+            )
+            self.per_layer_projection = nn.Linear(
+                self.hidden_size_per_layer_input, config.hidden_size, bias=False
+            )
+            self.post_per_layer_input_norm = nn.RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+        else:
+            self.per_layer_input_gate = None
+            self.per_layer_projection = None
+            self.post_per_layer_input_norm = None
+
+        # Layer scalar
+        self.layer_scalar = mx.ones((1,))
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        per_layer_input: Optional[mx.array] = None,
+        shared_kv: Optional[tuple] = None,
+        offset: Optional[Any] = None,
+    ) -> mx.array:
+        residual = x
+
+        h = self.input_layernorm(x)
+        h, shared_kv, offset = self.self_attn(
+            h, mask, cache, shared_kv=shared_kv, offset=offset
+        )
+        h = self.post_attention_layernorm(h)
+        h = residual + h
+
+        residual = h
+
+        if self.enable_moe:
+            h1 = self.pre_feedforward_layernorm(h)
+            h1 = self.mlp(h1)
+            h1 = self.post_feedforward_layernorm_1(h1)
+
+            top_k_indices, top_k_weights = self.router(h)
+            h2 = self.pre_feedforward_layernorm_2(h)
+            h2 = self.experts(h2, top_k_indices, top_k_weights)
+            h2 = self.post_feedforward_layernorm_2(h2)
+
+            h = h1 + h2
+        else:
+            h = self.pre_feedforward_layernorm(h)
+            h = self.mlp(h)
+
+        h = self.post_feedforward_layernorm(h)
+        h = residual + h
+
+        # Per-layer input gating
+        if (
+            self.per_layer_input_gate is not None
+            and self.per_layer_projection is not None
+            and self.post_per_layer_input_norm is not None
+            and per_layer_input is not None
+        ):
+            residual = h
+            gate = self.per_layer_input_gate(h)
+            gate = nn.gelu_approx(gate)
+            gate = mx.multiply(gate, per_layer_input)
+            gate = self.per_layer_projection(gate)
+            gate = self.post_per_layer_input_norm(gate)
+            h = residual + gate
+
+        if self.layer_scalar is not None:
+            h = h * self.layer_scalar
+
+        return h, shared_kv, offset
+
+
+class Gemma4TextModel(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.window_size = config.sliding_window
+        self.sliding_window_pattern = config.sliding_window_pattern
+        self.num_hidden_layers = config.num_hidden_layers
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.embed_scale = config.hidden_size**0.5
+        self.layers = [
+            DecoderLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # Per-layer input embeddings (2B/4B models)
+        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
+        if self.hidden_size_per_layer_input:
+            self.embed_tokens_per_layer = nn.Embedding(
+                config.vocab_size_per_layer_input,
+                config.num_hidden_layers * config.hidden_size_per_layer_input,
+            )
+            self.embed_tokens_per_layer_scale = config.hidden_size_per_layer_input**0.5
+            self.per_layer_input_scale = 2.0**-0.5
+            self.per_layer_projection_scale = config.hidden_size**-0.5
+            self.per_layer_model_projection = nn.Linear(
+                config.hidden_size,
+                config.num_hidden_layers * config.hidden_size_per_layer_input,
+                bias=False,
+            )
+            self.per_layer_projection_norm = nn.RMSNorm(
+                config.hidden_size_per_layer_input, eps=config.rms_norm_eps
+            )
+        else:
+            self.embed_tokens_per_layer = None
+            self.per_layer_input_scale = None
+            self.per_layer_projection_scale = None
+            self.per_layer_model_projection = None
+            self.per_layer_projection_norm = None
+
+        # Arrange for shared KVs
+        self.previous_kvs = list(range(len(self.layers)))
+        if config.num_kv_shared_layers > 0:
+            N = len(self.layers)
+            M = N - config.num_kv_shared_layers
+            kvs_by_type = {}
+            for i in range(M):
+                kvs_by_type[self.layers[i].layer_type] = i
+            for j in range(M, N):
+                self.previous_kvs[j] = kvs_by_type[self.layers[j].layer_type]
+
+    def _get_per_layer_inputs(
+        self,
+        input_ids: Optional[mx.array],
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        if input_ids is None:
+            if input_embeddings is None:
+                raise RuntimeError(
+                    "input_embeddings must be provided when input_ids are omitted."
+                )
+
+            # Split the sequence dimension if this still holds too much
+            # memory. 260k vocab means the distance tensor would be ~1GB
+            # per 2k tokens in bf16.
+            #
+            # If the embedding is quantized we have to dequantize it anyway to
+            # perform the match test.
+            norms_embedding = self.embed_tokens.weight.square().sum(-1)
+            norms_input = input_embeddings.square().sum(-1)
+            distance = _complete_square(
+                norms_embedding,
+                norms_input,
+                self.embed_tokens.as_linear(input_embeddings),
+            )
+
+            # Checks can be added if needed but they necessarily break the GPU
+            # pipelining and force an eval.
+            #
+            #   match_counts = (distance < eps).sum(-1)
+            #
+            input_ids = mx.argmin(distance, -1)
+
+        result = self.embed_tokens_per_layer(input_ids)
+        result = result * self.embed_tokens_per_layer_scale
+        return mx.unflatten(
+            result,
+            -1,
+            (self.config.num_hidden_layers, self.hidden_size_per_layer_input),
+        )
+
+    def _project_per_layer_inputs(
+        self,
+        input_embeddings: mx.array,
+        per_layer_inputs: Optional[mx.array] = None,
+    ) -> mx.array:
+        per_layer_projection = self.per_layer_model_projection(input_embeddings)
+        per_layer_projection = per_layer_projection * self.per_layer_projection_scale
+        per_layer_projection = mx.unflatten(
+            per_layer_projection,
+            -1,
+            (self.config.num_hidden_layers, self.hidden_size_per_layer_input),
+        )
+        per_layer_projection = self.per_layer_projection_norm(per_layer_projection)
+
+        if per_layer_inputs is None:
+            return per_layer_projection
+
+        return (per_layer_projection + per_layer_inputs) * self.per_layer_input_scale
+
+    def _make_masks(self, h, cache):
+        mask = {}
+        masks = []
+        for l, c in zip(self.layers, cache):
+            if l.layer_type not in mask:
+                if l.layer_type == "full_attention":
+                    mask["full_attention"] = create_attention_mask(h, c)
+                elif l.layer_type == "sliding_attention":
+                    mask["sliding_attention"] = create_attention_mask(
+                        h, c, window_size=self.window_size
+                    )
+            masks.append(mask[l.layer_type])
+        return masks
+
+    def __call__(
+        self,
+        inputs: mx.array = None,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+        per_layer_inputs: Optional[mx.array] = None,
+    ):
+        # Make the initial hidden state
+        if input_embeddings is None:
+            input_embeddings = self.embed_tokens(inputs)
+        h = input_embeddings
+        h = h * self.embed_scale
+
+        # Get the extra inputs per layer if we have per layer embeddings
+        if self.hidden_size_per_layer_input:
+            if per_layer_inputs is None:
+                per_layer_inputs = self._get_per_layer_inputs(inputs, input_embeddings)
+            per_layer_inputs = self._project_per_layer_inputs(h, per_layer_inputs)
+        if per_layer_inputs is not None:
+            per_layer_inputs = [
+                per_layer_inputs[:, :, i, :] for i, _ in enumerate(self.layers)
+            ]
+        else:
+            per_layer_inputs = [None] * len(self.layers)
+
+        # Make the kv cache list, be sure to append None for all the shared kv
+        # layers
+        if cache is None:
+            cache = [None] * len(self.layers)
+        else:
+            cache = cache + [None] * (len(self.layers) - len(cache))
+
+        # Apply each layer. We save all intermediate kvs and offset and grab
+        # the previous one for the shared kv layers.
+        masks = self._make_masks(h, cache)
+        intermediates = [(None, None)] * len(self.layers)
+        for idx, (layer, c, mask, prev_idx, per_layer_input) in enumerate(
+            zip(
+                self.layers,
+                cache,
+                masks,
+                self.previous_kvs,
+                per_layer_inputs,
+            )
+        ):
+            kvs, offset = intermediates[prev_idx]
+
+            h, kvs, offset = layer(
+                h,
+                mask,
+                c,
+                per_layer_input=per_layer_input,
+                shared_kv=kvs,
+                offset=offset,
+            )
+
+            intermediates[idx] = (kvs, offset)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Gemma4TextModel(args)
+        self.final_logit_softcapping = args.final_logit_softcapping
+        self.tie_word_embeddings = args.tie_word_embeddings
+        if not self.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+        per_layer_inputs: Optional[mx.array] = None,
+    ):
+        out = self.model(
+            inputs,
+            cache=cache,
+            input_embeddings=input_embeddings,
+            per_layer_inputs=per_layer_inputs,
+        )
+        if self.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        if self.final_logit_softcapping is not None:
+            out = logit_softcap(self.final_logit_softcapping, out)
+        return out
+
+    def sanitize(self, weights):
+        sanitized = {}
+        first_kv_shared = self.args.num_hidden_layers - self.args.num_kv_shared_layers
+        for k, v in weights.items():
+            if any(
+                s in k
+                for s in (
+                    "self_attn.rotary_emb",
+                    "input_max",
+                    "input_min",
+                    "output_max",
+                    "output_min",
+                )
+            ):
+                continue
+
+            # KV-shared layers reuse K/V from earlier layers — drop their projections
+            if any(
+                s in k
+                for s in (".self_attn.k_proj", ".self_attn.v_proj", ".self_attn.k_norm")
+            ):
+                try:
+                    layer_idx = int(k.split("layers.")[1].split(".")[0])
+                    if layer_idx >= first_kv_shared:
+                        continue
+                except (IndexError, ValueError):
+                    pass
+
+            if k.endswith(".experts.gate_up_proj"):
+                base = k.removesuffix(".gate_up_proj")
+                gate, up = map(mx.contiguous, mx.split(v, 2, axis=-2))
+                sanitized[f"{base}.switch_glu.gate_proj.weight"] = gate
+                sanitized[f"{base}.switch_glu.up_proj.weight"] = up
+                continue
+
+            if k.endswith(".experts.down_proj"):
+                base = k.removesuffix(".down_proj")
+                sanitized[f"{base}.switch_glu.down_proj.weight"] = v
+                continue
+
+            sanitized[k] = v
+
+        return sanitized
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("router.proj"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.args.head_dim
+
+    @property
+    def n_kv_heads(self):
+        return self.args.num_key_value_heads
+
+    def make_cache(self):
+        first_kv_shared = self.args.num_hidden_layers - self.args.num_kv_shared_layers
+        caches = []
+        for i in range(first_kv_shared):
+            if self.args.layer_types[i] == "full_attention":
+                caches.append(KVCache())
+            else:
+                caches.append(
+                    RotatingKVCache(
+                        max_size=self.args.sliding_window,
+                        keep=0,
+                    )
+                )
+        return caches

mlx_lm/models/glm.py

@@ -0,0 +1,188 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    head_dim: int
+    num_key_value_heads: int
+    max_position_embeddings: Optional[int] = None
+    attention_bias: bool = False
+    rope_theta: float = 10000
+    tie_word_embeddings: bool = True
+
+
+class GLMAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim or args.hidden_size // self.num_attention_heads
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            self.hidden_size,
+            self.num_attention_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.o_proj = nn.Linear(
+            self.num_attention_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.rope = nn.RoPE(dims=self.head_dim, traditional=True, base=args.rope_theta)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.num_attention_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+        keys = keys.reshape(B, L, self.num_key_value_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class GLMMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.gate_up_proj = nn.Linear(
+            args.hidden_size, 2 * args.intermediate_size, bias=False
+        )
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        x = self.gate_up_proj(x)
+        gate, x = mx.split(x, 2, axis=-1)
+        return self.down_proj(swiglu(gate, x))
+
+
+class GLMBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = GLMAttention(args)
+        self.mlp = GLMMLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class GLMModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [GLMBlock(args=args) for _ in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = GLMModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        weights = {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers

mlx_lm/models/glm4.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -38,7 +39,7 @@ class Glm4MLP(nn.Module):
     def __call__(self, x) -> mx.array:
         x = self.gate_up_proj(x)
         gate, up_states = mx.split(x, 2, axis=-1)
-        return self.down_proj(nn.silu(gate) * up_states)
+        return self.down_proj(swiglu(gate, up_states))
 
 
 class Glm4Attention(nn.Module):
@@ -144,17 +145,15 @@ class Glm4Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -172,10 +171,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
     @property

mlx_lm/models/glm4_moe.py

@@ -7,8 +7,11 @@ from typing import Any, Dict, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .pipeline import PipelineMixin
 from .switch_layers import SwitchGLU
 
 
@@ -121,7 +124,7 @@ class MLP(nn.Module):
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
 
     def __call__(self, x):
-        down_proj = self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
         return down_proj
 
 
@@ -140,7 +143,6 @@ def group_expert_select(
     orig_scores = scores
     scores = scores + e_score_correction_bias
     if n_group > 1:
-        k = top_k
         scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
         group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
         k = n_group - topk_group
@@ -205,13 +207,21 @@ class MoE(nn.Module):
                 config=config, intermediate_size=intermediate_size
             )
 
+        self.sharding_group = None
+
     def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
         inds, scores = self.gate(x)
         y = self.switch_mlp(x, inds)
         y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
         if self.config.n_shared_experts is not None:
             y = y + self.shared_experts(x)
 
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
         return y
 
 
@@ -244,7 +254,7 @@ class DecoderLayer(nn.Module):
         return h + r
 
 
-class LanguageModel(nn.Module):
+class LanguageModel(PipelineMixin, nn.Module):
     def __init__(self, config: ModelArgs):
         super().__init__()
         self.vocab_size = config.vocab_size
@@ -252,28 +262,38 @@ class LanguageModel(nn.Module):
         self.layers = [
             DecoderLayer(config, idx) for idx in range(config.num_hidden_layers)
         ]
-        self.start_idx = 0
-        self.end_idx = len(self.layers)
-        self.num_layers = self.end_idx
-
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
     def __call__(
         self,
         x: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ) -> mx.array:
         h = self.embed_tokens(x)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
 
         if cache is None:
-            cache = [None] * self.num_layers
+            cache = [None] * len(self.pipeline_layers)
+        mask = create_attention_mask(h, cache[0])
 
-        for i in range(self.num_layers):
-            h = self.layers[self.start_idx + i](h, mask, cache[i])
+        # Receive from the previous process in the pipeline
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1))
+
+        for l, c in zip(self.pipeline_layers, cache):
+            h = l(h, mask, cache=c)
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+            if cache[-1] is not None:
+                cache[-1].keys = mx.depends(cache[-1].keys, h)
+
+        # Broadcast h while keeping it in the graph
+        if pipeline_size > 1:
+            h = mx.distributed.all_gather(h)[: h.shape[0]]
 
         return self.norm(h)
 
@@ -290,9 +310,8 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, cache, mask)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
     def sanitize(self, weights):
@@ -317,9 +336,64 @@ class Model(nn.Module):
             if not k.startswith(f"model.layers.{mpt_layer}")
         }
 
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        for layer in self.model.layers:
+            # Shard the self attention
+            layer.self_attn.q_proj = shard_linear(
+                layer.self_attn.q_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.k_proj = shard_linear(
+                layer.self_attn.k_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.v_proj = shard_linear(
+                layer.self_attn.v_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.n_heads //= N
+            layer.self_attn.n_kv_heads //= N
+
+            # Shard the MLP
+            if isinstance(layer.mlp, MLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                layer.mlp.sharding_group = group
+                shard_inplace(
+                    layer.mlp.shared_experts.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+
     @property
     def layers(self):
-        return self.model.layers
+        return self.model.pipeline_layers
 
     @property
     def cast_predicate(self):

mlx_lm/models/glm4_moe_lite.py

@@ -0,0 +1,531 @@
+# Copyright © 2026 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .mla import MultiLinear
+from .pipeline import PipelineMixin
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "glm4_moe_lite"
+    vocab_size: int = 154880
+    hidden_size: int = 2048
+    intermediate_size: int = 10240
+    moe_intermediate_size: int = 1536
+    num_hidden_layers: int = 47
+    num_attention_heads: int = 20
+    num_key_value_heads: int = 20
+    n_shared_experts: Optional[int] = 1
+    n_routed_experts: Optional[int] = 64
+    routed_scaling_factor: float = 1.8
+    kv_lora_rank: int = 512
+    q_lora_rank: int = 768
+    qk_rope_head_dim: int = 64
+    qk_nope_head_dim: int = 192
+    v_head_dim: int = 256
+    topk_method: str = "noaux_tc"
+    scoring_func: str = "sigmoid"
+    norm_topk_prob: bool = True
+    n_group: int = 1
+    topk_group: int = 1
+    num_experts_per_tok: int = 4
+    moe_layer_freq: int = 1
+    first_k_dense_replace: int = 1
+    max_position_embeddings: int = 202752
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 1_000_000.0
+    rope_scaling: Optional[Dict] = None
+    attention_bias: bool = False
+    attention_dropout: float = 0.0
+    partial_rotary_factor: float = 1.0
+    tie_word_embeddings: bool = False
+    num_nextn_predict_layers: int = 1
+    quantization: Optional[Dict[str, Any]] = None
+
+
+class Glm4MoeLiteAttention(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        rope_params = config.rope_scaling
+        self.rope_theta = config.rope_theta
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
+
+        self.scale = self.q_head_dim**-0.5
+
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(
+                self.hidden_size, self.num_heads * self.q_head_dim, bias=False
+            )
+        else:
+            self.q_a_proj = nn.Linear(
+                self.hidden_size, self.q_lora_rank, bias=config.attention_bias
+            )
+            self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
+            self.q_b_proj = nn.Linear(
+                self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
+            )
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
+        head_dim = self.qk_nope_head_dim + self.v_head_dim
+        self.embed_q = MultiLinear(
+            self.qk_nope_head_dim, self.kv_lora_rank, self.num_heads
+        )
+        self.unembed_out = MultiLinear(
+            self.kv_lora_rank, self.v_head_dim, self.num_heads
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        if rope_params is not None:
+            mscale_all_dim = rope_params.get("mscale_all_dim", 0)
+            if mscale_all_dim:
+                scaling_factor = rope_params["factor"]
+                if scaling_factor > 1:
+                    s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0
+                    self.scale = self.scale * s * s
+
+        self.rope = initialize_rope(
+            dims=self.qk_rope_head_dim,
+            base=self.rope_theta,
+            traditional=True,
+            max_position_embeddings=self.max_position_embeddings,
+            scaling_config=rope_params,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(x)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
+
+        q = q.reshape(B, L, self.num_heads, self.q_head_dim).transpose(0, 2, 1, 3)
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+        kv_latent = self.kv_a_layernorm(compressed_kv)
+
+        offset = cache.offset if cache is not None else 0
+        q_pe = self.rope(q_pe, offset)
+        k_pe = self.rope(k_pe, offset)
+
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
+
+        if cache is not None:
+            kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe)
+
+        pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2)
+        if mask is not None:
+            pe_scores = mx.where(
+                mask,
+                pe_scores,
+                mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype),
+            )
+
+        if L == 1:
+            q_nope = self.embed_q(q_nope)
+            k = v = kv_latent
+        else:
+            k = self.embed_q(kv_latent, transpose=False)
+            v = self.unembed_out(kv_latent)
+        output = scaled_dot_product_attention(
+            q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores
+        )
+        if L == 1:
+            output = self.unembed_out(output)
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class Glm4MoeLiteMLP(nn.Module):
+    def __init__(
+        self, config: ModelArgs, hidden_size: int = None, intermediate_size: int = None
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x):
+        down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+        return down_proj
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / (denominator + 1e-20)
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
+class MoEGate(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, config.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert config.topk_method == "noaux_tc", "Unsupported topk method."
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class Glm4MoeLiteMoE(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.switch_mlp = SwitchGLU(
+            config.hidden_size,
+            config.moe_intermediate_size,
+            config.n_routed_experts,
+        )
+
+        self.gate = MoEGate(config)
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = Glm4MoeLiteMLP(
+                config=config, intermediate_size=intermediate_size
+            )
+
+        self.sharding_group = None
+
+    def __call__(self, x):
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
+        inds, scores = self.gate(x)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.config.n_shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y
+
+
+class Glm4MoeLiteDecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = Glm4MoeLiteAttention(config)
+        use_moe = (
+            config.n_routed_experts is not None
+            and layer_idx >= config.first_k_dense_replace
+            and layer_idx % config.moe_layer_freq == 0
+        )
+        self.mlp = Glm4MoeLiteMoE(config) if use_moe else Glm4MoeLiteMLP(config)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class Glm4MoeLiteModel(PipelineMixin, nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            Glm4MoeLiteDecoderLayer(config, idx)
+            for idx in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(x)
+
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
+
+        if cache is None:
+            cache = [None] * len(self.pipeline_layers)
+        mask = create_attention_mask(h, cache[0], return_array=True)
+
+        # Receive from the previous process in the pipeline
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1))
+
+        for l, c in zip(self.pipeline_layers, cache):
+            h = l(h, mask, cache=c)
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+            if cache[-1] is not None:
+                cache[-1].keys = mx.depends(cache[-1].keys, h)
+
+        # Broadcast h while keeping it in the graph
+        if pipeline_size > 1:
+            h = mx.distributed.all_gather(h)[: h.shape[0]]
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.model = Glm4MoeLiteModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+    def sanitize(self, weights):
+        def is_mpt_layer(key):
+            subkeys = key.split(".")
+            if len(subkeys) < 3:
+                return False
+            if (
+                subkeys[1] == "layers"
+                and int(subkeys[2]) >= self.args.num_hidden_layers
+            ):
+                return True
+            return False
+
+        new_weights = {}
+        for k, v in weights.items():
+            if is_mpt_layer(k):
+                continue
+            else:
+                new_weights[k] = v
+        weights = new_weights
+
+        # Stack experts
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+            prefix = f"model.layers.{l}.self_attn"
+            if f"{prefix}.kv_b_proj.weight" in weights:
+                layer = self.layers[l].self_attn.embed_q
+                quantized = f"{prefix}.kv_b_proj.scales" in weights
+                v = weights.pop(f"{prefix}.kv_b_proj.weight")
+                head_dim = self.args.qk_nope_head_dim + self.args.v_head_dim
+
+                if quantized:
+                    dims = self.args.kv_lora_rank
+                    scales = weights.pop(f"{prefix}.kv_b_proj.scales")
+                    biases = weights.pop(f"{prefix}.kv_b_proj.biases")
+                    # Try to infer bits and group size
+                    bits = (v.shape[-1] * 32) // dims
+                    group_size = dims // scales.shape[-1]
+                    v = mx.dequantize(
+                        v, scales, biases, bits=bits, group_size=group_size
+                    )
+                num_heads = self.args.num_attention_heads
+                v = v.reshape(num_heads, head_dim, -1)
+                wk = mx.contiguous(
+                    v[:, : self.args.qk_nope_head_dim, :].swapaxes(-1, -2)
+                )
+                wv = mx.contiguous(v[:, self.args.qk_nope_head_dim :, :])
+                if quantized:
+                    wk, wk_scales, wk_biases = mx.quantize(
+                        wk, bits=bits, group_size=group_size
+                    )
+                    wv, wv_scales, wv_biases = mx.quantize(
+                        wv, bits=bits, group_size=group_size
+                    )
+                    weights[f"{prefix}.embed_q.scales"] = wk_scales
+                    weights[f"{prefix}.unembed_out.scales"] = wv_scales
+                    weights[f"{prefix}.embed_q.biases"] = wk_biases
+                    weights[f"{prefix}.unembed_out.biases"] = wv_biases
+                weights[f"{prefix}.embed_q.weight"] = wk
+                weights[f"{prefix}.unembed_out.weight"] = wv
+
+        return weights
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        rank = group.rank()
+        N = group.size()
+        for layer in self.model.layers:
+            # Shard the self attention
+            if layer.self_attn.q_lora_rank is None:
+                layer.self_attn.q_proj = shard_linear(
+                    layer.self_attn.q_proj, "all-to-sharded", group=group
+                )
+            else:
+                layer.self_attn.q_b_proj = shard_linear(
+                    layer.self_attn.q_b_proj, "all-to-sharded", group=group
+                )
+            layer.self_attn.num_heads //= N
+            num_heads = layer.self_attn.num_heads
+            sh = rank * num_heads
+            eh = sh + num_heads
+
+            def shard_heads(w):
+                return w[sh:eh]
+
+            layer.self_attn.embed_q.apply(shard_heads)
+            layer.self_attn.unembed_out.apply(shard_heads)
+
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+
+            # Shard the MLP
+            if isinstance(layer.mlp, Glm4MoeLiteMLP):
+                layer.mlp.gate_proj = shard_linear(
+                    layer.mlp.gate_proj, "all-to-sharded", group=group
+                )
+                layer.mlp.down_proj = shard_linear(
+                    layer.mlp.down_proj, "sharded-to-all", group=group
+                )
+                layer.mlp.up_proj = shard_linear(
+                    layer.mlp.up_proj, "all-to-sharded", group=group
+                )
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                layer.mlp.sharding_group = group
+                if getattr(layer.mlp, "shared_experts", None) is not None:
+                    shard_inplace(
+                        layer.mlp.shared_experts.gate_proj,
+                        "all-to-sharded",
+                        group=group,
+                    )
+                    shard_inplace(
+                        layer.mlp.shared_experts.down_proj,
+                        "sharded-to-all",
+                        group=group,
+                    )
+                    shard_inplace(
+                        layer.mlp.shared_experts.up_proj,
+                        "all-to-sharded",
+                        group=group,
+                    )
+                shard_inplace(
+                    layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+
+    @property
+    def layers(self):
+        return self.model.pipeline_layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate

mlx_lm/models/glm_moe_dsa.py

@@ -0,0 +1,53 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+from .base import BaseModelArgs
+from .deepseek_v32 import Model as DSV32Model
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    index_head_dim: int
+    index_n_heads: int
+    index_topk: int
+    intermediate_size: int
+    moe_intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    n_shared_experts: Optional[int]
+    n_routed_experts: Optional[int]
+    routed_scaling_factor: float
+    kv_lora_rank: int
+    q_lora_rank: int
+    qk_rope_head_dim: int
+    v_head_dim: int
+    qk_nope_head_dim: int
+    topk_method: str
+    scoring_func: str
+    norm_topk_prob: bool
+    n_group: int
+    topk_group: int
+    num_experts_per_tok: int
+    moe_layer_freq: int
+    first_k_dense_replace: int
+    max_position_embeddings: int
+    rms_norm_eps: float
+    rope_parameters: Dict
+    attention_bias: bool
+    rope_scaling: Dict = None
+    rope_theta: Optional[float] = None
+
+    def __post_init__(self):
+        self.rope_scaling = self.rope_parameters
+        self.rope_theta = self.rope_parameters["rope_theta"]
+
+
+class Model(DSV32Model):
+    def __init__(self, config: ModelArgs):
+        super().__init__(config)

mlx_lm/models/gpt2.py

@@ -1,4 +1,4 @@
-# Copyright © 2023-2024 Apple Inc.
+# Copyright © 2023 - 2024 Apple Inc.
 
 from dataclasses import dataclass
 from typing import Any, Optional
@@ -125,26 +125,26 @@ class GPT2Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         _, L = inputs.shape
 
         hidden_states = self.wte(inputs)
 
-        offset = 0
-        if cache is not None and len(cache) > 0 and cache[0] is not None:
-            offset = cache[0].offset
-
-        position_ids = mx.arange(offset, offset + L)
-        hidden_states += self.wpe(position_ids)
-
-        if mask is None:
-            mask = create_attention_mask(hidden_states, cache)
-
         if cache is None:
             cache = [None] * len(self.h)
 
+        offset = 0
+        if cache[0] is not None:
+            offset = cache[0].offset
+
+        offset = mx.array(offset)
+        position_ids = mx.arange(L) + offset[..., None]
+
+        hidden_states += self.wpe(position_ids)
+
+        mask = create_attention_mask(hidden_states, cache[0])
+
         for layer, c in zip(self.h, cache):
             hidden_states = layer(hidden_states, mask, cache=c)
 
@@ -161,10 +161,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         out = self.model.wte.as_linear(out)
         return out
 

mlx_lm/models/gpt_bigcode.py

@@ -137,23 +137,20 @@ class GPTBigCodeModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         B, L = inputs.shape
 
         hidden_states = self.wte(inputs)
 
-        mask = None
-        if mask is not None and hidden_states.shape[1] > 1:
-            mask = create_attention_mask(hidden_states, cache)
-
         if cache is None:
             cache = [None] * len(self.h)
             position_ids = mx.array(np.arange(L))
         else:
             position_ids = mx.array(np.arange(cache[0].offset, cache[0].offset + L))
 
+        mask = create_attention_mask(hidden_states, cache[0])
+
         hidden_states += self.wpe(position_ids)
 
         for layer, c in zip(self.h, cache):
@@ -174,10 +171,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.transformer(inputs, mask, cache)
+        out = self.transformer(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.transformer.wte.as_linear(out)
         else:

mlx_lm/models/gpt_neox.py

@@ -23,6 +23,7 @@ class ModelArgs(BaseModelArgs):
     vocab_size: int
     rotary_emb_base: int
     rotary_pct: float
+    use_parallel_residual: bool = True
     num_key_value_heads: int = None
 
     def __post_init__(self):
@@ -107,6 +108,7 @@ class TransformerBlock(nn.Module):
         self.layer_norm_eps = args.layer_norm_eps
         self.attention = Attention(args)
         self.mlp = MLP(args)
+        self.use_parallel_residual = args.use_parallel_residual
         self.input_layernorm = nn.LayerNorm(
             self.hidden_size,
             eps=self.layer_norm_eps,
@@ -121,12 +123,20 @@ class TransformerBlock(nn.Module):
         mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ) -> mx.array:
-        residual = x
-        # NeoX runs attention and feedforward network in parallel.
-        attn = self.attention(self.input_layernorm(x), mask, cache)
-        ffn = self.mlp(self.post_attention_layernorm(x))
-        out = attn + ffn + residual
-        return out
+        if self.use_parallel_residual:
+            residual = x
+            # Run attention and feedforward network in parallel.
+            attn = self.attention(self.input_layernorm(x), mask, cache)
+            ffn = self.mlp(self.post_attention_layernorm(x))
+            out = attn + ffn + residual
+            return out
+        else:
+            # Run attention and feedforward network sequentially.
+            attn_output = self.attention(self.input_layernorm(x), mask, cache)
+            x = x + attn_output
+            ffn_output = self.mlp(self.post_attention_layernorm(x))
+            x = x + ffn_output
+            return x
 
 
 class GPTNeoXModel(nn.Module):
@@ -145,19 +155,17 @@ class GPTNeoXModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         _, L = inputs.shape
 
         hidden_states = self.embed_in(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(hidden_states, cache)
-
         if cache is None:
             cache = [None] * len(self.h)
 
+        mask = create_attention_mask(hidden_states, cache[0])
+
         for layer, c in zip(self.h, cache):
             hidden_states = layer(hidden_states, mask, cache=c)
 
@@ -177,10 +185,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         return out
 
     def sanitize(self, weights):

mlx_lm/models/gpt_oss.py

@@ -7,8 +7,9 @@ from typing import Any, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
 
-from .base import BaseModelArgs, create_causal_mask, scaled_dot_product_attention
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import KVCache, RotatingKVCache
 from .rope_utils import initialize_rope
 from .switch_layers import SwitchGLU
@@ -102,80 +103,6 @@ class AttentionBlock(nn.Module):
             scaling_config=config.rope_scaling,
         )
 
-        # Cache the mask so we don't have to create it every time
-        self._previous_mask = None
-
-    def get_causal_mask(self, x, cache):
-        _, L, _ = x.shape
-        offset = cache.offset if cache is not None else 0
-        offset = max(1, offset)
-
-        def _make_mask(L, offset):
-            zero = mx.array(0, dtype=x.dtype)
-            neginf = mx.array(-mx.inf, dtype=x.dtype)
-            mask = mx.where(create_causal_mask(L, offset - 1), zero, neginf)
-            mask = mask.reshape(1, 1, L, -1)
-            mask = mx.tile(mask, (1, self.num_attention_heads, 1, 1))
-            sinks = mx.tile(self.sinks.reshape(1, -1, 1, 1), (1, 1, L, 1))
-            mask = mx.concatenate([sinks, mask], axis=-1)
-            return mask
-
-        # When training re-create the mask so that gradients flow to the sinks.
-        # When L is large then recreate the mask because otherwise it will take
-        # a pretty significant chunk of memory.
-        if self.training or L > 8:
-            self._previous_mask = None
-            return _make_mask(L, offset)
-
-        # Create the mask once and try to reuse it. For this reason we round up
-        # to the closest multiple of 512 so we can reuse the mask several times.
-        length = ((L + offset + 511) // 512) * 512
-        if (
-            self._previous_mask is None
-            or self._previous_mask.shape[-1] < length
-            or self._previous_mask.shape[-2] != L
-        ):
-            self._previous_mask = _make_mask(L, length - L)
-
-        return self._previous_mask[..., : L + offset]
-
-    def get_sliding_window_mask(self, x, cache, window_size):
-        _, L, _ = x.shape
-        offset = cache.offset if cache is not None else 0
-        offset = max(1, offset)
-
-        def _make_mask(L, offset):
-            zero = mx.array(0, dtype=x.dtype)
-            neginf = mx.array(-mx.inf, dtype=x.dtype)
-            mask = create_causal_mask(L, offset - 1, window_size)
-            mask = mx.where(mask, zero, neginf)
-            mask = mask.reshape(1, 1, L, -1)
-            mask = mx.tile(mask, (1, self.num_attention_heads, 1, 1))
-            sinks = mx.tile(self.sinks.reshape(1, -1, 1, 1), (1, 1, L, 1))
-            mask = mx.concatenate([sinks, mask], axis=-1)
-            return mask
-
-        # If we are training then simply re-create the mask every time to make
-        # sure gradients flow to the sinks.
-        #
-        # For simplicity also re-create the mask if we have more than 1 query
-        # for now.
-        if self.training or L > 1:
-            self._previous_mask = None
-            return _make_mask(L, min(window_size + 1, offset))
-
-        # We are in inference so cache the mask and try to reuse it
-        if self._previous_mask is None:
-            self._previous_mask = _make_mask(L, window_size)
-
-        return self._previous_mask[..., : min(L + offset, window_size + 1)]
-
-    def get_mask(self, x, cache, window_size):
-        if window_size is not None:
-            return self.get_sliding_window_mask(x, cache, window_size)
-        else:
-            return self.get_causal_mask(x, cache)
-
     def __call__(self, x: mx.array, mask: mx.array, cache=None) -> mx.array:
         B, L, _ = x.shape
         D = self.head_dim
@@ -185,26 +112,17 @@ class AttentionBlock(nn.Module):
         k = self.k_proj(x).reshape(B, L, -1, D).swapaxes(1, 2)
         v = self.v_proj(x).reshape(B, L, -1, D).swapaxes(1, 2)
 
-        # If cache is None or the cache offset is 0 then we need to add a 0 key
-        # and value to make some space for the sink
-        if cache is None or cache.offset == 0:
+        if cache is not None:
+            q = self.rope(q, offset=cache.offset)
+            k = self.rope(k, offset=cache.offset)
+            k, v = cache.update_and_fetch(k, v)
+        else:
             q = self.rope(q)
             k = self.rope(k)
 
-            zeros = mx.zeros((B, Hk, 1, D), dtype=k.dtype)
-            k = mx.concatenate([zeros, k], axis=2)
-            v = mx.concatenate([zeros, v], axis=2)
-            if cache is not None:
-                k, v = cache.update_and_fetch(k, v)
-
-        # We have already put the 0 in the cache no need to do anything special
-        else:
-            q = self.rope(q, offset=cache.offset - 1)
-            k = self.rope(k, offset=cache.offset - 1)
-            k, v = cache.update_and_fetch(k, v)
-
-        # NOTE: mask should contain the sink weights already
-        v_hat = scaled_dot_product_attention(q, k, v, cache, self.sm_scale, mask=mask)
+        v_hat = scaled_dot_product_attention(
+            q, k, v, cache, self.sm_scale, mask=mask, sinks=self.sinks
+        )
 
         return self.o_proj(v_hat.swapaxes(1, 2).reshape(B, L, -1))
 
@@ -225,8 +143,12 @@ class MLPBlock(nn.Module):
             bias=True,
         )
         self.router = nn.Linear(config.hidden_size, config.num_local_experts, bias=True)
+        self.sharding_group = None
 
     def __call__(self, x: mx.array) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
         g = self.router(x)
         experts, indices = mlx_topk(g, k=self.num_experts_per_tok, axis=-1)
         expert_weights = mx.softmax(experts, axis=-1, precise=True)
@@ -235,7 +157,13 @@ class MLPBlock(nn.Module):
         x = self.experts(x, indices)
 
         x = x * mx.expand_dims(expert_weights, axis=-1)
-        return x.sum(axis=-2)
+
+        y = x.sum(axis=-2)
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y
 
 
 class TransformerBlock(nn.Module):
@@ -272,11 +200,12 @@ class GptOssMoeModel(nn.Module):
         ] * (args.num_hidden_layers // 2)
         self.layers = [TransformerBlock(args) for _ in range(args.num_hidden_layers)]
         self.window_size = args.sliding_window
+        self.swa_idx = self.layer_types.index("sliding_attention")
+        self.ga_idx = self.layer_types.index("full_attention")
 
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
@@ -288,18 +217,14 @@ class GptOssMoeModel(nn.Module):
         if cache is None:
             cache = [None] * len(self.layers)
 
-        if mask is None:
-            masks = [
-                l.self_attn.get_mask(
-                    x, c, self.window_size if lt == "sliding_attention" else None
-                )
-                for (l, c, lt) in zip(self.layers, cache, self.layer_types)
-            ]
-        else:
-            masks = [mask] * len(self.layers)
+        full_mask = create_attention_mask(x, cache[self.ga_idx])
+        swa_mask = create_attention_mask(
+            x, cache[self.swa_idx], window_size=self.window_size
+        )
 
-        for i, (layer, c, m) in enumerate(zip(self.layers, cache, masks)):
-            x = layer(x, m, c)
+        for layer, c, layer_type in zip(self.layers, cache, self.layer_types):
+            mask = full_mask if layer_type == "full_attention" else swa_mask
+            x = layer(x, mask, c)
         x = self.norm(x)
         return x
 
@@ -312,8 +237,8 @@ class Model(nn.Module):
         self.model = GptOssMoeModel(args)
         self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
 
-    def __call__(self, inputs: mx.array, mask: mx.array = None, cache=None):
-        return self.lm_head(self.model(inputs, mask, cache))
+    def __call__(self, inputs: mx.array, cache=None):
+        return self.lm_head(self.model(inputs, cache))
 
     def sanitize(self, weights):
         if any("gate_proj.weight" in k for k in weights.keys()):
@@ -354,6 +279,47 @@ class Model(nn.Module):
 
         return new_weights
 
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        R = group.rank()
+
+        for layer in self.model.layers:
+            layer.self_attn.q_proj = shard_linear(
+                layer.self_attn.q_proj, sharding="all-to-sharded", group=group
+            )
+            layer.self_attn.k_proj = shard_linear(
+                layer.self_attn.k_proj, sharding="all-to-sharded", group=group
+            )
+            layer.self_attn.v_proj = shard_linear(
+                layer.self_attn.v_proj, sharding="all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, sharding="sharded-to-all", group=group
+            )
+
+            layer.self_attn.num_attention_heads //= N
+            layer.self_attn.num_key_value_heads //= N
+            layer.self_attn.num_key_value_groups = (
+                layer.self_attn.num_attention_heads
+                // layer.self_attn.num_key_value_heads
+            )
+
+            layer.self_attn.sinks = layer.self_attn.sinks[
+                layer.self_attn.num_attention_heads
+                * R : layer.self_attn.num_attention_heads
+                * (R + 1)
+            ]
+
+            shard_inplace(layer.mlp.experts.gate_proj, "all-to-sharded", group=group)
+            shard_inplace(layer.mlp.experts.down_proj, "sharded-to-all", group=group)
+            layer.mlp.experts.down_proj.bias /= N
+            shard_inplace(
+                layer.mlp.experts.up_proj, sharding="all-to-sharded", group=group
+            )
+
+            layer.mlp.sharding_group = group
+
     @property
     def layers(self):
         return self.model.layers
@@ -373,7 +339,5 @@ class Model(nn.Module):
             if lt == "full_attention":
                 caches.append(KVCache())
             else:
-                caches.append(
-                    RotatingKVCache(max_size=self.args.sliding_window + 1, keep=1)
-                )
+                caches.append(RotatingKVCache(max_size=self.args.sliding_window))
         return caches

mlx_lm/models/granite.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .rope_utils import initialize_rope
 
@@ -104,7 +105,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -150,17 +151,15 @@ class GraniteModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs) * self.embedding_multiplier
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -180,10 +179,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/granitemoe.py

@@ -0,0 +1,235 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    logits_scaling: float
+    attention_multiplier: float
+    embedding_multiplier: float
+    residual_multiplier: float
+    max_position_embeddings: int
+    num_key_value_heads: int
+    attention_bias: bool
+    rope_theta: float
+    num_local_experts: int
+    num_experts_per_tok: int
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+
+class GraniteMoeAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.scale = args.attention_multiplier
+        attention_bias = args.attention_bias
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attention_bias)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            False,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class GraniteMoeTopKGating(nn.Module):
+    def __init__(self, input_size: int, num_experts: int, top_k: int):
+        super().__init__()
+        self.num_experts = num_experts
+        self.input_size = input_size
+        self.top_k = top_k
+        self.layer = nn.Linear(input_size, num_experts, bias=False)
+
+    def __call__(self, hidden_states: mx.array):
+        logits = self.layer(hidden_states)
+        top_k_idx = mx.argpartition(logits, kth=-self.top_k, axis=-1)[
+            ..., -self.top_k :
+        ]
+        top_k_logits = mx.take_along_axis(logits, top_k_idx, axis=-1)
+        top_k_gates = mx.softmax(top_k_logits.astype(mx.float32), axis=-1)
+        return top_k_idx, top_k_gates
+
+
+class GraniteMoeMoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        self.input_size = args.hidden_size
+        self.hidden_size = args.intermediate_size
+        self.switch_mlp = SwitchGLU(
+            self.input_size, self.hidden_size, args.num_local_experts
+        )
+        self.router = GraniteMoeTopKGating(
+            input_size=self.input_size,
+            num_experts=args.num_local_experts,
+            top_k=args.num_experts_per_tok,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        token_ids, gates = self.router(x)
+        y = self.switch_mlp(x, token_ids)
+        return (y * gates[..., None]).sum(axis=-2).astype(y.dtype)
+
+
+class GraniteMoeDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = GraniteMoeAttention(args)
+        self.block_sparse_moe = GraniteMoeMoE(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.residual_multiplier = args.residual_multiplier
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r * self.residual_multiplier
+        r = self.block_sparse_moe(self.post_attention_layernorm(h))
+        out = h + r * self.residual_multiplier
+        return out
+
+
+class GraniteMoEModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            GraniteMoeDecoderLayer(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.embedding_multiplier = args.embedding_multiplier
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs) * self.embedding_multiplier
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = GraniteMoEModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self.logits_scaling = args.logits_scaling
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out / self.logits_scaling
+
+    def sanitize(self, weights):
+        if "model.layers.0.block_sparse_moe.input_linear.weight" not in weights:
+            return weights
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}.block_sparse_moe"
+            key = f"{prefix}.input_linear.weight"
+            value = weights.pop(key)
+            gate_proj, up_proj = mx.split(value, 2, axis=1)
+            weights[key.replace("input_linear", "switch_mlp.gate_proj")] = gate_proj
+            weights[key.replace("input_linear", "switch_mlp.up_proj")] = up_proj
+            key = f"{prefix}.output_linear.weight"
+            weights[key.replace("output_linear", "switch_mlp.down_proj")] = weights.pop(
+                key
+            )
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("block_sparse_moe.router.layer"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def layers(self):
+        return self.model.layers

mlx_lm/models/granitemoehybrid.py

@@ -0,0 +1,559 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, KVCache
+from .rope_utils import initialize_rope
+from .ssm import ssm_update
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    # Required fields (no defaults)
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    max_position_embeddings: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    attention_bias: bool
+    embedding_multiplier: float
+    attention_multiplier: float
+    logits_scaling: float
+    residual_multiplier: float
+    layer_types: List[str]
+    rms_norm_eps: float
+    rope_theta: float
+
+    # Optional fields (with defaults)
+    # MoE parameters (optional for dense mode)
+    num_local_experts: Optional[int] = None
+    num_experts_per_tok: Optional[int] = None
+    shared_intermediate_size: Optional[int] = None
+
+    # Mamba parameters (optional for non-hybrid mode)
+    mamba_n_heads: Optional[int] = None
+    mamba_d_head: Optional[int] = None
+    mamba_proj_bias: Optional[bool] = None
+    mamba_d_state: Optional[int] = None
+    mamba_d_conv: Optional[int] = None
+    mamba_n_groups: Optional[int] = None
+    mamba_conv_bias: Optional[bool] = None
+
+    # Dense MLP parameters (for non-MoE mode)
+    mlp_bias: bool = False
+
+    # Other optional parameters
+    position_embedding_type: str = "rope"
+    tie_word_embeddings: bool = True
+    time_step_limit: Tuple[float, float] = (0.001, 100.0)
+
+    # Mode flags - inferred from num_local_experts
+    @property
+    def use_moe(self) -> bool:
+        return bool(self.num_local_experts)
+
+
+class GraniteMoeHybridRMSNormGated(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = mx.ones(hidden_size)
+
+    def __call__(self, hidden_states: mx.array, gate: mx.array = None) -> mx.array:
+        if gate is not None:
+            hidden_states = swiglu(gate, hidden_states)
+        return mx.fast.rms_norm(hidden_states, self.weight, self.eps)
+
+
+class GraniteMoeHybridMamba2Mixer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_heads = args.mamba_n_heads
+        self.hidden_size = args.hidden_size
+        self.ssm_state_size = args.mamba_d_state
+        self.conv_kernel_size = args.mamba_d_conv
+        self.intermediate_size = args.mamba_n_heads * args.mamba_d_head
+        self.n_groups = args.mamba_n_groups
+        self.head_dim = args.mamba_d_head
+        self.time_step_limit = args.time_step_limit
+        self.heads_per_group = self.num_heads // self.n_groups
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            kernel_size=args.mamba_d_conv,
+            padding=0,
+            groups=self.conv_dim,
+            bias=args.mamba_conv_bias,
+        )
+
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size, projection_size, bias=args.mamba_proj_bias
+        )
+
+        self.dt_bias = mx.ones(self.num_heads)
+        self.A_log = mx.log(mx.arange(1, self.num_heads + 1, dtype=mx.float32))
+        self.D = mx.ones(self.num_heads)
+
+        self.norm = GraniteMoeHybridRMSNormGated(
+            self.intermediate_size, eps=args.rms_norm_eps
+        )
+        self.out_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=args.mamba_proj_bias
+        )
+
+    def _conv(
+        self,
+        conv_input: mx.array,
+        cache: Optional[ArraysCache],
+        mask: Optional[mx.array],
+    ) -> mx.array:
+        if mask is not None:
+            conv_input = mx.where(mask[..., None], conv_input, 0)
+
+        if cache is not None:
+            if cache[0] is None:
+                conv_state = mx.zeros(
+                    (conv_input.shape[0], self.conv_kernel_size - 1, self.conv_dim),
+                    dtype=conv_input.dtype,
+                )
+            else:
+                conv_state = cache[0]
+            padded_input = mx.concatenate([conv_state, conv_input], axis=1)
+            n_keep = self.conv_kernel_size - 1
+            if cache.lengths is not None:
+                t = padded_input.shape[1]
+                ends = mx.clip(cache.lengths, 0, t - n_keep)
+                positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+                cache[0] = mx.take_along_axis(padded_input, positions, axis=1)
+            else:
+                cache[0] = padded_input[:, -n_keep:, :]
+        else:
+            padded_input = mx.pad(
+                conv_input, [(0, 0), (self.conv_kernel_size - 1, 0), (0, 0)]
+            )
+
+        conv_output = self.conv1d(padded_input)
+        return nn.silu(conv_output)
+
+    def _ssm(
+        self,
+        hidden_states: mx.array,
+        B: mx.array,
+        C: mx.array,
+        dt: mx.array,
+        cache: Optional[ArraysCache],
+        mask: Optional[mx.array],
+    ) -> mx.array:
+        batch_size, seq_len, _ = hidden_states.shape
+
+        hidden_states = hidden_states.reshape(
+            batch_size, seq_len, self.num_heads, self.head_dim
+        )
+        B = B.reshape(batch_size, seq_len, self.n_groups, self.ssm_state_size)
+        C = C.reshape(batch_size, seq_len, self.n_groups, self.ssm_state_size)
+        if cache:
+            state = cache[1]
+            lengths = cache.lengths
+        else:
+            state, lengths = None, None
+
+        y, state = ssm_update(
+            hidden_states,
+            self.A_log,
+            B,
+            C,
+            self.D.astype(hidden_states.dtype),
+            dt,
+            self.dt_bias,
+            state,
+            self.time_step_limit,
+            mask,
+        )
+        if cache:
+            cache[1] = state
+
+        return y.reshape(batch_size, seq_len, self.intermediate_size)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array],
+        cache: Optional[ArraysCache] = None,
+    ) -> mx.array:
+
+        projected = self.in_proj(hidden_states)
+
+        gate, conv_input, dt = mx.split(
+            projected,
+            [self.intermediate_size, self.intermediate_size + self.conv_dim],
+            axis=-1,
+        )
+        conv_output = self._conv(conv_input, cache, mask)
+        hidden_states_ssm, B, C = mx.split(
+            conv_output,
+            [
+                self.intermediate_size,
+                self.intermediate_size + self.n_groups * self.ssm_state_size,
+            ],
+            axis=-1,
+        )
+        y = self._ssm(hidden_states_ssm, B, C, dt, cache, mask)
+        if cache:
+            cache.advance(y.shape[1])
+        y = self.norm(y, gate)
+        return self.out_proj(y)
+
+
+class GraniteMoeHybridAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.scale = args.attention_multiplier
+        attention_bias = args.attention_bias
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attention_bias)
+
+        # Check if RoPE should be used based on position_embedding_type
+        # If position_embedding_type is "nope", don't use RoPE
+        use_rope = args.position_embedding_type != "nope"
+        if use_rope:
+            self.rope = initialize_rope(
+                self.head_dim,
+                args.rope_theta,
+                False,
+                None,  # rope_scaling
+                args.max_position_embeddings,
+            )
+        else:
+            self.rope = None
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        # Apply RoPE only if enabled
+        if self.rope is not None:
+            if cache is not None:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+            else:
+                queries = self.rope(queries)
+                keys = self.rope(keys)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class GraniteMoeHybridTopKGating(nn.Module):
+    def __init__(self, input_size: int, num_experts: int, top_k: int):
+        super().__init__()
+        self.num_experts = num_experts
+        self.input_size = input_size
+        self.top_k = top_k
+        self.layer = nn.Linear(input_size, num_experts, bias=False)
+
+    def __call__(self, hidden_states: mx.array):
+        logits = self.layer(hidden_states)
+        top_k_idx = mx.argpartition(logits, kth=-self.top_k, axis=-1)[
+            ..., -self.top_k :
+        ]
+        top_k_logits = mx.take_along_axis(logits, top_k_idx, axis=-1)
+        top_k_gates = mx.softmax(top_k_logits, precise=True, axis=-1)
+        return top_k_idx, top_k_gates
+
+
+class GraniteMoeHybridMoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        self.input_size = args.hidden_size
+        self.hidden_size = args.intermediate_size
+        self.switch_mlp = SwitchGLU(
+            self.input_size, self.hidden_size, args.num_local_experts
+        )
+        self.router = GraniteMoeHybridTopKGating(
+            input_size=self.input_size,
+            num_experts=args.num_local_experts,
+            top_k=args.num_experts_per_tok,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        token_ids, gates = self.router(x)
+        y = self.switch_mlp(x, token_ids)
+        return (y * gates[..., None]).sum(axis=-2)
+
+
+class GraniteMoeHybridSharedMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.input_linear = nn.Linear(
+            args.hidden_size, args.shared_intermediate_size * 2, bias=False
+        )
+        self.output_linear = nn.Linear(
+            args.shared_intermediate_size, args.hidden_size, bias=False
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        gate, up = mx.split(self.input_linear(x), 2, axis=-1)
+        return self.output_linear(swiglu(gate, up))
+
+
+class GraniteMoeHybridMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        hidden_dim = args.intermediate_size
+        mlp_bias = args.mlp_bias
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=mlp_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class GraniteMoeHybridLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_type: str):
+        super().__init__()
+        self.layer_type = layer_type
+        self.residual_multiplier = args.residual_multiplier
+        self.use_moe = args.use_moe
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+        if layer_type == "mamba":
+            self.mamba = GraniteMoeHybridMamba2Mixer(args)
+        elif layer_type == "attention":
+            self.self_attn = GraniteMoeHybridAttention(args)
+        else:
+            raise ValueError(f"Unknown layer type: {layer_type}")
+
+        # MoE or dense MLP after attention/mamba
+        if self.use_moe:
+            self.shared_mlp = GraniteMoeHybridSharedMLP(args)
+            self.block_sparse_moe = GraniteMoeHybridMoE(args)
+        else:
+            # Dense MLP mode
+            self.mlp = GraniteMoeHybridMLP(args)
+
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        # First block: either Mamba or Attention
+        residual = x
+        hidden_states = self.input_layernorm(x)
+
+        if self.layer_type == "mamba":
+            hidden_states = self.mamba(hidden_states, mask=mask, cache=cache)
+        else:
+            hidden_states = self.self_attn(hidden_states, mask=mask, cache=cache)
+
+        hidden_states = residual + hidden_states * self.residual_multiplier
+
+        # Second block: MoE + shared_mlp OR dense MLP
+        residual = hidden_states
+        normed = self.post_attention_layernorm(hidden_states)
+
+        if self.use_moe:
+            moe_out = self.block_sparse_moe(normed)
+            shared_out = self.shared_mlp(normed)
+            mlp_out = moe_out + shared_out
+        else:
+            mlp_out = self.mlp(normed)
+
+        hidden_states = residual + mlp_out * self.residual_multiplier
+
+        return hidden_states
+
+
+class GraniteMoeHybridModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            GraniteMoeHybridLayer(args, layer_type) for layer_type in args.layer_types
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.embedding_multiplier = args.embedding_multiplier
+
+        # Handle hybrid vs non-hybrid mode
+        self.fa_idx = (
+            args.layer_types.index("attention")
+            if "attention" in args.layer_types
+            else None
+        )
+        self.ssm_idx = (
+            args.layer_types.index("mamba") if "mamba" in args.layer_types else None
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        hidden_states = self.embed_tokens(inputs) * self.embedding_multiplier
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        # Create masks based on what layer types exist
+        attn_mask = None
+        mamba_mask = None
+
+        if self.fa_idx is not None:
+            attn_mask = create_attention_mask(hidden_states, cache[self.fa_idx])
+        if self.ssm_idx is not None:
+            mamba_mask = create_ssm_mask(hidden_states, cache[self.ssm_idx])
+
+        for layer, c in zip(self.layers, cache):
+            mask = attn_mask if layer.layer_type == "attention" else mamba_mask
+            hidden_states = layer(hidden_states, mask=mask, cache=c)
+
+        return self.norm(hidden_states)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = GraniteMoeHybridModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self.logits_scaling = args.logits_scaling
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache=cache)
+
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+
+        return out / self.logits_scaling
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        caches = []
+        for layer in self.layers:
+            if layer.layer_type == "mamba":
+                caches.append(ArraysCache(size=2))
+            elif layer.layer_type == "attention":
+                caches.append(KVCache())
+        return caches
+
+    def sanitize(self, weights):
+        # Handle conv1d weights
+        for k, v in weights.items():
+            if "conv1d.weight" in k and v.shape[-1] != 1:
+                weights[k] = v.moveaxis(2, 1)
+
+        # Handle MoE weight transformation to SwitchGLU format (only for MoE models)
+        if (
+            self.args.use_moe
+            and "model.layers.0.block_sparse_moe.input_linear.weight" in weights
+        ):
+            for l in range(self.args.num_hidden_layers):
+                prefix = f"model.layers.{l}.block_sparse_moe"
+
+                input_weight = weights.pop(f"{prefix}.input_linear.weight")
+                _, expert_hidden, _ = input_weight.shape
+
+                # Split into gate and up projections (each half of expert_hidden)
+                gate_proj = input_weight[:, : expert_hidden // 2, :]
+                up_proj = input_weight[:, expert_hidden // 2 :, :]
+                weights[f"{prefix}.switch_mlp.gate_proj.weight"] = gate_proj
+                weights[f"{prefix}.switch_mlp.up_proj.weight"] = up_proj
+
+                weights[f"{prefix}.switch_mlp.down_proj.weight"] = weights.pop(
+                    f"{prefix}.output_linear.weight"
+                )
+
+        # Handle dense MLP weight transformation (for dense models)
+        elif (
+            not self.args.use_moe
+            and "model.layers.0.shared_mlp.input_linear.weight" in weights
+        ):
+            for l in range(self.args.num_hidden_layers):
+                prefix = f"model.layers.{l}.shared_mlp"
+
+                # Transform shared_mlp weights to standard mlp weights
+                input_weight = weights.pop(f"{prefix}.input_linear.weight")
+                # Split into gate and up projections (each half)
+                gate_proj, up_proj = mx.split(input_weight, 2, axis=0)
+                weights[f"model.layers.{l}.mlp.gate_proj.weight"] = gate_proj
+                weights[f"model.layers.{l}.mlp.up_proj.weight"] = up_proj
+
+                weights[f"model.layers.{l}.mlp.down_proj.weight"] = weights.pop(
+                    f"{prefix}.output_linear.weight"
+                )
+
+        return weights
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if self.args.use_moe and path.endswith("router.layer"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate

mlx_lm/models/helium.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -92,7 +93,7 @@ class HeliumMLP(nn.Module):
         )
 
     def __call__(self, x: mx.array) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class HeliumDecoderLayer(nn.Module):
@@ -136,17 +137,15 @@ class HeliumModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ) -> mx.array:
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
+
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -170,10 +169,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ) -> mx.array:
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/hunyuan.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Tuple, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .switch_layers import SwitchGLU
 
@@ -148,7 +149,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class Gate(nn.Module):
@@ -259,17 +260,14 @@ class HunYuanModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
         for i, (layer, c) in enumerate(zip(self.layers, cache)):
             if (not self.args.use_cla) or i % self.args.cla_share_factor == 0:
                 shared_kv_states = None
@@ -288,10 +286,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         return self.model.embed_tokens.as_linear(out)
 
     def sanitize(self, weights):

mlx_lm/models/hunyuan_v1_dense.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -144,7 +145,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -186,17 +187,14 @@ class HunyuanV1DenseModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, c)
 
@@ -209,16 +207,25 @@ class Model(nn.Module):
         self.args = args
         self.model_type = args.model_type
         self.model = HunyuanV1DenseModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
 
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
-        return self.model.embed_tokens.as_linear(out)
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            return self.model.embed_tokens.as_linear(out)
+        else:
+            return self.lm_head(out)
 
     @property
     def layers(self):
         return self.model.layers
+
+    def sanitize(self, weights):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights

mlx_lm/models/internlm2.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -156,7 +157,7 @@ class MLP(nn.Module):
         self.w3 = nn.Linear(dim, hidden_dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.w2(nn.silu(self.w1(x)) * self.w3(x))
+        return self.w2(swiglu(self.w1(x), self.w3(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -193,17 +194,14 @@ class InternLM2Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.tok_embeddings(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -222,10 +220,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.tok_embeddings.as_linear(out)
         else:

mlx_lm/models/internlm3.py

@@ -6,6 +6,7 @@ from typing import Any, Dict, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 
 
@@ -154,7 +155,7 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=bias)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
@@ -193,17 +194,14 @@ class InternLM2Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        mask = create_attention_mask(h, cache[0])
         for layer, c in zip(self.layers, cache):
             h = layer(h, mask, cache=c)
 
@@ -222,10 +220,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:

mlx_lm/models/iquestloopcoder.py

@@ -0,0 +1,286 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_linear
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+from .rope_utils import initialize_rope
+
+
+@partial(mx.compile, shapeless=True)
+def _compute_gate(query: mx.array, weight: mx.array, bias: mx.array) -> mx.array:
+    gate_logits = query @ weight[:, None, :].swapaxes(-1, -2)
+    gate_logits = gate_logits + bias[..., None, None]
+    return mx.sigmoid(gate_logits)
+
+
+@partial(mx.compile, shapeless=True)
+def _silu_mul(gate: mx.array, up: mx.array) -> mx.array:
+    return nn.silu(gate) * up
+
+
+@partial(mx.compile, shapeless=True)
+def _mix_attention(
+    gate: mx.array, attn_global: mx.array, attn_local: mx.array
+) -> mx.array:
+    return gate * attn_global + (1 - gate) * attn_local
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    head_dim: int
+    num_key_value_heads: int
+    max_position_embeddings: int = 131072
+    attention_bias: bool = False
+    mlp_bias: bool = False
+    rope_theta: float = 500000.0
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+    loop_num: int = 2
+    loop_window_size: int = 64
+
+
+class LoopGateProjection(nn.Module):
+    def __init__(self, num_heads: int, head_dim: int):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+        self.weight = mx.zeros((num_heads, head_dim))
+        self.bias = mx.zeros((num_heads,))
+
+    def __call__(self, query: mx.array) -> mx.array:
+        return _compute_gate(query, self.weight, self.bias)
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+        self.head_dim = head_dim = args.head_dim
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.attention_bias)
+
+        self.rope = initialize_rope(
+            head_dim,
+            args.rope_theta,
+            traditional=False,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def get_qkv(
+        self, x: mx.array, offset: int = 0
+    ) -> Tuple[mx.array, mx.array, mx.array]:
+        B, L, _ = x.shape
+        queries = self.q_proj(x).reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = self.k_proj(x).reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = self.v_proj(x).reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        queries = self.rope(queries, offset=offset)
+        keys = self.rope(keys, offset=offset)
+
+        return queries, keys, values
+
+    def attention(
+        self,
+        queries: mx.array,
+        keys: mx.array,
+        values: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        return scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        hidden_dim = args.intermediate_size
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=args.mlp_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=args.mlp_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=args.mlp_bias)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(_silu_mul(self.gate_proj(x), self.up_proj(x)))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+
+class IQuestLoopCoderModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        assert args.loop_num == 2, f"Only loop_num=2 is supported, got {args.loop_num}"
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.gate_projections = [
+            LoopGateProjection(args.num_attention_heads, args.head_dim)
+            for _ in range(args.num_hidden_layers)
+        ]
+        self.loop_num = args.loop_num
+        self.loop_window_size = args.loop_window_size
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ):
+        B, L = inputs.shape[:2]
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * (2 * len(self.layers))
+
+        mask = create_attention_mask(h, cache[0])
+        window_mask = create_attention_mask(
+            h, cache[len(self.layers)], window_size=self.loop_window_size
+        )
+
+        loop1_kv = []
+        for layer, c in zip(self.layers, cache):
+            h_norm = layer.input_layernorm(h)
+            offset = c.offset if c is not None else 0
+            q1, k1, v1 = layer.self_attn.get_qkv(h_norm, offset)
+
+            if c is not None:
+                k1, v1 = c.update_and_fetch(k1, v1)
+            loop1_kv.append((k1, v1))
+
+            out = layer.self_attn.attention(q1, k1, v1, mask, cache=c)
+            r = layer.self_attn.o_proj(out.transpose(0, 2, 1, 3).reshape(B, L, -1))
+            h = h + r
+            r = layer.mlp(layer.post_attention_layernorm(h))
+            h = h + r
+
+        for layer, gate_proj, c, (k1, v1) in zip(
+            self.layers, self.gate_projections, cache[len(self.layers) :], loop1_kv
+        ):
+            h_norm = layer.input_layernorm(h)
+            offset = c.offset if c is not None else 0
+            q2, k2, v2 = layer.self_attn.get_qkv(h_norm, offset)
+            gate = gate_proj(q2)
+            attn_global = layer.self_attn.attention(q2, k1, v1, mask, cache=c)
+
+            if c is not None:
+                k2, v2 = c.update_and_fetch(k2, v2)
+            attn_local = layer.self_attn.attention(
+                q2,
+                k2,
+                v2,
+                window_mask,
+                cache=c,
+            )
+
+            mixed = _mix_attention(gate, attn_global, attn_local)
+            r = layer.self_attn.o_proj(mixed.transpose(0, 2, 1, 3).reshape(B, L, -1))
+            h = h + r
+            r = layer.mlp(layer.post_attention_layernorm(h))
+            h = h + r
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = IQuestLoopCoderModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        rank = group.rank()
+
+        for i, layer in enumerate(self.model.layers):
+            layer.self_attn.q_proj = shard_linear(
+                layer.self_attn.q_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.k_proj = shard_linear(
+                layer.self_attn.k_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.v_proj = shard_linear(
+                layer.self_attn.v_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.n_heads //= N
+            layer.self_attn.n_kv_heads //= N
+
+            layer.mlp.gate_proj = shard_linear(
+                layer.mlp.gate_proj, "all-to-sharded", group=group
+            )
+            layer.mlp.down_proj = shard_linear(
+                layer.mlp.down_proj, "sharded-to-all", group=group
+            )
+            layer.mlp.up_proj = shard_linear(
+                layer.mlp.up_proj, "all-to-sharded", group=group
+            )
+
+            gate_proj = self.model.gate_projections[i]
+            heads_per_rank = gate_proj.num_heads // N
+            start = rank * heads_per_rank
+            end = start + heads_per_rank
+            gate_proj.weight = gate_proj.weight[start:end, :]
+            gate_proj.bias = gate_proj.bias[start:end]
+            gate_proj.num_heads = heads_per_rank
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        return [KVCache() for _ in self.layers] + [
+            RotatingKVCache(max_size=self.args.loop_window_size) for _ in self.layers
+        ]

mlx_lm/models/jamba.py

@@ -0,0 +1,385 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, List, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, KVCache
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    attn_layer_offset: int
+    attn_layer_period: int
+    expert_layer_offset: int
+    expert_layer_period: int
+    mamba_d_conv: int
+    mamba_d_state: int
+    mamba_expand: int
+    num_experts: int
+    num_experts_per_tok: int
+    rms_norm_eps: float
+    max_position_embeddings: int
+    vocab_size: int
+    mamba_dt_rank: Union[str, int] = "auto"
+    mamba_proj_bias: bool = False
+    mamba_conv_bias: bool = True
+    layers_block_type: Optional[List[str]] = None
+    tie_word_embeddings: bool = True
+
+    def __post_init__(self):
+        if self.mamba_dt_rank == "auto":
+            self.mamba_dt_rank = math.ceil(self.hidden_size / 16)
+        if self.layers_block_type is None:
+            self.layers_block_type = [
+                (
+                    "attention"
+                    if i % self.attn_layer_period == self.attn_layer_offset
+                    else "mamba"
+                )
+                for i in range(self.num_hidden_layers)
+            ]
+
+
+class JambaMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.gate_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class JambaAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.head_dim = args.hidden_size // args.num_attention_heads
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size, args.num_attention_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size, args.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size, args.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            args.num_attention_heads * self.head_dim, args.hidden_size, bias=False
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.num_attention_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+        keys = keys.reshape(B, L, self.num_key_value_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+@mx.compile
+def fma(a, b, c):
+    return a * b + c
+
+
+class JambaMambaMixer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.ssm_state_size = args.mamba_d_state
+        self.conv_kernel_size = args.mamba_d_conv
+        self.intermediate_size = args.mamba_expand * args.hidden_size
+        self.time_step_rank = args.mamba_dt_rank
+        self.use_conv_bias = args.mamba_conv_bias
+        self.use_bias = args.mamba_proj_bias
+
+        self.in_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size * 2, bias=self.use_bias
+        )
+
+        self.conv1d = nn.Conv1d(
+            in_channels=self.intermediate_size,
+            out_channels=self.intermediate_size,
+            kernel_size=self.conv_kernel_size,
+            groups=self.intermediate_size,
+            bias=self.use_conv_bias,
+            padding=0,
+        )
+        self.x_proj = nn.Linear(
+            self.intermediate_size,
+            self.time_step_rank + self.ssm_state_size * 2,
+            bias=False,
+        )
+        self.dt_proj = nn.Linear(self.time_step_rank, self.intermediate_size, bias=True)
+
+        A = mx.repeat(
+            mx.arange(1.0, self.ssm_state_size + 1.0).reshape([1, self.ssm_state_size]),
+            repeats=self.intermediate_size,
+            axis=0,
+        )
+        self.A_log = mx.log(A)
+        self.D = mx.ones([self.intermediate_size])
+
+        self.out_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=self.use_bias
+        )
+
+        self.dt_layernorm = nn.RMSNorm(self.time_step_rank, eps=args.rms_norm_eps)
+        self.b_layernorm = nn.RMSNorm(self.ssm_state_size, eps=args.rms_norm_eps)
+        self.c_layernorm = nn.RMSNorm(self.ssm_state_size, eps=args.rms_norm_eps)
+
+    def ssm_step(self, x, A, state=None):
+        T = x.shape[1]
+        D = self.D
+        deltaBC = self.x_proj(x)
+        delta, B, C = mx.split(
+            deltaBC,
+            [self.time_step_rank, self.time_step_rank + self.ssm_state_size],
+            axis=-1,
+        )
+        delta, B, C = self.dt_layernorm(delta), self.b_layernorm(B), self.c_layernorm(C)
+        delta = nn.softplus(self.dt_proj(delta))
+        new_state = mx.expand_dims(delta * x, -1) * mx.expand_dims(B, -2)
+        dtA = mx.exp(mx.expand_dims(delta, -1) * A)
+
+        # TODO, speed up prefill with chunked scan
+        for t in range(T):
+            if state is not None:
+                new_state[:, t] = fma(state, dtA[:, t], new_state[:, t])
+            state = new_state[:, t]
+        y = (new_state @ mx.expand_dims(C, -1)).squeeze(-1)
+        y = y + D * x
+        return y, new_state[:, -1]
+
+    def _process_sequence(self, x, conv_state, ssm_state):
+        xz = self.in_proj(x)
+        x, z = xz.split(indices_or_sections=2, axis=-1)
+        K = self.conv_kernel_size
+        if conv_state is not None:
+            x_full = mx.concatenate([conv_state, x], axis=1)
+        else:
+            x_full = mx.pad(x, [(0, 0), (K - 1, 0), (0, 0)])
+        conv_out = self.conv1d(x_full)
+        conv_state = x_full[:, -(K - 1) :, :]
+        x = nn.silu(conv_out)
+        A = -mx.exp(self.A_log)
+        y, ssm_state = self.ssm_step(x, A, ssm_state)
+        z = self.out_proj(swiglu(z, y))
+        return z, (conv_state, ssm_state)
+
+    def __call__(self, x, cache):
+        if cache is None:
+            conv_state, ssm_state = None, None
+        else:
+            conv_state, ssm_state = cache[0], cache[1]
+
+        output, (conv_state, ssm_state) = self._process_sequence(
+            x, conv_state, ssm_state
+        )
+
+        if cache is not None:
+            cache[0] = conv_state
+            cache[1] = ssm_state
+
+        return output
+
+
+class JambaSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_experts_per_tok = args.num_experts_per_tok
+
+        self.router = nn.Linear(args.hidden_size, args.num_experts, bias=False)
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size, args.intermediate_size, args.num_experts
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        gates = self.router(x)
+        k = self.num_experts_per_tok
+        inds = mx.stop_gradient(mx.argpartition(-gates, kth=k - 1, axis=-1)[..., :k])
+        scores = mx.take_along_axis(gates, inds, axis=-1)
+        scores = mx.softmax(scores, axis=-1, precise=True)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2)
+        return y
+
+
+class JambaDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_type: str, layer_idx: int):
+        super().__init__()
+        self.is_attn = layer_type == "attention"
+        if self.is_attn:
+            self.self_attn = JambaAttention(args)
+        else:
+            self.mamba = JambaMambaMixer(args)
+        if (
+            args.num_experts > 1
+            and (layer_idx + args.expert_layer_offset) % args.expert_layer_period == 0
+        ):
+            ffn_layer_class = JambaSparseMoeBlock
+        else:
+            ffn_layer_class = JambaMLP
+        self.feed_forward = ffn_layer_class(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.pre_ff_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        if self.is_attn:
+            h = self.self_attn(self.input_layernorm(x), mask, cache)
+        else:
+            h = self.mamba(self.input_layernorm(x), cache)
+        r = x + h
+        out = r + self.feed_forward(self.pre_ff_layernorm(r))
+        return out
+
+
+class JambaModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+
+        self.layers = [
+            JambaDecoderLayer(args, t, idx)
+            for idx, t in enumerate(args.layers_block_type)
+        ]
+        self.final_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.attn_idx = args.layers_block_type.index("attention")
+        self.ssm_idx = args.layers_block_type.index("mamba")
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        attn_mask = create_attention_mask(h, cache[self.attn_idx])
+        ssm_mask = create_ssm_mask(h, cache[self.ssm_idx])
+
+        for layer, c in zip(self.layers, cache):
+            mask = attn_mask if layer.is_attn else ssm_mask
+            h = layer(h, mask=mask, cache=c)
+
+        return self.final_layernorm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.model_type = args.model_type
+        self.args = args
+        self.model = JambaModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def make_cache(self):
+        caches = []
+        for layer in self.model.layers:
+            if layer.is_attn:
+                caches.append(KVCache())
+            else:
+                caches.append(ArraysCache(size=2))
+        return caches
+
+    def sanitize(self, weights):
+        for k, v in list(weights.items()):
+            if "conv1d.weight" in k and v.shape[-1] != 1:
+                weights[k] = v.moveaxis(2, 1)
+
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        for l in range(self.args.num_hidden_layers):
+            base = f"model.layers.{l}.feed_forward"
+            if not any(key.startswith(f"{base}.experts.") for key in weights.keys()):
+                continue
+
+            for proj in ["gate_proj", "down_proj", "up_proj"]:
+                for name in ["weight", "bias", "scales", "biases"]:
+                    expert_tensors = [
+                        weights.pop(f"{base}.experts.{e}.{proj}.{name}")
+                        for e in range(len(weights))
+                        if f"{base}.experts.{e}.{proj}.{name}" in weights
+                    ]
+                    if expert_tensors:
+                        weights[f"{base}.switch_mlp.{proj}.{name}"] = mx.stack(
+                            expert_tensors
+                        )
+
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("router"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate

mlx_lm/models/kimi_k25.py

@@ -0,0 +1,83 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from .base import BaseModelArgs
+from .deepseek_v3 import DeepseekV3Model
+from .deepseek_v3 import Model as DeepseekV3LM
+from .deepseek_v3 import ModelArgs as TextConfig
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    text_config: Union[TextConfig, dict]
+    model_type: str = "kimi_k25"
+
+    def __post_init__(self):
+        if isinstance(self.text_config, dict):
+            self.text_config = TextConfig.from_dict(self.text_config)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.args = config
+        self.model = DeepseekV3Model(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.language_model = LanguageModel(config.text_config)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        return self.language_model(inputs, cache)
+
+    def sanitize(self, weights):
+        weights = tree_unflatten(list(weights.items()))
+        weights.pop("vision_tower", None)
+        weights.pop("vision_model", None)
+        weights.pop("multi_modal_projector", None)
+        weights.pop("mm_projector", None)
+        lm_weights = dict(tree_flatten(weights["language_model"]))
+        lm_weights = DeepseekV3LM.sanitize(self.language_model, lm_weights)
+        weights["language_model"] = tree_unflatten(list(lm_weights.items()))
+        return dict(tree_flatten(weights))
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        DeepseekV3LM.shard(self.language_model, group)
+
+    @property
+    def model(self):
+        return self.language_model.model
+
+    @property
+    def layers(self):
+        return self.language_model.model.pipeline_layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate

mlx_lm/models/kimi_linear.py

@@ -0,0 +1,611 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, KVCache
+from .gated_delta import gated_delta_update
+from .mla import MultiLinear
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    intermediate_size: int
+    head_dim: int
+    rope_theta: float
+    rms_norm_eps: float
+    linear_attn_config: Dict[str, Any]
+    model_max_length: int
+    num_experts: int
+    moe_intermediate_size: int
+    kv_lora_rank: int
+    rope_scaling: Optional[Dict[str, Any]] = None
+    tie_word_embeddings: bool = False
+    qk_nope_head_dim: Optional[int] = None
+    qk_rope_head_dim: Optional[int] = None
+    v_head_dim: Optional[int] = None
+    mla_use_nope: bool = False
+    num_experts_per_token: int = 1
+    num_shared_experts: int = 0
+    moe_router_activation_func: str = "sigmoid"
+    moe_renormalize: bool = True
+    routed_scaling_factor: float = 1.0
+    first_k_dense_replace: int = 0
+    moe_layer_freq: int = 1
+    use_grouped_topk: bool = True
+    num_expert_group: int = 1
+    topk_group: int = 1
+
+
+class KimiMLP(nn.Module):
+    def __init__(
+        self,
+        args: ModelArgs,
+        hidden_size: Optional[int] = None,
+        intermediate_size: Optional[int] = None,
+    ):
+        super().__init__()
+        dim = hidden_size or args.hidden_size
+        hidden = intermediate_size or args.intermediate_size
+        self.gate_proj = nn.Linear(dim, hidden, bias=False)
+        self.up_proj = nn.Linear(dim, hidden, bias=False)
+        self.down_proj = nn.Linear(hidden, dim, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+@mx.compile
+def _group_expert_select(
+    gates: mx.array,
+    bias: Optional[mx.array],
+    top_k: int,
+    n_group: int,
+    topk_group: int,
+    routed_scaling_factor: float,
+    renormalize: bool,
+    score_function: str,
+) -> Tuple[mx.array, mx.array]:
+    if score_function == "sigmoid":
+        scores = mx.sigmoid(gates)
+    elif score_function == "softmax":
+        scores = mx.softmax(gates, axis=-1, precise=True)
+    else:
+        raise ValueError(f"Unsupported MoE router activation '{score_function}'")
+
+    orig_scores = scores
+    if bias is not None:
+        scores = scores + bias.astype(scores.dtype)
+
+    if n_group > 1:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores,
+            mx.stop_gradient(group_idx),
+            mx.array(0.0, dtype=scores.dtype),
+            axis=-2,
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    inds = mx.argpartition(-scores, kth=top_k - 1, axis=-1)[..., :top_k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+
+    if top_k > 1 and renormalize:
+        denominator = scores.sum(axis=-1, keepdims=True) + 1e-20
+        scores = scores / denominator
+
+    return inds, scores * routed_scaling_factor
+
+
+class KimiSparseMoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        hidden = args.hidden_size
+        experts = args.num_experts
+        if experts is None:
+            raise ValueError("num_experts must be specified for MoE layers")
+
+        self.gate = nn.Linear(hidden, experts, bias=False)
+        self.switch_mlp = SwitchGLU(hidden, args.moe_intermediate_size, experts)
+        self.e_score_correction_bias = mx.zeros((experts,), dtype=mx.float32)
+
+        if args.num_shared_experts:
+            shared_hidden = args.moe_intermediate_size * args.num_shared_experts
+            self.shared_experts = KimiMLP(args, intermediate_size=shared_hidden)
+        else:
+            self.shared_experts = None
+
+    def __call__(self, x: mx.array) -> mx.array:
+        scores = self.gate(x)
+        inds, weights = _group_expert_select(
+            scores,
+            self.e_score_correction_bias,
+            self.args.num_experts_per_token,
+            self.args.num_expert_group,
+            self.args.topk_group,
+            self.args.routed_scaling_factor,
+            self.args.moe_renormalize,
+            self.args.moe_router_activation_func,
+        )
+        out = self.switch_mlp(x, inds)
+        out = (out * weights[..., None]).sum(axis=-2)
+        if self.shared_experts is not None:
+            out = out + self.shared_experts(x)
+        return out
+
+
+class KimiMLAAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.num_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.qk_nope_head_dim = args.qk_nope_head_dim or args.head_dim
+        self.qk_rope_head_dim = args.qk_rope_head_dim or 0
+        self.q_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+        self.v_head_dim = args.v_head_dim or args.head_dim
+        self.kv_lora_rank = args.kv_lora_rank
+        self.scale = self.q_head_dim**-0.5
+
+        hidden = args.hidden_size
+        self.q_proj = nn.Linear(hidden, self.num_heads * self.q_head_dim, bias=False)
+        self.kv_a_proj_with_mqa = nn.Linear(
+            hidden,
+            args.kv_lora_rank + self.qk_rope_head_dim,
+            bias=False,
+        )
+        self.kv_a_layernorm = nn.RMSNorm(args.kv_lora_rank, eps=args.rms_norm_eps)
+        self.embed_q = MultiLinear(
+            self.qk_nope_head_dim, args.kv_lora_rank, self.num_heads
+        )
+        self.unembed_out = MultiLinear(
+            args.kv_lora_rank, self.v_head_dim, self.num_heads
+        )
+        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, hidden, bias=False)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        q = self.q_proj(x).reshape(B, L, self.num_heads, self.q_head_dim)
+        q = q.transpose(0, 2, 1, 3)
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+        kv_latent = self.kv_a_layernorm(compressed_kv)
+
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
+
+        if cache is not None:
+            kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe)
+
+        pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2)
+        if mask is not None:
+            pe_scores = mx.where(
+                mask,
+                pe_scores,
+                mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype),
+            )
+
+        if L == 1:
+            q_nope = self.embed_q(q_nope)
+            k = v = kv_latent
+        else:
+            k = self.embed_q(kv_latent, transpose=False)
+            v = self.unembed_out(kv_latent)
+
+        output = scaled_dot_product_attention(
+            q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores
+        )
+
+        if L == 1:
+            output = self.unembed_out(output)
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class ShortConv1d(nn.Module):
+    def __init__(self, channels: int, kernel_size: int):
+        super().__init__()
+        self.kernel_size = kernel_size
+        self.conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            kernel_size=kernel_size,
+            bias=False,
+            groups=channels,
+            padding=0,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        state: Optional[mx.array],
+        mask: Optional[mx.array],
+        lengths: Optional[mx.array],
+    ) -> Tuple[mx.array, mx.array]:
+        if mask is not None:
+            x = mx.where(mask[..., None], x, 0)
+
+        if state is None:
+            state = mx.zeros(
+                (x.shape[0], self.kernel_size - 1, x.shape[-1]), dtype=x.dtype
+            )
+        conv_input = mx.concatenate([state, x], axis=1)
+        out = nn.silu(self.conv(conv_input))
+        n_keep = self.kernel_size - 1
+        if lengths is not None:
+            ends = mx.clip(lengths, 0, x.shape[1])
+            positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+            new_state = mx.take_along_axis(conv_input, positions, axis=1)
+        else:
+            new_state = mx.contiguous(conv_input[:, -n_keep:, :])
+
+        return out, new_state
+
+
+class KimiDeltaAttention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        cfg = args.linear_attn_config
+
+        self.layer_idx = layer_idx
+        self.num_heads = cfg["num_heads"]
+        self.head_dim = cfg["head_dim"]
+        self.conv_kernel = cfg.get("short_conv_kernel_size", 4)
+
+        self.projection_dim = self.num_heads * self.head_dim
+        hidden = args.hidden_size
+
+        self.scale = float(self.head_dim) ** -0.5
+
+        self.q_proj = nn.Linear(hidden, self.projection_dim, bias=False)
+        self.k_proj = nn.Linear(hidden, self.projection_dim, bias=False)
+        self.v_proj = nn.Linear(hidden, self.projection_dim, bias=False)
+
+        self.q_conv = ShortConv1d(self.projection_dim, self.conv_kernel)
+        self.k_conv = ShortConv1d(self.projection_dim, self.conv_kernel)
+        self.v_conv = ShortConv1d(self.projection_dim, self.conv_kernel)
+
+        self.f_a_proj = nn.Linear(hidden, self.head_dim, bias=False)
+        self.f_b_proj = nn.Linear(self.head_dim, self.projection_dim, bias=False)
+        self.b_proj = nn.Linear(hidden, self.num_heads, bias=False)
+
+        self.g_a_proj = nn.Linear(hidden, self.head_dim, bias=False)
+        self.g_b_proj = nn.Linear(self.head_dim, self.projection_dim, bias=False)
+
+        self.A_log = mx.expand_dims(
+            mx.log(mx.random.uniform(low=1.0, high=16.0, shape=(self.num_heads,))),
+            (0, 1, 3),
+        )
+        self.dt_bias = mx.zeros((self.projection_dim,))
+
+        self.o_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
+        self.o_proj = nn.Linear(self.projection_dim, hidden, bias=False)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, T, _ = x.shape
+        dtype = x.dtype
+
+        if cache is not None:
+            q_state, k_state, v_state, ssm_state = cache
+            lengths = cache.lengths
+        else:
+            q_state = None
+            k_state = None
+            v_state = None
+            ssm_state = None
+            lengths = None
+
+        if q_state is None:
+            s = mx.zeros((B, self.conv_kernel - 1, self.projection_dim), dtype=dtype)
+            q_state = s
+            k_state = s
+            v_state = s
+
+        q_conv, q_state = self.q_conv(self.q_proj(x), q_state, mask, lengths)
+        k_conv, k_state = self.k_conv(self.k_proj(x), k_state, mask, lengths)
+        v_conv, v_state = self.v_conv(self.v_proj(x), v_state, mask, lengths)
+
+        if cache is not None:
+            cache[0] = q_state
+            cache[1] = k_state
+            cache[2] = v_state
+
+        q = q_conv.reshape(B, T, self.num_heads, self.head_dim)
+        k = k_conv.reshape(B, T, self.num_heads, self.head_dim)
+        v = v_conv.reshape(B, T, self.num_heads, self.head_dim)
+
+        inv_scale = self.scale
+        q = (inv_scale**2) * mx.fast.rms_norm(q, None, 1e-6)
+        k = inv_scale * mx.fast.rms_norm(k, None, 1e-6)
+
+        a_logits = self.f_b_proj(self.f_a_proj(x)).reshape(
+            B, T, self.num_heads, self.head_dim
+        )
+        b_logits = self.b_proj(x).reshape(B, T, self.num_heads)
+
+        out, ssm_state = gated_delta_update(
+            q,
+            k,
+            v,
+            a_logits,
+            b_logits,
+            self.A_log.reshape(self.num_heads, 1),
+            self.dt_bias.reshape(self.num_heads, self.head_dim),
+            state=ssm_state,
+            mask=mask,
+            use_kernel=not self.training,
+        )
+
+        if cache is not None:
+            cache[3] = ssm_state
+            cache.advance(T)
+
+        gate = self.g_b_proj(self.g_a_proj(x)).reshape(
+            B, T, self.num_heads, self.head_dim
+        )
+        out = (
+            self.o_norm(out.reshape(B, T, self.num_heads, self.head_dim))
+            * mx.sigmoid(gate)
+        ).reshape(B, T, -1)
+        return self.o_proj(out)
+
+
+class KimiDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        kda_layers = args.linear_attn_config["kda_layers"]
+        self.is_linear = (layer_idx + 1) in kda_layers
+
+        if self.is_linear:
+            self.self_attn = KimiDeltaAttention(args, layer_idx)
+        else:
+            self.self_attn = KimiMLAAttention(args)
+
+        if (
+            args.num_experts > 0
+            and layer_idx >= args.first_k_dense_replace
+            and layer_idx % args.moe_layer_freq == 0
+        ):
+            self.mlp = KimiSparseMoE(args)
+        else:
+            self.mlp = KimiMLP(args)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        attn_cache = None if cache is None else cache
+        y = self.self_attn(self.input_layernorm(x), mask, attn_cache)
+        h = x + y
+        z = self.mlp(self.post_attention_layernorm(h))
+        return h + z
+
+
+class KimiLinearModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [KimiDecoderLayer(args, i) for i in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        kda_layers = args.linear_attn_config["kda_layers"]
+        self.ssm_idx = kda_layers[0] - 1
+        for i in range(len(self.layers)):
+            if (i + 1) not in kda_layers:
+                self.attn_idx = i
+                break
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        ssm_mask = create_ssm_mask(h, cache[self.ssm_idx])
+        attn_mask = create_attention_mask(h, cache[self.attn_idx], return_array=True)
+
+        for layer, layer_cache in zip(self.layers, cache):
+            mask = ssm_mask if layer.is_linear else attn_mask
+            h = layer(h, mask=mask, cache=layer_cache)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = KimiLinearModel(args)
+        if args.tie_word_embeddings:
+            self.lm_head = None
+        else:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache)
+        if self.lm_head is None:
+            return self.model.embed_tokens.as_linear(out)
+        return self.lm_head(out)
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        caches: List[Any] = []
+        for layer in self.layers:
+            if layer.is_linear:
+                caches.append(ArraysCache(size=4))
+            else:
+                caches.append(KVCache())
+        return caches
+
+    def sanitize(self, weights: Dict[str, mx.array]) -> Dict[str, mx.array]:
+        weights = {k: v for k, v in weights.items() if not k.startswith("model.mtp")}
+
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        for layer_idx, layer in enumerate(self.layers):
+            prefix = f"model.layers.{layer_idx}"
+
+            if isinstance(layer.mlp, KimiSparseMoE):
+                src_prefix = f"{prefix}.block_sparse_moe"
+                dst_prefix = f"{prefix}.mlp"
+                for src, dst in [
+                    ("w1", "gate_proj"),
+                    ("w2", "down_proj"),
+                    ("w3", "up_proj"),
+                ]:
+                    key = f"{src_prefix}.experts.0.{src}.weight"
+                    if key in weights:
+                        stacked = [
+                            weights.pop(f"{src_prefix}.experts.{i}.{src}.weight")
+                            for i in range(self.args.num_experts)
+                        ]
+                        weights[f"{dst_prefix}.switch_mlp.{dst}.weight"] = mx.stack(
+                            stacked
+                        )
+
+                for name in ("gate_proj", "up_proj", "down_proj"):
+                    src_key = f"{src_prefix}.shared_experts.{name}.weight"
+                    if src_key in weights:
+                        weights[f"{dst_prefix}.shared_experts.{name}.weight"] = (
+                            weights.pop(src_key)
+                        )
+
+                gate_key = f"{src_prefix}.gate.weight"
+                if gate_key in weights:
+                    weights[f"{dst_prefix}.gate.weight"] = weights.pop(gate_key)
+
+                bias_key = f"{src_prefix}.gate.e_score_correction_bias"
+                if bias_key in weights:
+                    weights[f"{dst_prefix}.e_score_correction_bias"] = weights.pop(
+                        bias_key
+                    )
+
+            attn = getattr(layer, "self_attn", None)
+            if isinstance(attn, KimiDeltaAttention):
+                attn_prefix = f"{prefix}.self_attn"
+                for src_name, dst_name in (
+                    ("q_conv1d", "q_conv"),
+                    ("k_conv1d", "k_conv"),
+                    ("v_conv1d", "v_conv"),
+                ):
+                    src_key = f"{attn_prefix}.{src_name}.weight"
+                    if src_key in weights:
+                        w = weights.pop(src_key)
+                        if w.ndim == 3:
+                            w = w.moveaxis(2, 1)
+                        weights[f"{attn_prefix}.{dst_name}.conv.weight"] = w
+                dt_key = f"{attn_prefix}.dt_bias"
+                if dt_key in weights:
+                    if weights[dt_key].ndim > 1:
+                        weights[dt_key] = mx.reshape(weights[dt_key], (-1,))
+
+            attn_prefix = f"{prefix}.self_attn"
+            kv_b_key = f"{attn_prefix}.kv_b_proj.weight"
+            if kv_b_key in weights:
+                qk_nope = self.args.qk_nope_head_dim or self.args.head_dim
+                v_head = self.args.v_head_dim or self.args.head_dim
+                head_dim = qk_nope + v_head
+                num_heads = self.args.num_attention_heads
+
+                quantized = f"{attn_prefix}.kv_b_proj.scales" in weights
+                v = weights.pop(kv_b_key)
+
+                if quantized:
+                    dims = self.args.kv_lora_rank
+                    scales = weights.pop(f"{attn_prefix}.kv_b_proj.scales")
+                    biases = weights.pop(f"{attn_prefix}.kv_b_proj.biases")
+                    bits = (v.shape[-1] * 32) // dims
+                    group_size = dims // scales.shape[-1]
+                    v = mx.dequantize(
+                        v, scales, biases, bits=bits, group_size=group_size
+                    )
+
+                v = v.reshape(num_heads, head_dim, -1)
+                wk = mx.contiguous(v[:, :qk_nope, :].swapaxes(-1, -2))
+                wv = mx.contiguous(v[:, qk_nope:, :])
+
+                if quantized:
+                    wk, wk_s, wk_b = mx.quantize(wk, bits=bits, group_size=group_size)
+                    wv, wv_s, wv_b = mx.quantize(wv, bits=bits, group_size=group_size)
+                    weights[f"{attn_prefix}.embed_q.scales"] = wk_s
+                    weights[f"{attn_prefix}.embed_q.biases"] = wk_b
+                    weights[f"{attn_prefix}.unembed_out.scales"] = wv_s
+                    weights[f"{attn_prefix}.unembed_out.biases"] = wv_b
+
+                weights[f"{attn_prefix}.embed_q.weight"] = wk
+                weights[f"{attn_prefix}.unembed_out.weight"] = wv
+
+        return weights
+
+    @property
+    def cast_predicate(self):
+        def predicate(path: str):
+            if "e_score_correction_bias" in path:
+                return False
+            if path.endswith("A_log") or path.endswith("dt_bias"):
+                return False
+            return True
+
+        return predicate
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("mlp.gate"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate

mlx_lm/models/kimi_vl.py

@@ -30,9 +30,9 @@ class TextArgs(BaseModelArgs):
     topk_method: str = "noaux_tc"
     scoring_func: str = "sigmoid"
     norm_topk_prob: bool = True
-    n_group: Optional[int] = None
-    topk_group: Optional[int] = None
-    num_experts_per_tok: Optional[int] = None
+    n_group: int = 1
+    topk_group: int = 1
+    num_experts_per_tok: int = 1
     moe_layer_freq: int = 1
     first_k_dense_replace: int = 0
     max_position_embeddings: int = 2048
@@ -62,9 +62,8 @@ class LanguageModel(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, cache, mask)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
 
@@ -79,9 +78,8 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache: Optional[Any] = None,
-        mask: Optional[mx.array] = None,
     ):
-        return self.language_model(inputs, cache, mask)
+        return self.language_model(inputs, cache)
 
     def sanitize(self, weights):
         def keep(key):

mlx_lm/models/lfm2-vl.py

@@ -18,11 +18,6 @@ class ModelArgs(BaseModelArgs):
 
     def __post_init__(self):
         self.text_config["tie_word_embeddings"] = False
-        self.text_config["full_attn_idxs"] = [
-            i
-            for i, layer_type in enumerate(self.text_config["layer_types"])
-            if layer_type == "full_attention"
-        ]
 
 
 class Model(nn.Module):
@@ -36,11 +31,10 @@ class Model(nn.Module):
         self,
         inputs: mx.array,
         cache=None,
-        mask: Optional[mx.array] = None,
         input_embeddings: Optional[mx.array] = None,
     ):
         return self.language_model(
-            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+            inputs, cache=cache, input_embeddings=input_embeddings
         )
 
     def sanitize(self, weights):

mlx_lm/models/lfm2.py

@@ -5,7 +5,13 @@ from typing import Any, List, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
-from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
 from .cache import ArraysCache, KVCache
 
 
@@ -26,8 +32,22 @@ class ModelArgs(BaseModelArgs):
     block_multiple_of: int
     block_ffn_dim_multiplier: float
     block_auto_adjust_ff_dim: bool
-    full_attn_idxs: List[int]
-    rope_theta: float
+    rope_theta: float = 1000000.0
+    rope_parameters: Optional[dict] = None
+    full_attn_idxs: Optional[List[int]] = None
+    layer_types: Optional[List[str]] = None
+
+    def __post_init__(self):
+        if self.rope_parameters is not None and "rope_theta" in self.rope_parameters:
+            self.rope_theta = self.rope_parameters["rope_theta"]
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+        if self.full_attn_idxs is None:
+            self.full_attn_idxs = [
+                i
+                for i, layer_type in enumerate(self.layer_types)
+                if layer_type == "full_attention"
+            ]
 
 
 class Attention(nn.Module):
@@ -114,24 +134,36 @@ class ShortConv(nn.Module):
     def __call__(
         self,
         x: mx.array,
+        mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ):
-        seqlen = x.shape[1]
         BCx = self.in_proj(x)
         B, C, x = mx.split(BCx, 3, axis=-1)
         Bx = B * x
+        if mask is not None:
+            Bx = mx.where(mask[..., None], Bx, 0)
 
-        state = None
         if cache is not None:
-            state = cache[0]
-        if state is None:
-            state = mx.zeros(
-                (Bx.shape[0], self.L_cache - 1, self.args.hidden_size), dtype=Bx.dtype
-            )
+            if cache[0] is None:
+                state = mx.zeros(
+                    (Bx.shape[0], self.L_cache - 1, self.args.hidden_size),
+                    dtype=Bx.dtype,
+                )
+            else:
+                state = cache[0]
+            Bx = mx.concatenate([state, Bx], axis=1)
+            n_keep = self.L_cache - 1
+            t = x.shape[1]
+            if cache.lengths is not None:
+                ends = mx.clip(cache.lengths, 0, t)
+                positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+                cache[0] = mx.take_along_axis(Bx, positions, axis=1)
+            else:
+                cache[0] = Bx[:, -n_keep:, :]
+            cache.advance(t)
+        else:
+            Bx = mx.pad(Bx, [(0, 0), (self.L_cache - 1, 0), (0, 0)])
 
-        Bx = mx.concatenate([state, Bx], axis=-2)
-        if cache is not None:
-            cache[0] = Bx[:, -(self.L_cache - 1) :]
         conv_out = self.conv(Bx)
 
         y = C * conv_out
@@ -159,7 +191,7 @@ class MLP(nn.Module):
         self.w2 = nn.Linear(ff_dim, dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.w2(nn.silu(self.w1(x)) * self.w3(x))
+        return self.w2(swiglu(self.w1(x), self.w3(x)))
 
 
 class Lfm2DecoderLayer(nn.Module):
@@ -194,6 +226,7 @@ class Lfm2DecoderLayer(nn.Module):
         else:
             r = self.conv(
                 self.operator_norm(x),
+                mask=mask,
                 cache=cache,
             )
         h = x + r
@@ -214,10 +247,17 @@ class Lfm2Model(nn.Module):
 
         self.embedding_norm = nn.RMSNorm(args.hidden_size, eps=args.norm_eps)
 
+        self.fa_idx = args.full_attn_idxs[0]
+        self.conv_idx = 0
+        for i in range(args.num_hidden_layers):
+            if i in args.full_attn_idxs:
+                self.conv_idx += 1
+            else:
+                break
+
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
@@ -226,15 +266,14 @@ class Lfm2Model(nn.Module):
         else:
             h = self.embed_tokens(inputs)
 
-        if mask is None:
-            first_attn_idx = self.args.full_attn_idxs[0]
-            c = [cache[first_attn_idx]] if cache is not None else None
-            mask = create_attention_mask(h, c)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        attn_mask = create_attention_mask(h, cache[self.fa_idx])
+        conv_mask = create_ssm_mask(h, cache[self.conv_idx])
+
         for layer, c in zip(self.layers, cache):
+            mask = attn_mask if layer.is_attention_layer else conv_mask
             h = layer(h, mask, cache=c)
 
         return self.embedding_norm(h)
@@ -250,11 +289,10 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, mask, cache, input_embeddings)
+        out = self.model(inputs, cache, input_embeddings)
         return self.model.embed_tokens.as_linear(out)
 
     def sanitize(self, weights):

mlx_lm/models/lfm2_moe.py

@@ -0,0 +1,387 @@
+# Copyright © 2025 Apple Inc.
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+    scaled_dot_product_attention,
+)
+from .cache import ArraysCache, KVCache
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    moe_intermediate_size: int
+    num_hidden_layers: int
+    num_experts: int
+    num_experts_per_tok: int
+    norm_topk_prob: bool
+    num_attention_heads: int
+    num_key_value_heads: int
+    max_position_embeddings: int
+    use_expert_bias: bool
+    num_dense_layers: int
+    norm_eps: float
+    conv_bias: bool
+    conv_L_cache: int
+    rope_theta: float = 1000000.0
+    rope_parameters: Optional[dict] = None
+    full_attn_idxs: Optional[List[int]] = None
+    layer_types: Optional[List[str]] = None
+
+    def __post_init__(self):
+        if self.rope_parameters is not None and "rope_theta" in self.rope_parameters:
+            self.rope_theta = self.rope_parameters["rope_theta"]
+        if self.full_attn_idxs is None:
+            self.full_attn_idxs = [
+                i
+                for i, layer_type in enumerate(self.layer_types)
+                if layer_type == "full_attention"
+            ]
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.scale = head_dim**-0.5
+
+        self.q_layernorm = nn.RMSNorm(head_dim, eps=args.norm_eps)
+        self.k_layernorm = nn.RMSNorm(head_dim, eps=args.norm_eps)
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.out_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.rope = nn.RoPE(
+            self.head_dim,
+            base=args.rope_theta,
+            traditional=False,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = self.q_layernorm(queries.reshape(B, L, self.n_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        keys = self.k_layernorm(keys.reshape(B, L, self.n_kv_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, mask=mask, scale=self.scale
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.out_proj(output)
+
+
+class ShortConv(nn.Module):
+    def __init__(
+        self,
+        args: ModelArgs,
+        layer_idx: int,
+    ):
+        super().__init__()
+        self.args = args
+        self.layer_idx = layer_idx
+        self.L_cache = args.conv_L_cache
+        self.bias = args.conv_bias
+
+        self.conv = nn.Conv1d(
+            in_channels=args.hidden_size,
+            out_channels=args.hidden_size,
+            kernel_size=self.L_cache,
+            groups=args.hidden_size,
+            bias=self.bias,
+        )
+        self.in_proj = nn.Linear(args.hidden_size, 3 * args.hidden_size, bias=self.bias)
+        self.out_proj = nn.Linear(args.hidden_size, args.hidden_size, bias=self.bias)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ):
+        BCx = self.in_proj(x)
+        B, C, x = mx.split(BCx, 3, axis=-1)
+        Bx = B * x
+        if mask is not None:
+            Bx = mx.where(mask[..., None], Bx, 0)
+
+        if cache is not None:
+            if cache[0] is None:
+                state = mx.zeros(
+                    (Bx.shape[0], self.L_cache - 1, self.args.hidden_size),
+                    dtype=Bx.dtype,
+                )
+            else:
+                state = cache[0]
+            Bx = mx.concatenate([state, Bx], axis=1)
+            n_keep = self.L_cache - 1
+            t = x.shape[1]
+            if cache.lengths is not None:
+                ends = mx.clip(cache.lengths, 0, t)
+                positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+                cache[0] = mx.take_along_axis(Bx, positions, axis=1)
+            else:
+                cache[0] = Bx[:, -n_keep:, :]
+            cache.advance(t)
+        else:
+            Bx = mx.pad(Bx, [(0, 0), (self.L_cache - 1, 0), (0, 0)])
+
+        conv_out = self.conv(Bx)
+
+        y = C * conv_out
+        return self.out_proj(y)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: ModelArgs, intermediate_size: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class Lfm2MoeSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        intermediate_size = args.moe_intermediate_size
+
+        self.num_experts = num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+        self.use_expert_bias = args.use_expert_bias
+
+        self.gate = nn.Linear(dim, num_experts, bias=False)
+        self.switch_mlp = SwitchGLU(dim, intermediate_size, num_experts)
+        if self.use_expert_bias:
+            self.expert_bias = mx.zeros((self.num_experts,))
+
+    def __call__(
+        self,
+        x: mx.array,
+    ):
+        gates = self.gate(x).astype(mx.float32)
+        gates = mx.softmax(gates, axis=-1)
+
+        if self.use_expert_bias:
+            gates += self.expert_bias
+
+        k = self.top_k
+        inds = mx.argpartition(gates, kth=-k, axis=-1)[..., -k:]
+
+        scores = mx.take_along_axis(gates, inds, axis=-1)
+        if self.norm_topk_prob:
+            scores /= mx.sum(scores, axis=-1, keepdims=True) + 1e-20
+        scores = scores.astype(x.dtype)
+
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2)
+
+        return y
+
+
+class Lfm2DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.is_attention_layer = layer_idx in args.full_attn_idxs
+
+        if self.is_attention_layer:
+            self.self_attn = Attention(args)
+        else:
+            self.conv = ShortConv(args, layer_idx)
+        self.feed_forward = (
+            MLP(
+                config=args,
+                intermediate_size=args.intermediate_size,
+            )
+            if layer_idx < args.num_dense_layers
+            else Lfm2MoeSparseMoeBlock(args)
+        )
+
+        self.operator_norm = nn.RMSNorm(args.hidden_size, eps=args.norm_eps)
+        self.ffn_norm = nn.RMSNorm(args.hidden_size, eps=args.norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+
+        if self.is_attention_layer:
+            r = self.self_attn(self.operator_norm(x), mask=mask, cache=cache)
+        else:
+            r = self.conv(
+                self.operator_norm(x),
+                mask=mask,
+                cache=cache,
+            )
+        h = x + r
+        out = h + self.feed_forward(self.ffn_norm(h))
+        return out
+
+
+class Lfm2Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            Lfm2DecoderLayer(args, layer_idx=i) for i in range(args.num_hidden_layers)
+        ]
+
+        self.embedding_norm = nn.RMSNorm(args.hidden_size, eps=args.norm_eps)
+
+        self.fa_idx = args.full_attn_idxs[0]
+        self.conv_idx = 0
+        for i in range(args.num_hidden_layers):
+            if i in args.full_attn_idxs:
+                self.conv_idx += 1
+            else:
+                break
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        if input_embeddings is not None:
+            h = input_embeddings
+        else:
+            h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        attn_mask = create_attention_mask(h, cache[self.fa_idx])
+        conv_mask = create_ssm_mask(h, cache[self.conv_idx])
+
+        for layer, c in zip(self.layers, cache):
+            mask = attn_mask if layer.is_attention_layer else conv_mask
+            h = layer(h, mask, cache=c)
+
+        return self.embedding_norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Lfm2Model(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        out = self.model(inputs, cache, input_embeddings)
+        return self.model.embed_tokens.as_linear(out)
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        for name, param in weights.items():
+            if "conv.weight" in name:
+                if param.shape[-1] > param.shape[1]:
+                    param = param.transpose(0, 2, 1)
+            replacements = {
+                "w1.weight": "gate_proj.weight",
+                "w2.weight": "down_proj.weight",
+                "w3.weight": "up_proj.weight",
+            }
+            for old, new in replacements.items():
+                if old in name:
+                    name = name.replace(old, new)
+            sanitized_weights[name] = param
+
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            # Only sanitize MoE layer weights
+            for n in ["gate_proj", "down_proj", "up_proj"]:
+                if f"{prefix}.feed_forward.experts.0.{n}.weight" in sanitized_weights:
+                    to_join = [
+                        sanitized_weights.pop(
+                            f"{prefix}.feed_forward.experts.{e}.{n}.weight"
+                        )
+                        for e in range(self.args.num_experts)
+                    ]
+                    sanitized_weights[
+                        f"{prefix}.feed_forward.switch_mlp.{n}.weight"
+                    ] = mx.stack(to_join)
+        return sanitized_weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        return [
+            KVCache() if l.is_attention_layer else ArraysCache(size=1)
+            for l in self.layers
+        ]
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("feed_forward.gate"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "expert_bias" not in k
+
+        return predicate

mlx_lm/models/lille-130m.py

@@ -0,0 +1,155 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    block_size: int
+    layer_norm_eps: float
+    n_embd: int
+    n_head: int
+    n_kv_heads: int
+    n_layer: int
+    rope_theta: float
+    vocab_size: int
+    tie_word_embeddings: bool = True
+
+
+class Lille130mAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.n_head = args.n_head
+        self.n_kv_heads = args.n_kv_heads
+        self.head_dim = args.n_embd // args.n_head
+        self.scale = self.head_dim**-0.5
+
+        self.qkv_proj = nn.Linear(
+            args.n_embd, (args.n_head + 2 * args.n_kv_heads) * self.head_dim, bias=False
+        )
+        self.out_proj = nn.Linear(args.n_head * self.head_dim, args.n_embd, bias=False)
+
+        self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps)
+
+        self.rope = nn.RoPE(args.n_embd // args.n_head, True, args.rope_theta)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        qkv = self.qkv_proj(self.norm(x))
+
+        q_size = self.n_head * self.head_dim
+        kv_size = self.n_kv_heads * self.head_dim
+
+        queries, keys, values = mx.split(qkv, [q_size, q_size + kv_size], axis=-1)
+
+        queries = queries.reshape(B, L, self.n_head, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.out_proj(output)
+
+
+class Lille130mMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        hidden_dim = 256 * round(int(8 * args.n_embd / 3) / 256)
+
+        self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps)
+        self.gate_proj = nn.Linear(args.n_embd, hidden_dim, bias=False)
+        self.up_proj = nn.Linear(args.n_embd, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, args.n_embd, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        h = self.norm(x)
+        return self.down_proj(swiglu(self.gate_proj(h), self.up_proj(h)))
+
+
+class Lille130Block(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.attention = Lille130mAttention(args)
+        self.feed_forward = Lille130mMLP(args)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = x + self.attention(x, mask, cache)
+        out = h + self.feed_forward(h)
+        return out
+
+
+class Lille130(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.tok_embeddings = nn.Embedding(args.vocab_size, args.n_embd)
+        self.layers = [Lille130Block(args=args) for _ in range(args.n_layer)]
+        self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.tok_embeddings(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.tok_embeddings.as_linear(self.norm(h))
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.transformer = Lille130(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        return self.transformer(inputs, cache=cache)
+
+    @property
+    def layers(self):
+        return self.transformer.layers
+
+    def sanitize(self, weights):
+        return {k: v for k, v in weights.items() if "rotary_emb" not in k}

mlx_lm/models/llama.py

@@ -1,12 +1,15 @@
 # Copyright © 2023-2024 Apple Inc.
 
 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Union
+from typing import Any, Dict, List, Optional, Union
 
 import mlx.core as mx
 import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_linear
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
 from .rope_utils import initialize_rope
 
 
@@ -28,11 +31,16 @@ class ModelArgs(BaseModelArgs):
     rope_traditional: bool = False
     rope_scaling: Optional[Dict[str, Union[float, str]]] = None
     tie_word_embeddings: bool = True
+    layer_types: Optional[List[str]] = None
+    sliding_window: Optional[int] = None
 
     def __post_init__(self):
         if self.num_key_value_heads is None:
             self.num_key_value_heads = self.num_attention_heads
 
+        if self.layer_types is None:
+            self.layer_types = ["full_attention"] * self.num_hidden_layers
+
 
 class Attention(nn.Module):
     def __init__(self, args: ModelArgs):
@@ -110,14 +118,15 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class TransformerBlock(nn.Module):
-    def __init__(self, args: ModelArgs):
+    def __init__(self, args: ModelArgs, use_sliding: bool = False):
         super().__init__()
         self.num_attention_heads = args.num_attention_heads
         self.hidden_size = args.hidden_size
+        self.use_sliding = use_sliding
         self.self_attn = Attention(args)
         self.mlp = MLP(args)
         self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
@@ -145,17 +154,25 @@ class LlamaModel(nn.Module):
         self.args = args
         self.vocab_size = args.vocab_size
         self.num_hidden_layers = args.num_hidden_layers
+        self.layer_types = args.layer_types
+        self.sliding_window = args.sliding_window
         assert self.vocab_size > 0
         self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
         self.layers = [
-            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+            TransformerBlock(args=args, use_sliding=layer_type == "sliding_attention")
+            for layer_type in self.layer_types
         ]
         self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.fa_idx = self.layer_types.index("full_attention")
+        self.swa_idx = None
+        for e, l in enumerate(self.layers):
+            if l.use_sliding:
+                self.swa_idx = e
+                break
 
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
@@ -164,14 +181,18 @@ class LlamaModel(nn.Module):
         else:
             h = self.embed_tokens(inputs)
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-
         if cache is None:
             cache = [None] * len(self.layers)
 
-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, cache=c)
+        fa_mask = create_attention_mask(h, cache[self.fa_idx])
+        if self.swa_idx is not None:
+            swa_mask = create_attention_mask(
+                h, cache[self.swa_idx], window_size=self.sliding_window
+            )
+
+        for layer, cache in zip(self.layers, cache):
+            mask = swa_mask if layer.use_sliding else fa_mask
+            h = layer(h, mask, cache=cache)
 
         return self.norm(h)
 
@@ -188,11 +209,10 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
     ):
-        out = self.model(inputs, mask, cache, input_embeddings)
+        out = self.model(inputs, cache, input_embeddings)
         if self.args.tie_word_embeddings:
             out = self.model.embed_tokens.as_linear(out)
         else:
@@ -208,6 +228,47 @@ class Model(nn.Module):
             weights.pop("lm_head.weight", None)
         return weights
 
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        for layer in self.model.layers:
+            # Shard the self attention
+            layer.self_attn.q_proj = shard_linear(
+                layer.self_attn.q_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.k_proj = shard_linear(
+                layer.self_attn.k_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.v_proj = shard_linear(
+                layer.self_attn.v_proj, "all-to-sharded", group=group
+            )
+            layer.self_attn.o_proj = shard_linear(
+                layer.self_attn.o_proj, "sharded-to-all", group=group
+            )
+            layer.self_attn.n_heads //= N
+            layer.self_attn.n_kv_heads //= N
+
+            # Shard the MLP
+            layer.mlp.gate_proj = shard_linear(
+                layer.mlp.gate_proj, "all-to-sharded", group=group
+            )
+            layer.mlp.down_proj = shard_linear(
+                layer.mlp.down_proj, "sharded-to-all", group=group
+            )
+            layer.mlp.up_proj = shard_linear(
+                layer.mlp.up_proj, "all-to-sharded", group=group
+            )
+
     @property
     def layers(self):
         return self.model.layers
+
+    def make_cache(self):
+        return [
+            (
+                RotatingKVCache(max_size=self.model.sliding_window)
+                if layer.use_sliding
+                else KVCache()
+            )
+            for layer in self.layers
+        ]

mlx_lm/models/llama4.py

@@ -6,6 +6,7 @@ from typing import Any, Optional, Union
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import ChunkedKVCache, KVCache
 from .rope_utils import initialize_rope
@@ -17,7 +18,6 @@ class TextArgs(BaseModelArgs):
     attention_bias: bool
     attention_chunk_size: int
     head_dim: int
-    hidden_act: str
     hidden_size: int
     interleave_moe_layer_step: int
     intermediate_size: int
@@ -146,13 +146,14 @@ class MLP(nn.Module):
         self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
 
     def __call__(self, x) -> mx.array:
-        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
 
 
 class MoE(nn.Module):
     def __init__(self, args):
         super().__init__()
         self.top_k = args.num_experts_per_tok
+        assert self.top_k == 1, "Only 1 expert per token supported"
         self.num_experts = args.num_local_experts
         self.experts = SwitchGLU(
             args.hidden_size, args.intermediate_size, self.num_experts
@@ -219,7 +220,6 @@ class LlamaModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
         h = self.embed_tokens(inputs)
@@ -242,21 +242,15 @@ class LlamaModel(nn.Module):
         token_pos = linds <= rinds
         chunk_mask = (block_pos == 0) & token_pos
 
-        if mask is None:
-            mask = create_attention_mask(h, cache)
-        else:
-            chunk_mask &= mask
-
         if cache is None:
             cache = [None] * len(self.layers)
 
+        global_mask = create_attention_mask(h, cache[3])
+
         for idx, (layer, c) in enumerate(zip(self.layers, cache)):
             use_chunked_attention = (idx + 1) % 4 != 0
-            if use_chunked_attention:
-                local_mask = chunk_mask
-            else:
-                local_mask = mask
-            h = layer(h, local_mask, cache=c)
+            mask = chunk_mask if use_chunked_attention else global_mask
+            h = layer(h, mask, cache=c)
 
         return self.norm(h)
 
@@ -274,10 +268,9 @@ class LanguageModel(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        out = self.model(inputs, mask, cache)
+        out = self.model(inputs, cache)
         return self.lm_head(out)
 
 
@@ -291,10 +284,9 @@ class Model(nn.Module):
     def __call__(
         self,
         inputs: mx.array,
-        mask: mx.array = None,
         cache=None,
     ):
-        return self.language_model(inputs, mask, cache)
+        return self.language_model(inputs, cache)
 
     def sanitize(self, weights):
         def to_remove(k):

mlx_lm/models/llama4_text.py

@@ -0,0 +1,182 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_attention_heads: int
+    num_hidden_layers: int
+    vocab_size: int
+    intermediate_size: int
+    intermediate_size_mlp: int
+    num_key_value_heads: int
+    rms_norm_eps: float
+    rope_theta: float
+    head_dim: int
+    tie_word_embeddings: bool
+    no_rope_layers: list
+    use_qk_norm: bool
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, use_rope):
+        super().__init__()
+        self.args = args
+        self.n_heads = args.num_attention_heads
+        self.n_kv_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size, self.n_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size, self.n_kv_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.n_heads * self.head_dim, args.hidden_size, bias=False
+        )
+        self.use_rope = use_rope
+        if use_rope:
+            self.rope = nn.RoPE(self.head_dim, traditional=True, base=args.rope_theta)
+        self.use_qk_norm = args.use_qk_norm
+        self.rms_norm_eps = args.rms_norm_eps
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1)
+        if self.use_qk_norm:
+            queries = mx.fast.rms_norm(queries, None, self.rms_norm_eps)
+            keys = mx.fast.rms_norm(keys, None, self.rms_norm_eps)
+        queries = queries.transpose(0, 2, 1, 3)
+        keys = keys.transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if self.use_rope:
+            offset = cache.offset if cache is not None else 0
+            queries = self.rope(queries, offset=offset)
+            keys = self.rope(keys, offset=offset)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, intermediate_size, activation=nn.silu):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(dim, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, dim, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs, use_rope):
+        super().__init__()
+        self.self_attn = Attention(args, use_rope)
+
+        self.feed_forward = MLP(
+            args.hidden_size,
+            args.intermediate_size_mlp,
+        )
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.feed_forward(self.post_attention_layernorm(h))
+        return h + r
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args, use_rope=args.no_rope_layers[i])
+            for i in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        mask = create_attention_mask(h, cache[0])
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = LanguageModel(args)
+
+        self.tie_word_embeddings = args.tie_word_embeddings
+        if not self.tie_word_embeddings:
+            self.output = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.model(inputs, cache)
+        if self.tie_word_embeddings:
+            return h @ self.model.embed_tokens.weight.T
+        else:
+            return self.output(h)
+
+    @property
+    def layers(self):
+        return self.model.layers

mlx_lm/models/longcat_flash.py

@@ -0,0 +1,493 @@
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+from .activations import swiglu
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import CacheList, KVCache
+from .mla import MultiLinear
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    attention_method: str
+    zero_expert_type: str
+    hidden_size: int
+    ffn_hidden_size: int
+    moe_topk: int
+    expert_ffn_hidden_size: int
+    n_routed_experts: int
+    zero_expert_num: int
+    num_layers: int
+    vocab_size: int
+    max_position_embeddings: int
+    num_attention_heads: int
+    kv_lora_rank: int
+    q_lora_rank: int
+    qk_rope_head_dim: int
+    qk_nope_head_dim: int
+    v_head_dim: int
+    routed_scaling_factor: float
+    rms_norm_eps: float
+    rope_theta: float
+    mla_scale_q_lora: bool
+    mla_scale_kv_lora: bool
+    attention_bias: bool
+    norm_topk_prob: bool = False
+    router_bias: bool = False
+    rope_scaling: Optional[Dict] = None
+
+
+class LongcatFlashMLA(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.qk_rope_head_dim = args.qk_rope_head_dim
+        self.qk_nope_head_dim = args.qk_nope_head_dim
+        self.kv_lora_rank = args.kv_lora_rank
+        self.q_lora_rank = args.q_lora_rank
+        self.v_head_dim = args.v_head_dim
+
+        self.qk_head_dim = args.qk_nope_head_dim + args.qk_rope_head_dim
+        self.scale = self.qk_head_dim**-0.5
+
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(
+                args.hidden_size,
+                self.num_attention_heads * self.qk_head_dim,
+                bias=False,
+            )
+        else:
+            self.q_a_proj = nn.Linear(
+                args.hidden_size, self.q_lora_rank, bias=args.attention_bias
+            )
+            self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank)
+            self.q_b_proj = nn.Linear(
+                self.q_lora_rank,
+                self.num_attention_heads * self.qk_head_dim,
+                bias=False,
+            )
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            args.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=args.attention_bias,
+        )
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank)
+        self.embed_q = MultiLinear(
+            self.qk_nope_head_dim, self.kv_lora_rank, self.num_attention_heads
+        )
+        self.unembed_out = MultiLinear(
+            self.kv_lora_rank, self.v_head_dim, self.num_attention_heads
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_attention_heads * args.v_head_dim,
+            args.hidden_size,
+            bias=args.attention_bias,
+        )
+
+        if args.mla_scale_q_lora:
+            self.mla_scale_q_lora = (args.hidden_size / self.q_lora_rank) ** 0.5
+        if args.mla_scale_kv_lora:
+            self.mla_scale_kv_lora = (args.hidden_size / self.kv_lora_rank) ** 0.5
+
+        if args.rope_scaling is not None:
+            mscale_all_dim = args.rope_scaling.get("mscale_all_dim", 0)
+            if mscale_all_dim:
+                scaling_factor = args.rope_scaling["factor"]
+                if scaling_factor > 1:
+                    s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0
+                    self.scale = self.scale * s * s
+
+        self.rope = initialize_rope(
+            dims=self.qk_rope_head_dim,
+            base=args.rope_theta,
+            traditional=True,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(x)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
+
+        q = q.reshape(B, L, self.num_attention_heads, self.qk_head_dim).transpose(
+            0, 2, 1, 3
+        )
+
+        if self.mla_scale_q_lora is not None:
+            q = q * self.mla_scale_q_lora
+
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+        kv_latent = self.kv_a_layernorm(compressed_kv)
+
+        if self.mla_scale_kv_lora is not None:
+            kv_latent = kv_latent * self.mla_scale_kv_lora
+
+        offset = cache.offset if cache is not None else 0
+        q_pe = self.rope(q_pe, offset)
+        k_pe = self.rope(k_pe, offset)
+
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
+
+        if cache is not None:
+            kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe)
+
+        pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2)
+        if mask is not None:
+            pe_scores = mx.where(
+                mask,
+                pe_scores,
+                mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype),
+            )
+
+        if L == 1:
+            q_nope = self.embed_q(q_nope)
+            k = v = kv_latent
+        else:
+            k = self.embed_q(kv_latent, transpose=False)
+            v = self.unembed_out(kv_latent)
+
+        output = scaled_dot_product_attention(
+            q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores
+        )
+        if L == 1:
+            output = self.unembed_out(output)
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class LongcatFlashMLP(nn.Module):
+    def __init__(self, args: ModelArgs, is_expert: bool = False):
+        super().__init__()
+        hidden_size = args.expert_ffn_hidden_size if is_expert else args.ffn_hidden_size
+
+        self.gate_proj = nn.Linear(args.hidden_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(args.hidden_size, hidden_size, bias=False)
+        self.down_proj = nn.Linear(hidden_size, args.hidden_size, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class LongcatFlashTopkRouter(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.config = args
+        self.top_k = args.moe_topk
+        self.n_routed_experts = args.n_routed_experts + args.zero_expert_num
+        self.routed_scaling_factor = args.routed_scaling_factor
+        self.norm_topk_prob = args.norm_topk_prob
+        self.router_bias = args.router_bias
+
+        self.classifier = nn.Linear(
+            args.hidden_size, self.n_routed_experts, bias=self.router_bias
+        )
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+
+    def __call__(self, hidden_states: mx.array) -> Tuple[mx.array, mx.array]:
+
+        dtype = hidden_states.dtype
+        router_logits = self.classifier(hidden_states)
+        scores = mx.softmax(router_logits, axis=-1)
+
+        corrected_scores = scores + self.e_score_correction_bias
+        topk_indices = mx.argpartition(corrected_scores, kth=-self.top_k, axis=-1)[
+            ..., -self.top_k :
+        ]
+        topk_weights = mx.take_along_axis(scores, topk_indices, axis=-1)
+
+        if self.norm_topk_prob:
+            denominator = mx.sum(topk_weights, axis=-1, keepdims=True) + 1e-20
+            topk_weights = topk_weights / denominator
+
+        topk_weights = topk_weights * self.routed_scaling_factor
+
+        return topk_indices, topk_weights.astype(dtype)
+
+
+class LongcatFlashMoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.config = args
+        self.num_experts_per_tok = args.moe_topk
+        self.n_routed_experts = args.n_routed_experts
+        self.zero_expert_num = args.zero_expert_num
+        self.zero_expert_type = args.zero_expert_type
+
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size,
+            args.expert_ffn_hidden_size,
+            args.n_routed_experts,
+        )
+
+        self.router = LongcatFlashTopkRouter(args)
+        self.sharding_group = None
+
+    def __call__(self, hidden_states):
+        if self.sharding_group is not None:
+            hidden_states = sum_gradients(self.sharding_group)(hidden_states)
+
+        topk_indices, topk_weights = self.router(hidden_states)
+
+        # Process all regular experts at once
+        mask = topk_indices >= self.n_routed_experts
+        topk_indices = mx.where(mask, 0, topk_indices)
+        regular_weights = mx.where(mask, 0.0, topk_weights)
+
+        regular_outputs = self.switch_mlp(hidden_states, topk_indices)
+
+        weighted_outputs = regular_outputs * regular_weights[..., None]
+        final_output = mx.sum(weighted_outputs, axis=-2)
+
+        if self.sharding_group is not None:
+            final_output = mx.distributed.all_sum(
+                final_output, group=self.sharding_group
+            )
+
+        # Add identity expert contribution after all_sum to avoid summing it N times
+        assert self.zero_expert_type == "identity"
+        identity_weights_sum = mx.sum(
+            mx.where(mask, topk_weights, 0.0), axis=-1, keepdims=True
+        )
+        final_output = final_output + hidden_states * identity_weights_sum
+
+        return final_output
+
+
+class LongcatFlashDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.mlp = LongcatFlashMoE(args)
+
+        self.self_attn = [LongcatFlashMLA(args) for _ in range(2)]
+        self.mlps = [LongcatFlashMLP(args, False) for _ in range(2)]
+        self.input_layernorm = [
+            nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) for _ in range(2)
+        ]
+        self.post_attention_layernorm = [
+            nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) for _ in range(2)
+        ]
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        hidden_states = x
+        shortcut_mlp_output = None
+
+        if cache is None:
+            cache = (None, None)
+
+        for i in range(2):
+            residual = hidden_states
+
+            hidden_states = self.input_layernorm[i](hidden_states)
+            hidden_states = self.self_attn[i](hidden_states, mask=mask, cache=cache[i])
+            hidden_states = residual + hidden_states
+
+            residual = hidden_states
+            hidden_states = self.post_attention_layernorm[i](hidden_states)
+
+            if i == 0:
+                shortcut_mlp_output = self.mlp(hidden_states)
+
+            hidden_states = self.mlps[i](hidden_states)
+            hidden_states = residual + hidden_states
+
+            if i == 1:
+                hidden_states = hidden_states + shortcut_mlp_output
+
+        return hidden_states
+
+
+class LongcatFlashModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_layers = args.num_layers
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [LongcatFlashDecoderLayer(args) for idx in range(args.num_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(x)
+
+        if cache is None:
+            cache = [(None, None)] * self.num_layers
+
+        mask = create_attention_mask(h, cache[0][0], return_array=True)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = LongcatFlashModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("classifier"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
+
+    def sanitize(self, weights):
+        for l in range(self.args.num_layers):
+            prefix = f"model.layers.{l}"
+            for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+
+        for l in range(self.args.num_layers):
+            for i in range(2):
+                prefix = f"model.layers.{l}.self_attn.{i}"
+                kv_b_key = f"{prefix}.kv_b_proj.weight"
+                if kv_b_key in weights:
+                    num_heads = self.args.num_attention_heads
+                    head_dim = self.args.qk_nope_head_dim + self.args.v_head_dim
+                    quantized = f"{prefix}.kv_b_proj.scales" in weights
+                    v = weights.pop(kv_b_key)
+
+                    if quantized:
+                        dims = self.args.kv_lora_rank
+                        scales = weights.pop(f"{prefix}.kv_b_proj.scales")
+                        biases = weights.pop(f"{prefix}.kv_b_proj.biases")
+                        bits = (v.shape[-1] * 32) // dims
+                        group_size = dims // scales.shape[-1]
+                        v = mx.dequantize(
+                            v, scales, biases, bits=bits, group_size=group_size
+                        )
+
+                    v = v.reshape(num_heads, head_dim, -1)
+                    wk = mx.contiguous(
+                        v[:, : self.args.qk_nope_head_dim, :].swapaxes(-1, -2)
+                    )
+                    wv = mx.contiguous(v[:, self.args.qk_nope_head_dim :, :])
+
+                    if quantized:
+                        wk, wk_s, wk_b = mx.quantize(
+                            wk, bits=bits, group_size=group_size
+                        )
+                        wv, wv_s, wv_b = mx.quantize(
+                            wv, bits=bits, group_size=group_size
+                        )
+                        weights[f"{prefix}.embed_q.scales"] = wk_s
+                        weights[f"{prefix}.embed_q.biases"] = wk_b
+                        weights[f"{prefix}.unembed_out.scales"] = wv_s
+                        weights[f"{prefix}.unembed_out.biases"] = wv_b
+
+                    weights[f"{prefix}.embed_q.weight"] = wk
+                    weights[f"{prefix}.unembed_out.weight"] = wv
+
+        new_weights = {}
+        for k, v in weights.items():
+            if k.startswith("model.mtp"):
+                continue
+            new_weights[k] = v
+        return new_weights
+
+    def make_cache(self):
+        return [CacheList(KVCache(), KVCache()) for _ in self.model.layers]
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        rank = group.rank()
+
+        for layer in self.model.layers:
+            for attn in layer.self_attn:
+                if attn.q_lora_rank is None:
+                    attn.q_proj = shard_linear(
+                        attn.q_proj, "all-to-sharded", group=group
+                    )
+                else:
+                    attn.q_b_proj = shard_linear(
+                        attn.q_b_proj, "all-to-sharded", group=group
+                    )
+                attn.o_proj = shard_linear(attn.o_proj, "sharded-to-all", group=group)
+                attn.num_attention_heads //= N
+                num_heads = attn.num_attention_heads
+                sh = rank * num_heads
+                eh = sh + num_heads
+
+                def shard_heads(w):
+                    return w[sh:eh]
+
+                attn.embed_q.apply(shard_heads)
+                attn.unembed_out.apply(shard_heads)
+
+            for mlp in layer.mlps:
+                mlp.gate_proj = shard_linear(
+                    mlp.gate_proj, "all-to-sharded", group=group
+                )
+                mlp.up_proj = shard_linear(mlp.up_proj, "all-to-sharded", group=group)
+                mlp.down_proj = shard_linear(
+                    mlp.down_proj, "sharded-to-all", group=group
+                )
+
+            layer.mlp.sharding_group = group
+            shard_inplace(layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group)
+            shard_inplace(layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group)
+            shard_inplace(layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group)

mlx_lm/models/longcat_flash_ngram.py

@@ -0,0 +1,214 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask
+from .cache import ArraysCache, CacheList, KVCache
+from .longcat_flash import LongcatFlashDecoderLayer
+from .longcat_flash import Model as LongcatFlashLM
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    ffn_hidden_size: int
+    moe_topk: int
+    expert_ffn_hidden_size: int
+    n_routed_experts: int
+    zero_expert_num: int
+    num_layers: int
+    vocab_size: int
+    max_position_embeddings: int
+    num_attention_heads: int
+    kv_lora_rank: int
+    q_lora_rank: int
+    qk_rope_head_dim: int
+    qk_nope_head_dim: int
+    v_head_dim: int
+    routed_scaling_factor: float
+    rms_norm_eps: float
+    rope_theta: float
+    mla_scale_q_lora: bool
+    mla_scale_kv_lora: bool
+    attention_bias: bool = False
+    zero_expert_type: str = "identity"
+    ngram_vocab_size_ratio: int = 78
+    emb_neighbor_num: int = 4
+    emb_split_num: int = 4
+    norm_topk_prob: bool = False
+    router_bias: bool = False
+    rope_scaling: Optional[Dict] = None
+
+
+class NgramEmbedding(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.vocab_size = args.vocab_size
+        self.hidden_size = args.hidden_size
+        self.m = args.ngram_vocab_size_ratio * args.vocab_size
+        self.k = args.emb_split_num
+        self.n = args.emb_neighbor_num
+
+        self.word_embeddings = nn.Embedding(args.vocab_size, args.hidden_size)
+
+        num_embedders = self.k * (self.n - 1)
+        emb_dim = args.hidden_size // num_embedders
+
+        self.embedders = []
+        self.post_projs = []
+        for i in range(num_embedders):
+            emb_vocab_size = int(self.m + i * 2 + 1)
+            self.embedders.append(nn.Embedding(emb_vocab_size, emb_dim))
+            self.post_projs.append(nn.Linear(emb_dim, args.hidden_size, bias=False))
+        self._compute_vocab_mods()
+
+    def _compute_vocab_mods(self):
+        vocab_mods = {}
+        for i in range(2, self.n + 1):
+            for j in range(self.k):
+                index = (i - 2) * self.k + j
+                emb_vocab_dim = int(self.m + index * 2 + 1)
+                mods = []
+                power_mod = 1
+                for _ in range(i - 1):
+                    power_mod = (power_mod * self.vocab_size) % emb_vocab_dim
+                    mods.append(power_mod)
+                vocab_mods[(i, j)] = mods
+        self._vocab_mods = vocab_mods
+
+    def _shift_right(self, x: mx.array, n: int) -> mx.array:
+        if n <= 0:
+            return x
+        batch_size, seq_len = x.shape
+        if seq_len <= n:
+            return mx.zeros_like(x)
+        return mx.concatenate(
+            [mx.zeros((batch_size, n), dtype=x.dtype), x[..., :-n]], axis=-1
+        )
+
+    def _get_ngram_ids(
+        self,
+        input_ids: mx.array,
+        shifted_ids: Dict[int, mx.array],
+        vocab_mods: List[int],
+        ngram: int,
+    ) -> mx.array:
+        ngram_ids = input_ids
+        for k in range(2, ngram + 1):
+            ngram_ids = ngram_ids + shifted_ids[k] * vocab_mods[k - 2]
+        return ngram_ids
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        seq_len = input_ids.shape[-1]
+
+        input_ids = input_ids.astype(mx.int64)
+        if cache is not None:
+            context = cache[0]
+            if context is None:
+                context = input_ids
+            else:
+                context = mx.concatenate([context, input_ids], axis=-1)
+            cache[0] = context[..., max(0, context.shape[-1] - self.n + 1) :]
+        else:
+            context = input_ids
+
+        x = self.word_embeddings(input_ids)
+        vocab_mods = self._vocab_mods
+
+        shifted_ids = {}
+        for i in range(2, self.n + 1):
+            shifted_ids[i] = self._shift_right(context, i - 1)
+
+        for i in range(2, self.n + 1):
+            for j in range(self.k):
+                index = (i - 2) * self.k + j
+                emb_vocab_dim = int(self.m + index * 2 + 1)
+                ngram_ids = self._get_ngram_ids(
+                    context, shifted_ids, vocab_mods[(i, j)], ngram=i
+                )
+                new_ids = (ngram_ids % emb_vocab_dim)[..., -seq_len:]
+                x_ngram = self.embedders[index](new_ids)
+                x_proj = self.post_projs[index](x_ngram)
+                x = x + x_proj
+
+        return x / (1 + self.k * (self.n - 1))
+
+
+class LongcatFlashNgramModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_layers = args.num_layers
+        self.ngram_embeddings = NgramEmbedding(args)
+        self.layers = [LongcatFlashDecoderLayer(args) for _ in range(args.num_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, args.rms_norm_eps)
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        if cache is None:
+            cache = [None] + [(None, None)] * self.num_layers
+
+        h = self.ngram_embeddings(input_ids, cache=cache[0])
+
+        mask = create_attention_mask(h, cache[1][0], return_array=True)
+
+        for layer, c in zip(self.layers, cache[1:]):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = LongcatFlashNgramModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, cache)
+        return self.lm_head(out)
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def quant_predicate(self):
+        return LongcatFlashLM.quant_predicate.fget(self)
+
+    @property
+    def cast_predicate(self):
+        return LongcatFlashLM.cast_predicate.fget(self)
+
+    def sanitize(self, weights):
+        weights = LongcatFlashLM.sanitize(self, weights)
+        if "model.embed_tokens.weight" in weights:
+            weights["model.ngram_embeddings.word_embeddings.weight"] = weights.pop(
+                "model.embed_tokens.weight"
+            )
+        return weights
+
+    def make_cache(self):
+        return [ArraysCache(size=1)] + [
+            CacheList(KVCache(), KVCache()) for _ in self.model.layers
+        ]
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        LongcatFlashLM.shard(self, group)

mlx_lm/models/mamba.py

@@ -6,8 +6,9 @@ from dataclasses import dataclass
 import mlx.core as mx
 import mlx.nn as nn
 
+from .activations import swiglu
 from .base import BaseModelArgs
-from .cache import MambaCache
+from .cache import ArraysCache
 
 
 @dataclass
@@ -50,32 +51,6 @@ class ModelArgs(BaseModelArgs):
             self.use_bcdt_rms = True
 
 
-class DepthWiseConv1d(nn.Module):
-    def __init__(self, channels, kernel_size, bias=True, padding=0):
-        super().__init__()
-        self.channels = channels
-        self.kernel_size = kernel_size
-        self.padding = padding
-        self.weight = mx.random.normal((self.channels, kernel_size, 1))
-        self.bias = mx.zeros((channels,)) if bias else None
-
-    def __call__(self, x, cache=None):
-        B, L, C = x.shape
-        groups, K, _ = self.weight.shape
-
-        if cache is not None:
-            x = mx.concatenate([cache, x], axis=1)
-        else:
-            x = mx.pad(x, [(0, 0), (K - 1, 0), (0, 0)])
-
-        y = mx.conv_general(x, self.weight, groups=groups)
-
-        if self.bias is not None:
-            y = y + self.bias
-
-        return y, x[:, -K + 1 :, :]
-
-
 class MambaBlock(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
@@ -97,11 +72,13 @@ class MambaBlock(nn.Module):
             self.hidden_size, self.intermediate_size * 2, bias=args.use_bias
         )
 
-        self.conv1d = DepthWiseConv1d(
-            channels=self.intermediate_size,
+        self.conv1d = nn.Conv1d(
+            in_channels=self.intermediate_size,
+            out_channels=self.intermediate_size,
             kernel_size=self.conv_kernel_size,
+            groups=self.intermediate_size,
             bias=self.use_conv_bias,
-            padding=self.conv_kernel_size - 1,
+            padding=0,
         )
 
         self.x_proj = nn.Linear(
@@ -148,20 +125,22 @@ class MambaBlock(nn.Module):
         B, T, D = x.shape
         xz = self.in_proj(x)
         x, z = xz.split(indices_or_sections=2, axis=-1)
-
-        conv_out, new_conv_cache = self.conv1d(x, conv_cache)
+        K = self.conv_kernel_size
+        if conv_cache is not None:
+            x_full = mx.concatenate([conv_cache, x], axis=1)
+        else:
+            x_full = mx.pad(x, [(0, 0), (K - 1, 0), (0, 0)])
+        conv_out = self.conv1d(x_full)
+        new_conv_cache = x_full[:, -(K - 1) :, :]
         x = nn.silu(conv_out)
-
         A = -mx.exp(self.A_log)
-
-        outputs = []
         current_state = state_cache
         y = []
         for t in range(T):
             y_t, current_state = self.ssm_step(x[:, t], A, current_state)
             y.append(y_t)
         y = mx.stack(y, axis=1)
-        z = self.out_proj(nn.silu(z) * y)
+        z = self.out_proj(swiglu(z, y))
         return z, (new_conv_cache, current_state)
 
     def __call__(self, x, cache):
@@ -174,7 +153,7 @@ class MambaBlock(nn.Module):
             x, conv_cache, state_cache
         )
 
-        if isinstance(cache, MambaCache):
+        if isinstance(cache, ArraysCache):
             cache[0] = new_conv_cache
             cache[1] = new_state_cache
 
@@ -228,15 +207,15 @@ class Model(nn.Module):
 
         return logits
 
+    def make_cache(self):
+        return [ArraysCache(size=2) for _ in range(len(self.layers))]
+
+    @property
+    def layers(self):
+        return self.backbone.layers
+
     def sanitize(self, weights):
         for k, v in weights.items():
             if "conv1d.weight" in k and v.shape[-1] != 1:
                 weights[k] = v.moveaxis(2, 1)
         return weights
-
-    def make_cache(self):
-        return [MambaCache() for _ in range(len(self.layers))]
-
-    @property
-    def layers(self):
-        return self.backbone.layers