sharded gemma 4

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

.github/workflows/pull_request.yml

@@ -40,5 +40,5 @@ jobs:
         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
-          HF_HOME="." python -m xmlrunner discover -v tests -o test-results/
+          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

ACKNOWLEDGMENTS.md

@@ -10,7 +10,7 @@ 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` and 
-`Mamba v2`, Z.ai & THUKEG's `GLM`, `GLM4`, Rednote `dots.llm1`, Baidu's `Ernie4.5 MoE`, 
+`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`, 
@@ -26,4 +26,8 @@ Added support for the following other features:
   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`.
+ 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)`

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:

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/_version.py

@@ -1,3 +1,3 @@
 # Copyright © 2023-2025 Apple Inc.
 
-__version__ = "0.30.6"
+__version__ = "0.31.3"

mlx_lm/benchmark.py

@@ -1,6 +1,7 @@
 # Copyright © 2025 Apple Inc.
 
 import argparse
+import time
 
 import mlx.core as mx
 
@@ -60,6 +61,18 @@ def setup_arg_parser():
         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
 
 
@@ -103,14 +116,22 @@ def main():
 
     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
 
     def batch_bench():
         return batch_generate(
-            model, tokenizer, prompts, max_tokens=generation_tokens
+            model,
+            tokenizer,
+            prompts,
+            max_tokens=generation_tokens,
+            prefill_step_size=args.prefill_step_size,
         ).stats
 
     if batch_size == 1:
@@ -125,10 +146,15 @@ def main():
     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]
+        results.append(f"total_time={toc - tic:.3f}")
         rprint(f"Trial {i+1}:  " + ", ".join(results))
 
     def avg(k):

mlx_lm/cli.py

@@ -22,6 +22,7 @@ def main():
         "gptq",
         "server",
         "upload",
+        "share",
     )
     subpackages = {
         "awq": "quant",

mlx_lm/convert.py

@@ -72,7 +72,7 @@ def mixed_quant_predicate_builder(
         if "lm_head" in path:
             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
 
@@ -86,8 +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,
@@ -128,6 +128,8 @@ def convert(
 
     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)

mlx_lm/examples/batch_generate_response.py

@@ -27,7 +27,7 @@ prompts = [
 
 # Set `verbose=True` to see generation statistics
 result = batch_generate(
-    model, tokenizer, prompts, verbose=False, return_prompt_caches=True
+    model, tokenizer, prompts, verbose=False, return_prompt_caches=True, max_tokens=2048
 )
 print(result.texts[-1])
 

mlx_lm/lora.py

@@ -21,7 +21,7 @@ from .tuner.utils import (
     load_adapters,
     print_trainable_parameters,
 )
-from .utils import load, save_config
+from .utils import _parse_size, load, save_config
 
 yaml_loader = yaml.SafeLoader
 yaml_loader.add_implicit_resolver(
@@ -69,6 +69,7 @@ CONFIG_DEFAULTS = {
     "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,
@@ -190,6 +191,12 @@ def build_parser():
         help="Use gradient checkpointing to reduce memory use.",
         default=None,
     )
+    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",
         type=str,

mlx_lm/models/apertus.py

@@ -167,7 +167,8 @@ class Model(nn.Module):
         self.args = args
         self.model_type = args.model_type
         self.model = ApertusModel(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,
@@ -175,12 +176,18 @@ class Model(nn.Module):
         cache: Optional[Any] = None,
     ) -> mx.array:
         out = self.model(inputs, cache)
-        return self.lm_head(out)
+        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

mlx_lm/models/cache.py

@@ -1,11 +1,13 @@
 # Copyright © 2023-2024 Apple Inc.
 
 import copy
+from collections import deque
+from dataclasses import dataclass
 from typing import Any, Dict, List, Optional
 
 import mlx.core as mx
 import mlx.nn as nn
-from mlx.utils import tree_flatten, tree_map, tree_unflatten
+from mlx.utils import tree_flatten, tree_map, tree_reduce, tree_unflatten
 
 from .base import create_causal_mask
 
@@ -153,6 +155,11 @@ class _BaseCache:
         """
         return 0
 
+    @property
+    def nbytes(self):
+        """Return the size of this cache in bytes"""
+        raise NotImplementedError("Cache sub-class must implement nbytes")
+
     def empty(self):
         """
         Return if the cache is empty or not.
@@ -215,6 +222,12 @@ class ConcatenateKVCache(_BaseCache):
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
 
 class QuantizedKVCache(_BaseCache):
     step = 256
@@ -304,6 +317,10 @@ class QuantizedKVCache(_BaseCache):
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        return tree_reduce(lambda a, x: a + x.nbytes, (self.keys, self.values), 0)
+
 
 class KVCache(_BaseCache):
     step = 256
@@ -383,6 +400,12 @@ class KVCache(_BaseCache):
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
 
 class RotatingKVCache(_BaseCache):
     step = 256
@@ -561,6 +584,12 @@ class RotatingKVCache(_BaseCache):
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
 
 class ArraysCache(_BaseCache):
     def __new__(cls, *args, **kwargs):
@@ -574,6 +603,18 @@ class ArraysCache(_BaseCache):
         if left_padding:
             self.left_padding = mx.array(left_padding)
 
+    @property
+    def batch_size(self):
+        for c in self.cache:
+            if c is not None:
+                return c.shape[0]
+        if self.left_padding is not None:
+            return self.left_padding.size
+        elif self.lengths is not None:
+            return self.lengths.size
+        else:
+            return 1
+
     def __setitem__(self, idx, value):
         self.cache[idx] = value
 
@@ -592,13 +633,42 @@ class ArraysCache(_BaseCache):
         """
         In-place filter to keep just the given indices in the cache.
         """
-        self.cache = [c[batch_indices] for c in self.cache]
+        self.cache = [c[batch_indices] if c is not None else None for c in self.cache]
+        if self.left_padding is not None:
+            self.left_padding = self.left_padding[batch_indices]
+        if self.lengths is not None:
+            self.lengths = self.lengths[batch_indices]
 
     def extend(self, other):
         """
         In-place extend this cache with the other cache.
         """
-        self.cache = [mx.concatenate([c, o]) for c, o in zip(self.cache, other.cache)]
+
+        a_batch = self.batch_size
+        b_batch = other.batch_size
+
+        def cat(a, b):
+            shape = dtype = None
+            if a is not None:
+                shape = a.shape
+                dtype = a.dtype
+            if b is not None:
+                shape = b.shape
+                dtype = b.dtype
+
+            if shape is None:
+                return None
+
+            if a is None:
+                a = mx.zeros((a_batch,) + shape[1:], dtype=dtype)
+            if b is None:
+                b = mx.zeros((b_batch,) + shape[1:], dtype=dtype)
+
+            return mx.concatenate([a, b])
+
+        self.cache = [cat(c, o) for c, o in zip(self.cache, other.cache)]
+        self.left_padding = cat(self.left_padding, other.left_padding)
+        self.lengths = cat(self.lengths, other.lengths)
 
     def extract(self, idx):
         cache = ArraysCache(len(self.cache))
@@ -633,6 +703,12 @@ class ArraysCache(_BaseCache):
         n_state = len(caches[0].cache)
         B = len(caches)
         cache = cls(n_state)
+
+        # All caches are empty so return early
+        if all(c.empty() for c in caches):
+            cache.left_padding = mx.array([0] * B)
+            return cache
+
         for e in range(n_state):
             c_init = next(iter(c[e] for c in caches if c[e] is not None))
             shape = list(c_init.shape)
@@ -647,6 +723,10 @@ class ArraysCache(_BaseCache):
     def empty(self):
         return self.cache[0] is None
 
+    @property
+    def nbytes(self):
+        return sum(c.nbytes for c in self.cache if c is not None)
+
 
 class ChunkedKVCache(_BaseCache):
     step = 256
@@ -724,6 +804,12 @@ class ChunkedKVCache(_BaseCache):
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
 
 class CacheList(_BaseCache):
     def __init__(self, *caches):
@@ -742,16 +828,24 @@ class CacheList(_BaseCache):
 
     @property
     def state(self):
-        return [s for c in self.caches for s in c.state]
+        return [c.state for c in self.caches]
 
     @state.setter
     def state(self, v):
-        state_lens = [len(c.state) for c in self.caches]
-        start = 0
-        for c in self.caches:
-            l = len(c.state)
-            c.state = v[start : start + l]
-            start += l
+        for c, s in zip(self.caches, v):
+            c.state = s
+
+    @property
+    def meta_state(self):
+        return (
+            [type(c).__name__ for c in self.caches],
+            [c.meta_state for c in self.caches],
+        )
+
+    @meta_state.setter
+    def meta_state(self, v):
+        for c, m in zip(self.caches, v[1]):
+            c.meta_state = m
 
     def filter(self, batch_indices):
         """
@@ -793,6 +887,18 @@ class CacheList(_BaseCache):
     def empty(self):
         return self.caches[0].empty()
 
+    @property
+    def nbytes(self):
+        return sum(c.nbytes for c in self.caches)
+
+    @classmethod
+    def from_state(cls, state, meta_state):
+        obj = cls.__new__(cls)
+        obj.caches = [
+            globals()[c].from_state(s, m) for s, c, m in zip(state, *meta_state)
+        ]
+        return obj
+
 
 def dynamic_roll(x, shifts, axis):
     n = x.shape[axis]
@@ -911,16 +1017,18 @@ class BatchKVCache(_BaseCache):
         """
         In-place filter to keep just the given indices in the cache.
         """
-        self.keys = self.keys[batch_indices]
-        self.values = self.values[batch_indices]
+        if self.keys is not None:
+            self.keys = self.keys[batch_indices]
+            self.values = self.values[batch_indices]
         self.offset = self.offset[batch_indices]
         self.left_padding = self.left_padding[batch_indices]
 
         # Shift left to reduce padding
         min_left_pad = self.left_padding.min().item()
         if min_left_pad > 0:
-            self.keys = self.keys[..., min_left_pad:, :]
-            self.values = self.values[..., min_left_pad:, :]
+            if self.keys is not None:
+                self.keys = self.keys[..., min_left_pad:, :]
+                self.values = self.values[..., min_left_pad:, :]
             self._idx -= min_left_pad
             self.left_padding -= min_left_pad
 
@@ -928,15 +1036,31 @@ class BatchKVCache(_BaseCache):
         """
         In-place extend this cache with the other cache.
         """
+        if self.keys is None and other.keys is None:
+            self.left_padding = mx.concatenate([self.left_padding, other.left_padding])
+            self.offset = mx.concatenate([self.offset, other.offset])
+            return
+
         max_idx = max(self._idx, other._idx)
-        max_size = max(self.keys.shape[2], other.keys.shape[2])
+        L1 = L2 = 0
+        if self.keys is not None:
+            B, H, L1, D = self.keys.shape
+            M = self.values.shape[3]
+        if other.keys is not None:
+            B, H, L2, D = other.keys.shape
+            M = other.values.shape[3]
+        max_size = max(L1, L2)
 
         # Pad the keys and values so they are right-justified
         # with the index and the same size
         def pad(c):
-            left = max_idx - c._idx
-            right = max_size - c.keys.shape[2] - left
             k, v = c.keys, c.values
+            if k is None:
+                Bc = c.offset.shape[0]
+                k = mx.array([]).reshape(Bc, H, 0, D)
+                v = mx.array([]).reshape(Bc, H, 0, M)
+            left = max_idx - c._idx
+            right = max_size - k.shape[2] - left
             if right < 0:
                 k = k[..., :right, :]
                 v = v[..., :right, :]
@@ -965,6 +1089,11 @@ class BatchKVCache(_BaseCache):
     def merge(cls, caches):
         lengths = [c.size() for c in caches]
         max_length = max(lengths)
+
+        # No cache has content so make an empty one
+        if max_length == 0:
+            return BatchKVCache([0] * len(caches))
+
         padding = [max_length - l for l in lengths]
         B = len(caches)
         H = max(c.keys.shape[1] for c in caches if c.keys is not None)
@@ -988,9 +1117,18 @@ class BatchKVCache(_BaseCache):
 
         return cache
 
+    def size(self):
+        return self._idx
+
     def empty(self):
         return self.keys is None
 
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
 
 class BatchRotatingKVCache(_BaseCache):
     step = 256
@@ -1061,6 +1199,10 @@ class BatchRotatingKVCache(_BaseCache):
         self.offset += keys.shape[2]
         self._offset += keys.shape[2]
         self._idx = self.keys.shape[2]
+
+        # Make sure left_padding and offset are evaluated
+        self.keys = mx.depends(self.keys, (self.left_padding, self.offset))
+
         return self.keys, self.values
 
     def _update_in_place(self, keys, values):
@@ -1111,6 +1253,9 @@ class BatchRotatingKVCache(_BaseCache):
         self.offset += S
         self._idx += S
 
+        # Make sure left_padding and offset are evaluated
+        self.keys = mx.depends(self.keys, (self.left_padding, self.offset))
+
         # If the buffer is not full, slice off the end
         if self._offset < self.max_size:
             return (
@@ -1215,8 +1360,9 @@ class BatchRotatingKVCache(_BaseCache):
         """
         In-place filter to keep just the given indices in the cache.
         """
-        self.keys = self.keys[batch_indices]
-        self.values = self.values[batch_indices]
+        if self.keys is not None:
+            self.keys = self.keys[batch_indices]
+            self.values = self.values[batch_indices]
         self.offset = self.offset[batch_indices]
         self.left_padding = self.left_padding[batch_indices]
 
@@ -1224,17 +1370,33 @@ class BatchRotatingKVCache(_BaseCache):
         """
         In-place extend this cache with the other cache.
         """
+        if self.keys is None and other.keys is None:
+            self.left_padding = mx.concatenate([self.left_padding, other.left_padding])
+            self.offset = mx.concatenate([self.offset, other.offset])
+            return
+
         if (self.rotated != other.rotated) or self._idx != other._idx:
             self._temporal_order()
             other._temporal_order()
 
         max_idx = max(self._idx, other._idx)
-        max_size = max(self.keys.shape[2], other.keys.shape[2])
+        L1 = L2 = 0
+        if self.keys is not None:
+            B, H, L1, D = self.keys.shape
+            M = self.values.shape[3]
+        if other.keys is not None:
+            B, H, L2, D = other.keys.shape
+            M = other.values.shape[3]
+        max_size = max(L1, L2)
 
         def pad(c):
             left = max_idx - c._idx
-            right = max_size - c.keys.shape[2] - left
             k, v = c.keys, c.values
+            if k is None:
+                Bc = c.offset.shape[0]
+                k = mx.array([]).reshape(Bc, H, 0, D)
+                v = mx.array([]).reshape(Bc, H, 0, M)
+            right = max_size - k.shape[2] - left
             if right < 0:
                 k = k[..., :right, :]
                 v = v[..., :right, :]
@@ -1253,9 +1415,10 @@ class BatchRotatingKVCache(_BaseCache):
         self._offset = max(self._offset, other._offset)
 
     def extract(self, idx):
+        mx.eval(self.left_padding, self.offset)
         cache = RotatingKVCache(self.max_size)
-        padding = self.left_padding[idx].item()
-        offset = self.offset[idx].item()
+        padding = max(0, self.left_padding.tolist()[idx])
+        offset = self.offset.tolist()[idx]
         cache.keys = self.keys[idx : idx + 1]
         cache.values = self.values[idx : idx + 1]
         cache._idx = self._idx
@@ -1279,6 +1442,11 @@ class BatchRotatingKVCache(_BaseCache):
         offsets = [c.offset for c in caches]
         lengths = [c.size() for c in caches]
         max_length = max(lengths)
+
+        # No cache has content so make an empty one
+        if max_length == 0:
+            return cls(caches[0].max_size, [0] * len(caches))
+
         padding = [max_length - l for l in lengths]
         B = len(caches)
         H = max(c.keys.shape[1] for c in caches if c.keys is not None)
@@ -1288,11 +1456,11 @@ class BatchRotatingKVCache(_BaseCache):
 
         keys = mx.zeros((B, H, max_length, Dk), dtype=dt)
         values = mx.zeros((B, H, max_length, Dv), dtype=dt)
-        for i, (p, c) in enumerate(zip(padding, caches)):
+        for i, (p, l, c) in enumerate(zip(padding, lengths, caches)):
             if c.keys is None:
                 continue
-            keys[i : i + 1, :, p : p + c._idx] = c._temporal_order(c.keys)
-            values[i : i + 1, :, p : p + c._idx] = c._temporal_order(c.values)
+            keys[i : i + 1, :, p : p + l] = c._temporal_order(c.keys)[..., -l:, :]
+            values[i : i + 1, :, p : p + l] = c._temporal_order(c.values)[..., -l:, :]
 
         cache = cls(caches[0].max_size, padding)
         cache.keys = keys
@@ -1303,5 +1471,293 @@ class BatchRotatingKVCache(_BaseCache):
 
         return cache
 
+    def size(self):
+        return min(self._offset, self.max_size)
+
     def empty(self):
         return self.keys is None
+
+    @property
+    def nbytes(self):
+        if self.keys is None:
+            return 0
+        return self.keys.nbytes + self.values.nbytes
+
+
+class TokenBuffer:
+    """A simple token buffer that can be efficiently appended to in a similar
+    fashion to the KVCache.
+
+    Perhaps these could share some logic in the future.
+    """
+
+    step = 256
+
+    def __init__(self, tokens=[]):
+        self._buffer = mx.array(tokens, dtype=mx.int32)
+        self._size = len(tokens)
+
+    def update_and_fetch(self, tokens):
+        start = self._size
+        end = start + len(tokens)
+
+        new_size = ((end + self.step - 1) // self.step) * self.step
+        if new_size > self._buffer.size:
+            self._buffer = mx.concatenate(
+                [self._buffer, mx.zeros(new_size - self._buffer.size, dtype=mx.int32)]
+            )
+        self._buffer[start:end] = tokens
+        self._size = end
+
+        return self._buffer[:end]
+
+    @property
+    def state(self):
+        return self._buffer
+
+    @property
+    def tokens(self):
+        return self._buffer[: self._size]
+
+
+@dataclass
+class PromptTrieResult:
+    model: Any
+    exact: Optional[List[int]]  # Exact match found
+    shorter: Optional[List[int]]  # Longest prefix with a value
+    longer: Optional[List[int]]  # Shortest value that extends beyond tokens
+    common_prefix: int  # Length of common prefix with any path
+
+
+class PromptTrie:
+    def __init__(self):
+        self._trie = {}
+
+    def add(self, model: Any, tokens: List[int], value: Any):
+        if model not in self._trie:
+            self._trie[model] = {}
+
+        current = self._trie[model]
+        for tok in tokens:
+            if tok not in current:
+                current[tok] = {}
+            current = current[tok]
+        prev = current.get("__value__", None)
+        current["__value__"] = value
+        return prev
+
+    def get(self, model: Any, tokens: List[int]):
+        current = self._trie[model]
+        for tok in tokens:
+            current = current[tok]
+        return current["__value__"]
+
+    def pop(self, model: Any, tokens: List[int]):
+        path = [self._trie[model]]
+        for tok in tokens:
+            path.append(path[-1][tok])
+        value = path[-1].pop("__value__")
+        for i in range(len(tokens), 0, -1):
+            node = path[i]
+            parent = path[i - 1]
+            tok = tokens[i - 1]
+            if len(node) > 0:
+                break
+            del parent[tok]
+        return value
+
+    def pop_prefixes(self, model: Any, tokens: List[int]):
+        values = []
+        current = self._trie[model]
+        for i, tok in enumerate(tokens):
+            if "__value__" in current:
+                values.append((i, current.pop("__value__")))
+            current = current[tok]
+        return values
+
+    def search(self, model: Any, tokens: List[int]) -> PromptTrieResult:
+        if model not in self._trie:
+            return PromptTrieResult(model, None, None, None, 0)
+
+        current = self._trie[model]
+
+        if not tokens and "__value__" in current:
+            return PromptTrieResult(model, [], None, None, 0)
+
+        # Walk the tokens as far as we can
+        last_index = -1
+        index = 0
+        while index < len(tokens) and tokens[index] in current:
+            current = current[tokens[index]]
+            if "__value__" in current:
+                last_index = index
+            index += 1
+
+        # Got an exact match
+        if last_index == len(tokens) - 1 >= 0:
+            return PromptTrieResult(model, tokens, None, None, 0)
+
+        # Check if we found a prefix at any point
+        shorter = None
+        if last_index > 0:
+            shorter = tokens[: last_index + 1]
+
+        # Check for sequences that are longer
+        longer = None
+        common_prefix = index
+        if index > 0:
+            best = None
+            stack = [(current, [])]
+            while stack:
+                current, extra = stack.pop()
+                if "__value__" in current:
+                    if best is None or len(extra) < len(best):
+                        best = extra
+                elif best is None or len(extra) < len(best):
+                    for tok in current:
+                        stack.append((current[tok], extra + [tok]))
+            longer = tokens[:index] + best
+        return PromptTrieResult(model, None, shorter, longer, common_prefix)
+
+
+class LRUPromptCache:
+    @dataclass
+    class CacheEntry:
+        prompt_cache: List[Any]
+        nbytes: int
+        cache_type: str
+
+    class CacheOrder:
+        def __init__(self, ordering: List[str] = ["assistant", "user", "system"]):
+            self._ordering = ordering
+            self._lrus = {k: deque() for k in ordering}
+
+        def __len__(self):
+            return sum(len(lru) for lru in self._lrus.values())
+
+        def push(self, model: Any, tokens: List[Any], cache_type: str = "assistant"):
+            self._lrus[cache_type].append((model, tokens))
+
+        def remove(self, model: Any, tokens: List[Any]):
+            for cache_type in self._ordering:
+                try:
+                    self._lrus[cache_type].remove((model, tokens))
+                    break
+                except ValueError:
+                    pass
+
+        def pop(self):
+            i = 0
+            while i + 1 < len(self._ordering):
+                lru_a = self._lrus[self._ordering[i]]
+                lru_b = self._lrus[self._ordering[i + 1]]
+                if lru_a and len(lru_a) >= len(lru_b):
+                    return lru_a.popleft()
+                i += 1
+            return lru_b.popleft()
+
+    def __init__(self, max_size: int = 10, max_bytes: int = 1 << 63):
+        self.max_size = max_size
+        self.max_bytes = max_bytes
+        self._trie = PromptTrie()
+        self._lru = LRUPromptCache.CacheOrder()
+        self._n_bytes = 0
+        self._n_bytes_by_type = {k: 0 for k in self._lru._ordering}
+
+    def __len__(self):
+        return len(self._lru)
+
+    @property
+    def nbytes(self):
+        return self._n_bytes
+
+    def fetch_nearest_cache(self, model: Any, tokens: List[int]):
+        result = self._trie.search(model, tokens)
+        if result.exact is not None:
+            cache_entry = self._trie.get(result.model, result.exact)
+            return copy.deepcopy(cache_entry.prompt_cache), []
+
+        short_length = len(result.shorter) if result.shorter is not None else 0
+        if result.longer is not None and result.common_prefix > short_length:
+            cache_entry = self._trie.get(result.model, result.longer)
+            if can_trim_prompt_cache(cache_entry.prompt_cache):
+                cache = copy.deepcopy(cache_entry.prompt_cache)
+                prefix = min(len(tokens) - 1, result.common_prefix)
+                num_to_trim = len(result.longer) - prefix
+                trim_prompt_cache(cache, num_to_trim)
+                return cache, tokens[prefix:]
+
+        if short_length > 0:
+            cache_entry = self._trie.get(result.model, result.shorter)
+            return copy.deepcopy(cache_entry.prompt_cache), tokens[short_length:]
+
+        return None, tokens
+
+    def insert_cache(
+        self,
+        model: Any,
+        tokens: List[int],
+        prompt_cache: List[Any],
+        *,
+        cache_type: str = "assistant",
+    ):
+        # Make the cache entry
+        entry = LRUPromptCache.CacheEntry(
+            prompt_cache, sum(c.nbytes for c in prompt_cache), cache_type
+        )
+
+        # Insert into the trie and update the byte counter and lru position
+        self._n_bytes += entry.nbytes
+        self._n_bytes_by_type[cache_type] += entry.nbytes
+        prev = self._trie.add(model, tokens, entry)
+        if prev is not None:
+            self._n_bytes -= prev.nbytes
+            self._n_bytes_by_type[prev.cache_type] -= prev.nbytes
+            self._lru.remove(model, tokens)
+        self._lru.push(model, tokens, cache_type)
+
+        # If it is a trimmable cache remove all prefixes cause they just take
+        # space
+        if can_trim_prompt_cache(prompt_cache):
+            for prefix_len, entry in self._trie.pop_prefixes(model, tokens):
+                self._n_bytes -= entry.nbytes
+                self._n_bytes_by_type[entry.cache_type] -= entry.nbytes
+                self._lru.remove(model, tokens[:prefix_len])
+
+        # Ensure we match the constraints
+        if len(self._lru) > self.max_size:
+            model, tokens = self._lru.pop()
+            entry = self._trie.pop(model, tokens)
+            self._n_bytes -= entry.nbytes
+            self._n_bytes_by_type[entry.cache_type] -= entry.nbytes
+        while self._n_bytes > self.max_bytes:
+            model, tokens = self._lru.pop()
+            entry = self._trie.pop(model, tokens)
+            self._n_bytes -= entry.nbytes
+            self._n_bytes_by_type[entry.cache_type] -= entry.nbytes
+
+    def trim_to(
+        self, *, n_sequences: Optional[int] = None, n_bytes: Optional[int] = None
+    ):
+        n_sequences = max(0, n_sequences) if n_sequences is not None else 1 << 63
+        n_bytes = max(0, n_bytes) if n_bytes is not None else 1 << 63
+
+        while len(self._lru) > n_sequences:
+            model, tokens = self._lru.pop()
+            entry = self._trie.pop(model, tokens)
+            self._n_bytes -= entry.nbytes
+            self._n_bytes_by_type[entry.cache_type] -= entry.nbytes
+        while self._n_bytes > n_bytes:
+            model, tokens = self._lru.pop()
+            entry = self._trie.pop(model, tokens)
+            self._n_bytes -= entry.nbytes
+            self._n_bytes_by_type[entry.cache_type] -= entry.nbytes
+
+    def stats_by_type(self):
+        result = {}
+        for cache_type in self._lru._ordering:
+            result[cache_type] = {
+                "n_sequences": len(self._lru._lrus[cache_type]),
+                "n_bytes": self._n_bytes_by_type[cache_type],
+            }
+        return result

mlx_lm/models/deepseek_v32.py

@@ -71,7 +71,7 @@ class Indexer(nn.Module):
         self.rope = initialize_rope(
             dims=args.qk_rope_head_dim,
             base=args.rope_theta,
-            traditional=False,
+            traditional=True,
             max_position_embeddings=args.max_position_embeddings,
             scaling_config=args.rope_scaling,
         )
@@ -87,19 +87,15 @@ class Indexer(nn.Module):
         b, s, _ = x.shape
         q = self.wq_b(qr)
         q = q.reshape(b, s, self.n_heads, self.head_dim).swapaxes(1, 2)
-        q_pe, q_nope = mx.split(q, [self.rope_head_dim], axis=-1)
-
-        offset = cache.offset if cache is not None else 0
-
-        q_pe = self.rope(q_pe, offset=offset)
-        q = mx.concatenate([q_pe, q_nope], axis=-1)
-
         k = self.wk(x)
         k = self.k_norm(k)
         k = mx.reshape(k, (b, 1, s, self.head_dim))
-        k_pe, k_nope = mx.split(k, [self.rope_head_dim], axis=-1)
-        k_pe = self.rope(k_pe, offset=offset)
-        k = mx.concatenate([k_pe, k_nope], axis=-1)
+
+        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:
@@ -209,15 +205,30 @@ class DeepseekV32Attention(nn.Module):
 
         topk_indices = self.indexer(x, qr, mask, cache=cache[1])
         if topk_indices is not None:
-            shape = list(topk_indices.shape)
-            shape[-1] = keys.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
+            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:
@@ -481,6 +492,16 @@ class Model(nn.Module):
         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)
@@ -521,6 +542,7 @@ 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
@@ -557,12 +579,7 @@ class Model(nn.Module):
                 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 {
-            k: v
-            for k, v in weights.items()
-            if not k.startswith("model.layers.61") and "rotary_emb.inv_freq" not in k
-        }
+        return weights
 
     def shard(self, group: Optional[mx.distributed.Group] = None):
         group = group or mx.distributed.init()

mlx_lm/models/gated_delta.py

@@ -7,9 +7,7 @@ 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)).astype(
-        A_log.dtype
-    )
+    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):
@@ -83,6 +81,8 @@ def _make_gated_delta_kernel(has_mask=False, vectorized=False):
             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;
@@ -94,7 +94,7 @@ def _make_gated_delta_kernel(has_mask=False, vectorized=False):
         }}
         for (int i = 0; i < n_per_t; ++i) {{
           auto s_idx = n_per_t * dk_idx + i;
-          o_state[s_idx] = static_cast<InT>(state[i]);
+          o_state[s_idx] = static_cast<StT>(state[i]);
         }}
     """
     inputs = ["q", "k", "v", "g", "beta", "state_in", "T"]
@@ -165,7 +165,7 @@ def _gated_delta_step_ops(
     if mask is not None:
         mask = mx.expand_dims(mask, axis=(1, 2, 3))
         state = mx.where(mask, state, old_state)
-    return y, state
+    return y.astype(q.dtype), state
 
 
 def gated_delta_kernel(
@@ -180,6 +180,7 @@ def gated_delta_kernel(
     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]
@@ -197,6 +198,7 @@ def gated_delta_kernel(
         inputs=inputs,
         template=[
             ("InT", input_type),
+            ("StT", state_type),
             ("Dk", Dk),
             ("Dv", Dv),
             ("Hk", Hk),
@@ -205,7 +207,7 @@ def gated_delta_kernel(
         grid=(32, Dv, B * Hv),
         threadgroup=(32, 4, 1),
         output_shapes=[(B, T, Hv, Dv), state.shape],
-        output_dtypes=[input_type, input_type],
+        output_dtypes=[input_type, state_type],
     )
 
 
@@ -235,7 +237,7 @@ def gated_delta_ops(
     B, T, Hk, Dk = q.shape
     Hv, Dv = v.shape[-2:]
     if state is None:
-        state = mx.zeros((B, Hv, Dv, Dk), dtype=q.dtype)
+        state = mx.zeros((B, Hv, Dv, Dk), dtype=mx.float32)
 
     if (repeat_factor := Hv // Hk) > 1:
         q = mx.repeat(q, repeat_factor, -2)
@@ -269,13 +271,12 @@ def gated_delta_update(
     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=q.dtype)
+        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)

mlx_lm/models/gemma4.py

@@ -0,0 +1,94 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+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()
+
+    def shard(self, group: Optional[mx.distributed.Group] = None):
+        self.language_model.shard(group)

mlx_lm/models/gemma4_text.py

@@ -0,0 +1,728 @@
+# 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 mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+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,
+        )
+
+        self.sharding_group = None
+
+    def __call__(
+        self, x: mx.array, top_k_indices: mx.array, top_k_weights: mx.array
+    ) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
+        w = mx.expand_dims(top_k_weights, -1)
+        y = self.switch_glu(x, top_k_indices)
+
+        y = (w * y).sum(-2)
+
+        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, 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 = {}
+        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
+
+            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
+
+    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
+            )
+            if hasattr(layer.self_attn, "v_proj"):
+                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
+            )
+
+            if layer.enable_moe:
+                layer.experts.sharding_group = group
+                shard_inplace(
+                    layer.experts.switch_glu.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.experts.switch_glu.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.experts.switch_glu.up_proj, "all-to-sharded", group=group
+                )

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/kimi_linear.py

@@ -15,7 +15,7 @@ from .base import (
 )
 from .cache import ArraysCache, KVCache
 from .gated_delta import gated_delta_update
-from .rope_utils import initialize_rope
+from .mla import MultiLinear
 from .switch_layers import SwitchGLU
 
 
@@ -165,6 +165,7 @@ class KimiMLAAttention(nn.Module):
         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
@@ -175,23 +176,14 @@ class KimiMLAAttention(nn.Module):
             bias=False,
         )
         self.kv_a_layernorm = nn.RMSNorm(args.kv_lora_rank, eps=args.rms_norm_eps)
-        self.kv_b_proj = nn.Linear(
-            args.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, 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)
 
-        rope_dim = self.qk_rope_head_dim or self.q_head_dim
-        self.rope = initialize_rope(
-            rope_dim,
-            base=args.rope_theta,
-            traditional=False,
-            scaling_config=args.rope_scaling,
-            max_position_embeddings=args.model_max_length,
-        )
-
     def __call__(
         self,
         x: mx.array,
@@ -199,51 +191,45 @@ class KimiMLAAttention(nn.Module):
         cache: Optional[KVCache] = None,
     ) -> mx.array:
         B, L, _ = x.shape
-        q_states = self.q_proj(x).reshape(B, L, self.num_heads, self.q_head_dim)
-        q_pass, q_rot = mx.split(q_states, [self.qk_nope_head_dim], axis=-1)
 
-        compressed = self.kv_a_proj_with_mqa(x)
-        k_pass, k_rot = mx.split(
-            compressed, [compressed.shape[-1] - self.qk_rope_head_dim], axis=-1
-        )
-        k_pass = self.kv_a_layernorm(k_pass)
-        kv = self.kv_b_proj(k_pass)
-        kv = kv.reshape(
-            B,
-            L,
-            self.num_heads,
-            self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim,
-        )
-        k_pass, v_states = mx.split(kv, [self.qk_nope_head_dim], axis=-1)
+        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)
 
-        if self.qk_rope_head_dim:
-            k_rot = mx.reshape(k_rot, (B, L, 1, self.qk_rope_head_dim))
-            k_rot = mx.broadcast_to(k_rot, (*k_pass.shape[:-1], self.qk_rope_head_dim))
-        else:
-            k_rot = mx.zeros((*k_pass.shape[:-1], 0), dtype=k_pass.dtype)
+        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)
 
-        queries = mx.concatenate([q_pass, q_rot], axis=-1).transpose(0, 2, 1, 3)
-        keys = mx.concatenate([k_pass, k_rot], axis=-1).transpose(0, 2, 1, 3)
-        values = v_states.transpose(0, 2, 1, 3)
+        kv_latent = mx.expand_dims(kv_latent, axis=1)
 
         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)
+            kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe)
 
-        out = scaled_dot_product_attention(
-            queries,
-            keys,
-            values,
-            cache,
-            scale=self.scale,
-            mask=mask,
+        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
         )
-        out = out.transpose(0, 2, 1, 3).reshape(B, L, -1)
-        return self.o_proj(out)
+
+        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):
@@ -277,11 +263,11 @@ class ShortConv1d(nn.Module):
         out = nn.silu(self.conv(conv_input))
         n_keep = self.kernel_size - 1
         if lengths is not None:
-            ends = mx.clip(cache.lengths, 0, x.shape[1])
+            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 = conv_input[:, -n_keep:, :]
+            new_state = mx.contiguous(conv_input[:, -n_keep:, :])
 
         return out, new_state
 
@@ -335,39 +321,37 @@ class KimiDeltaAttention(nn.Module):
         dtype = x.dtype
 
         if cache is not None:
-            conv_state, ssm_state = cache
+            q_state, k_state, v_state, ssm_state = cache
             lengths = cache.lengths
         else:
-            conv_state = None
+            q_state = None
+            k_state = None
+            v_state = None
             ssm_state = None
             lengths = None
 
-        if conv_state is 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
-        else:
-            q_state, k_state, v_state = conv_state
 
         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, k_state, v_state)
+            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)
 
-        def _l2norm(x, eps=1e-6):
-            norm = mx.linalg.norm(x, axis=-1, keepdims=True)
-            return x / (norm + eps)
-
-        q = _l2norm(q)
-        k = _l2norm(k)
-        q = q * self.scale
+        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
@@ -388,7 +372,7 @@ class KimiDeltaAttention(nn.Module):
         )
 
         if cache is not None:
-            cache[1] = ssm_state
+            cache[3] = ssm_state
             cache.advance(T)
 
         gate = self.g_b_proj(self.g_a_proj(x)).reshape(
@@ -462,7 +446,7 @@ class KimiLinearModel(nn.Module):
             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])
+        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
@@ -500,7 +484,7 @@ class Model(nn.Module):
         caches: List[Any] = []
         for layer in self.layers:
             if layer.is_linear:
-                caches.append(ArraysCache(size=2))
+                caches.append(ArraysCache(size=4))
             else:
                 caches.append(KVCache())
         return caches
@@ -568,6 +552,42 @@ class Model(nn.Module):
                     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

mlx_lm/models/lfm2.py

@@ -32,11 +32,14 @@ class ModelArgs(BaseModelArgs):
     block_multiple_of: int
     block_ffn_dim_multiplier: float
     block_auto_adjust_ff_dim: bool
-    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:

mlx_lm/models/lfm2_moe.py

@@ -35,11 +35,14 @@ class ModelArgs(BaseModelArgs):
     norm_eps: float
     conv_bias: bool
     conv_L_cache: 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.full_attn_idxs is None:
             self.full_attn_idxs = [
                 i

mlx_lm/models/longcat_flash.py

@@ -9,6 +9,7 @@ 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
 
@@ -80,10 +81,11 @@ class LongcatFlashMLA(nn.Module):
             bias=args.attention_bias,
         )
         self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank)
-        self.kv_b_proj = nn.Linear(
-            self.kv_lora_rank,
-            self.num_attention_heads * (self.qk_nope_head_dim + args.v_head_dim),
-            bias=False,
+        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(
@@ -122,56 +124,59 @@ class LongcatFlashMLA(nn.Module):
         B, L, _ = x.shape
 
         if self.q_lora_rank is None:
-            q_states = self.q_proj(x)
+            q = self.q_proj(x)
         else:
-            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
 
-        q_states = q_states.reshape(B, L, -1, self.qk_head_dim).transpose(0, 2, 1, 3)
-
-        if self.mla_scale_q_lora is not None:
-            q_states = q_states * self.mla_scale_q_lora
-
-        q_pass, q_rot = mx.split(q_states, [self.qk_nope_head_dim], axis=-1)
-
-        compressed_kv = self.kv_a_proj_with_mqa(x)
-        k_pass, k_rot = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
-        k_pass = self.kv_a_layernorm(k_pass)
-
-        if self.mla_scale_kv_lora is not None:
-            k_pass = k_pass * self.mla_scale_kv_lora
-
-        key_shape = (B, L, -1, self.qk_nope_head_dim + self.v_head_dim)
-        k_pass = self.kv_b_proj(k_pass).reshape(*key_shape).transpose(0, 2, 1, 3)
-        k_pass, value_states = mx.split(k_pass, [self.qk_nope_head_dim], axis=-1)
-
-        k_rot = k_rot.reshape(B, 1, L, self.qk_rope_head_dim)
-
-        if cache is not None:
-            q_rot = self.rope(q_rot, cache.offset)
-            k_rot = self.rope(k_rot, cache.offset)
-        else:
-            q_rot = self.rope(q_rot)
-            k_rot = self.rope(k_rot)
-
-        k_rot = mx.broadcast_to(k_rot, (*k_pass.shape[:-1], k_rot.shape[-1]))
-
-        query_states = mx.concatenate([q_pass, q_rot], axis=-1)
-        key_states = mx.concatenate([k_pass, k_rot], axis=-1)
-
-        if cache is not None:
-            key_states, value_states = cache.update_and_fetch(key_states, value_states)
-
-        attn_output = scaled_dot_product_attention(
-            query_states,
-            key_states,
-            value_states,
-            cache=cache,
-            scale=self.scale,
-            mask=mask,
+        q = q.reshape(B, L, self.num_attention_heads, self.qk_head_dim).transpose(
+            0, 2, 1, 3
         )
 
-        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(B, L, -1)
-        return self.o_proj(attn_output)
+        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):
@@ -339,7 +344,7 @@ class LongcatFlashModel(nn.Module):
         if cache is None:
             cache = [(None, None)] * self.num_layers
 
-        mask = create_attention_mask(h, cache[0][0])
+        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)
@@ -395,6 +400,47 @@ class Model(nn.Module):
                         ]
                         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"):
@@ -408,6 +454,7 @@ class Model(nn.Module):
     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:
@@ -419,11 +466,17 @@ class Model(nn.Module):
                     attn.q_b_proj = shard_linear(
                         attn.q_b_proj, "all-to-sharded", group=group
                     )
-                attn.kv_b_proj = shard_linear(
-                    attn.kv_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(

mlx_lm/models/longcat_flash_ngram.py

@@ -161,7 +161,7 @@ class LongcatFlashNgramModel(nn.Module):
 
         h = self.ngram_embeddings(input_ids, cache=cache[0])
 
-        mask = create_attention_mask(h, cache[1][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)

mlx_lm/models/minimax.py

@@ -1,6 +1,7 @@
 # Copyright © 2025 Apple Inc.
 
 from dataclasses import dataclass
+from functools import lru_cache
 from typing import Any, List, Optional
 
 import mlx.core as mx
@@ -33,6 +34,55 @@ class ModelArgs(BaseModelArgs):
     use_qk_norm: bool = True
 
 
+@lru_cache
+def sharded_rms_norm(group):
+    @mx.compile
+    def _cast_square_sum(x):
+        return x.astype(mx.float32).square().sum(-1, keepdims=True)
+
+    @mx.compile
+    def _normalize(x, norm2, w, eps):
+        norm2 = mx.distributed.all_sum(norm2, group=group)
+        norm = mx.rsqrt(norm2 / (x.shape[-1] * group.size()) + eps)
+        return (x.astype(mx.float32) * norm * w).astype(x.dtype)
+
+    # Split the compile so that x upcasting doesn't break the compile and we
+    # have 2 kernels generated 1 for f(x) = square(upcast(x)) and another
+    # g(x) = downcast(upcast(x) * norm * w)
+    def _inner_sharded_rms_norm(x, w, eps):
+        return _normalize(x, _cast_square_sum(x), w, eps)
+
+    return _inner_sharded_rms_norm
+
+
+class ShardedRMSNorm(nn.Module):
+    def __init__(
+        self, dims: int, eps: float = 1e-5, group: Optional[mx.distributed.Group] = None
+    ):
+        super().__init__()
+        group = group or mx.distributed.init()
+        self.weight = mx.ones((dims // group.size(),))
+        self.group = group
+        self.eps = eps
+
+    def _extra_repr(self):
+        return f"{self.weight.shape[0] * self.group.size()}, eps={self.eps}"
+
+    def __call__(self, x):
+        return sharded_rms_norm(self.group)(x, self["weight"], self.eps)
+
+    @classmethod
+    def from_rms_norm(
+        cls, norm_module, *, group: Optional[mx.distributed.Group] = None
+    ):
+        sn = cls(norm_module.weight.shape[0], norm_module.eps, group=group)
+        sn.weight = mx.contiguous(
+            mx.split(norm_module.weight, group.size(), axis=-1)[group.rank()]
+        )
+
+        return sn
+
+
 class MiniMaxAttention(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
@@ -295,12 +345,12 @@ class Model(nn.Module):
                 layer.self_attn.o_proj, "sharded-to-all", group=group
             )
             if layer.self_attn.use_qk_norm:
-                layer.self_attn.q_norm.weight = layer.self_attn.q_norm.weight.split(
-                    N, axis=-1
-                )[rank]
-                layer.self_attn.k_norm.weight = layer.self_attn.k_norm.weight.split(
-                    N, axis=-1
-                )[rank]
+                layer.self_attn.q_norm = ShardedRMSNorm.from_rms_norm(
+                    layer.self_attn.q_norm, group=group
+                )
+                layer.self_attn.k_norm = ShardedRMSNorm.from_rms_norm(
+                    layer.self_attn.k_norm, group=group
+                )
 
             layer.self_attn.num_attention_heads //= N
             layer.self_attn.num_key_value_heads //= N

mlx_lm/models/mixtral.py

@@ -1,4 +1,4 @@
-# Copyright © 2023-2024 Apple Inc.
+# Copyright © 2026 Apple Inc.
 
 from dataclasses import dataclass
 from typing import Any, Dict, Optional, Union
@@ -25,6 +25,7 @@ class ModelArgs(BaseModelArgs):
     rope_theta: float = 1e6
     rope_traditional: bool = False
     rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
 
     def __post_init__(self):
         if self.num_key_value_heads is None:
@@ -162,8 +163,12 @@ class MixtralModel(nn.Module):
         self,
         inputs: mx.array,
         cache=None,
-    ):
-        h = self.embed_tokens(inputs)
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        if input_embeddings is not None:
+            h = input_embeddings
+        else:
+            h = self.embed_tokens(inputs)
 
         if cache is None:
             cache = [None] * len(self.layers)
@@ -179,20 +184,27 @@ class MixtralModel(nn.Module):
 class Model(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
+        self.args = args
         self.model_type = args.model_type
         self.model = MixtralModel(args)
-        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
-        self.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: mx.array,
         cache=None,
-    ):
-        out = self.model(inputs, cache)
-        return self.lm_head(out)
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache, input_embeddings)
+        if self.args.tie_word_embeddings:
+            return self.model.embed_tokens.as_linear(out)
+        else:
+            return self.lm_head(out)
 
     def sanitize(self, weights):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
         if "model.layers.0.block_sparse_moe.experts.0.w1.weight" not in weights:
             return weights
         for l in range(self.args.num_hidden_layers):

mlx_lm/models/mla.py

@@ -26,7 +26,7 @@ class MultiLinear(nn.Module):
         self,
         group_size: int,
         bits: int,
-        mode: str,
+        mode: str = "affine",
     ):
         num_heads, output_dims, input_dims = self.weight.shape
         ql = QuantizedMultiLinear(

mlx_lm/models/nemotron_h.py

@@ -40,10 +40,12 @@ class ModelArgs(BaseModelArgs):
     layer_norm_epsilon: float
     use_bias: bool
     use_conv_bias: bool
-    hybrid_override_pattern: List[str]
+    hybrid_override_pattern: Optional[List[str]] = None
+    layers_block_type: Optional[List[str]] = None
     head_dim: Optional[int] = None
     moe_intermediate_size: Optional[int] = None
     moe_shared_expert_intermediate_size: Optional[int] = None
+    moe_latent_size: Optional[int] = None
     n_group: Optional[int] = None
     n_routed_experts: Optional[int] = None
     n_shared_experts: Optional[int] = None
@@ -55,13 +57,20 @@ class ModelArgs(BaseModelArgs):
     time_step_min: Optional[float] = None
     time_step_max: Optional[float] = None
 
+    # Map from layers_block_type names to single-char pattern codes
+    _block_type_to_char = {"mamba": "M", "attention": "*", "moe": "E", "mlp": "-"}
+
     def __post_init__(self):
-        if (
-            self.time_step_limit is None
-            and self.time_step_min is not None
-            and self.time_step_max is not None
-        ):
-            self.time_step_limit = (self.time_step_min, self.time_step_max)
+        if self.time_step_limit is None:
+            self.time_step_limit = (0.0, float("inf"))
+
+        # Normalize to hybrid_override_pattern (single-char list)
+        if self.hybrid_override_pattern is None and self.layers_block_type is not None:
+            self.hybrid_override_pattern = [
+                self._block_type_to_char[t] for t in self.layers_block_type
+            ]
+        if self.hybrid_override_pattern is not None:
+            self.num_hidden_layers = len(self.hybrid_override_pattern)
 
 
 class MambaRMSNormGated(nn.Module):
@@ -365,8 +374,16 @@ class NemotronHMoE(nn.Module):
         super().__init__()
         self.config = config
         self.num_experts_per_tok = config.num_experts_per_tok
+        self.moe_latent_size = config.moe_latent_size
+
+        # When latent projection is used, experts operate on the latent dim
+        expert_input_dim = (
+            config.moe_latent_size
+            if config.moe_latent_size is not None
+            else config.hidden_size
+        )
         self.switch_mlp = SwitchMLP(
-            config.hidden_size,
+            expert_input_dim,
             config.moe_intermediate_size,
             config.n_routed_experts,
             activation=nn.ReLU2(),
@@ -379,12 +396,30 @@ class NemotronHMoE(nn.Module):
                 config, intermediate_size=intermediate_size
             )
 
+        # Latent projection layers for dimensionality reduction before/after experts
+        if config.moe_latent_size is not None:
+            self.fc1_latent_proj = nn.Linear(
+                config.hidden_size, config.moe_latent_size, bias=config.mlp_bias
+            )
+            self.fc2_latent_proj = nn.Linear(
+                config.moe_latent_size, config.hidden_size, bias=config.mlp_bias
+            )
+
     def __call__(self, x):
+        residuals = x
         inds, scores = self.gate(x)
+
+        if self.moe_latent_size is not None:
+            x = self.fc1_latent_proj(x)
+
         y = self.switch_mlp(x, inds)
         y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+
+        if self.moe_latent_size is not None:
+            y = self.fc2_latent_proj(y)
+
         if self.config.n_shared_experts is not None:
-            y = y + self.shared_experts(x)
+            y = y + self.shared_experts(residuals)
 
         return y
 
@@ -501,6 +536,7 @@ class Model(nn.Module):
         return caches
 
     def sanitize(self, weights):
+        weights = {k: v for (k, v) in weights.items() if not k.startswith("mtp.")}
         for k, v in weights.items():
             if "conv1d.weight" in k and v.shape[-1] != 1:
                 weights[k] = v.moveaxis(2, 1)

mlx_lm/models/qwen3_5.py

@@ -0,0 +1,531 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass, field
+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_inplace, shard_linear, sum_gradients
+from mlx.utils import tree_map
+
+from .base import (
+    BaseModelArgs,
+    create_attention_mask,
+    create_ssm_mask,
+)
+from .cache import ArraysCache, KVCache
+from .gated_delta import gated_delta_update
+from .qwen3_next import Qwen3NextAttention as Attention
+from .qwen3_next import Qwen3NextMLP as MLP
+from .qwen3_next import Qwen3NextRMSNormGated as RMSNormGated
+from .qwen3_next import Qwen3NextSparseMoeBlock as SparseMoeBlock
+
+
+@dataclass
+class TextModelArgs(BaseModelArgs):
+    model_type: str = ""
+    hidden_size: int = 4096
+    intermediate_size: int = 14336
+    num_hidden_layers: int = 32
+    num_attention_heads: int = 32
+    rms_norm_eps: float = 1e-6
+    vocab_size: int = 151936
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 131072
+    linear_num_value_heads: int = 64
+    linear_num_key_heads: int = 16
+    linear_key_head_dim: int = 192
+    linear_value_head_dim: int = 128
+    linear_conv_kernel_dim: int = 4
+    tie_word_embeddings: bool = False
+    attention_bias: bool = False
+    head_dim: Optional[int] = None
+    full_attention_interval: int = 4
+
+    # MoE fields (optional, for Qwen3_5MoeForConditionalGeneration)
+    num_experts: int = 0
+    num_experts_per_tok: int = 0
+    decoder_sparse_step: int = 1
+    shared_expert_intermediate_size: int = 0
+    moe_intermediate_size: int = 0
+    norm_topk_prob: bool = True
+
+    # Rope parameters
+    rope_parameters: Optional[Dict[str, Union[float, str, bool, List[int]]]] = field(
+        default_factory=lambda: {
+            "type": "default",
+            "mrope_section": [11, 11, 10],
+            "rope_theta": 100000,
+            "partial_rotary_factor": 0.25,
+        }
+    )
+
+    # Derived from rope_parameters (set in __post_init__)
+    partial_rotary_factor: float = 0.25
+    rope_theta: float = 100000.0
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+    def __post_init__(self):
+        if self.head_dim is None:
+            self.head_dim = self.hidden_size // self.num_attention_heads
+
+        if self.rope_parameters:
+            if (
+                "type" not in self.rope_parameters
+                and "rope_type" in self.rope_parameters
+            ):
+                self.rope_parameters["type"] = self.rope_parameters.pop("rope_type")
+
+            self.partial_rotary_factor = self.rope_parameters.get(
+                "partial_rotary_factor", 0.25
+            )
+            self.rope_theta = self.rope_parameters.get("rope_theta", 100000.0)
+            self.rope_scaling = self.rope_parameters
+
+
+class GatedDeltaNet(nn.Module):
+    def __init__(self, config: TextModelArgs):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_v_heads = config.linear_num_value_heads
+        self.num_k_heads = config.linear_num_key_heads
+        self.head_k_dim = config.linear_key_head_dim
+        self.head_v_dim = config.linear_value_head_dim
+        self.key_dim = self.head_k_dim * self.num_k_heads
+        self.value_dim = self.head_v_dim * self.num_v_heads
+        if self.num_v_heads % self.num_k_heads != 0:
+            raise ValueError(
+                f"num_v_heads ({self.num_v_heads}) must be divisible by num_k_heads ({self.num_k_heads})"
+            )
+
+        self.conv_kernel_size = config.linear_conv_kernel_dim
+        self.layer_norm_epsilon = config.rms_norm_eps
+
+        self.conv_dim = self.key_dim * 2 + self.value_dim
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=False,
+            kernel_size=self.conv_kernel_size,
+            groups=self.conv_dim,
+            padding=0,
+        )
+
+        self.in_proj_qkv = nn.Linear(
+            self.hidden_size, self.key_dim * 2 + self.value_dim, bias=False
+        )
+        self.in_proj_z = nn.Linear(self.hidden_size, self.value_dim, bias=False)
+        self.in_proj_b = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)
+        self.in_proj_a = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)
+
+        self.dt_bias = mx.ones(self.num_v_heads)
+
+        A = mx.random.uniform(low=0, high=16, shape=(self.num_v_heads,))
+        self.A_log = mx.log(A)
+
+        self.norm = RMSNormGated(self.head_v_dim, eps=self.layer_norm_epsilon)
+
+        self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
+
+        self.sharding_group = None
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, S, _ = inputs.shape
+
+        if self.sharding_group is not None:
+            inputs = sum_gradients(self.sharding_group)(inputs)
+
+        qkv = self.in_proj_qkv(inputs)
+        z = self.in_proj_z(inputs).reshape(B, S, self.num_v_heads, self.head_v_dim)
+        b = self.in_proj_b(inputs)
+        a = self.in_proj_a(inputs)
+
+        if cache is not None and cache[0] is not None:
+            conv_state = cache[0]
+        else:
+            conv_state = mx.zeros(
+                (B, self.conv_kernel_size - 1, self.conv_dim),
+                dtype=inputs.dtype,
+            )
+
+        if mask is not None:
+            qkv = mx.where(mask[..., None], qkv, 0)
+        conv_input = mx.concatenate([conv_state, qkv], axis=1)
+        if cache is not None:
+            n_keep = self.conv_kernel_size - 1
+            if cache.lengths is not None:
+                ends = mx.clip(cache.lengths, 0, S)
+                positions = (ends[:, None] + mx.arange(n_keep))[..., None]
+                cache[0] = mx.take_along_axis(conv_input, positions, axis=1)
+            else:
+                cache[0] = mx.contiguous(conv_input[:, -n_keep:, :])
+        conv_out = nn.silu(self.conv1d(conv_input))
+
+        q, k, v = [
+            t.reshape(B, S, h, d)
+            for t, h, d in zip(
+                mx.split(conv_out, [self.key_dim, 2 * self.key_dim], -1),
+                [self.num_k_heads, self.num_k_heads, self.num_v_heads],
+                [self.head_k_dim, self.head_k_dim, self.head_v_dim],
+            )
+        ]
+
+        state = cache[1] if cache else None
+        inv_scale = k.shape[-1] ** -0.5
+        q = (inv_scale**2) * mx.fast.rms_norm(q, None, 1e-6)
+        k = inv_scale * mx.fast.rms_norm(k, None, 1e-6)
+
+        out, state = gated_delta_update(
+            q,
+            k,
+            v,
+            a,
+            b,
+            self.A_log,
+            self.dt_bias,
+            state,
+            mask,
+            use_kernel=not self.training,
+        )
+
+        if cache is not None:
+            cache[1] = state
+            cache.advance(S)
+
+        out = self.norm(out, z)
+        out = self.out_proj(out.reshape(B, S, -1))
+
+        if self.sharding_group is not None:
+            out = mx.distributed.all_sum(out, group=self.sharding_group)
+
+        return out
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: TextModelArgs, layer_idx: int):
+        super().__init__()
+        self.is_linear = (layer_idx + 1) % args.full_attention_interval != 0
+        if self.is_linear:
+            self.linear_attn = GatedDeltaNet(args)
+        else:
+            self.self_attn = Attention(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
+        )
+
+        if args.num_experts > 0:
+            self.mlp = SparseMoeBlock(args)
+        else:
+            self.mlp = MLP(args.hidden_size, args.intermediate_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        if self.is_linear:
+            r = self.linear_attn(self.input_layernorm(x), mask, cache)
+        else:
+            r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        out = h + self.mlp(self.post_attention_layernorm(h))
+        return out
+
+
+class Qwen3_5TextModel(nn.Module):
+    def __init__(self, args: TextModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            DecoderLayer(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)
+        self.ssm_idx = 0
+        self.fa_idx = args.full_attention_interval - 1
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        if input_embeddings is not None:
+            hidden_states = input_embeddings
+        else:
+            hidden_states = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        fa_mask = create_attention_mask(hidden_states, cache[self.fa_idx])
+        ssm_mask = create_ssm_mask(hidden_states, cache[self.ssm_idx])
+
+        for layer, c in zip(self.layers, cache):
+            mask = ssm_mask if layer.is_linear else fa_mask
+            hidden_states = layer(hidden_states, mask=mask, cache=c)
+
+        return self.norm(hidden_states)
+
+
+class TextModel(nn.Module):
+    def __init__(self, args: TextModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Qwen3_5TextModel(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,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        out = self.model(inputs, cache, input_embeddings=input_embeddings)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        return [ArraysCache(size=2) if l.is_linear else KVCache() for l in self.layers]
+
+    def sanitize(self, weights):
+        has_mtp_weights = any("mtp." in k for k in weights)
+        has_unsanitized_conv1d = any(
+            "conv1d.weight" in k and v.shape[-1] != 1 for k, v in weights.items()
+        )
+        should_shift_norm_weights = has_mtp_weights or has_unsanitized_conv1d
+        weights = {k: v for k, v in weights.items() if "mtp." not in k}
+
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+
+        norm_keys = (
+            ".input_layernorm.weight",
+            ".post_attention_layernorm.weight",
+            "model.norm.weight",
+            ".q_norm.weight",
+            ".k_norm.weight",
+        )
+        for k, v in weights.items():
+            if "conv1d.weight" in k and v.shape[-1] != 1:
+                weights[k] = v.moveaxis(2, 1)
+            if should_shift_norm_weights and any(k.endswith(sfx) for sfx in norm_keys):
+                if v.ndim == 1:
+                    weights[k] = v + 1.0
+        return weights
+
+    @property
+    def quant_predicate(self):
+        if self.args.num_experts <= 0:
+            return None
+
+        def predicate(path, _):
+            if path.endswith("mlp.gate") or path.endswith("shared_expert_gate"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate
+
+    @property
+    def cast_predicate(self):
+        def predicate(path: str):
+            if path.endswith("A_log"):
+                return False
+            return True
+
+        return predicate
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+
+    @classmethod
+    def from_dict(cls, params):
+        if "text_config" not in params:
+            return cls(model_type=params["model_type"], text_config=params)
+        return super().from_dict(params)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = TextModel(TextModelArgs.from_dict(args.text_config))
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs, cache=cache, input_embeddings=input_embeddings
+        )
+
+    def sanitize(self, weights):
+        sanitized = {}
+        for key, value in weights.items():
+            if key.startswith("vision_tower") or key.startswith("model.visual"):
+                continue
+            if key.startswith("model.visual"):
+                continue
+            if key.startswith("model.language_model"):
+                key = key.replace("model.language_model", "language_model.model")
+            elif key.startswith("language_model."):
+                pass
+            else:
+                key = "language_model." + key
+            sanitized[key] = value
+        return self.language_model.sanitize(sanitized)
+
+    def shard(self, group=None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        rank = group.rank()
+
+        # A sharding factory for the convolution in gated delta net
+        def conv_sharding(key_dim):
+            return lambda p, w: (0, [key_dim, 2 * key_dim])
+
+        def repeat_kv_layer_inplace(layer, h):
+            # No repeat needed cause we have more heads than nodes
+            if N <= h:
+                return
+
+            # Repeat function to apply to the layer weights
+            def _repeat(p):
+                s = p.shape
+                p = p.reshape(h, s[0] // h, *s[1:])
+                p = mx.repeat(p, N // h, axis=0)
+                p = p.reshape(-1, *s[1:])
+                return p
+
+            layer.update(tree_map(_repeat, layer.parameters()))
+
+        for layer in self.layers:
+            # Linear attention
+            if layer.is_linear:
+                kd = layer.linear_attn.key_dim
+                layer.linear_attn.sharding_group = group
+                shard_inplace(layer.linear_attn.conv1d, conv_sharding(kd), group=group)
+                layer.linear_attn.conv1d.groups //= N
+                shard_inplace(
+                    layer.linear_attn.in_proj_qkv,
+                    "all-to-sharded",
+                    segments=[kd, 2 * kd],
+                    group=group,
+                )
+                shard_inplace(
+                    layer.linear_attn.in_proj_z, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.linear_attn.in_proj_b, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.linear_attn.in_proj_a, "all-to-sharded", group=group
+                )
+                layer.linear_attn.dt_bias = mx.contiguous(
+                    mx.split(layer.linear_attn.dt_bias, N)[rank]
+                )
+                layer.linear_attn.A_log = mx.contiguous(
+                    mx.split(layer.linear_attn.A_log, N)[rank]
+                )
+                shard_inplace(layer.linear_attn.out_proj, "sharded-to-all", group=group)
+                layer.linear_attn.num_k_heads //= N
+                layer.linear_attn.num_v_heads //= N
+                layer.linear_attn.key_dim //= N
+                layer.linear_attn.value_dim //= N
+                layer.linear_attn.conv_dim //= N
+
+            # Softmax attention
+            else:
+                layer.self_attn.o_proj = shard_linear(
+                    layer.self_attn.o_proj, "sharded-to-all", group=group
+                )
+                layer.self_attn.q_proj = shard_linear(
+                    layer.self_attn.q_proj, "all-to-sharded", group=group
+                )
+                repeat_kv_layer_inplace(
+                    layer.self_attn.k_proj, layer.self_attn.num_key_value_heads
+                )
+                repeat_kv_layer_inplace(
+                    layer.self_attn.v_proj, layer.self_attn.num_key_value_heads
+                )
+                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.num_attention_heads //= N
+                layer.self_attn.num_key_value_heads = max(
+                    1, layer.self_attn.num_key_value_heads // N
+                )
+
+            # 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
+                )
+
+            # MoE
+            else:
+                layer.mlp.sharding_group = group
+                shard_inplace(
+                    layer.mlp.shared_expert.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_expert.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    layer.mlp.shared_expert.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.language_model.model.layers
+
+    def make_cache(self):
+        return self.language_model.make_cache()
+
+    @property
+    def quant_predicate(self):
+        return self.language_model.quant_predicate
+
+    @property
+    def cast_predicate(self):
+        return self.language_model.cast_predicate

mlx_lm/models/qwen3_5_moe.py

@@ -0,0 +1,52 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+
+from .base import BaseModelArgs
+from .qwen3_5 import Model as Qwen3_5Model
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+
+    @classmethod
+    def from_dict(cls, params):
+        if "text_config" not in params:
+            return cls(model_type=params["model_type"], text_config=params)
+        return super().from_dict(params)
+
+
+class Model(Qwen3_5Model):
+
+    def sanitize(self, weights):
+        new_weights = {}
+        for key, value in weights.items():
+            if key.startswith("vision_tower") or key.startswith("model.visual"):
+                continue
+            if key.startswith("model.language_model"):
+                key = key.replace("model.language_model", "language_model.model")
+            elif key.startswith("language_model."):
+                pass
+            else:
+                key = "language_model." + key
+            new_weights[key] = value
+
+        for l in range(self.language_model.args.num_hidden_layers):
+            prefix = f"language_model.model.layers.{l}.mlp"
+            gate_up_key = f"{prefix}.experts.gate_up_proj"
+            if gate_up_key in new_weights:
+                gate_up = new_weights.pop(gate_up_key)
+                mid = gate_up.shape[-2] // 2
+                new_weights[f"{prefix}.switch_mlp.gate_proj.weight"] = gate_up[
+                    ..., :mid, :
+                ]
+                new_weights[f"{prefix}.switch_mlp.up_proj.weight"] = gate_up[
+                    ..., mid:, :
+                ]
+                new_weights[f"{prefix}.switch_mlp.down_proj.weight"] = new_weights.pop(
+                    f"{prefix}.experts.down_proj"
+                )
+
+        return self.language_model.sanitize(new_weights)

mlx_lm/models/qwen3_moe.py

@@ -1,4 +1,4 @@
-# Copyright © 2025 Apple Inc.
+# Copyright © 2026 Apple Inc.
 
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Union
@@ -123,7 +123,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
     def __call__(
         self,
         x: mx.array,
-    ):
+    ) -> mx.array:
         gates = self.gate(x)
         gates = mx.softmax(gates, axis=-1, precise=True)
 
@@ -190,7 +190,7 @@ class Qwen3MoeModel(nn.Module):
         inputs: mx.array,
         cache=None,
         input_embeddings: Optional[mx.array] = None,
-    ):
+    ) -> mx.array:
         if input_embeddings is not None:
             h = input_embeddings
         else:
@@ -213,15 +213,25 @@ class Model(nn.Module):
         self.args = args
         self.model_type = args.model_type
         self.model = Qwen3MoeModel(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, cache=None, input_embeddings: Optional[mx.array] = None
-    ):
+        self,
+        inputs: mx.array,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
         out = self.model(inputs, cache, input_embeddings)
-        return self.lm_head(out)
+        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):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
         if "model.layers.0.mlp.experts.0.up_proj.weight" not in weights:
             return weights
         for l in range(self.args.num_hidden_layers):

mlx_lm/models/qwen3_next.py

@@ -3,10 +3,12 @@
 from __future__ import annotations
 
 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 sum_gradients
 
 from .activations import swiglu
 from .base import (
@@ -53,6 +55,13 @@ class ModelArgs(BaseModelArgs):
     full_attention_interval: int = 4
 
 
+@partial(mx.compile, shapeless=True)
+def _precise_swiglu(h, gate, x):
+    gate = nn.silu(gate.astype(mx.float32))
+    x = x.astype(mx.float32)
+    return (gate * x).astype(h.dtype)
+
+
 class Qwen3NextRMSNormGated(nn.Module):
     def __init__(self, hidden_size: int, eps: float = 1e-6):
         super().__init__()
@@ -64,8 +73,9 @@ class Qwen3NextRMSNormGated(nn.Module):
     ) -> mx.array:
         x = mx.fast.rms_norm(hidden_states, self.weight, self.eps)
         if gate is not None:
-            x = swiglu(gate, x)
-        return x
+            return _precise_swiglu(hidden_states, gate, x)
+        else:
+            return x.astype(hidden_states.dtype)
 
 
 class Qwen3NextAttention(nn.Module):
@@ -256,7 +266,7 @@ class Qwen3NextGatedDeltaNet(nn.Module):
                 positions = (ends[:, None] + mx.arange(n_keep))[..., None]
                 cache[0] = mx.take_along_axis(conv_input, positions, axis=1)
             else:
-                cache[0] = conv_input[:, -n_keep:, :]
+                cache[0] = mx.contiguous(conv_input[:, -n_keep:, :])
 
         conv_out = nn.silu(self.conv1d(conv_input))
 
@@ -312,10 +322,15 @@ class Qwen3NextSparseMoeBlock(nn.Module):
         self.shared_expert = Qwen3NextMLP(dim, shared_expert_intermediate_size)
         self.shared_expert_gate = nn.Linear(dim, 1, bias=False)
 
+        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)
+
         gates = self.gate(x)
         gates = mx.softmax(gates, axis=-1, precise=True)
 
@@ -331,7 +346,12 @@ class Qwen3NextSparseMoeBlock(nn.Module):
         shared_y = self.shared_expert(x)
         shared_y = mx.sigmoid(self.shared_expert_gate(x)) * shared_y
 
-        return y + shared_y
+        y = y + shared_y
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y
 
 
 class Qwen3NextDecoderLayer(nn.Module):

mlx_lm/models/rope_utils.py

@@ -58,6 +58,7 @@ class SuScaledRoPE(nn.Module):
         self._scale = long_mscale or (1.0 if factor <= 1.0 else default_scale(factor))
 
     def __call__(self, x, offset: Union[int, mx.array] = 0):
+        x = x[...]
         x[..., : self.dim] = self._scale * x[..., : self.dim]
         return mx.fast.rope(
             x,
@@ -71,7 +72,6 @@ class SuScaledRoPE(nn.Module):
 
 
 class Llama3RoPE(nn.Module):
-
     def __init__(
         self,
         dims: int,
@@ -183,6 +183,7 @@ class YarnRoPE(nn.Module):
 
     def __call__(self, x, offset=0):
         if self.mscale != 1.0:
+            x = x[...]
             x[..., : self.dims] = self.mscale * x[..., : self.dims]
         return mx.fast.rope(
             x,
@@ -195,6 +196,42 @@ class YarnRoPE(nn.Module):
         )
 
 
+class ProportionalRoPE(nn.Module):
+    def __init__(
+        self,
+        dims: int,
+        rotated_dims: int,
+        traditional: bool = False,
+        base: float = 10000.0,
+        factor: float = 1.0,
+    ):
+        super().__init__()
+        self.dims = dims
+        self.traditional = traditional
+
+        if rotated_dims > dims:
+            raise ValueError("rotated_dims should be smaller than dims")
+
+        exponents = mx.arange(0, rotated_dims, 2, dtype=mx.float32) / dims
+        self._freqs = mx.concatenate(
+            [
+                factor * (base**exponents),
+                mx.full(((dims - rotated_dims) // 2,), mx.inf),
+            ]
+        )
+
+    def __call__(self, x, offset=0):
+        return mx.fast.rope(
+            x,
+            self.dims,
+            traditional=self.traditional,
+            base=None,
+            scale=1.0,
+            offset=offset,
+            freqs=self._freqs,
+        )
+
+
 def initialize_rope(
     dims,
     base,
@@ -253,6 +290,14 @@ def initialize_rope(
             short_factor=scaling_config["short_factor"],
             long_factor=scaling_config["long_factor"],
         )
+    elif rope_type == "proportional":
+        return ProportionalRoPE(
+            dims=dims,
+            rotated_dims=int(dims * scaling_config.get("partial_rotary_factor", 1.0)),
+            traditional=traditional,
+            base=base,
+            factor=scaling_config.get("factor", 1.0),
+        )
     elif rope_type == "mrope":
         mrope_section = scaling_config.get("mrope_section", [])
         assert (

mlx_lm/models/ssm.py

@@ -6,6 +6,7 @@ import mlx.nn as nn
 
 @mx.compile
 def compute_dt(dt, dt_bias, time_step_limit):
+    dt = dt.astype(mx.float32)
     dt = nn.softplus(dt + dt_bias)
     return mx.clip(dt, time_step_limit[0], time_step_limit[1])
 
@@ -44,7 +45,7 @@ def make_ssm_kernel():
             auto idx = d_idx * Ds + s_idx;
             auto dB_by_x = x_ * dt_ * static_cast<float>(B_[s_idx]);
             auto state = dA * i_state[idx] + dB_by_x;
-            o_state[idx] = static_cast<T>(state);
+            o_state[idx] = static_cast<U>(state);
             acc += state * C_[s_idx];
         }
         acc = simd_sum(acc);
@@ -76,15 +77,23 @@ def ssm_update_kernel(
 ):
     n, _, h, d = hidden_states.shape
     input_type = hidden_states.dtype
+    state_type = state.dtype
     hb, ds = B.shape[-2:]
     dt = compute_dt(dt, dt_bias, time_step_limit)
     return _ssm_kernel(
         inputs=[hidden_states, A_log, B, C, D, dt, state],
-        template=[("T", input_type), ("Dh", d), ("Ds", ds), ("H", h), ("G", h // hb)],
+        template=[
+            ("T", input_type),
+            ("U", state_type),
+            ("Dh", d),
+            ("Ds", ds),
+            ("H", h),
+            ("G", h // hb),
+        ],
         grid=(32, d, h * n),
         threadgroup=(32, 8, 1),
         output_shapes=[(n, 1, h, d), state.shape],
-        output_dtypes=[input_type, input_type],
+        output_dtypes=[input_type, state_type],
     )
 
 
@@ -186,7 +195,7 @@ def ssm_attn(
                 mx.expand_dims(lengths < 0, (1, 2, 3)), state, next_state
             )
 
-        return y, next_state
+        return y.astype(x.dtype), next_state
 
     ys = []
     for i in range(0, l, step):

mlx_lm/models/step3p5.py

@@ -10,7 +10,7 @@ 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
+from .cache import KVCache, RotatingKVCache
 from .rope_utils import initialize_rope
 from .switch_layers import SwiGLU, SwitchGLU
 
@@ -394,7 +394,14 @@ class Model(nn.Module):
         return self.model.layers
 
     def make_cache(self):
-        return [KVCache() for _ in self.layers]
+        return [
+            (
+                RotatingKVCache(max_size=self.args.sliding_window)
+                if layer.is_sliding
+                else KVCache()
+            )
+            for layer in self.layers
+        ]
 
     def sanitize(self, weights):
         remappings = [

mlx_lm/perplexity.py

@@ -106,6 +106,11 @@ def main():
         required=True,
         help="Path to model or Hugging Face model ID",
     )
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Enable trusting remote code for tokenizer/model loading from Hugging Face.",
+    )
     parser.add_argument(
         "--batch-size", type=int, default=8, help="Batch size for evaluation"
     )
@@ -139,7 +144,8 @@ def main():
 
     # Load model
     print(f"Loading model from {args.model}...")
-    model, tokenizer = load(args.model)
+    tokenizer_config = {"trust_remote_code": True if args.trust_remote_code else None}
+    model, tokenizer = load(args.model, tokenizer_config=tokenizer_config)
 
     # Count parameters
     total_params = get_total_parameters(model)

mlx_lm/quant/dwq.py

@@ -314,7 +314,11 @@ def main():
 
     if args.target_dir is not None:
         target_dir = Path(args.target_dir)
-        has_targets = target_dir.exists()
+        has_targets = (
+            target_dir.is_dir()
+            and any((target_dir / "train").glob("*.safetensors"))
+            and any((target_dir / "valid").glob("*.safetensors"))
+        )
     else:
         has_targets = False
         target_dir = None
@@ -383,7 +387,7 @@ def main():
         del model
 
     if mx.metal.is_available():
-        max_rec_size = mx.metal.device_info()["max_recommended_working_set_size"]
+        max_rec_size = mx.device_info()["max_recommended_working_set_size"]
         mx.set_wired_limit(max_rec_size)
 
     opt = optimizers.Adam(learning_rate=args.learning_rate, bias_correction=True)

mlx_lm/sample_utils.py

@@ -73,15 +73,28 @@ def make_logits_processors(
     logit_bias: Optional[Dict[int, float]] = None,
     repetition_penalty: Optional[float] = None,
     repetition_context_size: Optional[int] = 20,
+    presence_penalty: Optional[float] = None,
+    presence_context_size: Optional[int] = 20,
+    frequency_penalty: Optional[float] = None,
+    frequency_context_size: Optional[int] = 20,
 ):
     """
     Make logits processors for use with ``generate_step``.
 
     Args:
-        repetition_penalty (float, optional): The penalty factor for repeating
-          tokens.
+        repetition_penalty (float, optional): A (sign-aware) multiplicative
+          penalty for repeating tokens.
         repetition_context_size (int, optional): The number of tokens to
           consider for repetition penalty. Default: ``20``.
+        presence_penalty (float, optional): An additive penalty to reduce
+          repeating tokens.
+        presence_context_size (int, optional): The number of tokens to consider
+          for the presence penalty. Default: ``20``.
+        frequency_penalty (float, optional): An additive penalty to reduce
+          repeating tokens. The tokens are penalized proportionally to their
+          frequency.
+        frequency_context_size (int, optional): The number of tokens to consider
+          for the frequency penalty. Default: ``20``.
         logit_bias (dictionary, optional): Additive logit bias.
 
     Returns:
@@ -96,15 +109,20 @@ def make_logits_processors(
         values = mx.array(list(logit_bias.values()))
 
         def logit_bias_processor(_, logits):
-            logits[:, indices] += values
-            return logits
+            return logits.at[:, indices].add(values)
 
         logits_processors.append(logit_bias_processor)
 
-    if repetition_penalty and repetition_penalty != 0.0:
-        logits_processors.append(
-            make_repetition_penalty(repetition_penalty, repetition_context_size)
-        )
+    repetition_penalties = [
+        (make_repetition_penalty, repetition_penalty, repetition_context_size),
+        (make_presence_penalty, presence_penalty, presence_context_size),
+        (make_frequency_penalty, frequency_penalty, frequency_context_size),
+    ]
+
+    for make_penalty, penalty, context_size in repetition_penalties:
+        if penalty is not None and penalty != 0:
+            logits_processors.append(make_penalty(penalty, context_size))
+
     return logits_processors
 
 
@@ -163,39 +181,24 @@ def apply_min_p(
         raise ValueError(
             f"`min_tokens_to_keep` has to be a positive integer, but is {min_tokens_to_keep}"
         )
-    # reference implementation: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L531-L605
 
-    # Indices sorted in decreasing order
-    sorted_indices = mx.argsort(-logprobs, axis=-1)
-    sorted_logprobs = mx.take_along_axis(logprobs, sorted_indices, axis=-1)
-
-    # Top probability
-    top_logprobs = sorted_logprobs[:, 0:1]
-
-    # Calculate the min_p threshold
+    # Mask tokens that have a probability less than the max(p) * min_p
+    top_logprobs = mx.max(logprobs, axis=-1, keepdims=True)
     scaled_min_p = top_logprobs + math.log(min_p)
+    tokens_to_remove = logprobs < scaled_min_p
 
-    # Mask tokens that have a probability less than the scaled min_p
-    tokens_to_remove = sorted_logprobs < scaled_min_p
-    tokens_to_remove[..., :min_tokens_to_keep] = False
+    # Ensure at least min_tokens_to_keep survive the filter
+    if min_tokens_to_keep > 1:
+        top_indices = mx.argpartition(logprobs, kth=-min_tokens_to_keep, axis=-1)
+        top_indices = top_indices[..., -min_tokens_to_keep:]
+        tokens_to_remove = mx.put_along_axis(
+            tokens_to_remove,
+            top_indices,
+            False,
+            axis=-1,
+        )
 
-    # Create pool of tokens with probability less than scaled min_p
-    selected_logprobs = mx.where(tokens_to_remove, -float("inf"), sorted_logprobs)
-
-    # Create a mapping to rearrange back to original indices
-    inverse_indices = mx.put_along_axis(
-        mx.zeros_like(sorted_indices),
-        sorted_indices,
-        mx.arange(sorted_indices.shape[-1], dtype=sorted_indices.dtype),
-        axis=-1,
-    )
-
-    # Rearrange selected_logprobs back to original order
-    original_order_logprobs = mx.take_along_axis(
-        selected_logprobs, inverse_indices, axis=-1
-    )
-
-    return original_order_logprobs
+    return mx.where(tokens_to_remove, -float("inf"), logprobs)
 
 
 @partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
@@ -307,3 +310,58 @@ def make_repetition_penalty(penalty: float, context_size: int = 20):
         return logits
 
     return repetition_penalty_processor
+
+
+def make_presence_penalty(penalty: float, context_size: int = 20):
+    """
+    Make a presence penalty processor.
+
+    Corresponds to the OpenAI option with the same name. Namely, subtracts
+    ``penalty`` from a logit if the token has occured at least once in the
+    ``context_size`` previous tokens.
+
+    Args:
+        penalty (float): The presence penalty to be applied.
+        context_size (int): The number of previous tokens to use.
+            Default: ``20``.
+
+    Returns:
+        Callable[[mx.array, List[int]], mx.array]
+    """
+
+    def presence_penalty_processor(tokens, logits):
+        if len(tokens) > 0:
+            tokens = tokens[-context_size:]
+            logits[:, tokens] -= penalty
+        return logits
+
+    return presence_penalty_processor
+
+
+def make_frequency_penalty(penalty: float, context_size: int = 20):
+    """
+    Make a frequency penalty processor.
+
+    Corresponds to the OpenAI option with the same name. Namely, subtracts
+    ``penalty`` from a logit for every time that the token has occured in the
+    ``context_size`` previous tokens.
+
+    The difference with the presence penalty is that the more often a token
+    occurs the more it will be penalized.
+
+    Args:
+        penalty (float): The frequency penalty to be applied.
+        context_size (int): The number of previous tokens to use.
+            Default: ``20``.
+
+    Returns:
+        Callable[[mx.array, List[int]], mx.array]
+    """
+
+    def frequency_penalty_processor(tokens, logits):
+        if len(tokens) > 0:
+            tokens = tokens[-context_size:]
+            logits = logits.at[:, tokens].subtract(penalty)
+        return logits
+
+    return frequency_penalty_processor

mlx_lm/share.py

@@ -0,0 +1,290 @@
+# Copyright © 2026 Apple Inc.
+
+import argparse
+import os
+import pickle
+import sys
+import time
+from dataclasses import dataclass
+from functools import partial, total_ordering
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from typing import Literal, Optional
+
+import mlx.core as mx
+from huggingface_hub.errors import LocalEntryNotFoundError
+from mlx._distributed_utils.common import Hostfile
+from mlx._distributed_utils.launch import launch_jaccl, launch_ring
+from tqdm import tqdm
+
+from .utils import hf_repo_to_path
+
+CHUNK_SIZE = 100 * 1024 * 1024
+
+
+@total_ordering
+@dataclass
+class DirectoryEntry:
+    entry_type: Literal["directory", "symlink", "file"]
+    path: str
+    dst: Optional[str]
+
+    def __lt__(self, other):
+        order_type = dict(directory=0, symlink=1, file=2)
+        o1 = order_type[self.entry_type]
+        o2 = order_type[other.entry_type]
+        return o1 < o2 or (o1 == o2 and self.path < other.path)
+
+    def __eq__(self, other):
+        return (
+            self.entry_type == other.entry_type
+            and self.path == other.path
+            and self.dst == other.dst
+        )
+
+    @classmethod
+    def from_path(cls, root, path):
+        entry_type = {
+            (True, False): "directory",
+            (False, True): "symlink",
+            (False, False): "file",
+        }[path.is_dir(), path.is_symlink()]
+        dst = path.readlink() if path.is_symlink() else None
+
+        return cls(entry_type, str(path.relative_to(root)), str(dst))
+
+
+def error(*args, **kwargs):
+    kwargs["file"] = sys.stderr
+    print("\033[31m[ERROR]", *args, "\033[0m", **kwargs)
+
+
+def launch(args):
+    if args.hostfile is None:
+        raise ValueError("No hostfile provided")
+
+    hostfile = Hostfile.from_file(args.hostfile)
+    if hostfile.backend == "":
+        raise ValueError("Backend needs to be defined in the hostfile.")
+    if len(hostfile.hosts) == 1:
+        raise ValueError("More than one node needs to be in the hostfile")
+
+    launch_args = argparse.Namespace(
+        backend=hostfile.backend,
+        cwd=str(Path.cwd()),
+        env=hostfile.envs,
+        verbose=False,
+        python=None,
+        starting_port=32323,
+        connections_per_ip=1,
+    )
+    cmd = [
+        sys.executable,
+        "-m",
+        "mlx_lm",
+        "share",
+    ]
+    if args.path is not None:
+        cmd += ["--path", args.path]
+    if args.model is not None:
+        cmd += ["--model", args.model]
+    if args.tmpdir is not None:
+        cmd += ["--tmpdir", args.tmpdir]
+    if args.dst is not None:
+        cmd += ["--dst", args.dst]
+
+    if hostfile.backend == "ring":
+        launch_ring(None, hostfile.hosts, launch_args, cmd)
+    elif hostfile.backend == "jaccl" or hostfile.backend == "jaccl-ring":
+        launch_jaccl(None, hostfile.hosts, launch_args, cmd)
+    else:
+        raise ValueError("Only ring, jaccl and jaccl-ring backends are supported.")
+
+
+def get_files(path):
+    if not path.is_dir():
+        return path.parent, [DirectoryEntry.from_path(path.parent, path)]
+
+    files = [DirectoryEntry.from_path(path, f) for f in path.rglob("*")]
+    return path, sorted(files)
+
+
+def format_bw(x):
+    if x >= 1e9:
+        return f"{x / 1e9:.2} GB/s"
+    if x >= 1e6:
+        return f"{x / 1e6:.2} MB/s"
+    if x >= 1e3:
+        return f"{x / 1e3:.2} KB/s"
+    return f"{x:.2} B/s"
+
+
+def share_file(path, file, src, group=None):
+    group = group or mx.distributed.init()
+    all_sum = partial(mx.distributed.all_sum, group=group)
+    total_size = 0
+    start_time = time.time()
+
+    if group.rank() == src:
+        with open(path / file, "rb") as f:
+            f.seek(0, 2)
+            total_size = f.tell()
+            f.seek(0)
+
+            pbar = tqdm(
+                total=total_size,
+                unit="B",
+                unit_scale=True,
+                desc=file,
+                position=1,
+                leave=False,
+            )
+            while True:
+                data = f.read(CHUNK_SIZE)
+                if not data:
+                    mx.eval(all_sum(0))
+                    break
+
+                mx.eval(all_sum(len(data)))
+                mx.async_eval(all_sum(data))
+                pbar.update(len(data))
+            pbar.close()
+
+    else:
+        with open(path / file, "wb") as f:
+            data = None
+            chunk_size = all_sum(0).item()
+            if chunk_size > 0:
+                data = all_sum(mx.zeros(chunk_size, dtype=mx.uint8))
+                mx.eval(data)
+
+            while chunk_size > 0:
+                next_data = None
+                chunk_size = all_sum(0).item()
+                if chunk_size > 0:
+                    next_data = all_sum(mx.zeros(chunk_size, dtype=mx.uint8))
+                    mx.async_eval(next_data)
+
+                f.write(bytes(data))
+                data = next_data
+
+    return total_size, time.time() - start_time
+
+
+def share_files(path, files, src, group=None):
+    group = group or mx.distributed.init()
+    all_sum = partial(mx.distributed.all_sum, group=group)
+
+    if group.rank() == src:
+        # Share the list first
+        file_list = pickle.dumps(files)
+        mx.eval(all_sum(len(file_list)))
+        mx.eval(all_sum(file_list))
+
+    else:
+        # Get the list first
+        file_list_size = all_sum(0).item()
+        data = all_sum(mx.zeros(file_list_size, dtype=mx.uint8))
+        files = pickle.loads(bytes(data))
+
+        # Make the directories and symlinks
+        for file in files:
+            if file.entry_type == "directory":
+                (path / file.path).mkdir()
+            elif file.entry_type == "symlink":
+                (path / file.path).symlink_to(file.dst)
+
+    # Everybody shares the files
+    total_size = 0
+    total_time = 1e-6
+    pbar = tqdm(total=len(files), desc="Files", position=0, disable=group.rank() != src)
+    for file in files:
+        if file.entry_type == "file":
+            s, t = share_file(path, file.path, src, group)
+            total_size += s
+            total_time += t
+        pbar.update(1)
+        pbar.set_postfix(speed=format_bw(total_size / total_time))
+    pbar.close()
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Distribute a model to other nodes using MLX distributed."
+    )
+    parser.add_argument("--path", type=str, help="Path to a file or folder to share.")
+    parser.add_argument(
+        "--model", type=str, help="The path to a local model or Hugging Face repo"
+    )
+    parser.add_argument(
+        "--hostfile",
+        type=str,
+        help="The file containing the hosts and connection information",
+    )
+    parser.add_argument(
+        "--dst",
+        type=str,
+        help="The destination path in other nodes (defaults to --path or --model)",
+    )
+    parser.add_argument(
+        "--tmpdir",
+        type=str,
+        help="Intermediate temporary directory to ensure successfull transfer",
+    )
+
+    args = parser.parse_args()
+
+    if args.path is args.model is None:
+        parser.error("One of --path or --model must be provided")
+
+    mx.set_default_device(mx.cpu)
+    world = mx.distributed.init()
+
+    if world.size() == 1:
+        launch(args)
+        return
+
+    # Check if any node has the data
+    path = None
+    files = []
+    if args.path is not None and (path := Path(args.path)).exists():
+        path, files = get_files(path)
+    elif args.model is not None:
+        try:
+            path = hf_repo_to_path(args.model)
+            if path.parent.name != "snapshots":
+                raise ValueError(
+                    f"The model repository appears to be corrupted, it resolved to {str(path)}"
+                )
+            path, files = get_files(path.parent.parent)
+        except Exception as e:
+            pass
+    has_file = mx.distributed.all_gather(len(files) > 0)
+    src = has_file.argmax().item()
+    has_file = has_file.any().item()
+
+    if not has_file:
+        error("The --path needs to exist in at least one node.")
+        error("If it is a remote repository download it first with `hf download`")
+        sys.exit(1)
+
+    # Share the path that is resolved
+    if args.dst is None:
+        if world.rank() == src:
+            data = str(path).encode("utf-8")
+            mx.eval(mx.distributed.all_sum(len(data)))
+            mx.eval(mx.distributed.all_sum(data))
+        else:
+            data_size = mx.distributed.all_sum(0).item()
+            data = mx.distributed.all_sum(mx.zeros(data_size, dtype=mx.uint8))
+            path = Path(bytes(data).decode("utf-8"))
+    elif world.rank() != src:
+        path = Path(args.dst)
+
+    with TemporaryDirectory(dir=args.tmpdir) as tmp:
+        if world.rank() == src:
+            share_files(path, files, src, world)
+        else:
+            share_files(Path(tmp), files, src, world)
+            path.mkdir(parents=True, exist_ok=True)
+            os.rename(tmp, path)

mlx_lm/tokenizer_utils.py

@@ -253,6 +253,37 @@ class BPEStreamingDetokenizer(StreamingDetokenizer):
         cls._byte_decoder = char_to_bytes
 
 
+def _infer_thinking(tokenizer):
+    vocab = tokenizer.get_vocab()
+    THINK_TOKENS = [
+        ("<think>", "</think>"),
+        ("<longcat_think>", "</longcat_think>"),
+    ]
+
+    # Single token thinking modes
+    for think_start, think_end in THINK_TOKENS:
+        if think_start in vocab and think_end in vocab:
+            return (
+                think_start,
+                think_end,
+                (vocab[think_start],),
+                (vocab[think_end],),
+            )
+
+    # Multi token thinking modes
+    if "<|channel>" in vocab and "<channel|>" in vocab:
+        think_start = "<|channel>thought"
+        think_end = "<channel|>"
+        return (
+            think_start,
+            think_end,
+            tuple(tokenizer.encode(think_start, add_special_tokens=False)),
+            tuple(tokenizer.encode(think_end, add_special_tokens=False)),
+        )
+
+    return (None, None, None, None)
+
+
 class TokenizerWrapper:
     """A wrapper that combines an HF tokenizer and a detokenizer.
 
@@ -277,10 +308,12 @@ class TokenizerWrapper:
             if eos_token_ids is not None
             else {tokenizer.eos_token_id}
         )
-        self._think_start = None
-        self._think_end = None
-        self._think_start_id = None
-        self._think_end_id = None
+        (
+            self._think_start,
+            self._think_end,
+            self._think_start_tokens,
+            self._think_end_tokens,
+        ) = _infer_thinking(tokenizer)
 
         self._chat_template = chat_template
         self.has_chat_template = (
@@ -289,29 +322,20 @@ class TokenizerWrapper:
         self._tool_parser = tool_parser
         self._tool_call_start = tool_call_start
         self._tool_call_end = tool_call_end
-
-        vocab = tokenizer.get_vocab()
-        THINK_TOKENS = [
-            ("<think>", "</think>"),
-            ("<longcat_think>", "</longcat_think>"),
-        ]
-        for think_start, think_end in THINK_TOKENS:
-            if think_start in vocab and think_end in vocab:
-                self._think_start = think_start
-                self._think_end = think_end
-                self._think_start_id = vocab[think_start]
-                self._think_end_id = vocab[think_end]
-                break
-
-        # Disable tool calling if tool call tokens aren't in vocab
-        if (tool_call_start and tool_call_start not in vocab) or (
-            tool_call_end and tool_call_end not in vocab
-        ):
-            self._tool_call_start = None
-            self._tool_call_end = None
-            self._tool_parser = None
+        self._tool_call_start_tokens = None
+        self._tool_call_end_tokens = None
+        if tool_call_start is not None:
+            self._tool_call_start_tokens = tuple(
+                tokenizer.encode(tool_call_start, add_special_tokens=False)
+            )
+            self._tool_call_end_tokens = tuple(
+                tokenizer.encode(tool_call_end, add_special_tokens=False)
+            )
 
     def apply_chat_template(self, *args, tokenize=True, **kwargs):
+        if "enable_thinking" not in kwargs:
+            kwargs["enable_thinking"] = self.has_thinking
+
         if self._chat_template is not None:
             out = self._chat_template(*args, **kwargs)
             if tokenize:
@@ -333,6 +357,36 @@ class TokenizerWrapper:
 
         self._eos_token_ids.add(token_id)
 
+    def _find(self, tokens, sequence, start=None, end=None, reverse=False):
+        start = start or 0
+        end = end or len(tokens)
+        outer_loop = (
+            range(end - len(sequence), start - 1, -1)
+            if reverse
+            else range(start, end - len(sequence) + 1)
+        )
+        for i in outer_loop:
+            if tokens[i] == sequence[0]:
+                if all(tokens[i + j] == sequence[j] for j in range(1, len(sequence))):
+                    return i
+        return -1
+
+    def find_think_start(self, tokens, start=None, end=None):
+        return self._find(tokens, self._think_start_tokens, start=start, end=end)
+
+    def rfind_think_start(self, tokens, start=None, end=None):
+        return self._find(
+            tokens, self._think_start_tokens, start=start, end=end, reverse=True
+        )
+
+    def find_think_end(self, tokens, start=None, end=None):
+        return self._find(tokens, self._think_end_tokens, start=start, end=end)
+
+    def rfind_think_end(self, tokens, start=None, end=None):
+        return self._find(
+            tokens, self._think_end_tokens, start=start, end=end, reverse=True
+        )
+
     @property
     def has_thinking(self):
         return self._think_start is not None
@@ -343,7 +397,15 @@ class TokenizerWrapper:
 
     @property
     def think_start_id(self):
-        return self._think_start_id
+        if self._think_start_tokens is None:
+            return None
+        if len(self._think_start_tokens) > 1:
+            raise ValueError("The start thinking sequence is more than 1 token")
+        return self._think_start_tokens[0]
+
+    @property
+    def think_start_tokens(self):
+        return self._think_start_tokens
 
     @property
     def think_end(self):
@@ -351,7 +413,15 @@ class TokenizerWrapper:
 
     @property
     def think_end_id(self):
-        return self._think_end_id
+        if self._think_end_tokens is None:
+            return None
+        if len(self._think_end_tokens) > 1:
+            raise ValueError("The end thinking sequence is more than 1 token")
+        return self._think_end_tokens[0]
+
+    @property
+    def think_end_tokens(self):
+        return self._think_end_tokens
 
     @property
     def has_tool_calling(self):
@@ -361,10 +431,18 @@ class TokenizerWrapper:
     def tool_call_start(self):
         return self._tool_call_start
 
+    @property
+    def tool_call_start_tokens(self):
+        return self._tool_call_start_tokens
+
     @property
     def tool_call_end(self):
         return self._tool_call_end
 
+    @property
+    def tool_call_end_tokens(self):
+        return self._tool_call_end_tokens
+
     @property
     def tool_parser(self):
         return self._tool_parser
@@ -473,12 +551,16 @@ def _infer_tool_parser(chat_template):
         return None
     elif "<minimax:tool_call>" in chat_template:
         return "minimax_m2"
+    elif "<|tool_call>" in chat_template and "<tool_call|>" in chat_template:
+        return "gemma4"
     elif "<start_function_call>" in chat_template:
         return "function_gemma"
     elif "<longcat_tool_call>" in chat_template:
         return "longcat"
     elif "<arg_key>" in chat_template:
         return "glm47"
+    elif "<|tool_list_start|>" in chat_template:
+        return "pythonic"
     elif (
         "<tool_call>\\n<function=" in chat_template
         or "<tool_call>\n<function=" in chat_template
@@ -486,6 +568,8 @@ def _infer_tool_parser(chat_template):
         return "qwen3_coder"
     elif "<|tool_calls_section_begin|>" in chat_template:
         return "kimi_k2"
+    elif "[TOOL_CALLS]" in chat_template:
+        return "mistral"
     elif "<tool_call>" in chat_template and "tool_call.name" in chat_template:
         return "json_tools"
     return None

mlx_lm/tool_parsers/gemma4.py

@@ -0,0 +1,65 @@
+# Copyright © 2025 Apple Inc.
+
+import json
+from typing import Any, Optional
+
+import regex as re
+
+# Matches <|"|>...<|"|> string literals (Gemma 4's string delimiter).
+_GEMMA4_STR = r'<\|"\|>(?:(?!<\|"\|>)[\s\S])*?<\|"\|>'
+
+# Matches call:name{...} with balanced braces via the regex module's
+# recursive (?R)-style support.  The inner alternatives handle:
+#   [^{}<]          – any char that is not a brace or start of <|"|>
+#   <(?!\|"\|>)     – a lone '<' that is NOT the start of <|"|>
+#   <|"|>...<|"|>   – a complete string literal (braces inside are ignored)
+#   (?2)            – recursively balanced nested brace group
+_tool_call_regex = re.compile(
+    r"call:([\w-]+)(\{(?:[^{}<]|<(?!\|\"\|>)|" + _GEMMA4_STR + r"|(?2))*\})",
+    re.DOTALL,
+)
+
+
+def _gemma4_args_to_json(text: str) -> str:
+    """Convert Gemma 4 tool call args to valid JSON.
+
+    Gemma 4 uses unquoted keys and <|"|> as string delimiters
+    instead of standard double quotes.
+    """
+    strings = []
+
+    def _capture(m):
+        strings.append(m.group(1))
+        return f"\x00{len(strings) - 1}\x00"
+
+    # Extract <|"|>-delimited strings and replace with placeholders
+    text = re.sub(r'<\|"\|>(.*?)<\|"\|>', _capture, text, flags=re.DOTALL)
+    # Quote bare keys
+    text = re.sub(r"(?<=[{,])(\w+):", r'"\1":', text)
+    # Restore captured strings as properly escaped JSON strings
+    for i, s in enumerate(strings):
+        text = text.replace(f"\x00{i}\x00", json.dumps(s))
+
+    return text
+
+
+def _parse_single(match: re.Match) -> dict:
+    """Parse a single call:name{args} regex match into a tool call dict."""
+    func_name = match.group(1)
+    args_str = match.group(2)
+    json_str = _gemma4_args_to_json(args_str)
+    arguments = json.loads(json_str)
+    return dict(name=func_name, arguments=arguments)
+
+
+def parse_tool_call(text: str, _: Optional[Any] = None):
+    matches = list(_tool_call_regex.finditer(text))
+    if not matches:
+        raise ValueError("No function provided.")
+    if len(matches) == 1:
+        return _parse_single(matches[0])
+    return [_parse_single(m) for m in matches]
+
+
+tool_call_start = "<|tool_call>"
+tool_call_end = "<tool_call|>"

mlx_lm/tool_parsers/minimax_m2.py

@@ -157,43 +157,44 @@ def _get_param_types_from_config(param_name: str, param_config: dict) -> list[st
 
 
 def parse_tool_call(text: str, tools: list | None = None):
-    invoke_match = _invoke_complete_regex.findall(text)
-    if not invoke_match:
+    invoke_matches = _invoke_complete_regex.findall(text)
+    if not invoke_matches:
         raise ValueError("No tool call found")
-    invoke_text = invoke_match[0]
 
-    name_match = re.search(r"^([^>]+)", invoke_text)
-    if not name_match:
-        return None
-
-    function_name = _extract_name(name_match.group(1))
-
-    # Get parameter configuration
-    param_config = {}
+    param_config_for = {}
     if tools:
         for tool in tools:
             if func := tool.get("function", False):
-                if func["name"] != function_name:
-                    continue
                 if params := func.get("parameters", False):
-                    param_config = params.get("properties", {})
+                    param_config_for[func["name"]] = params.get("properties", {})
 
-    # Extract parameters
-    param_dict = {}
-    for match in _parameter_complete_regex.findall(invoke_text):
-        param_match = re.search(r"^([^>]+)>(.*)", match, re.DOTALL)
-        if param_match:
-            param_name = _extract_name(param_match.group(1))
-            param_value = param_match.group(2).strip()
-            if param_value.startswith("\n"):
-                param_value = param_value[1:]
-            if param_value.endswith("\n"):
-                param_value = param_value[:-1]
+    calls = []
+    for invoke_text in invoke_matches:
+        name_match = re.search(r"^([^>]+)", invoke_text)
+        if not name_match:
+            continue
+        function_name = _extract_name(name_match.group(1))
+        param_config = param_config_for.get(function_name, {})
 
-            param_type = _get_param_types_from_config(param_name, param_config)
+        param_dict = {}
+        for match in _parameter_complete_regex.findall(invoke_text):
+            param_match = re.search(r"^([^>]+)>(.*)", match, re.DOTALL)
+            if param_match:
+                param_name = _extract_name(param_match.group(1))
+                param_value = param_match.group(2).strip()
+                if param_value.startswith("\n"):
+                    param_value = param_value[1:]
+                if param_value.endswith("\n"):
+                    param_value = param_value[:-1]
 
-            param_dict[param_name] = _convert_param_value_with_types(
-                param_value, param_type
-            )
+                param_type = _get_param_types_from_config(param_name, param_config)
 
-    return dict(name=function_name, arguments=param_dict)
+                param_dict[param_name] = _convert_param_value_with_types(
+                    param_value, param_type
+                )
+
+        calls.append(dict(name=function_name, arguments=param_dict))
+
+    if len(calls) == 1:
+        return calls[0]
+    return calls

mlx_lm/tool_parsers/mistral.py

@@ -0,0 +1,20 @@
+# Copyright © 2026 Apple Inc.
+
+import json
+from typing import Any
+
+import regex as re
+
+_tool_call_regex = re.compile(r"\s*(\w+)\[ARGS\]\s*(\{.*\})", re.DOTALL)
+
+tool_call_start = "[TOOL_CALLS]"
+tool_call_end = ""
+
+
+def parse_tool_call(text: str, tools: Any | None = None):
+    match = _tool_call_regex.search(text)
+    if match is None:
+        raise ValueError(f"Could not parse tool call from: {text}")
+    func_name = match.group(1)
+    func_args = json.loads(match.group(2))
+    return dict(name=func_name, arguments=func_args)

mlx_lm/tool_parsers/pythonic.py

@@ -0,0 +1,49 @@
+# Copyright © 2026 Apple Inc.
+
+import ast
+from typing import Any, Dict, List
+
+import regex as re
+
+"""
+Tool parser for Pythonic function call formats.
+
+Parses assistant responses containing tool calls in formats like:
+<|tool_call_start|>[function_name(arg1="value1", arg2=2)]<|tool_call_end|>
+"""
+
+
+_tool_call_regex = re.compile(r"\[(\w+)\((.*?)\)\]", re.DOTALL)
+_tool_args_regex = re.compile(r'(\w+)=(?:"([^"]*)"|([^,]+))(?:,\s*|$)', re.DOTALL)
+
+
+def parse_tool_call(text: str, tools: Any | None = None):
+    match = _tool_call_regex.search(text)
+    if not match:
+        raise ValueError("No function provided.")
+
+    func_name = match.group(1)
+    args_str = match.group(2)
+
+    arguments = {}
+    if args_str:
+        matches = _tool_args_regex.findall(args_str)
+        for pair in matches:
+            key = pair[0].strip()
+            # pair[1] is quoted value, pair[2] is unquoted value
+            value = pair[1] if pair[1] else pair[2].strip()
+
+            # Try to parse the value using ast.literal_eval
+            try:
+                value = ast.literal_eval(value)
+            except (ValueError, SyntaxError):
+                # If parsing fails, keep as string
+                pass
+
+            arguments[key] = value
+
+    return dict(name=func_name, arguments=arguments)
+
+
+tool_call_start = "<|tool_call_start|>"
+tool_call_end = "<|tool_call_end|>"

mlx_lm/tool_parsers/qwen3_coder.py

@@ -4,6 +4,7 @@
 Modified from:
 https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/blob/main/qwen3coder_tool_parser.py
 """
+
 import ast
 import json
 from typing import Any, Optional
@@ -70,7 +71,10 @@ def _convert_param_value(param_value: str, param_name: str, param_config: dict)
             or param_type.startswith("dict")
             or param_type.startswith("list")
         ):
-            return json.loads(param_value)
+            try:
+                return json.loads(param_value)
+            except json.JSONDecodeError:
+                return ast.literal_eval(param_value)
 
         return ast.literal_eval(param_value)
 

mlx_lm/tuner/datasets.py

@@ -116,7 +116,7 @@ class CompletionsDataset:
         if self.mask_prompt:
             offset = len(
                 self.tokenizer.apply_chat_template(
-                    messages[0],
+                    messages[:-1],
                     tools=tools,
                     add_generation_prompt=True,
                     return_dict=False,
@@ -322,8 +322,8 @@ def load_dataset(args, tokenizer: PreTrainedTokenizer):
             "Training set not found or empty. Must provide training set for fine-tuning."
         )
     if args.train and len(valid) == 0:
-        raise ValueError(
-            "Validation set not found or empty. Must provide validation set for fine-tuning."
+        print(
+            "Warning: Validation set not found or empty. Training will proceed without validation."
         )
     if args.test and len(test) == 0:
         raise ValueError(

mlx_lm/tuner/trainer.py

@@ -17,6 +17,11 @@ from .callbacks import TrainingCallback
 from .datasets import CacheDataset
 
 
+def _clear_cache(threshold: int):
+    if mx.get_cache_memory() > threshold:
+        mx.clear_cache()
+
+
 def grad_checkpoint(layer):
     """
     Update all instances of type(layer) to use gradient checkpointing.
@@ -70,6 +75,12 @@ class TrainingArgs:
             "help": "Number of steps to accumulate gradients before applying an optimizer update."
         },
     )
+    clear_cache_threshold: int = field(
+        default=0,
+        metadata={
+            "help": "Clear the allocator cache between steps if it grows too large."
+        },
+    )
 
 
 def default_loss(model, batch, lengths):
@@ -170,6 +181,7 @@ def evaluate(
     max_seq_length=2048,
     loss: callable = default_loss,
     iterate_batches: callable = iterate_batches,
+    clear_cache_threshold: int = 0,
 ):
     model.eval()
     all_losses = mx.array(0.0)
@@ -194,25 +206,27 @@ def evaluate(
         all_losses += losses * toks
         ntokens += toks
         mx.eval(all_losses, ntokens)
+        _clear_cache(clear_cache_threshold)
 
     all_losses = mx.distributed.all_sum(all_losses, stream=mx.cpu)
     ntokens = mx.distributed.all_sum(ntokens, stream=mx.cpu)
+    avg_loss = (all_losses / ntokens).item()
 
-    return (all_losses / ntokens).item()
+    return avg_loss
 
 
 def train(
     model,
     optimizer,
     train_dataset,
-    val_dataset,
+    val_dataset=None,
     args: TrainingArgs = TrainingArgs(),
     loss: callable = default_loss,
     iterate_batches: callable = iterate_batches,
     training_callback: TrainingCallback = None,
 ):
     if mx.metal.is_available():
-        mx.set_wired_limit(mx.metal.device_info()["max_recommended_working_set_size"])
+        mx.set_wired_limit(mx.device_info()["max_recommended_working_set_size"])
     print(f"Starting training..., iters: {args.iters}")
     world = mx.distributed.init()
     world_size = world.size()
@@ -269,7 +283,9 @@ def train(
         tic = time.perf_counter()
         # Report validation loss if needed, the first validation loss
         # is always measured before any training.
-        if it == 1 or it % args.steps_per_eval == 0 or it == args.iters:
+        if val_dataset and (
+            it == 1 or it % args.steps_per_eval == 0 or it == args.iters
+        ):
             tic = time.perf_counter()
             val_loss = evaluate(
                 model=model,
@@ -310,6 +326,7 @@ def train(
         n_tokens += toks
         steps += 1
         mx.eval(state, losses, n_tokens, grad_accum)
+        _clear_cache(args.clear_cache_threshold)
         train_time += time.perf_counter() - tic
 
         # Report training loss if needed

mlx_lm/utils.py

@@ -47,6 +47,7 @@ MODEL_REMAPPING = {
     "llava": "mistral3",
     "phi-msft": "phixtral",
     "falcon_mamba": "mamba",
+    "joyai_llm_flash": "deepseek_v3",
     "kimi_k2": "deepseek_v3",
     "qwen2_5_vl": "qwen2_vl",
     "minimax_m2": "minimax",
@@ -56,6 +57,18 @@ MODEL_REMAPPING = {
 MAX_FILE_SIZE_GB = 5
 
 
+def _parse_size(x):
+    sizes = {"M": 1e6, "G": 1e9, "MB": 1e6, "GB": 1e9, "": 1}
+    split = 0
+    for xi in x:
+        if not (xi.isdigit() or xi == "."):
+            break
+        split += 1
+    digits = float(x[:split])
+    size = (x[split:]).strip().upper()
+    return int(digits * sizes[size])
+
+
 def _unpack_awq_weights(qweight: mx.array) -> mx.array:
     bits = 4
     pack_factor = 32 // bits
@@ -494,6 +507,8 @@ def sharded_load(
     pipeline_group: Optional[mx.distributed.Group] = None,
     tensor_group: Optional[mx.distributed.Group] = None,
     return_config: bool = False,
+    *,
+    tokenizer_config: Optional[Dict[str, Any]] = None,
 ):
     # Get model path with everything but weight safetensors
     model_path = _download(
@@ -514,7 +529,7 @@ def sharded_load(
     # weights we need to download.
     model, config = load_model(model_path, lazy=True, strict=False)
 
-    has_pipelining = hasattr(model.model, "pipeline")
+    has_pipelining = hasattr(model, "model") and hasattr(model.model, "pipeline")
     has_tensor_parallel = hasattr(model, "shard")
 
     if pipeline_group is not None and not has_pipelining:
@@ -558,7 +573,7 @@ def sharded_load(
     # Load and shard the model, and load the weights
     tokenizer = load_tokenizer(
         model_path,
-        {"trust_remote_code": True},
+        tokenizer_config or {"trust_remote_code": True},
         eos_token_ids=config.get("eos_token_id", None),
     )
     model, _ = load_model(model_path, lazy=True, strict=False)

setup.py

@@ -10,7 +10,7 @@ sys.path.append(str(package_dir))
 
 from _version import __version__
 
-MIN_MLX_VERSION = "0.30.4"
+MIN_MLX_VERSION = "0.31.2"
 
 setup(
     name="mlx-lm",
@@ -66,6 +66,7 @@ setup(
             "mlx_lm.lora = mlx_lm.lora:main",
             "mlx_lm.perplexity = mlx_lm.perplexity:main",
             "mlx_lm.server = mlx_lm.server:main",
+            "mlx_lm.share = mlx_lm.share:main",
             "mlx_lm.manage = mlx_lm.manage:main",
             "mlx_lm.upload = mlx_lm.upload:main",
         ]

tests/test_datsets.py

@@ -61,6 +61,37 @@ class TestDatasets(unittest.TestCase):
         self.assertTrue(len(valid[0]) > 0)
         self.assertTrue(isinstance(train, datasets.CompletionsDataset))
 
+    def test_completions_mask_prompt(self):
+        data = {"prompt": "What is the capital of France?", "completion": "Paris."}
+        self.save_data(4 * [data])
+        args = types.SimpleNamespace(
+            train=True, test=False, data=self.test_dir, mask_prompt=True
+        )
+        tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_PATH, local_files_only=True)
+        train, valid, test = datasets.load_dataset(args, tokenizer)
+        self.assertEqual(len(train), 4)
+        self.assertEqual(len(valid), 4)
+        self.assertEqual(len(test), 0)
+        expected_prompt_tokens = tokenizer.apply_chat_template(
+            [{"role": "user", "content": data["prompt"]}],
+            add_generation_prompt=True,
+            return_dict=False,
+        )
+        expected_offset = len(expected_prompt_tokens)
+
+        train_tokens, train_offset = train.process(train[0])
+        valid_tokens, valid_offset = valid.process(valid[0])
+
+        self.assertTrue(len(train_tokens) > 0)
+        self.assertTrue(len(valid_tokens) > 0)
+        self.assertEqual(train_offset, expected_offset)
+        self.assertEqual(valid_offset, expected_offset)
+        self.assertLess(train_offset, len(train_tokens))
+        self.assertLess(valid_offset, len(valid_tokens))
+        self.assertEqual(train_tokens[:train_offset], expected_prompt_tokens)
+        self.assertEqual(valid_tokens[:valid_offset], expected_prompt_tokens)
+        self.assertTrue(isinstance(train, datasets.CompletionsDataset))
+
     def test_chat(self):
         data = {
             "messages": [

tests/test_generate.py

@@ -9,12 +9,13 @@ import mlx.core as mx
 from mlx_lm.generate import (
     BatchGenerator,
     GenerationResponse,
+    SequenceStateMachine,
     batch_generate,
     generate,
     generate_step,
     stream_generate,
 )
-from mlx_lm.models.cache import RotatingKVCache
+from mlx_lm.models.cache import KVCache, RotatingKVCache
 from mlx_lm.sample_utils import make_logits_processors, make_sampler
 from mlx_lm.utils import load
 
@@ -199,7 +200,7 @@ class TestGenerate(unittest.TestCase):
             self.model, stop_tokens=self.tokenizer.eos_token_ids, max_tokens=1
         )
         uids = gen.insert(prompts)
-        batch_responses = {r.uid: r for r in gen.next()}
+        batch_responses = {r.uid: r for r in gen.next_generated()}
 
         # Do a test for each prompt the logits are close
         for e, prompt in enumerate(prompts):
@@ -241,7 +242,7 @@ class TestGenerate(unittest.TestCase):
         batch_responses = {}
         not_in = True
         iters = 0
-        while responses := gen.next():
+        while responses := gen.next_generated():
             for r in responses:
                 not_in &= r.uid not in batch_responses
                 batch_responses[r.uid] = r
@@ -289,7 +290,7 @@ class TestGenerate(unittest.TestCase):
         num_toks = [2, 3, 4, 5]
         uids = gen.insert(prompts, max_tokens=num_toks)
         batch_responses = {uid: [] for uid in uids}
-        while responses := gen.next():
+        while responses := gen.next_generated():
             for r in responses:
                 batch_responses[r.uid].append(r.token)
 
@@ -337,7 +338,7 @@ class TestGenerate(unittest.TestCase):
         )
         uids = batch_gen.insert(prompts)
         batch_responses = {uid: [] for uid in uids}
-        while responses := batch_gen.next():
+        while responses := batch_gen.next_generated():
             for r in responses:
                 batch_responses[r.uid].append(r.logprobs)
 
@@ -370,7 +371,7 @@ class TestGenerate(unittest.TestCase):
         )
         prompt = self.tokenizer.encode("hello")
         uids = batch_gen.insert([prompt])
-        response = batch_gen.next()[0]
+        response = batch_gen.next_generated()[0]
         logprobs = response.logprobs
         self.assertEqual(logprobs[0].item(), 0.0)
         self.assertEqual(logprobs.argmin().item(), 1)
@@ -395,12 +396,48 @@ class TestGenerate(unittest.TestCase):
         processors = make_logits_processors(logit_bias)
         (uid2,) = batch_gen.insert([prompt], logits_processors=[processors])
 
-        responses = batch_gen.next()
+        responses = batch_gen.next_generated()
         responses = {response.uid: response for response in responses}
         self.assertEqual(responses[uid0].logprobs[0].item(), 0.0)
         self.assertEqual(responses[uid1].logprobs[1].item(), 0.0)
         self.assertEqual(responses[uid2].logprobs[2].item(), 0.0)
 
+    def test_batch_generate_processor_tokens_match_prompt_on_first_step(self):
+        prompt = self.tokenizer.encode("hello")
+        seen = []
+
+        def processor(tokens, logits):
+            seen.append(tokens)
+            return logits
+
+        batch_gen = BatchGenerator(
+            self.model,
+            max_tokens=1,
+            logits_processors=[processor],
+        )
+        batch_gen.insert([prompt])
+        batch_gen.next_generated()
+
+        self.assertTrue(hasattr(seen[0], "shape"))
+        self.assertEqual(seen[0].tolist(), prompt)
+
+    def test_batch_generate_function_with_logits_processors(self):
+        """Test that batch_generate function with logits_processors produces correct results."""
+        logit_bias = {0: 2000.0, 1: -2000.0}
+        processors = make_logits_processors(logit_bias)
+
+        prompts = [self.tokenizer.encode("hello")]
+        response = batch_generate(
+            self.model,
+            self.tokenizer,
+            prompts,
+            max_tokens=1,
+            logits_processors=processors,
+        )
+        self.assertEqual(len(response.texts), 1)
+        generated_token = self.tokenizer.encode(response.texts[0])[0]
+        self.assertEqual(generated_token, 0)
+
     def test_batch_generate_with_samplers(self):
         """Test that batch_generate with logits_processors produces correct results."""
         batch_gen = BatchGenerator(
@@ -410,7 +447,7 @@ class TestGenerate(unittest.TestCase):
         )
         prompt = self.tokenizer.encode("hello")
         uids = batch_gen.insert([prompt])
-        response = batch_gen.next()[0]
+        response = batch_gen.next_generated()[0]
         self.assertEqual(response.token, 1)
 
         del batch_gen
@@ -427,12 +464,47 @@ class TestGenerate(unittest.TestCase):
             samplers=[lambda _: mx.array([2]), lambda _: mx.array([3])],
         )
 
-        responses = batch_gen.next()
+        responses = batch_gen.next_generated()
         responses = {response.uid: response for response in responses}
         self.assertEqual(responses[uid0].token, 1)
         self.assertEqual(responses[uid1].token, 2)
         self.assertEqual(responses[uid2].token, 3)
 
+    def test_batch_generate_with_state_machines(self):
+        """Test that batch_generate with per-sequence state_machines stops on different tokens."""
+        batch_gen = BatchGenerator(
+            self.model,
+            max_tokens=10,
+        )
+        prompt = self.tokenizer.encode("hello")
+
+        sm_0 = SequenceStateMachine({"normal": [([0], None)]}, initial="normal")
+        sm_1 = SequenceStateMachine({"normal": [([1], None)]}, initial="normal")
+        sm_2 = SequenceStateMachine({"normal": [([2], None)]}, initial="normal")
+
+        processor_0 = make_logits_processors({0: 2000.0})
+        processor_1 = make_logits_processors({1: 2000.0})
+        processor_2 = make_logits_processors({2: 2000.0})
+
+        uid0, uid1, uid2 = batch_gen.insert(
+            [prompt, prompt, prompt],
+            logits_processors=[processor_0, processor_1, processor_2],
+            state_machines=[sm_0, sm_1, sm_2],
+        )
+
+        responses = batch_gen.next_generated()
+        responses = {response.uid: response for response in responses}
+
+        self.assertEqual(responses[uid0].token, 0)
+        self.assertEqual(responses[uid1].token, 1)
+        self.assertEqual(responses[uid2].token, 2)
+        self.assertEqual(responses[uid0].finish_reason, "stop")
+        self.assertEqual(responses[uid1].finish_reason, "stop")
+        self.assertEqual(responses[uid2].finish_reason, "stop")
+        self.assertEqual(responses[uid0].match_sequence, (0,))
+        self.assertEqual(responses[uid1].match_sequence, (1,))
+        self.assertEqual(responses[uid2].match_sequence, (2,))
+
     def test_batch_continued_generation(self):
         for rotating in [False, True]:
             if rotating:
@@ -481,7 +553,7 @@ class TestGenerate(unittest.TestCase):
             )
             uids = batch_gen.insert(prompts_a)
             caches = {uid: None for uid in uids}
-            while responses := batch_gen.next():
+            while responses := batch_gen.next_generated():
                 for r in responses:
                     if r.finish_reason is not None:
                         caches[r.uid] = r.prompt_cache
@@ -490,7 +562,7 @@ class TestGenerate(unittest.TestCase):
             # Generate the 2nd time
             uids = batch_gen.insert(prompts_b, caches=caches)
             batch_responses = {uid: [] for uid in uids}
-            while responses := batch_gen.next():
+            while responses := batch_gen.next_generated():
                 for r in responses:
                     batch_responses[r.uid].append(r.logprobs)
 
@@ -543,7 +615,7 @@ class TestGenerate(unittest.TestCase):
 
         uids = batch_gen.insert(prompts_a)
         caches = {uid: None for uid in uids}
-        while responses := batch_gen.next():
+        while responses := batch_gen.next_generated():
             for r in responses:
                 if r.finish_reason is not None:
                     caches[r.uid] = r.prompt_cache
@@ -553,7 +625,7 @@ class TestGenerate(unittest.TestCase):
         # Generate the 2nd time
         uids = batch_gen.insert(prompts_b, caches=caches)
         batch_responses = {uid: [] for uid in uids}
-        while responses := batch_gen.next():
+        while responses := batch_gen.next_generated():
             for r in responses:
                 batch_responses[r.uid].append(r.logprobs)
 
@@ -632,6 +704,108 @@ class TestGenerate(unittest.TestCase):
         model = qwen3_next.Model(args)
         self._continued_generation_test_helper(model)
 
+    def test_extend_cache_with_empty(self):
+        from mlx_lm.generate import _extend_cache
+        from mlx_lm.models.cache import make_prompt_cache
+
+        cache_a = make_prompt_cache(self.model)
+
+        prompt = mx.array([[1, 2, 3]])
+        self.model(prompt, cache=cache_a)
+        mx.eval([c.state for c in cache_a])
+
+        result = _extend_cache(cache_a, [])
+        self.assertEqual(len(result), len(cache_a))
+        for c in result:
+            self.assertGreater(c.offset, 0)
+
+        result = _extend_cache([], cache_a)
+        self.assertEqual(len(result), len(cache_a))
+        for c in result:
+            self.assertGreater(c.offset, 0)
+
+    def test_remove_prompt_batch_updates_currently_processing(self):
+        prompt_a = self.tokenizer.encode("Write a long story about a cat")
+        prompt_b = self.tokenizer.encode("Write a long story about a dog")
+
+        gen = BatchGenerator(
+            self.model,
+            max_tokens=5,
+            prefill_batch_size=2,
+            prefill_step_size=4,
+            completion_batch_size=4,
+        )
+        uid_a, uid_b = gen.insert([prompt_a, prompt_b])
+
+        gen.next()
+
+        found = gen._find_uids([uid_a, uid_b])
+        for uid in [uid_a, uid_b]:
+            self.assertIn(uid, found)
+            self.assertEqual(found[uid][0], 1)
+
+        gen.remove([uid_a])
+
+        self.assertEqual(len(gen._currently_processing), len(gen._prompt_batch))
+
+        found = gen._find_uids([uid_b])
+        self.assertIn(uid_b, found)
+
+        while responses := gen.next_generated():
+            if all(r.finish_reason is not None for r in responses):
+                break
+
+    def test_batch_max_kv_size_creates_rotating_cache(self):
+        max_kv_size = 256
+        gen = BatchGenerator(
+            self.model,
+            max_tokens=1,
+            max_kv_size=max_kv_size,
+        )
+
+        prompt = self.tokenizer.encode("Write a long story about a cat")
+        gen.insert([prompt])
+
+        for r in gen.next_generated():
+            if r.finish_reason is not None:
+                for cache in r.prompt_cache:
+                    self.assertIsInstance(cache, RotatingKVCache)
+                    self.assertEqual(cache.max_size, max_kv_size)
+
+    def test_batch_max_kv_size_limits_cache_growth(self):
+        max_kv_size = 5
+        gen = BatchGenerator(
+            self.model,
+            max_tokens=10,
+            max_kv_size=max_kv_size,
+            prefill_batch_size=1,
+            prefill_step_size=128,
+            completion_batch_size=1,
+        )
+
+        prompt = self.tokenizer.encode("Write a long story about a cat")
+        gen.insert([prompt])
+
+        for r in gen.next_generated():
+            if r.finish_reason is not None:
+                for cache in r.prompt_cache:
+                    self.assertLessEqual(cache.keys.shape[2], max_kv_size)
+
+    def test_batch_max_kv_size_none_creates_regular_cache(self):
+        gen = BatchGenerator(
+            self.model,
+            max_tokens=1,
+            max_kv_size=None,
+        )
+
+        prompt = self.tokenizer.encode("Write a long story about a cat")
+        gen.insert([prompt])
+
+        for r in gen.next_generated():
+            if r.finish_reason is not None:
+                for cache in r.prompt_cache:
+                    self.assertIsInstance(cache, KVCache)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_gguf.py

@@ -35,6 +35,9 @@ class TestConvertToGGUFWithoutMocks(unittest.TestCase):
         mock_tokenizer.get_vocab.return_value = {"<pad>": 0, "hello": 1, "world": 2}
         mock_tokenizer.all_special_tokens = ["<pad>"]
         mock_tokenizer.all_special_ids = [0]
+        mock_tokenizer.bos_token_id = None
+        mock_tokenizer.eos_token_id = None
+        mock_tokenizer.unk_token_id = None
         mock_from_pretrained.return_value = mock_tokenizer
 
         model_path = Path(self.test_dir)

tests/test_losses.py

@@ -13,21 +13,21 @@ class TestLosses(unittest.TestCase):
         self.assertTrue(can_run_metal())
         mx.random.seed(0)
 
-        logits_q = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        logits_p = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
+        logits_q = mx.random.normal((2, 4, 4000))
+        logits_p = mx.random.normal((2, 4, 4000))
 
         with mx.stream(mx.cpu):
             expected = kl_div_loss(logits_q, logits_p)
         kl = kl_div_loss(logits_q, logits_p)
 
-        self.assertTrue(mx.allclose(kl, expected, rtol=1e-4))
+        self.assertTrue(mx.allclose(kl, expected))
 
     def test_js_div_loss(self):
         self.assertTrue(can_run_metal())
         mx.random.seed(0)
 
-        logits_q = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        logits_p = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
+        logits_q = mx.random.normal((2, 4, 4000))
+        logits_p = mx.random.normal((2, 4, 4000))
 
         with mx.stream(mx.cpu):
             expected = js_div_loss(logits_q, logits_p)
@@ -39,9 +39,9 @@ class TestLosses(unittest.TestCase):
         self.assertTrue(can_run_metal())
         mx.random.seed(0)
 
-        logits_q = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        logits_p = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        cotan = mx.random.uniform(shape=(4, 8), dtype=mx.float32)
+        logits_q = mx.random.normal((2, 4, 4000))
+        logits_p = mx.random.normal((2, 4, 4000))
+        cotan = mx.random.normal((2, 4))
 
         with mx.stream(mx.cpu):
             expected = mx.vjp(kl_div_loss, [logits_q, logits_p], [cotan])[1][0]
@@ -53,9 +53,9 @@ class TestLosses(unittest.TestCase):
         self.assertTrue(can_run_metal())
         mx.random.seed(0)
 
-        logits_q = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        logits_p = mx.random.uniform(shape=(4, 8, 4000), dtype=mx.float32)
-        cotan = mx.random.uniform(shape=(4, 8), dtype=mx.float32)
+        logits_q = mx.random.normal((2, 4, 4000))
+        logits_p = mx.random.normal((2, 4, 4000))
+        cotan = mx.random.normal((2, 4))
 
         with mx.stream(mx.cpu):
             expected = mx.vjp(js_div_loss, [logits_q, logits_p], [cotan])[1][0]

tests/test_models.py

@@ -5,12 +5,16 @@ import unittest
 
 import mlx.core as mx
 import mlx.nn as nn
-from mlx.utils import tree_map
+from mlx.utils import tree_flatten, tree_map
 
 from mlx_lm.models import rope_utils
 from mlx_lm.models.base import create_causal_mask, scaled_dot_product_attention
 from mlx_lm.models.cache import KVCache, RotatingKVCache, make_prompt_cache
-from mlx_lm.models.gated_delta import gated_delta_kernel, gated_delta_ops
+from mlx_lm.models.gated_delta import (
+    gated_delta_kernel,
+    gated_delta_ops,
+    gated_delta_update,
+)
 from mlx_lm.models.ssm import ssm_attn, ssm_update
 
 
@@ -238,6 +242,67 @@ class TestModels(unittest.TestCase):
         )
         self.assertTrue(isinstance(rope, rope_utils.Llama3RoPE))
 
+        rope = rope_utils.initialize_rope(
+            16,
+            base=100.0,
+            traditional=False,
+            scaling_config={
+                "rope_type": "proportional",
+                "partial_rotary_factor": 0.5,
+            },
+        )
+        self.assertTrue(isinstance(rope, rope_utils.ProportionalRoPE))
+        expected_freqs = 100.0 ** (mx.arange(0, 8, 2, dtype=mx.float32) / 16)
+        self.assertTrue(mx.allclose(rope._freqs[:4], expected_freqs))
+        self.assertTrue(mx.all(mx.isinf(rope._freqs[4:])))
+
+        x = mx.arange(16, dtype=mx.float32).reshape(1, 1, 1, 16)
+        y = rope(x, offset=1)
+        expected_rotated = mx.fast.rope(
+            mx.concatenate([x[..., :4], x[..., 8:12]], axis=-1),
+            8,
+            traditional=False,
+            base=None,
+            scale=1.0,
+            offset=1,
+            freqs=expected_freqs,
+        )
+        expected = mx.concatenate(
+            [
+                expected_rotated[..., :4],
+                x[..., 4:8],
+                expected_rotated[..., 4:],
+                x[..., 12:],
+            ],
+            axis=-1,
+        )
+        mx.eval(y, expected)
+        self.assertTrue(mx.allclose(y, expected))
+
+    def test_su_scaled_rope_no_mutation(self):
+        rope = rope_utils.SuScaledRoPE(
+            dims=8,
+            max_position_embeddings=131072,
+            original_max_position_embeddings=4096,
+            long_factor=[1.0] * 4,
+        )
+        x = mx.ones((1, 2, 4, 8))
+        rope(x)
+        mx.eval(x)
+        self.assertTrue((x == 1).all())
+
+    def test_yarn_rope_no_mutation(self):
+        rope = rope_utils.YarnRoPE(
+            dims=8,
+            scaling_factor=2.0,
+            mscale=1.0,
+            mscale_all_dim=0,
+        )
+        x = mx.ones((1, 2, 4, 8))
+        rope(x)
+        mx.eval(x)
+        self.assertTrue((x == 1).all())
+
     def test_quantized_sdpa(self):
         cache = KVCache()
 
@@ -531,6 +596,252 @@ class TestModels(unittest.TestCase):
             model, args.model_type, args.vocab_size, args.num_hidden_layers
         )
 
+    def test_qwen3_5_family_convert_then_load_norm_not_shift_twice(self):
+        text_config = {
+            "hidden_size": 8,
+            "intermediate_size": 16,
+            "num_hidden_layers": 1,
+            "num_attention_heads": 1,
+            "num_key_value_heads": 1,
+            "rms_norm_eps": 1e-5,
+            "vocab_size": 32,
+            "linear_num_value_heads": 1,
+            "linear_num_key_heads": 1,
+            "linear_key_head_dim": 4,
+            "linear_value_head_dim": 4,
+            "linear_conv_kernel_dim": 1,
+            "full_attention_interval": 1,
+            "tie_word_embeddings": False,
+            "max_position_embeddings": 64,
+        }
+        hf_norm_key = "model.language_model.layers.0.input_layernorm.weight"
+        mlx_norm_key = "language_model.model.layers.0.input_layernorm.weight"
+
+        for model_type, hf_mtp_key in (
+            ("qwen3_5", "mtp.fc.weights"),
+            ("qwen3_5_moe", "mtp.fc.weight"),
+        ):
+            module = importlib.import_module(f"mlx_lm.models.{model_type}")
+            args = module.ModelArgs.from_dict(
+                {
+                    "model_type": model_type,
+                    "text_config": {"model_type": model_type, **text_config},
+                }
+            )
+            model = module.Model(args)
+
+            base = mx.arange(8, dtype=mx.float32)
+
+            # Simulate convert sanitize on HF-style keys.
+            converted = model.sanitize(
+                {
+                    hf_norm_key: base,
+                    hf_mtp_key: mx.zeros((1,), dtype=mx.float32),
+                }
+            )
+            self.assertIn(mlx_norm_key, converted)
+            self.assertTrue(mx.array_equal(converted[mlx_norm_key], base + 1.0))
+            self.assertFalse(any("mtp." in k for k in converted))
+
+            # Simulate load sanitize on already-converted keys.
+            loaded = model.sanitize(converted)
+            self.assertTrue(
+                mx.array_equal(loaded[mlx_norm_key], converted[mlx_norm_key])
+            )
+
+    def test_gemma4_convert_then_load_keeps_language_model_prefix(self):
+        from mlx_lm.models import gemma4
+
+        args = gemma4.ModelArgs.from_dict(
+            {
+                "model_type": "gemma4",
+                "vocab_size": 32,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 8,
+                    "num_hidden_layers": 1,
+                    "intermediate_size": 16,
+                    "num_attention_heads": 1,
+                    "num_key_value_heads": 1,
+                    "num_global_key_value_heads": 1,
+                    "head_dim": 8,
+                    "global_head_dim": 8,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 1,
+                    "layer_types": ["full_attention"],
+                    "hidden_size_per_layer_input": 0,
+                    "num_kv_shared_layers": 0,
+                    "tie_word_embeddings": True,
+                },
+            }
+        )
+        model = gemma4.Model(args)
+
+        base = mx.arange(8, dtype=mx.float32)
+        hf_norm_key = "model.language_model.layers.0.input_layernorm.weight"
+        mlx_norm_key = "language_model.model.layers.0.input_layernorm.weight"
+
+        converted = model.sanitize(
+            {
+                hf_norm_key: base,
+                "model.vision_tower.stub": mx.zeros((1,), dtype=mx.float32),
+            }
+        )
+        self.assertIn(mlx_norm_key, converted)
+        self.assertNotIn(
+            "language_model.model.model.layers.0.input_layernorm.weight", converted
+        )
+        self.assertTrue(mx.array_equal(converted[mlx_norm_key], base))
+        self.assertFalse(any("vision_tower" in k for k in converted))
+
+        loaded = model.sanitize({mlx_norm_key: base})
+        self.assertIn(mlx_norm_key, loaded)
+        self.assertNotIn(
+            "language_model.model.model.layers.0.input_layernorm.weight", loaded
+        )
+        self.assertTrue(mx.array_equal(loaded[mlx_norm_key], base))
+
+    def test_gemma4_raw_hf_language_model_prefixes_model(self):
+        from mlx_lm.models import gemma4
+
+        args = gemma4.ModelArgs.from_dict(
+            {
+                "model_type": "gemma4",
+                "vocab_size": 32,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 8,
+                    "num_hidden_layers": 1,
+                    "intermediate_size": 16,
+                    "num_attention_heads": 1,
+                    "num_key_value_heads": 1,
+                    "num_global_key_value_heads": 1,
+                    "head_dim": 8,
+                    "global_head_dim": 8,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 1,
+                    "layer_types": ["full_attention"],
+                    "hidden_size_per_layer_input": 0,
+                    "num_kv_shared_layers": 0,
+                    "tie_word_embeddings": True,
+                },
+            }
+        )
+        model = gemma4.Model(args)
+
+        base = mx.arange(8, dtype=mx.float32)
+        hf_norm_key = "model.language_model.layers.0.input_layernorm.weight"
+        mlx_norm_key = "language_model.model.layers.0.input_layernorm.weight"
+
+        converted = model.sanitize({hf_norm_key: base})
+        self.assertIn(mlx_norm_key, converted)
+        self.assertTrue(mx.array_equal(converted[mlx_norm_key], base))
+
+    def test_gemma4_raw_hf_moe_expert_weights_split_for_switch_glu(self):
+        from mlx_lm.models import gemma4
+
+        args = gemma4.ModelArgs.from_dict(
+            {
+                "model_type": "gemma4",
+                "vocab_size": 32,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 8,
+                    "num_hidden_layers": 1,
+                    "intermediate_size": 16,
+                    "num_attention_heads": 1,
+                    "num_key_value_heads": 1,
+                    "num_global_key_value_heads": 1,
+                    "head_dim": 8,
+                    "global_head_dim": 8,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 1,
+                    "layer_types": ["full_attention"],
+                    "hidden_size_per_layer_input": 0,
+                    "num_kv_shared_layers": 0,
+                    "tie_word_embeddings": True,
+                    "enable_moe_block": True,
+                    "num_experts": 2,
+                    "top_k_experts": 1,
+                    "moe_intermediate_size": 3,
+                },
+            }
+        )
+        model = gemma4.Model(args)
+
+        gate_up = mx.arange(2 * 6 * 8, dtype=mx.float32).reshape(2, 6, 8)
+        down = mx.arange(2 * 8 * 3, dtype=mx.float32).reshape(2, 8, 3)
+
+        converted = model.sanitize(
+            {
+                "model.language_model.layers.0.experts.gate_up_proj": gate_up,
+                "model.language_model.layers.0.experts.down_proj": down,
+            }
+        )
+
+        gate_key = "language_model.model.layers.0.experts.switch_glu.gate_proj.weight"
+        up_key = "language_model.model.layers.0.experts.switch_glu.up_proj.weight"
+        down_key = "language_model.model.layers.0.experts.switch_glu.down_proj.weight"
+
+        self.assertIn(gate_key, converted)
+        self.assertIn(up_key, converted)
+        self.assertIn(down_key, converted)
+        self.assertTrue(mx.array_equal(converted[gate_key], gate_up[:, :3, :]))
+        self.assertTrue(mx.array_equal(converted[up_key], gate_up[:, 3:, :]))
+        self.assertTrue(mx.array_equal(converted[down_key], down))
+        self.assertFalse(any("gate_up_proj" in k for k in converted))
+
+    def test_gemma4_moe_router_quantizes_to_8bit(self):
+        from mlx_lm.models import gemma4
+        from mlx_lm.models.switch_layers import QuantizedSwitchLinear
+        from mlx_lm.utils import quantize_model
+
+        args = gemma4.ModelArgs.from_dict(
+            {
+                "model_type": "gemma4",
+                "vocab_size": 64,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 64,
+                    "num_hidden_layers": 1,
+                    "intermediate_size": 128,
+                    "moe_intermediate_size": 128,
+                    "num_attention_heads": 1,
+                    "num_key_value_heads": 1,
+                    "num_global_key_value_heads": 1,
+                    "head_dim": 64,
+                    "global_head_dim": 64,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 1,
+                    "layer_types": ["full_attention"],
+                    "hidden_size_per_layer_input": 0,
+                    "num_kv_shared_layers": 0,
+                    "tie_word_embeddings": True,
+                    "enable_moe_block": True,
+                    "num_experts": 8,
+                    "top_k_experts": 2,
+                },
+            }
+        )
+        model = gemma4.Model(args)
+        model, config = quantize_model(
+            model,
+            {"model_type": "gemma4", "text_config": copy.deepcopy(args.text_config)},
+            group_size=64,
+            bits=4,
+        )
+
+        layer = model.language_model.model.layers[0]
+        self.assertIsInstance(layer.router.proj, nn.QuantizedLinear)
+        self.assertEqual(layer.router.proj.bits, 8)
+        self.assertIsInstance(layer.experts.switch_glu.gate_proj, QuantizedSwitchLinear)
+        self.assertEqual(layer.experts.switch_glu.gate_proj.bits, 4)
+        self.assertEqual(
+            config["quantization"]["language_model.model.layers.0.router.proj"]["bits"],
+            8,
+        )
+        self.assertEqual(config["quantization"]["bits"], 4)
+
     def test_qwen2_moe(self):
         from mlx_lm.models import qwen2_moe
 
@@ -709,6 +1020,104 @@ class TestModels(unittest.TestCase):
             model, args.model_type, args.vocab_size, args.num_hidden_layers
         )
 
+    def test_step3p5_make_cache_uses_rotating_for_sliding_layers(self):
+        from mlx_lm.models import step3p5
+
+        args = step3p5.ModelArgs(
+            model_type="step3p5",
+            hidden_size=256,
+            num_hidden_layers=4,
+            vocab_size=1024,
+            num_attention_heads=4,
+            num_attention_groups=2,
+            head_dim=64,
+            intermediate_size=512,
+            rms_norm_eps=1e-5,
+            rope_theta=[10000.0, 10000.0, 10000.0, 10000.0],
+            sliding_window=4,
+            layer_types=[
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+            ],
+            partial_rotary_factors=[0.5, 1.0, 1.0, 0.5],
+            attention_other_setting={
+                "num_attention_heads": 8,
+                "num_attention_groups": 2,
+            },
+            use_head_wise_attn_gate=True,
+            moe_num_experts=4,
+            moe_top_k=2,
+            moe_intermediate_size=256,
+            share_expert_dim=256,
+            moe_layers_enum="1,2,3",
+        )
+        model = step3p5.Model(args)
+
+        caches = model.make_cache()
+        self.assertIsInstance(caches[0], KVCache)
+        self.assertIsInstance(caches[1], RotatingKVCache)
+        self.assertIsInstance(caches[2], RotatingKVCache)
+        self.assertIsInstance(caches[3], KVCache)
+
+        tokens = mx.array([[1, 2, 3, 4, 5, 6, 7]], dtype=mx.int32)
+        step = model(tokens[:, :3], cache=caches)
+        mx.eval(step)
+        for i in range(3, 7):
+            step = model(tokens[:, i : i + 1], cache=caches)
+            mx.eval(step)
+
+        self.assertEqual(caches[0].size(), 7)
+        self.assertEqual(caches[1].size(), args.sliding_window)
+        self.assertEqual(caches[2].size(), args.sliding_window)
+        self.assertEqual(caches[3].size(), 7)
+
+    def test_step3p5_make_cache_uses_fallback_sliding_pattern(self):
+        from mlx_lm.models import step3p5
+
+        args = step3p5.ModelArgs(
+            model_type="step3p5",
+            hidden_size=256,
+            num_hidden_layers=5,
+            vocab_size=1024,
+            num_attention_heads=4,
+            num_attention_groups=2,
+            head_dim=64,
+            intermediate_size=512,
+            rms_norm_eps=1e-5,
+            rope_theta=10000.0,
+            sliding_window=4,
+            partial_rotary_factors=[1.0] * 5,
+            use_head_wise_attn_gate=True,
+            moe_num_experts=4,
+            moe_top_k=2,
+            moe_intermediate_size=256,
+            share_expert_dim=256,
+            moe_layers_enum="1,2,3,4",
+        )
+        model = step3p5.Model(args)
+
+        caches = model.make_cache()
+        self.assertIsInstance(caches[0], RotatingKVCache)
+        self.assertIsInstance(caches[1], KVCache)
+        self.assertIsInstance(caches[2], RotatingKVCache)
+        self.assertIsInstance(caches[3], KVCache)
+        self.assertIsInstance(caches[4], RotatingKVCache)
+
+        tokens = mx.array([[1, 2, 3, 4, 5, 6]], dtype=mx.int32)
+        step = model(tokens[:, :2], cache=caches)
+        mx.eval(step)
+        for i in range(2, 6):
+            step = model(tokens[:, i : i + 1], cache=caches)
+            mx.eval(step)
+
+        self.assertEqual(caches[0].size(), args.sliding_window)
+        self.assertEqual(caches[1].size(), 6)
+        self.assertEqual(caches[2].size(), args.sliding_window)
+        self.assertEqual(caches[3].size(), 6)
+        self.assertEqual(caches[4].size(), args.sliding_window)
+
     def test_cohere(self):
         from mlx_lm.models import cohere
 
@@ -1052,6 +1461,206 @@ class TestModels(unittest.TestCase):
             model, args.model_type, args.vocab_size, args.num_hidden_layers
         )
 
+    def test_gemma4_text(self):
+        from mlx_lm.models import gemma4_text
+
+        args = gemma4_text.ModelArgs(
+            model_type="gemma4_text",
+            hidden_size=128,
+            num_hidden_layers=10,
+            intermediate_size=256,
+            num_attention_heads=4,
+            head_dim=32,
+            global_head_dim=64,
+            rms_norm_eps=1e-6,
+            vocab_size=1000,
+            vocab_size_per_layer_input=1000,
+            num_key_value_heads=1,
+            num_kv_shared_layers=4,
+            hidden_size_per_layer_input=32,
+            sliding_window=8,
+            sliding_window_pattern=5,
+            final_logit_softcapping=30.0,
+            layer_types=[
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+            ],
+            rope_parameters={
+                "full_attention": {
+                    "partial_rotary_factor": 0.25,
+                    "rope_theta": 1000000.0,
+                },
+                "sliding_attention": {
+                    "rope_theta": 10000.0,
+                },
+            },
+        )
+        model = gemma4_text.Model(args)
+        self.model_test_runner(
+            model, args.model_type, args.vocab_size, args.num_hidden_layers
+        )
+
+    def test_gemma4_quantized_embedding_preserves_lookup_scale(self):
+        from mlx_lm.models import gemma4_text
+
+        args = gemma4_text.ModelArgs(
+            model_type="gemma4_text",
+            hidden_size=32,
+            num_hidden_layers=1,
+            intermediate_size=64,
+            num_attention_heads=2,
+            num_key_value_heads=1,
+            num_global_key_value_heads=1,
+            head_dim=16,
+            global_head_dim=16,
+            sliding_window=8,
+            sliding_window_pattern=1,
+            layer_types=["full_attention"],
+            hidden_size_per_layer_input=0,
+            vocab_size=4,
+            num_kv_shared_layers=0,
+        )
+        model = gemma4_text.Gemma4TextModel(args)
+        model.embed_tokens.weight = mx.ones((4, 32), dtype=mx.float32)
+        model.embed_tokens = model.embed_tokens.to_quantized(group_size=32, bits=8)
+
+        token_ids = mx.array([[0, 1]], dtype=mx.int32)
+        lookup = model.embed_tokens(token_ids) * model.embed_scale
+        logits = model.embed_tokens.as_linear(mx.ones((1, 1, 32), dtype=mx.float32))
+        mx.eval(lookup, logits)
+
+        self.assertTrue(
+            mx.allclose(
+                lookup,
+                mx.ones((1, 2, 32), dtype=mx.float32) * (32.0**0.5),
+            )
+        )
+        self.assertTrue(
+            mx.allclose(logits, mx.ones((1, 1, 4), dtype=mx.float32) * 32.0)
+        )
+
+    def test_gemma4_kv_shared_layers_omit_kv_projections(self):
+        """KV-shared layers must not create k_proj/v_proj/k_norm/v_norm so that
+        models saved without redundant weights (e.g. via transformers
+        save_pretrained) can be loaded with strict=True."""
+        from mlx_lm.models import gemma4_text
+
+        args = gemma4_text.ModelArgs(
+            model_type="gemma4_text",
+            hidden_size=128,
+            num_hidden_layers=10,
+            intermediate_size=256,
+            num_attention_heads=4,
+            head_dim=32,
+            global_head_dim=64,
+            rms_norm_eps=1e-6,
+            vocab_size=1000,
+            vocab_size_per_layer_input=1000,
+            num_key_value_heads=1,
+            num_kv_shared_layers=4,
+            hidden_size_per_layer_input=32,
+            sliding_window=8,
+            sliding_window_pattern=5,
+            final_logit_softcapping=30.0,
+            layer_types=[
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+            ],
+            rope_parameters={
+                "full_attention": {
+                    "partial_rotary_factor": 0.25,
+                    "rope_theta": 1000000.0,
+                },
+                "sliding_attention": {
+                    "rope_theta": 10000.0,
+                },
+            },
+        )
+        model = gemma4_text.Model(args)
+
+        # Non-shared layers (0-5) should have KV projections
+        for i in range(6):
+            attn = model.model.layers[i].self_attn
+            self.assertTrue(attn.has_kv)
+            self.assertTrue(hasattr(attn, "k_proj"))
+            self.assertTrue(hasattr(attn, "k_norm"))
+
+        # Shared layers (6-9) should NOT have KV projections
+        for i in range(6, 10):
+            attn = model.model.layers[i].self_attn
+            self.assertFalse(attn.has_kv)
+            self.assertFalse(hasattr(attn, "k_proj"))
+            self.assertFalse(hasattr(attn, "k_norm"))
+            self.assertFalse(hasattr(attn, "v_proj"))
+
+        # Verify the model can load weights that omit shared-layer KV params
+        weights = dict(tree_flatten(model.parameters()))
+        kv_keys = [
+            k for k in weights if "k_proj" in k or "v_proj" in k or "k_norm" in k
+        ]
+        for k in kv_keys:
+            # All KV keys should belong to non-shared layers (0-5)
+            layer_idx = int(k.split("layers.")[1].split(".")[0])
+            self.assertLess(layer_idx, 6)
+
+    def test_gemma4_input_embeddings_reconstruct_per_layer_inputs(self):
+        from mlx_lm.models import gemma4_text
+
+        args = gemma4_text.ModelArgs(
+            model_type="gemma4_text",
+            hidden_size=32,
+            num_hidden_layers=2,
+            intermediate_size=64,
+            num_attention_heads=2,
+            num_key_value_heads=1,
+            num_global_key_value_heads=1,
+            head_dim=16,
+            global_head_dim=16,
+            sliding_window=8,
+            sliding_window_pattern=1,
+            layer_types=["full_attention", "full_attention"],
+            hidden_size_per_layer_input=8,
+            vocab_size=32,
+            vocab_size_per_layer_input=32,
+            num_kv_shared_layers=0,
+        )
+        model = gemma4_text.Model(args)
+        tokens = mx.array([[1, 2, 3]], dtype=mx.int32)
+        embeddings = model.model.embed_tokens(tokens)
+        per_layer_inputs = model.model._get_per_layer_inputs(tokens)
+
+        direct = model(tokens)
+        from_embeddings = model(None, input_embeddings=embeddings)
+        explicit = model(
+            None,
+            input_embeddings=embeddings,
+            per_layer_inputs=per_layer_inputs,
+        )
+        mx.eval(direct, from_embeddings, explicit)
+
+        self.assertTrue(
+            mx.allclose(direct.astype(mx.float32), from_embeddings.astype(mx.float32))
+        )
+        self.assertTrue(
+            mx.allclose(direct.astype(mx.float32), explicit.astype(mx.float32))
+        )
+
     def test_gpt_bigcode(self):
         from mlx_lm.models import gpt_bigcode
 
@@ -1485,6 +2094,50 @@ class TestModels(unittest.TestCase):
                 "sliding_window": 8,
                 "sliding_window_pattern": "LLGL",
             },
+            {
+                "model_type": "gemma4",
+                "num_hidden_layers": 10,
+                "vocab_size": 1000,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 128,
+                    "num_hidden_layers": 10,
+                    "intermediate_size": 128,
+                    "num_attention_heads": 4,
+                    "head_dim": 32,
+                    "global_head_dim": 64,
+                    "rms_norm_eps": 1e-6,
+                    "vocab_size": 1000,
+                    "vocab_size_per_layer_input": 1000,
+                    "num_key_value_heads": 1,
+                    "num_kv_shared_layers": 4,
+                    "hidden_size_per_layer_input": 32,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 5,
+                    "final_logit_softcapping": 30.0,
+                    "layer_types": [
+                        "sliding_attention",
+                        "sliding_attention",
+                        "sliding_attention",
+                        "sliding_attention",
+                        "full_attention",
+                        "sliding_attention",
+                        "sliding_attention",
+                        "sliding_attention",
+                        "sliding_attention",
+                        "full_attention",
+                    ],
+                    "rope_parameters": {
+                        "full_attention": {
+                            "partial_rotary_factor": 0.25,
+                            "rope_theta": 1000000.0,
+                        },
+                        "sliding_attention": {
+                            "rope_theta": 10000.0,
+                        },
+                    },
+                },
+            },
             {
                 "model_type": "gemma3n",
                 "num_hidden_layers": 4,
@@ -1532,7 +2185,7 @@ class TestModels(unittest.TestCase):
                 "rms_norm_eps": 1e-5,
                 "vocab_size": 1000,
                 "num_key_value_heads": 2,
-                "partial_rotary_factor": 0,
+                "partial_rotary_factor": 0.5,
                 "rope_theta": 1000,
             },
             {
@@ -1560,7 +2213,7 @@ class TestModels(unittest.TestCase):
                 "use_qk_norm": True,
                 "tie_word_embeddings": False,
                 "attention_bias": False,
-                "partial_rotary_factor": 0.0,
+                "partial_rotary_factor": 0.5,
             },
             {
                 "model_type": "glm4_moe_lite",
@@ -2121,6 +2774,47 @@ class TestModels(unittest.TestCase):
                 "partial_rotary_factor": 0.5,
                 "max_position_embeddings": 1000,
             },
+            {
+                "model_type": "qwen3_5",
+                "hidden_size": 128,
+                "num_hidden_layers": 4,
+                "intermediate_size": 128,
+                "num_attention_heads": 8,
+                "num_key_value_heads": 4,
+                "vocab_size": 1000,
+                "linear_num_value_heads": 4,
+                "linear_num_key_heads": 4,
+                "linear_key_head_dim": 32,
+                "linear_value_head_dim": 32,
+                "linear_conv_kernel_dim": 3,
+                "rms_norm_eps": 1e-5,
+                "head_dim": 64,
+                "rope_theta": 1000.0,
+                "partial_rotary_factor": 0.5,
+                "max_position_embeddings": 1000,
+            },
+            {
+                "model_type": "qwen3_5_moe",
+                "hidden_size": 128,
+                "num_hidden_layers": 4,
+                "num_attention_heads": 8,
+                "num_key_value_heads": 4,
+                "vocab_size": 1000,
+                "linear_num_value_heads": 4,
+                "linear_num_key_heads": 4,
+                "linear_key_head_dim": 32,
+                "linear_value_head_dim": 32,
+                "linear_conv_kernel_dim": 3,
+                "num_experts": 4,
+                "num_experts_per_tok": 2,
+                "shared_expert_intermediate_size": 128,
+                "moe_intermediate_size": 128,
+                "rms_norm_eps": 1e-5,
+                "head_dim": 64,
+                "rope_theta": 1000.0,
+                "partial_rotary_factor": 0.5,
+                "max_position_embeddings": 1000,
+            },
             {
                 "model_type": "kimi_linear",
                 "vocab_size": 1000,
@@ -2141,6 +2835,9 @@ class TestModels(unittest.TestCase):
                 "num_experts": 2,
                 "moe_intermediate_size": 128,
                 "kv_lora_rank": 8,
+                "qk_nope_head_dim": 16,
+                "qk_rope_head_dim": 16,
+                "v_head_dim": 16,
             },
             {
                 "model_type": "afmoe",
@@ -2435,6 +3132,60 @@ class TestModels(unittest.TestCase):
                 self.assertTrue(mx.allclose(y_op, y_c, rtol=1e-4, atol=1e-4))
                 self.assertTrue(mx.allclose(st_op, st_c, rtol=1e-4, atol=1e-4))
 
+    def test_gated_delta_precision(self):
+        mx.random.seed(42)
+
+        N_STEPS = 512
+        B = 1
+        Hk = 4
+        Hv = 4
+        Dk = 64
+        Dv = 64
+
+        A_log = mx.zeros((Hv,))
+        dt_bias = mx.ones((Hv,))
+
+        all_q = mx.random.normal(shape=(N_STEPS, B, 1, Hk, Dk)) * 0.1
+        all_k = mx.random.normal(shape=(N_STEPS, B, 1, Hk, Dk)) * 0.1
+        all_v = mx.random.normal(shape=(N_STEPS, B, 1, Hv, Dv)) * 0.1
+        all_a = -7.0 + mx.random.normal(shape=(N_STEPS, B, 1, Hv)) * 0.3
+        all_b = mx.random.normal(shape=(N_STEPS, B, 1, Hv))
+        mx.eval(all_q, all_k, all_v, all_a, all_b, A_log, dt_bias)
+
+        state_ref = mx.zeros((B, Hv, Dv, Dk), dtype=mx.float32)
+        for t in range(N_STEPS):
+            y_ref, state_ref = gated_delta_update(
+                all_q[t],
+                all_k[t],
+                all_v[t],
+                all_a[t],
+                all_b[t],
+                A_log,
+                dt_bias,
+                state_ref,
+                use_kernel=False,
+            )
+            mx.eval(y_ref, state_ref)
+
+        for use_kernel in (False, True):
+            state_lo = mx.zeros((B, Hv, Dv, Dk), dtype=mx.bfloat16)
+            for t in range(N_STEPS):
+                y_lo, state_lo = gated_delta_update(
+                    all_q[t].astype(mx.bfloat16),
+                    all_k[t].astype(mx.bfloat16),
+                    all_v[t].astype(mx.bfloat16),
+                    all_a[t].astype(mx.bfloat16),
+                    all_b[t].astype(mx.bfloat16),
+                    A_log,
+                    dt_bias,
+                    state_lo,
+                    use_kernel=use_kernel,
+                )
+                mx.eval(y_lo, state_lo)
+
+            self.assertTrue(mx.allclose(state_lo, state_ref, rtol=0.05, atol=0.01))
+            self.assertTrue(mx.allclose(y_lo, y_ref, rtol=0.05, atol=0.01))
+
     def test_gated_delta_masked(self):
         B = 1
         T = 3

tests/test_prompt_cache.py

@@ -132,6 +132,41 @@ class TestPromptCache(unittest.TestCase):
                 self.assertTrue(mx.array_equal(k, lk))
                 self.assertTrue(mx.array_equal(v, lv))
 
+    def test_save_load_cache_list(self):
+        cache_file = os.path.join(self.test_dir, "prompt_cache.safetensors")
+
+        cache = [
+            ArraysCache(size=2),
+            KVCache(),
+            RotatingKVCache(8),
+            ArraysCache(size=2),
+            ChunkedKVCache(256),
+        ]
+        for c in cache:
+            if isinstance(c, ArraysCache):
+                c[0] = mx.random.uniform(shape=(4, 4, 4))
+                c[1] = mx.random.uniform(shape=(4, 4, 4))
+            else:
+                x = mx.random.uniform(shape=(4, 4, 7, 4))
+                y = mx.random.uniform(shape=(4, 4, 7, 4))
+                c.update_and_fetch(x, y)
+        cache = [CacheList(*cache)]
+
+        save_prompt_cache(cache_file, cache)
+        loaded_cache = load_prompt_cache(cache_file)
+        for c, lc in zip(cache[0].caches, loaded_cache[0].caches):
+            if isinstance(c, ArraysCache):
+                self.assertTrue(mx.array_equal(c[0], lc[0]))
+                self.assertTrue(mx.array_equal(c[1], lc[1]))
+            else:
+                x = mx.random.uniform(shape=(4, 4, 1, 4))
+                y = mx.random.uniform(shape=(4, 4, 1, 4))
+                k, v = c.update_and_fetch(x, y)
+                lk, lv = lc.update_and_fetch(x, y)
+                self.assertEqual(c.offset, lc.offset)
+                self.assertTrue(mx.array_equal(k, lk))
+                self.assertTrue(mx.array_equal(v, lv))
+
     def test_save_load_arrays_cache(self):
         cache_file = os.path.join(self.test_dir, "prompt_cache.safetensors")
 
@@ -627,6 +662,102 @@ class TestPromptCache(unittest.TestCase):
         c_out = KVCache.merge((c1, c2))
         self.assertEqual(c_out.keys.shape, (2, 4, 4, 4))
 
+    def test_extend_with_empty_and_nonempty_batch_caches(self):
+        """Extending a batch cache when one side has keys=None should use the
+        correct batch size for the placeholder, not the batch size from the
+        non-None side. Regression test for broadcast error in dynamic_roll."""
+        H, D = 8, 64
+        max_size = 512
+
+        # -- BatchRotatingKVCache --
+        # Create 2 caches with content and 3 empty caches
+        c1 = RotatingKVCache(max_size=max_size)
+        c2 = RotatingKVCache(max_size=max_size)
+        c1.update_and_fetch(mx.ones((1, H, 5, D)), mx.ones((1, H, 5, D)))
+        c2.update_and_fetch(mx.ones((1, H, 3, D)), mx.ones((1, H, 3, D)))
+        batch_full = BatchRotatingKVCache.merge([c1, c2])
+
+        empty_caches = [RotatingKVCache(max_size=max_size) for _ in range(3)]
+        batch_empty = BatchRotatingKVCache.merge(empty_caches)
+
+        # Extend non-empty with empty (different batch sizes)
+        batch_full.extend(batch_empty)
+        self.assertEqual(batch_full.keys.shape[0], 5)
+        self.assertEqual(batch_full.offset.shape[0], 5)
+
+        # Prompt processing with right padding should not crash
+        batch_full.prepare(lengths=[10, 8, 12, 7, 11], right_padding=[2, 4, 0, 5, 1])
+        new_kv = mx.ones((5, H, 12, D))
+        batch_full.update_and_fetch(new_kv, new_kv)
+
+        # Also test empty extending non-empty
+        batch_full2 = BatchRotatingKVCache.merge(
+            [RotatingKVCache(max_size=max_size) for _ in range(3)]
+        )
+        c3 = RotatingKVCache(max_size=max_size)
+        c4 = RotatingKVCache(max_size=max_size)
+        c3.update_and_fetch(mx.ones((1, H, 4, D)), mx.ones((1, H, 4, D)))
+        c4.update_and_fetch(mx.ones((1, H, 6, D)), mx.ones((1, H, 6, D)))
+        batch_content = BatchRotatingKVCache.merge([c3, c4])
+        batch_full2.extend(batch_content)
+        self.assertEqual(batch_full2.keys.shape[0], 5)
+        self.assertEqual(batch_full2.offset.shape[0], 5)
+
+        # -- BatchKVCache --
+        c1 = KVCache()
+        c2 = KVCache()
+        c1.update_and_fetch(mx.ones((1, H, 5, D)), mx.ones((1, H, 5, D)))
+        c2.update_and_fetch(mx.ones((1, H, 3, D)), mx.ones((1, H, 3, D)))
+        batch_full = BatchKVCache.merge([c1, c2])
+
+        empty_caches = [KVCache() for _ in range(3)]
+        batch_empty = BatchKVCache.merge(empty_caches)
+
+        batch_full.extend(batch_empty)
+        self.assertEqual(batch_full.keys.shape[0], 5)
+        self.assertEqual(batch_full.offset.shape[0], 5)
+
+    def test_arrays_cache_extend_with_empty(self):
+        # test simple merge
+        c1 = ArraysCache(2)
+        c2 = ArraysCache(2)
+        c1[0] = mx.zeros((1, 4, 8))
+        c1[1] = mx.zeros((1, 4))
+        c2[0] = mx.zeros((1, 4, 8))
+        c2[1] = mx.zeros((1, 4))
+        full = ArraysCache.merge((c1, c2))
+        self.assertEqual(full[0].shape, (2, 4, 8))
+
+        # extend with empty
+        empty = ArraysCache.merge((ArraysCache(2),))
+        full.extend(empty)
+        self.assertEqual(full[0].shape, (3, 4, 8))
+        self.assertEqual(full[1].shape, (3, 4))
+        self.assertTrue(mx.all(full[0][2:] == 0))
+
+        # making an empty cache with 2 sequences and merging it with
+        # another one with 2 sequences
+        empty2 = ArraysCache.merge((ArraysCache(2), ArraysCache(2)))
+        content = ArraysCache.merge((c1, c2))
+        empty2.extend(content)
+        self.assertEqual(empty2[0].shape, (4, 4, 8))
+        self.assertEqual(empty2[1].shape, (4, 4))
+
+        # Extend content with empty
+        content = ArraysCache.merge((c1, c2))
+        empty2 = ArraysCache.merge((ArraysCache(2), ArraysCache(2)))
+        content.extend(empty2)
+        self.assertEqual(content[0].shape, (4, 4, 8))
+        self.assertEqual(content[1].shape, (4, 4))
+        self.assertEqual(content.make_mask(10).shape, (4, 10))
+
+        # multiple empty extensions accumulate correctly
+        stepwise = ArraysCache.merge((c1,))
+        stepwise.extend(ArraysCache(2))
+        stepwise.extend(ArraysCache.merge((ArraysCache(2), ArraysCache(2))))
+        self.assertEqual(stepwise[0].shape, (4, 4, 8))
+        self.assertEqual(stepwise[1].shape, (4, 4))
+
     def test_window_mask_with_full_kv_cache(self):
         c = KVCache()
         kv = mx.zeros((1, 1, 32, 128))

tests/test_sample_utils.py

@@ -116,6 +116,64 @@ class TestSampleUtils(unittest.TestCase):
         new_probs = mx.softmax(apply_xtc(mx.log(probs), 0, 0.1, [0]), -1)
         self.assertTrue(mx.allclose(new_probs, probs))
 
+    def test_presence_penalty(self):
+        from mlx_lm.sample_utils import make_presence_penalty
+
+        # Token appears multiple times - penalty applied once
+        tokens = mx.array([0, 0, 0, 1, 1])
+        logits = mx.zeros((1, 4))
+        processor = make_presence_penalty(0.5, context_size=5)
+        result = processor(tokens, logits)
+        # Token 0 appears 3 times, token 1 appears 2 times - both penalized once
+        self.assertAlmostEqual(result[0, 0].item(), -0.5)
+        self.assertAlmostEqual(result[0, 1].item(), -0.5)
+        # Tokens not in context not penalized
+        self.assertAlmostEqual(result[0, 2].item(), 0.0)
+        self.assertAlmostEqual(result[0, 3].item(), 0.0)
+
+    def test_frequency_penalty(self):
+        from mlx_lm.sample_utils import make_frequency_penalty
+
+        # Token appears multiple times - penalty applied proportionally
+        tokens = mx.array([0, 0, 0, 1, 1])
+        logits = mx.zeros((1, 4))
+        processor = make_frequency_penalty(0.5, context_size=5)
+        result = processor(tokens, logits)
+        # Token 0 appears 3 times -> 3 * 0.5 = 1.5 penalty
+        self.assertAlmostEqual(result[0, 0].item(), -1.5)
+        # Token 1 appears 2 times -> 2 * 0.5 = 1.0 penalty
+        self.assertAlmostEqual(result[0, 1].item(), -1.0)
+        # Tokens not in context not penalized
+        self.assertAlmostEqual(result[0, 2].item(), 0.0)
+        self.assertAlmostEqual(result[0, 3].item(), 0.0)
+
+    def test_make_logits_processors(self):
+        from mlx_lm.sample_utils import make_logits_processors
+
+        # Create processors with all three penalty types
+        tokens = mx.array([0, 0, 0, 1, 1])
+        # Use non-zero logits so repetition penalty has effect
+        logits = mx.array([[1.0, 0.5, 0.0, -0.5]])
+        processors = make_logits_processors(
+            repetition_penalty=1.5,
+            repetition_context_size=5,
+            presence_penalty=0.5,
+            presence_context_size=5,
+            frequency_penalty=0.25,
+            frequency_context_size=5,
+        )
+        # Apply all processors
+        for processor in processors:
+            logits = processor(tokens, logits)
+        # Token 0 (appears 3x): 1.0/1.5 - 0.5 - 0.75 = -0.5833
+        # Token 1 (appears 2x): 0.5/1.5 - 0.5 - 0.5 = -0.6667
+        # Token 2 (not in context): 0.0 (no penalty)
+        # Token 3 (not in context): -0.5 (no penalty)
+        self.assertAlmostEqual(logits[0, 0].item(), -0.5833, places=4)
+        self.assertAlmostEqual(logits[0, 1].item(), -0.6667, places=4)
+        self.assertAlmostEqual(logits[0, 2].item(), 0.0, places=4)
+        self.assertAlmostEqual(logits[0, 3].item(), -0.5, places=4)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_server.py

@@ -4,13 +4,20 @@ import http
 import io
 import json
 import threading
+import types
 import unittest
 
 import mlx.core as mx
 import requests
 
 from mlx_lm.models.cache import KVCache
-from mlx_lm.server import APIHandler, LRUPromptCache, ResponseGenerator
+from mlx_lm.server import (
+    APIHandler,
+    LRUPromptCache,
+    Response,
+    ResponseGenerator,
+    _process_control_tokens,
+)
 from mlx_lm.utils import load
 
 
@@ -43,6 +50,11 @@ class DummyModelProvider:
                 "model": None,
                 "decode_concurrency": 32,
                 "prompt_concurrency": 8,
+                "prefill_step_size": 2048,
+                "prompt_cache_size": 10,
+                "prompt_cache_bytes": 1 << 63,
+                "prompt_cache_total_bytes": None,
+                "allowed_origins": ["*"],
             },
         )
 
@@ -56,6 +68,94 @@ class DummyModelProvider:
         assert model in ["default_model", "chat_model"]
         return self.model, self.tokenizer
 
+    def load_default(self):
+        return self.load("default_model", None, "default_model")
+
+
+class MockCache:
+    def __init__(self, value, is_trimmable: bool = True):
+        self.value = value
+        self._is_trimmable = is_trimmable
+
+    @property
+    def nbytes(self):
+        return len(self.value)
+
+    def __eq__(self, other):
+        return other.value == self.value
+
+    def is_trimmable(self):
+        return self._is_trimmable
+
+    def trim(self, n):
+        assert self._is_trimmable
+        return n
+
+
+class TestProcessControlTokens(unittest.TestCase):
+    @staticmethod
+    def _r(text, state, match=None):
+        return Response(text, 0, state, match, 0.0, None, ())
+
+    def test_single_tool_call_passes_body_with_open_and_close_crossings(self):
+        r = self._r
+        stream = [
+            r("hi ", "normal"),
+            r("<tool_call>", "tool", match=(0,)),
+            r("body", "tool"),
+            r("</tool_call>", "normal", match=(1,)),
+            r(" bye", "normal"),
+        ]
+        ctx = types.SimpleNamespace(
+            sequences={(0,): "<tool_call>", (1,): "</tool_call>"}
+        )
+        out = list(_process_control_tokens(ctx, iter(stream)))
+
+        self.assertEqual("".join(t.text for t in out), "hi body bye")
+        states = [t.state for t in out]
+        self.assertEqual(sum(1 for a, b in zip(states, states[1:]) if a != b), 2)
+
+    def test_back_to_back_tool_calls_emit_state_crossings(self):
+        r = self._r
+        stream = [
+            r("<tool_call>", "tool", match=(0,)),
+            r("call1_body", "tool"),
+            r("</tool_call>", "normal", match=(1,)),
+            r("<tool_call>", "tool", match=(0,)),
+            r("call2_body", "tool"),
+            r("</tool_call>", "normal", match=(1,)),
+        ]
+        ctx = types.SimpleNamespace(
+            sequences={(0,): "<tool_call>", (1,): "</tool_call>"}
+        )
+        out = list(_process_control_tokens(ctx, iter(stream)))
+
+        self.assertEqual("".join(t.text for t in out), "call1_bodycall2_body")
+        states = [t.state for t in out]
+        crossings = sum(
+            1 for a, b in zip(states, states[1:]) if a == "tool" and b == "normal"
+        )
+        self.assertEqual(crossings, 2)
+
+    def test_multi_token_match_preserves_order(self):
+        r = self._r
+        match = (10, 11, 12)
+        stream = [
+            r("body", "tool"),
+            r("</", "tool"),
+            r("tool", "tool"),
+            r("_call>", "normal", match=match),
+            r(" ok", "normal"),
+        ]
+        ctx = types.SimpleNamespace(sequences={match: "</tool_call>"})
+        out = list(_process_control_tokens(ctx, iter(stream)))
+
+        self.assertEqual([t.text for t in out], ["body", "", "", "", " ok"])
+        self.assertEqual(
+            [t.state for t in out],
+            ["tool", "tool", "tool", "normal", "normal"],
+        )
+
 
 class TestServer(unittest.TestCase):
     @classmethod
@@ -180,6 +280,33 @@ class TestServer(unittest.TestCase):
         self.assertIn("id", response_body)
         self.assertIn("choices", response_body)
 
+    def test_make_state_machine_empty_tool_call_end(self):
+        class FakeTokenizer:
+            has_thinking = False
+            has_tool_calling = True
+            tool_call_start = "[TOOL_CALLS]"
+            tool_call_end = ""
+            tool_call_start_tokens = (100,)
+            tool_call_end_tokens = ()
+            eos_token_ids = [2]
+
+            def convert_ids_to_tokens(self, t):
+                return f"<eos{t}>"
+
+        sm, _ = self.response_generator._make_state_machine(
+            ("fake-empty-end", None, None),
+            FakeTokenizer(),
+            stop_words=[],
+        )
+        state = sm.make_state()
+        state, _, s = sm.match(state, 100)
+        self.assertEqual(s, "tool")
+        for tok in [42, 43, 44]:
+            state, _, s = sm.match(state, tok)
+            self.assertEqual(s, "tool")
+        state, _, s = sm.match(state, 2)
+        self.assertIsNone(s)
+
     def test_handle_models(self):
         url = f"http://localhost:{self.port}/v1/models"
         response = requests.get(url)
@@ -193,18 +320,6 @@ class TestServer(unittest.TestCase):
         self.assertEqual(model["object"], "model")
         self.assertIn("created", model)
 
-    def test_sequence_overlap(self):
-        from mlx_lm.server import sequence_overlap
-
-        self.assertTrue(sequence_overlap([1], [1]))
-        self.assertTrue(sequence_overlap([1, 2], [1, 2]))
-        self.assertTrue(sequence_overlap([1, 3], [3, 4]))
-        self.assertTrue(sequence_overlap([1, 2, 3], [2, 3]))
-
-        self.assertFalse(sequence_overlap([1], [2]))
-        self.assertFalse(sequence_overlap([1, 2], [3, 4]))
-        self.assertFalse(sequence_overlap([1, 2, 3], [4, 1, 2, 3]))
-
 
 class TestServerWithDraftModel(unittest.TestCase):
     @classmethod
@@ -354,7 +469,6 @@ class TestServerWithDraftModel(unittest.TestCase):
 
 
 class TestKeepalive(unittest.TestCase):
-
     def test_keepalive_callback(self):
         """Test keepalive callback sends SSE comments and handles errors"""
         from unittest.mock import Mock
@@ -404,7 +518,6 @@ class TestKeepalive(unittest.TestCase):
 
 
 class TestLRUPromptCache(unittest.TestCase):
-
     def test_caching(self):
         cache = LRUPromptCache(max_size=10)
 
@@ -423,18 +536,23 @@ class TestLRUPromptCache(unittest.TestCase):
         c[0].update_and_fetch(*get_kv(24))
         cache.insert_cache(model, t, c)
 
+        # Fetching a cache that is strictly a prefix doesn't remove it from the
+        # lru cache
         tokens = tokens + [20] * 5
         c, t = cache.fetch_nearest_cache(model, tokens)
         k, v = c[0].state
         self.assertTrue((k == v).all().item())
         self.assertTrue((k.flatten() == mx.arange(24)).all().item())
         self.assertEqual(t, [20] * 5)
-        self.assertEqual(len(cache._lru), 0)
+        self.assertEqual(len(cache), 1)
 
+        # Inserting a trimmable cache with shared prefix removes the prefixes
         tokens = tokens + [30] * 3
         c[0].update_and_fetch(*get_kv(8))
         cache.insert_cache(model, tokens, c)
+        self.assertEqual(len(cache), 1)
 
+        # Fetching a cache with a shared prefix doesn't remove it either
         tokens = tokens[:26] + [40] * 8
         c, t = cache.fetch_nearest_cache(model, tokens)
         k, v = c[0].state
@@ -443,38 +561,130 @@ class TestLRUPromptCache(unittest.TestCase):
             (k.flatten() == mx.concatenate([mx.arange(24), mx.arange(2)])).all().item()
         )
         self.assertEqual(t, [40] * 8)
-        self.assertEqual(len(cache._lru), 1)
+        self.assertEqual(len(cache), 1)
+
+        # Inserting a diverged cache actually creates another entry
+        c[0].update_and_fetch(*get_kv(8))
+        cache.insert_cache(model, tokens, c)
+        self.assertEqual(len(cache), 2)
 
     def test_lru(self):
         cache = LRUPromptCache(max_size=2)
         model = ("test", None, None)
-        cache.insert_cache(model, [1, 2], ["test1"])
-        cache.insert_cache(model, [1, 2], ["test1"])
+        cache.insert_cache(model, [1, 2], [MockCache("test1")])
+        cache.insert_cache(model, [2, 3], [MockCache("test2")])
 
         c, t = cache.fetch_nearest_cache(model, [1, 2])
-        self.assertEqual(c, ["test1"])
+        self.assertEqual(c, [MockCache("test1")])
         self.assertEqual(t, [])
-        c, t = cache.fetch_nearest_cache(model, [1, 2])
-        self.assertEqual(c, ["test1"])
-        self.assertEqual(t, [])
-        c, t = cache.fetch_nearest_cache(model, [1, 2])
-        self.assertEqual(c, None)
-        self.assertEqual(t, [1, 2])
+        c, t = cache.fetch_nearest_cache(model, [1])
+        self.assertEqual(c, [MockCache("test1")])
+        self.assertEqual(t, [1])
+        c, t = cache.fetch_nearest_cache(model, [1, 3, 4])
+        self.assertEqual(c, [MockCache("test1")])
+        self.assertEqual(t, [3, 4])
+        c, t = cache.fetch_nearest_cache(model, [2, 3, 4])
+        self.assertEqual(c, [MockCache("test2")])
+        self.assertEqual(t, [4])
+        c, t = cache.fetch_nearest_cache(model, [2, 4, 5])
+        self.assertEqual(c, [MockCache("test2")])
+        self.assertEqual(t, [4, 5])
 
-        cache.insert_cache(model, [1, 2], ["test1"])
-        cache.insert_cache(model, [2, 3], ["test2"])
-        cache.insert_cache(model, [3, 4], ["test3"])
+        cache.insert_cache(model, [1, 2], [MockCache("test1")])
+        cache.insert_cache(model, [2, 3], [MockCache("test2")])
+        cache.insert_cache(model, [3, 4], [MockCache("test3")])
 
         c, t = cache.fetch_nearest_cache(model, [1, 2])
         self.assertEqual(c, None)
         self.assertEqual(t, [1, 2])
         c, t = cache.fetch_nearest_cache(model, [2, 3])
-        self.assertEqual(c, ["test2"])
+        self.assertEqual(c, [MockCache("test2")])
         self.assertEqual(t, [])
         c, t = cache.fetch_nearest_cache(model, [3, 4])
-        self.assertEqual(c, ["test3"])
+        self.assertEqual(c, [MockCache("test3")])
         self.assertEqual(t, [])
 
+        cache.insert_cache(model, [4, 5], [MockCache("test4")], cache_type="user")
+        c, t = cache.fetch_nearest_cache(model, [2, 3])
+        self.assertEqual(c, None)
+        self.assertEqual(t, [2, 3])
+        c, t = cache.fetch_nearest_cache(model, [3, 4])
+        self.assertEqual(c, [MockCache("test3")])
+        self.assertEqual(t, [])
+        c, t = cache.fetch_nearest_cache(model, [4, 5])
+        self.assertEqual(c, [MockCache("test4")])
+        self.assertEqual(t, [])
+
+        cache.insert_cache(model, [5, 6], [MockCache("test5")])
+        cache.insert_cache(model, [6, 7], [MockCache("test6")])
+        c, t = cache.fetch_nearest_cache(model, [5, 6])
+        self.assertEqual(c, None)
+        self.assertEqual(t, [5, 6])
+        c, t = cache.fetch_nearest_cache(model, [6, 7])
+        self.assertEqual(c, [MockCache("test6")])
+        self.assertEqual(t, [])
+        c, t = cache.fetch_nearest_cache(model, [4, 5])
+        self.assertEqual(c, [MockCache("test4")])
+        self.assertEqual(t, [])
+
+    def test_insert_trimmable_cache_removes_immediate_prefix(self):
+        cache = LRUPromptCache(max_size=10)
+        model = ("test", None, None)
+
+        cache.insert_cache(model, [1, 2], [MockCache("ab")])
+        self.assertEqual(len(cache), 1)
+        self.assertEqual(cache.nbytes, 2)
+
+        cache.insert_cache(model, [1, 2, 3], [MockCache("abc")])
+        self.assertEqual(len(cache), 1)
+        self.assertEqual(cache.nbytes, 3)
+
+    def test_insert_empty_tokens_does_not_self_destruct(self):
+        cache = LRUPromptCache(max_size=10)
+        model = ("test", None, None)
+
+        cache.insert_cache(model, [], [MockCache("root")])
+        self.assertEqual(len(cache), 1)
+        self.assertEqual(cache.nbytes, 4)
+
+        c, t = cache.fetch_nearest_cache(model, [])
+        self.assertIsNotNone(c)
+        self.assertEqual(t, [])
+
+    def test_fetch_empty_tokens_after_root_eviction(self):
+        cache = LRUPromptCache(max_size=10)
+        model = ("test", None, None)
+
+        cache.insert_cache(model, [], [MockCache("root")])
+        cache.insert_cache(model, [1], [MockCache("a")])
+
+        c, t = cache.fetch_nearest_cache(model, [])
+        self.assertIsNone(c)
+        self.assertEqual(t, [])
+
+    def test_lru_bytes(self):
+        cache = LRUPromptCache(max_size=100, max_bytes=10)
+        model = ("test", None, None)
+
+        cache.insert_cache(model, [1, 2], [MockCache("aaa")])
+        cache.insert_cache(model, [3, 4], [MockCache("bbb")])
+        cache.insert_cache(model, [4, 5], [MockCache("ccc")])
+        cache.insert_cache(model, [6, 7], [MockCache("ddd")])
+
+        self.assertEqual(len(cache), 3)
+        self.assertEqual(cache.nbytes, 9)
+
+        cache.trim_to(n_bytes=7)
+        self.assertEqual(len(cache), 2)
+        self.assertEqual(cache.nbytes, 6)
+
+        c, t = cache.fetch_nearest_cache(model, [1, 2])
+        self.assertEqual(c, None)
+        self.assertEqual(t, [1, 2])
+        c, t = cache.fetch_nearest_cache(model, [3, 4])
+        self.assertEqual(c, None)
+        self.assertEqual(t, [3, 4])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_tokenizers.py

@@ -101,6 +101,14 @@ class TestTokenizers(unittest.TestCase):
         self.assertEqual(tokenizer.think_start, "<think>")
         self.assertEqual(tokenizer.think_end, "</think>")
 
+        tokenizer_repo = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+        tokenizer = load_tokenizer(tokenizer_repo)
+        self.assertFalse(tokenizer.has_thinking)
+        self.assertIsNone(tokenizer.think_start)
+        self.assertIsNone(tokenizer.think_end)
+        self.assertIsNone(tokenizer.think_start_id)
+        self.assertIsNone(tokenizer.think_end_id)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_tool_parsing.py

@@ -3,20 +3,23 @@ from pathlib import Path
 
 from mlx_lm.tool_parsers import (
     function_gemma,
+    gemma4,
     glm47,
     json_tools,
     kimi_k2,
     longcat,
     minimax_m2,
+    mistral,
+    pythonic,
     qwen3_coder,
 )
 
 
 class TestToolParsing(unittest.TestCase):
-
     def test_parsers(self):
         test_cases = [
             ("call:multiply{a:12234585,b:48838483920}", function_gemma),
+            ("call:multiply{a:12234585,b:48838483920}", gemma4),
             (
                 '{"name": "multiply", "arguments": {"a": 12234585, "b": 48838483920}}',
                 glm47,
@@ -46,6 +49,14 @@ class TestToolParsing(unittest.TestCase):
                 '{"name": "multiply", "arguments": {"a": 12234585, "b": 48838483920}}',
                 longcat,
             ),
+            (
+                "[multiply(a=12234585, b=48838483920)]",
+                pythonic,
+            ),
+            (
+                'multiply[ARGS]{"a": 12234585, "b": 48838483920}',
+                mistral,
+            ),
         ]
 
         tools = [
@@ -80,6 +91,10 @@ class TestToolParsing(unittest.TestCase):
                 "call:get_current_temperature{location:<escape>London<escape>}",
                 function_gemma,
             ),
+            (
+                'call:get_current_temperature{location:<|"|>London<|"|>}',
+                gemma4,
+            ),
             (
                 'get_current_temperature<arg_key>location</arg_key><arg_value>"London"</arg_value>',
                 glm47,
@@ -104,6 +119,14 @@ class TestToolParsing(unittest.TestCase):
                 '{"name": "get_current_temperature", "arguments": {"location": "London"}}',
                 longcat,
             ),
+            (
+                '[get_current_temperature(location="London")]',
+                pythonic,
+            ),
+            (
+                'get_current_temperature[ARGS]{"location": "London"}',
+                mistral,
+            ),
         ]
         tools = [
             {
@@ -131,6 +154,127 @@ class TestToolParsing(unittest.TestCase):
                 }
                 self.assertEqual(tool_call, expected)
 
+    def test_qwen3_coder_single_quoted_params(self):
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "search",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "filters": {"type": "object"},
+                            "tags": {"type": "array"},
+                        },
+                    },
+                },
+            }
+        ]
+
+        # single-quoted dict (python-style, not valid JSON)
+        test_case = (
+            "<function=search>"
+            "<parameter=filters>{'category': 'books', 'in_stock': True}</parameter>"
+            "<parameter=tags>['fiction', 'new']</parameter>"
+            "</function>"
+        )
+        tool_call = qwen3_coder.parse_tool_call(test_case, tools)
+        self.assertEqual(tool_call["name"], "search")
+        self.assertEqual(
+            tool_call["arguments"]["filters"],
+            {"category": "books", "in_stock": True},
+        )
+        self.assertEqual(tool_call["arguments"]["tags"], ["fiction", "new"])
+
+        # valid JSON (double-quoted) should still work
+        test_case = (
+            "<function=search>"
+            '<parameter=filters>{"category": "books"}</parameter>'
+            '<parameter=tags>["fiction", "new"]</parameter>'
+            "</function>"
+        )
+        tool_call = qwen3_coder.parse_tool_call(test_case, tools)
+        self.assertEqual(tool_call["arguments"]["filters"], {"category": "books"})
+        self.assertEqual(tool_call["arguments"]["tags"], ["fiction", "new"])
+
+    def test_gemma4(self):
+        # Nested object
+        test_case = 'call:configure{settings:{enabled:true,name:<|"|>test<|"|>}}'
+        tool_call = gemma4.parse_tool_call(test_case, None)
+        self.assertEqual(tool_call["name"], "configure")
+        self.assertEqual(
+            tool_call["arguments"],
+            {"settings": {"enabled": True, "name": "test"}},
+        )
+
+        # Array of strings
+        test_case = 'call:tag{items:[<|"|>foo<|"|>,<|"|>bar<|"|>]}'
+        tool_call = gemma4.parse_tool_call(test_case, None)
+        self.assertEqual(tool_call["name"], "tag")
+        self.assertEqual(tool_call["arguments"], {"items": ["foo", "bar"]})
+
+        # Mixed types
+        test_case = 'call:search{query:<|"|>hello world<|"|>,limit:10,verbose:false}'
+        tool_call = gemma4.parse_tool_call(test_case, None)
+        self.assertEqual(tool_call["name"], "search")
+        self.assertEqual(
+            tool_call["arguments"],
+            {"query": "hello world", "limit": 10, "verbose": False},
+        )
+
+        # Multiple tool calls in a single block (no delimiter between them)
+        test_case = (
+            'call:glob{pattern:<|"|>README*.md<|"|>}'
+            'call:glob{pattern:<|"|>CONTRIBUTING.md<|"|>}'
+        )
+        tool_calls = gemma4.parse_tool_call(test_case, None)
+        self.assertIsInstance(tool_calls, list)
+        self.assertEqual(len(tool_calls), 2)
+        self.assertEqual(tool_calls[0]["name"], "glob")
+        self.assertEqual(tool_calls[0]["arguments"], {"pattern": "README*.md"})
+        self.assertEqual(tool_calls[1]["name"], "glob")
+        self.assertEqual(tool_calls[1]["arguments"], {"pattern": "CONTRIBUTING.md"})
+
+        # Multiple tool calls with nested args
+        test_case = (
+            'call:search{query:<|"|>weather<|"|>,limit:5}'
+            'call:configure{settings:{enabled:true,name:<|"|>test<|"|>}}'
+        )
+        tool_calls = gemma4.parse_tool_call(test_case, None)
+        self.assertIsInstance(tool_calls, list)
+        self.assertEqual(len(tool_calls), 2)
+        self.assertEqual(tool_calls[0]["name"], "search")
+        self.assertEqual(
+            tool_calls[0]["arguments"],
+            {"query": "weather", "limit": 5},
+        )
+        self.assertEqual(tool_calls[1]["name"], "configure")
+        self.assertEqual(
+            tool_calls[1]["arguments"],
+            {"settings": {"enabled": True, "name": "test"}},
+        )
+
+        # Hyphenated function name (e.g. manim-video)
+        test_case = (
+            'call:manim-video{mode:<|"|>plan<|"|>,prompt:<|"|>explain KV caching<|"|>}'
+        )
+        tool_call = gemma4.parse_tool_call(test_case, None)
+        self.assertEqual(tool_call["name"], "manim-video")
+        self.assertEqual(
+            tool_call["arguments"],
+            {"mode": "plan", "prompt": "explain KV caching"},
+        )
+
+        # Braces inside a string argument (e.g. code snippets or markdown in content)
+        test_case = (
+            'call:skill_manage{action:<|"|>create<|"|>,'
+            'content:<|"|>use a dict like {key: value} in your code<|"|>}'
+        )
+        tool_call = gemma4.parse_tool_call(test_case, None)
+        self.assertEqual(tool_call["name"], "skill_manage")
+        self.assertEqual(tool_call["arguments"]["action"], "create")
+        self.assertIn("{", tool_call["arguments"]["content"])
+
     def test_kimi_k2(self):
         # Single tool call
         test_case = (
@@ -169,6 +313,22 @@ class TestToolParsing(unittest.TestCase):
         ]
         self.assertEqual(tool_calls, expected)
 
+    def test_minimax_m2(self):
+        test_case = (
+            '<invoke name="search">\n'
+            '<parameter name="query">weather</parameter>\n'
+            "</invoke>\n"
+            '<invoke name="read_file">\n'
+            '<parameter name="path">/tmp/test.txt</parameter>\n'
+            "</invoke>"
+        )
+        expected = [
+            {"name": "search", "arguments": {"query": "weather"}},
+            {"name": "read_file", "arguments": {"path": "/tmp/test.txt"}},
+        ]
+        tool_calls = minimax_m2.parse_tool_call(test_case, None)
+        self.assertEqual(expected, tool_calls)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_utils.py

@@ -3,6 +3,7 @@
 import os
 import tempfile
 import unittest
+from pathlib import Path
 
 import mlx.core as mx
 import mlx.nn as nn
@@ -123,6 +124,65 @@ class TestUtils(unittest.TestCase):
         self.assertEqual(model.custom_attribute, "This is a custom model")
         self.assertTrue(hasattr(model, "qwenWeights"))
 
+    def test_load_model_gemma4_with_per_layer_projection_quantization(self):
+        from mlx_lm.models import gemma4
+
+        args = gemma4.ModelArgs.from_dict(
+            {
+                "model_type": "gemma4",
+                "vocab_size": 32,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 32,
+                    "num_hidden_layers": 2,
+                    "intermediate_size": 64,
+                    "num_attention_heads": 2,
+                    "num_key_value_heads": 1,
+                    "num_global_key_value_heads": 1,
+                    "head_dim": 16,
+                    "global_head_dim": 16,
+                    "sliding_window": 8,
+                    "sliding_window_pattern": 1,
+                    "layer_types": ["full_attention", "full_attention"],
+                    "hidden_size_per_layer_input": 32,
+                    "vocab_size_per_layer_input": 32,
+                    "num_kv_shared_layers": 0,
+                    "tie_word_embeddings": True,
+                },
+            }
+        )
+        model = gemma4.Model(args)
+        model, config = utils.quantize_model(
+            model,
+            {
+                "model_type": "gemma4",
+                "vocab_size": args.vocab_size,
+                "text_config": args.text_config,
+            },
+            group_size=32,
+            bits=4,
+        )
+
+        config["quantization"]["language_model.model.per_layer_model_projection"] = {
+            "group_size": 32,
+            "bits": 4,
+        }
+
+        with tempfile.TemporaryDirectory(dir=self.test_dir) as mlx_path:
+            utils.save_model(mlx_path, model)
+            utils.save_config(config, os.path.join(mlx_path, "config.json"))
+
+            loaded, loaded_config = utils.load_model(Path(mlx_path))
+
+            self.assertIn(
+                "language_model.model.per_layer_model_projection",
+                loaded_config["quantization"],
+            )
+
+            logits = loaded(mx.array([[1, 2, 3]], dtype=mx.int32))
+            mx.eval(logits)
+            self.assertEqual(logits.shape, (1, 3, args.vocab_size))
+
 
 if __name__ == "__main__":
     unittest.main()