Replace recursion with an explicit stack

Add DP search across multiple quant options
Fix the formating
2025-07-06 16:31:49 -07:00 · 2025-07-06 16:31:47 -07:00 · 2025-07-06 16:11:38 -07:00 · 2025-07-06 15:54:58 -07:00 · 2025-07-04 19:18:31 -07:00 · 2025-07-02 16:34:07 -07:00
121 changed files with 12873 additions and 1677 deletions
@@ -0,0 +1,66 @@
+version: 2.1
+
+orbs:
+  apple: ml-explore/pr-approval@0.1.0
+ 
+jobs:
+  linux_build_and_test:
+    docker:
+      - image: cimg/python:3.9
+
+    steps:
+      - checkout
+      - run:
+          name: Run style checks
+          command: |
+            pip install pre-commit
+            pre-commit run --all
+            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
+
+  mlx_lm_build_and_test:
+    macos:
+      xcode: "15.2.0"
+    resource_class: macos.m1.large.gen1
+    steps:
+      - checkout
+      - run:
+          name: Install dependencies
+          command: |
+            brew install python@3.9
+            python3.9 -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
+            pip install unittest-xml-reporting
+            pip install -e ".[test]"
+      - run:
+          name: Run Python tests
+          command: |
+            source env/bin/activate
+            python -m xmlrunner discover -v tests -o test-results/
+      - store_test_results:
+          path: test-results
+
+workflows:
+  build_and_test:
+    when:
+      matches:
+        pattern: "^(?!pull/)[-\\w]+$"
+        value: << pipeline.git.branch >>
+    jobs:
+      - mlx_lm_build_and_test
+      - linux_build_and_test
+
+  prb:
+    when:
+      matches:
+        pattern: "^pull/\\d+(/head)?$"
+        value: << pipeline.git.branch >>
+    jobs:
+      - hold:
+          type: approval
+      - apple/authenticate:
+          context: pr-approval
+      - mlx_lm_build_and_test:
+          requires: [ hold ]
+      - linux_build_and_test:
+          requires: [ hold ]
@@ -0,0 +1,139 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Vim
+*.swp
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# IDE files
+.idea/
+.vscode/
+
+# .DS_Store files
+.DS_Store
@@ -0,0 +1,11 @@
+repos:
+-   repo: https://github.com/psf/black-pre-commit-mirror
+    rev: 25.1.0
+    hooks:
+    -   id: black
+-   repo: https://github.com/pycqa/isort
+    rev: 6.0.0
+    hooks:
+    -   id: isort
+        args:
+            - --profile=black
@@ -5,13 +5,8 @@ with a short description of your contribution(s) below. For example:

 - Jane Smith: Added the `foo` example.

-MLX Examples was developed with contributions from the following individuals:
+MLX LM was developed with contributions from the following individuals:

- Juarez Bochi: Added support for T5 models.
- Sarthak Yadav: Added the `cifar` and `speechcommands` examples.
 - Shunta Saito: Added support for PLaMo models.
- Gabrijel Boduljak: Implemented `CLIP`.
- Markus Enzweiler: Added the `cvae` examples.
- Prince Canuma: Helped add support for `Starcoder2` models.
- Shiyu Li: Added the `Segment Anything Model`.
- Gökdeniz Gülmez: Added support for `MiniCPM`, `Mamba` and support for `full-fine-tuning`.
+- Gökdeniz Gülmez: Added support for the following architectures: OpenBMB's `MiniCPM` and `MiniCPM3`, Kyutai's `Helium`, State-Space's`Mamba v1`, Z.ai & THUKEG's `GLM4`, Rednote `dots.llm1`, and Allenai's `OLMoE`; Added support for the following training algorithms: `full-fine-tuning`; Added support for the following other features: `Multiple Optimizers to choose for training`, and `reporting training metrics to WandB (Weights & Biases)`.
+- Prince Canuma: Helped add support for the following model architectures: HuggingFace's `Starcoder2`, Cohere's `Cohere (1 and 2)`, Alibaba Qwen's `Qwen (2, 3 and MoE)`, Microsoft's `Phi (3 and 3.5 MoE)`, `BitNet1.58`, Meta's `Llama (3 and 4)`, Google DeepMind's `Gemma 3`, and InterLM's `InternLM 2.5`.
@@ -1,11 +1,54 @@
 # Contributing to MLX LM 

+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+
+1. Fork and submit pull requests to the repo. 
+2. If you've added code that should be tested, add tests.
+3. Every PR should have passing tests and at least one review. 
+4. For code formatting install `pre-commit` using something like `pip install pre-commit` and run `pre-commit install`.
+   This should install hooks for running `black` and `clang-format` to ensure
+   consistent style for C++ and python code.
+ 
+   You can also run the formatters manually as follows on individual files:
+ 
+     ```bash
+     clang-format -i file.cpp
+     ```
+ 
+     ```bash
+     black file.py
+     ```
+
+     or,
+
+     ```bash
+     # single file
+     pre-commit run --files file1.py 
+
+     # specific files
+     pre-commit run --files file1.py file2.py
+     ```
+ 
+   or run `pre-commit run --all-files` to check all files in the repo.
+
+## Issues
+
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+## License
+
+By contributing to mlx-lm, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.
+
+## Adding New Models
+
 Below are some tips to port LLMs available on Hugging Face to MLX.

-Before starting checkout the [general contribution
-guidelines](https://github.com/ml-explore/mlx-examples/blob/main/CONTRIBUTING.md).
-
-Next, from this directory, do an editable install:
+From this directory, do an editable install:

 ```shell
 pip install -e .
@@ -17,7 +60,7 @@ Then check if the model has weights in the
 convert it.

 After that, add the model file to the
-[`mlx_lm/models`](https://github.com/ml-explore/mlx-examples/tree/main/llms/mlx_lm/models)
+[`mlx_lm/models`](https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/models)
 directory. You can see other examples there. We recommend starting from a model
 that is similar to the model you are porting.

@@ -35,12 +78,12 @@ To determine the model layer names, we suggest either:
  in the Hugging Face repo.

 To add LoRA support edit
-[`mlx_lm/tuner/utils.py`](https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/tuner/utils.py#L27-L60)
+[`mlx_lm/tuner/utils.py`](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/tuner/utils.py#L27-L60)

 Finally, add a test for the new modle type to the [model
-tests](https://github.com/ml-explore/mlx-examples/blob/main/llms/tests/test_models.py).
+tests](https://github.com/ml-explore/mlx-lm/blob/main/tests/test_models.py).

-From the `llms/` directory, you can run the tests with:
+You can run the tests with:

 ```shell
 python -m unittest discover tests/
@@ -1,2 +1,2 @@
-include mlx_lm/requirements.txt
+include requirements.txt
 recursive-include mlx_lm/ *.py
@@ -1,4 +1,17 @@
-## Generate Text with LLMs and MLX
+## MLX LM 
+
+MLX LM is a Python package for generating text and fine-tuning large language
+models on Apple silicon with MLX.
+
+Some key features include:
+
+* Integration with the Hugging Face Hub to easily use thousands of LLMs with a
+  single command. 
+* Support for quantizing and uploading models to the Hugging Face Hub.
+* [Low-rank and full model
+  fine-tuning](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/LORA.md)
+  with support for quantized models.
+* Distributed inference and fine-tuning with `mx.distributed`

 The easiest way to get started is to install the `mlx-lm` package:

@@ -14,18 +27,12 @@ pip install mlx-lm
 conda install -c conda-forge mlx-lm
 ```

-The `mlx-lm` package also has:
-
- [LoRA, QLoRA, and full fine-tuning](https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/LORA.md)
- [Merging models](https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/MERGE.md)
- [HTTP model serving](https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/SERVER.md)
-
 ### Quick Start

 To generate text with an LLM use:

 ```bash
-mlx_lm.generate --prompt "Hi!"
+mlx_lm.generate --prompt "How tall is Mt Everest?"
 ```

 To chat with an LLM use:
@@ -58,7 +65,7 @@ prompt = "Write a story about Einstein"

 messages = [{"role": "user", "content": prompt}]
 prompt = tokenizer.apply_chat_template(
-    messages, tokenize=False, add_generation_prompt=True
+    messages, add_generation_prompt=True
 )

 text = generate(model, tokenizer, prompt=prompt, verbose=True)
@@ -71,7 +78,7 @@ To see a description of all the arguments you can do:
 ```

 Check out the [generation
-example](https://github.com/ml-explore/mlx-examples/tree/main/llms/mlx_lm/examples/generate_response.py)
+example](https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/examples/generate_response.py)
 to see how to use the API in more detail.

 The `mlx-lm` package also comes with functionality to quantize and optionally
@@ -115,7 +122,7 @@ prompt = "Write a story about Einstein"

 messages = [{"role": "user", "content": prompt}]
 prompt = tokenizer.apply_chat_template(
-    messages, tokenize=False, add_generation_prompt=True
+    messages, add_generation_prompt=True
 )

 for response in stream_generate(model, tokenizer, prompt, max_tokens=512):
@@ -123,6 +130,18 @@ for response in stream_generate(model, tokenizer, prompt, max_tokens=512):
 print()
 ```

+#### Sampling
+
+The `generate` and `stream_generate` functions accept `sampler` and
+`logits_processors` keyword arguments. A sampler is any callable which accepts
+a possibly batched logits array and returns an array of sampled tokens.  The
+`logits_processors` must be a list of callables which take the token history
+and current logits as input and return the processed logits. The logits
+processors are applied in order.
+
+Some standard sampling functions and logits processors are provided in
+`mlx_lm.sample_utils`.
+
 ### Command Line

 You can also use `mlx-lm` from the command line with:
@@ -164,7 +183,7 @@ mlx_lm.convert \
 ```

 Models can also be converted and quantized directly in the
-[mlx-my-repo]https://huggingface.co/spaces/mlx-community/mlx-my-repo) Hugging
+[mlx-my-repo](https://huggingface.co/spaces/mlx-community/mlx-my-repo) Hugging
 Face Space.

 ### Long Prompts and Generations 
@@ -201,17 +220,17 @@ The cached prompt is treated as a prefix to the supplied prompt. Also notice
 when using a cached prompt, the model to use is read from the cache and need
 not be supplied explicitly.

-Prompt caching can also be used in the Python API in order to to avoid
+Prompt caching can also be used in the Python API in order to avoid
 recomputing the prompt. This is useful in multi-turn dialogues or across
 requests that use the same context. See the
-[example](https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/examples/chat.py)
+[example](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/examples/chat.py)
 for more usage details.

 ### Supported Models

 `mlx-lm` supports thousands of Hugging Face format LLMs. If the model you want to
 run is not supported, file an
-[issue](https://github.com/ml-explore/mlx-examples/issues/new) or better yet,
+[issue](https://github.com/ml-explore/mlx-lm/issues/new) or better yet,
 submit a pull request.

 Here are a few examples of Hugging Face models that work with this example:
@@ -0,0 +1,149 @@
+# Learned Quantization 
+
+To reduce the quality loss from quantization MLX LM has several options:
+
+- Distilled Weight Quantization (DWQ)
+- Activation-aware Weight Quantization (AWQ)[^1]
+- Dynamic quantization
+
+All methods use calibration data to tune parameters or hyper-parameters of the
+model. DWQ fine-tunes non-quantized parameters (including quantization scales
+and biases) using the non-quantized model as a teacher. AWQ scales and clips
+the weights prior to quantization. Dynamic quantization estimates the
+sensitivity of a model's outputs to each layer and uses a higher precision for
+layers which have higher sensitivity.
+
+Dynamic quantization is the fastest to run. DWQ takes longer but typically
+yields better results. You can also cascade methods. For example a dynamically
+quantized model can be further refined with DWQ.
+
+To get started, first install the requirements:
+
+```
+pip install mlx-lm[quant]
+```
+
+### DWQ
+
+Use `mlx_lm.dwq` to run DWQ on a given model. For example:
+
+```bash
+mlx_lm.dwq --model mistralai/Mistral-7B-Instruct-v0.3
+```
+
+Some important options, along with their default values are:
+
+- `--mlx-path mlx_model`: The location to save the DWQ model.
+- `--bits 4`: Precision of the quantization.
+- `--num-samples 1024`: Number of samples to use. Using more samples can lead to
+  better results but takes longer.
+- `--batch-size 8`: Use a smaller batch size to reduce the memory footprint.
+
+For a full list of options run:
+
+```bash
+mlx_lm.dwq --help
+```
+
+#### Tips
+
+- DWQ works best distilling to lower precision, anywhere from 2-bit to 4-bit
+  models.
+- Distilling 16-bit precision to 8-bit and even 6-bit often doesn't work well.
+  The loss starts out so low that it's difficult to reduce further.
+- Decreasing the quantization group size (e.g. `--group-size 32`) doubles the
+  number of tunable parameters and can work much better.
+- If the loss is oscillating and not going down consistently, try reducing the
+  learning rate. If it is decreasing but slowly, try increasing the learning
+  rate.
+- As a rule of thumb, lower precision can benefit from a higher learning rate
+  since the loss starts out higher. Conversely, higher precision needs a lower
+  learning rate.
+
+
+#### Memory Use
+
+A few options to reduce memory use for DWQ:
+
+- Distill from an 8-bit model instead of a 16-bit model. The 8-bit
+  models are usually as good as 16-bit precision models.
+- Use a shorter maximum sequence length. The default is 2048. Using
+  `--max-seq-length 512` reduces the memory and still gets good results.
+- Use a smaller batch size, e.g. `--batch-size 1`
+
+### Dynamic Quantization
+
+Use `mlx_lm.dynamic_quant` to generate a dynamic quantization of given model.
+For example:
+
+```bash
+mlx_lm.dynamic_quant --model mistralai/Mistral-7B-Instruct-v0.3
+```
+
+The script will estimate the sensitivity for each quantizable layer in the
+model. It will then quantize the model using higher precision (default 5 bits)
+for the more sensitive layers and lower precision (default 4 bits) for the
+rest. The script also saves a JSON file with each layer's sensitivities which
+saves needing to compute it multiple times to make different precision quants
+of the same model.
+
+Some important options are:
+
+- `--target-bpw`: The target bits-per-weight. For a given set of quantization
+  parameters only certain ranges are possible. For example, with the default
+  parameters a BPW in the range `[4.5, 5.5]` is achievable.
+- `--sensitivities`: A path to a precomputed sensitivities file.
+- `--low-bits`: The number of bits to use for the less sensitive layers.
+- `--high-bits`: The number of bits to use for the more sensitive layers.
+
+### AWQ 
+
+Use `mlx_lm.awq` to run AWQ on a given model. For example:
+
+```bash
+mlx_lm.awq --model mistralai/Mistral-7B-Instruct-v0.3
+```
+
+The script can take anywhere form a few minutes to several hours to run
+depending on the model size and the number of samples.
+
+Some important options, along with their default values, are:
+
+- `--mlx-path mlx_model`: The location to save the AWQ model.
+- `--bits 4`: Precision of the quantization.
+- `--num-samples 32`: Number of samples to use. Using more samples can lead to
+  better results but takes longer.
+- `--n-grid 10`: The granularity of the AWQ search. A larger grid can lead to
+  better results but takes longer.
+
+For a full list of options run:
+
+```bash
+mlx_lm.awq --help
+```
+
+### Evaluate
+
+Once the training script finishes, you can evaluate the quality of the model
+on downstream tasks using `mlx_lm.evaluate`. For example:
+
+```bash
+mlx_lm.evaluate \
+    --model mlx_model \
+    --tasks winogrande boolq arc_challenge arc_easy hellaswag openbookqa piqa social_iqa                     
+```
+
+### Upload to Hugging Face
+
+Use `mlx_lm.upload` to upload the quantized model to the Hugging Face Hub. For
+example:
+
+```bash
+mlx_lm.upload \
+    --path mlx_model \
+    --upload-repo mlx-community/Mistral-7B-Instruct-v0.3-3bit-DWQ
+```
+
+[^1]: Refer to the [paper](https://arxiv.org/abs/2306.00978)
+and [github repository](https://github.com/mit-han-lab/llm-awq) for more
+details.
@@ -76,6 +76,19 @@ You can specify the output location with `--adapter-path`.
 You can resume fine-tuning with an existing adapter with
 `--resume-adapter-file <path_to_adapters.safetensors>`.

+#### Logging
+
+You can log training metrics to Weights & Biases by passing a project name with
+the `--wandb` flag. Make sure to install wandb with `pip install wandb`.
+
+#### Prompt Masking
+
+The default training computes a loss for every token in the sample. You can
+ignore the prompt and compute loss for just the completion by passing
+`--mask-prompt`. Note this is only supported for `chat` and `completion`
+datasets. For `chat` datasets the final message in the message list is
+considered the completion. See the [dataset section](#Data) for more details. 
+
 ### Evaluate

 To compute test set perplexity use:
@@ -241,14 +254,25 @@ Refer to the documentation for the model you are fine-tuning for more details.
 {"prompt": "What is the capital of France?", "completion": "Paris."}
 ```

+For the `completions` data format, a different key can be used for the prompt
+and completion by specifying the following in the YAML config:
+
+```yaml
+prompt_feature: "input"
+completion_feature: "output"
+```
+
+Here, `"input"` is the expected key instead of the default `"prompt"`, and
+`"output"` is the expected key instead of `"completion"`. 
+
 `text`:

 ```jsonl
 {"text": "This is an example for the model."}
 ```

-Note, the format is automatically determined by the dataset. Note also, keys in
-each line not expected by the loader will be ignored.
+Note, the format is automatically determined by the dataset. Note also, keys
+in each line not expected by the loader will be ignored.

 > [!NOTE]
 > Each example in the datasets must be on a single line. Do not put more than
@@ -270,20 +294,36 @@ Otherwise, provide a mapping of keys in the dataset to the features MLX LM
 expects. Use a YAML config to specify the Hugging Face dataset arguments. For
 example:

-```
+```yaml
 hf_dataset:
-  name: "billsum"
+  path: "billsum"
  prompt_feature: "text"
  completion_feature: "summary"
 ```

 - Use `prompt_feature` and `completion_feature` to specify keys for a
  `completions` dataset. Use `text_feature` to specify the key for a `text`
-  dataset. 
+  dataset. Use `chat_feature` to specify the key for a chat dataset.

 - To specify the train, valid, or test splits, set the corresponding
  `{train,valid,test}_split` argument. 

+You can specify a list of Hugging Face datasets with a list of records each
+with the same structure as above. For example:
+
+```yaml
+hf_dataset:
+  - path: "Open-Orca/OpenOrca"
+    train_split: "train[:90%]"
+    valid_split: "train[-10%:]"
+    prompt_feature: "question"
+    completion_feature: "response"
+  - path: "trl-lib/ultrafeedback_binarized"
+    train_split: "train[:90%]"
+    valid_split: "train[-10%:]"
+    chat_feature: "chosen"
+```
+
 - Arguments specified in `config` will be passed as keyword arguments to
  [`datasets.load_dataset`](https://huggingface.co/docs/datasets/v2.20.0/en/package_reference/loading_methods#datasets.load_dataset).

@@ -344,7 +384,7 @@ mlx_lm.lora \
    --train \
    --batch-size 1 \
    --num-layers 4 \
-    --data wikisql
+    --data  mlx-community/wikisql
 ```

 The above command on an M1 Max with 32 GB runs at about 250
@@ -1,50 +0,0 @@
-# Model Merging
-
-You can use `mlx-lm` to merge models and upload them to the Hugging
-Face hub or save them locally for LoRA fine tuning.
-
-The main command is `mlx_lm.merge`:
-
-```shell
-mlx_lm.merge --config config.yaml 
-```
-
-The merged model will be saved by default in `mlx_merged_model`. To see a
-full list of options run:
-
-```shell
-mlx_lm.merge --help
-```
-
-Here is an example `config.yaml`:
-
-```yaml
-models:
-  - OpenPipe/mistral-ft-optimized-1218
-  - mlabonne/NeuralHermes-2.5-Mistral-7B
-method: slerp
-parameters:
-  t:
-    - filter: self_attn
-      value: [0, 0.5, 0.3, 0.7, 1]
-    - filter: mlp
-      value: [1, 0.5, 0.7, 0.3, 0]
-    - value: 0.5
-```
-
-The `models` field is a list of Hugging Face repo ids. The first model in the
-list is treated as the base model into which the remaining models are merged.
-
-The `method` field is the merging method. Right now `slerp` is the only
-supported method.
-
-The `parameters` are the corresponding parameters for the given `method`.
-Each parameter is a list with `filter` determining which layer the parameter
-applies to and `value` determining the actual value used. The last item in
-the list without a `filter` field is the default.
-
-If `value` is a list, it specifies the start and end values for the
-corresponding segment of blocks. In the example above, the models have 32
-blocks. For blocks 1-8, the layers with `self_attn` in the name will use the
-values `np.linspace(0, 0.5, 8)`, the same layers in the next 8 blocks (9-16)
-will use `np.linspace(0.5, 0.3, 8)`, and so on.
@@ -54,18 +54,24 @@ curl localhost:8080/v1/chat/completions \
  sequences of tokens on which the generation should stop.

 - `max_tokens`: (Optional) An integer specifying the maximum number of tokens
-  to generate. Defaults to `100`.
+  to generate. Defaults to `512`.

 - `stream`: (Optional) A boolean indicating if the response should be
  streamed. If true, responses are sent as they are generated. Defaults to
  false.

 - `temperature`: (Optional) A float specifying the sampling temperature.
-  Defaults to `1.0`.
+  Defaults to `0.0`.

 - `top_p`: (Optional) A float specifying the nucleus sampling parameter.
  Defaults to `1.0`.

+- `top_k`: (Optional) An integer specifying the top-k sampling parameter.
+  Defaults to `0` (disabled).
+
+- `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`.

@@ -86,6 +92,12 @@ curl localhost:8080/v1/chat/completions \
 - `adapters`: (Optional) A string path to low-rank adapters. The path must be
  relative to the directory the server was started in.

+- `draft_model`: (Optional) Specifies a smaller model to use for speculative
+  decoding. Set to `null` to unload.
+
+- `num_draft_tokens`: (Optional) The number of draft tokens the draft model
+  should predict at once. Defaults to `3`.
+
 ### Response Fields

 - `id`: A unique identifier for the chat.
@@ -6,4 +6,6 @@ from ._version import __version__

 os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"

-from .utils import convert, generate, load, stream_generate
+from .convert import convert
+from .generate import generate, stream_generate
+from .utils import load
@@ -0,0 +1,28 @@
+# Copyright © 2025 Apple Inc.
+
+import importlib
+import sys
+
+if __name__ == "__main__":
+    subcommands = {
+        "quant.awq",
+        "quant.dwq",
+        "quant.dynamic_quant",
+        "cache_prompt",
+        "chat",
+        "convert",
+        "evaluate",
+        "fuse",
+        "generate",
+        "lora",
+        "server",
+        "manage",
+        "upload",
+    }
+    if len(sys.argv) < 2:
+        raise ValueError(f"CLI requires a subcommand in {subcommands}")
+    subcommand = sys.argv.pop(1)
+    if subcommand not in subcommands:
+        raise ValueError(f"CLI requires a subcommand in {subcommands}")
+    submodule = importlib.import_module(f"mlx_lm.{subcommand}")
+    submodule.main()
@@ -1,3 +1,3 @@
 # Copyright © 2023-2024 Apple Inc.

-__version__ = "0.20.4"
+__version__ = "0.25.3"
@@ -7,8 +7,9 @@ import time

 import mlx.core as mx

+from .generate import generate_step
 from .models.cache import make_prompt_cache, save_prompt_cache
-from .utils import generate_step, load
+from .utils import load

 DEFAULT_QUANTIZED_KV_START = 5000

@@ -110,29 +111,17 @@ def main():
        if tokenizer.chat_template is None:
            tokenizer.chat_template = tokenizer.default_chat_template

-    if not args.ignore_chat_template and (
-        hasattr(tokenizer, "apply_chat_template")
-        and tokenizer.chat_template is not None
-    ):
+    if not args.ignore_chat_template and tokenizer.chat_template is not None:
        messages = [{"role": "user", "content": args.prompt}]
        prompt = tokenizer.apply_chat_template(
-            messages, tokenize=False, add_generation_prompt=True
+            messages, add_generation_prompt=False, continue_final_message=True
        )

-        # Treat the prompt as a prefix assuming that the suffix will be
-        # provided at generation time.
-        test_prompt = tokenizer.apply_chat_template(
-            [{"role": "user", "content": "<query>"}],
-            tokenize=False,
-            add_generation_prompt=True,
-        )
-        n = len(test_prompt) - test_prompt.index("<query>") - len("<query>")
-        prompt = prompt[:-n]
    else:
-        prompt = args.prompt
+        prompt = tokenizer.encode(args.prompt)

    cache = make_prompt_cache(model, args.max_kv_size)
-    y = mx.array(tokenizer.encode(prompt))
+    y = mx.array(prompt)

    # Process the prompt
    start = time.time()
@@ -159,15 +148,19 @@ def main():
        pass

    print()
-    print(f"Peak memory: {mx.metal.get_peak_memory() / 1e9:.3f} GB")
+    print(f"Peak memory: {mx.get_peak_memory() / 1e9:.3f} GB")

    print("Saving...")
    metadata = {}
    metadata["model"] = args.model
-    metadata["chat_template"] = tokenizer.chat_template
+    metadata["chat_template"] = json.dumps(tokenizer.chat_template)
    metadata["tokenizer_config"] = json.dumps(tokenizer_config)
    save_prompt_cache(args.prompt_cache_file, cache, metadata)


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.cache_prompt...` directly is deprecated."
+        " Use `mlx_lm.cache_prompt...` or `python -m mlx_lm cache_prompt ...` instead."
+    )
    main()
@@ -1,17 +1,19 @@
 # Copyright © 2023-2024 Apple Inc.

 import argparse
-import json

 import mlx.core as mx

+from .generate import stream_generate
 from .models.cache import make_prompt_cache
 from .sample_utils import make_sampler
-from .utils import load, stream_generate
+from .utils import load

 DEFAULT_TEMP = 0.0
 DEFAULT_TOP_P = 1.0
-DEFAULT_SEED = 0
+DEFAULT_XTC_PROBABILITY = 0.0
+DEFAULT_XTC_THRESHOLD = 0.0
+DEFAULT_SEED = None
 DEFAULT_MAX_TOKENS = 256
 DEFAULT_MODEL = "mlx-community/Llama-3.2-3B-Instruct-4bit"

@@ -36,7 +38,24 @@ def setup_arg_parser():
    parser.add_argument(
        "--top-p", type=float, default=DEFAULT_TOP_P, help="Sampling top-p"
    )
-    parser.add_argument("--seed", type=int, default=DEFAULT_SEED, help="PRNG seed")
+    parser.add_argument(
+        "--xtc-probability",
+        type=float,
+        default=DEFAULT_XTC_PROBABILITY,
+        help="Probability of XTC sampling to happen each next token",
+    )
+    parser.add_argument(
+        "--xtc-threshold",
+        type=float,
+        default=0.0,
+        help="Thresold the probs of each next token candidate to be sampled by XTC",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=DEFAULT_SEED,
+        help="PRNG seed",
+    )
    parser.add_argument(
        "--max-kv-size",
        type=int,
@@ -57,7 +76,8 @@ def main():
    parser = setup_arg_parser()
    args = parser.parse_args()

-    mx.random.seed(args.seed)
+    if args.seed is not None:
+        mx.random.seed(args.seed)

    model, tokenizer = load(
        args.model,
@@ -65,22 +85,41 @@ def main():
        tokenizer_config={"trust_remote_code": True},
    )

-    print(f"[INFO] Starting chat session with {args.model}. To exit, enter 'q'.")
+    def print_help():
+        print("The command list:")
+        print("- 'q' to exit")
+        print("- 'r' to reset the chat")
+        print("- 'h' to display these commands")
+
+    print(f"[INFO] Starting chat session with {args.model}.")
+    print_help()
    prompt_cache = make_prompt_cache(model, args.max_kv_size)
    while True:
        query = input(">> ")
        if query == "q":
            break
+        if query == "r":
+            prompt_cache = make_prompt_cache(model, args.max_kv_size)
+            continue
+        if query == "h":
+            print_help()
+            continue
        messages = [{"role": "user", "content": query}]
-        prompt = tokenizer.apply_chat_template(
-            messages, tokenize=False, add_generation_prompt=True
-        )
+        prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
        for response in stream_generate(
            model,
            tokenizer,
            prompt,
            max_tokens=args.max_tokens,
-            sampler=make_sampler(args.temp, args.top_p),
+            sampler=make_sampler(
+                args.temp,
+                args.top_p,
+                xtc_threshold=args.xtc_threshold,
+                xtc_probability=args.xtc_probability,
+                xtc_special_tokens=(
+                    tokenizer.encode("\n") + list(tokenizer.eos_token_ids)
+                ),
+            ),
            prompt_cache=prompt_cache,
        ):
            print(response.text, flush=True, end="")
@@ -88,4 +127,8 @@ def main():


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.chat...` directly is deprecated."
+        " Use `mlx_lm.chat...` or `python -m mlx_lm chat ...` instead."
+    )
    main()
@@ -1,8 +1,174 @@
 # Copyright © 2023-2024 Apple Inc.

 import argparse
+from pathlib import Path
+from typing import Callable, Optional, Union

-from .utils import convert
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_map_with_path
+
+from .utils import (
+    dequantize_model,
+    fetch_from_hub,
+    get_model_path,
+    quantize_model,
+    save,
+    upload_to_hub,
+)
+
+
+def mixed_quant_predicate_builder(
+    recipe: str, model: nn.Module
+) -> Callable[[str, nn.Module, dict], Union[bool, dict]]:
+
+    high_bits = 6
+    group_size = 64
+
+    if recipe == "mixed_2_6":
+        low_bits = 2
+    elif recipe == "mixed_3_4":
+        low_bits = 3
+        high_bits = 4
+    elif recipe == "mixed_3_6":
+        low_bits = 3
+    elif recipe == "mixed_4_6":
+        low_bits = 4
+    else:
+        raise ValueError("Invalid quant recipe {recipe}")
+
+    down_keys = [k for k, _ in model.named_modules() if "down_proj" in k]
+    if len(down_keys) == 0:
+        raise ValueError("Model does not have expected keys for mixed quant.")
+
+    # Look for the layer index location in the path:
+    for layer_location, k in enumerate(down_keys[0].split(".")):
+        if k.isdigit():
+            break
+    num_layers = len(model.layers)
+
+    def mixed_quant_predicate(
+        path: str,
+        module: nn.Module,
+        config: dict,
+    ) -> Union[bool, dict]:
+        """Implements mixed quantization predicates with similar choices to, for example, llama.cpp's Q4_K_M.
+        Ref: https://github.com/ggerganov/llama.cpp/blob/917786f43d0f29b7c77a0c56767c0fa4df68b1c5/src/llama.cpp#L5265
+        By Alex Barron: https://gist.github.com/barronalex/84addb8078be21969f1690c1454855f3
+        """
+
+        if not hasattr(module, "to_quantized"):
+            return False
+        if module.weight.shape[1] % group_size != 0:
+            return False
+
+        index = (
+            int(path.split(".")[layer_location])
+            if len(path.split(".")) > layer_location
+            else 0
+        )
+        use_more_bits = (
+            index < num_layers // 8
+            or index >= 7 * num_layers // 8
+            or (index - num_layers // 8) % 3 == 2
+        )
+        if "v_proj" in path and use_more_bits:
+            return {"group_size": group_size, "bits": high_bits}
+        if "down_proj" in path and use_more_bits:
+            return {"group_size": group_size, "bits": high_bits}
+        if "lm_head" in path:
+            return {"group_size": group_size, "bits": high_bits}
+
+        return {"group_size": group_size, "bits": low_bits}
+
+    return mixed_quant_predicate
+
+
+QUANT_RECIPES = ["mixed_2_6", "mixed_3_4", "mixed_3_6", "mixed_4_6"]
+
+MODEL_CONVERSION_DTYPES = ["float16", "bfloat16", "float32"]
+
+
+def convert(
+    hf_path: str,
+    mlx_path: str = "mlx_model",
+    quantize: bool = False,
+    q_group_size: int = 64,
+    q_bits: int = 4,
+    dtype: Optional[str] = None,
+    upload_repo: str = None,
+    revision: Optional[str] = None,
+    dequantize: bool = False,
+    quant_predicate: Optional[
+        Union[Callable[[str, nn.Module, dict], Union[bool, dict]], str]
+    ] = None,
+):
+    # Check the save path is empty
+    if isinstance(mlx_path, str):
+        mlx_path = Path(mlx_path)
+
+    if mlx_path.exists():
+        raise ValueError(
+            f"Cannot save to the path {mlx_path} as it already exists."
+            " Please delete the file/directory or specify a new path to save to."
+        )
+
+    print("[INFO] Loading")
+    model_path, hf_path = get_model_path(hf_path, revision=revision)
+    model, config, tokenizer = fetch_from_hub(model_path, lazy=True)
+
+    def base_quant_predicate(path, module, config):
+        if not hasattr(module, "to_quantized"):
+            return False
+        if module.weight.shape[1] % q_group_size != 0:
+            return False
+        return True
+
+    if isinstance(quant_predicate, str):
+        quant_predicate = mixed_quant_predicate_builder(quant_predicate, model)
+    quant_predicate = quant_predicate or base_quant_predicate
+
+    if dtype is None:
+        dtype = config.get("torch_dtype", None)
+    if dtype in MODEL_CONVERSION_DTYPES:
+        print("[INFO] Using dtype:", dtype)
+        dtype = getattr(mx, dtype)
+        cast_predicate = getattr(model, "cast_predicate", lambda _: True)
+
+        def set_dtype(k, v):
+            if cast_predicate(k) and mx.issubdtype(v.dtype, mx.floating):
+                return v.astype(dtype)
+            else:
+                return v
+
+        model.update(tree_map_with_path(set_dtype, model.parameters()))
+
+    if quantize and dequantize:
+        raise ValueError("Choose either quantize or dequantize, not both.")
+
+    if quantize:
+        print("[INFO] Quantizing")
+        model, config = quantize_model(
+            model, config, q_group_size, q_bits, quant_predicate=quant_predicate
+        )
+
+    if dequantize:
+        print("[INFO] Dequantizing")
+        config.pop("quantization", None)
+        config.pop("quantization_config", None)
+        model = dequantize_model(model)
+
+    save(
+        mlx_path,
+        model_path,
+        model,
+        tokenizer,
+        config,
+        hf_repo=hf_path,
+    )
+
+    if upload_repo is not None:
+        upload_to_hub(mlx_path, upload_repo)


 def configure_parser() -> argparse.ArgumentParser:
@@ -30,11 +196,18 @@ def configure_parser() -> argparse.ArgumentParser:
        "--q-bits", help="Bits per weight for quantization.", type=int, default=4
    )
    parser.add_argument(
-        "--dtype",
-        help="Type to save the non-quantized parameters.",
+        "--quant-predicate",
+        help=f"Mixed-bit quantization recipe.",
+        choices=QUANT_RECIPES,
        type=str,
-        choices=["float16", "bfloat16", "float32"],
-        default="float16",
+        required=False,
+    )
+    parser.add_argument(
+        "--dtype",
+        help="Type to save the non-quantized parameters. Defaults to config.json's `torch_dtype` or the current model weights dtype.",
+        type=str,
+        choices=MODEL_CONVERSION_DTYPES,
+        default=None,
    )
    parser.add_argument(
        "--upload-repo",
@@ -59,4 +232,8 @@ def main():


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.convert ...` directly is deprecated."
+        " Use `mlx_lm.convert ...` or `python -m mlx_lm convert ...` instead."
+    )
    main()
@@ -1,12 +1,18 @@
-# Adapted from a PyTorch implementation by David Grangier
+# Copyright © 2024 Apple Inc.
+
+"""
+Adapted from a PyTorch implementation by David Grangier
+"""

 import argparse
+import collections
+import copy
 import json
 import logging
 import os
 from importlib.metadata import version
 from pathlib import Path
-from typing import Optional
+from typing import Any, Optional

 import lm_eval
 import mlx.core as mx
@@ -16,19 +22,10 @@ from lm_eval.api.model import LM
 from lm_eval.api.registry import register_model
 from tqdm import tqdm

+from .generate import stream_generate
+from .models.base import create_causal_mask
 from .models.cache import make_prompt_cache
-from .utils import load, stream_generate
-
-PAD = 0
-
-
-def _len_longest_common_prefix(a, b):
-    l = 0
-    for item_a, item_b in zip(a, b):
-        if item_a != item_b:
-            break
-        l += 1
-    return l
+from .utils import common_prefix_len, load


 def _rstrip_until(s, untils):
@@ -39,71 +36,85 @@ def _rstrip_until(s, untils):
    return s[: min(f)]


-def _pad_inputs(
-    inputs,
-    maxlen,
-    genlen=0,
-    pad_left=False,
-    pad_multiple=32,
-    truncate=False,
-):
-    # pad the prompts to the left with at least genlen tokens.
-    actual_maxlen = max(len(p) for p in inputs) + genlen
-    if actual_maxlen > maxlen:
-        if not truncate:
-            raise ValueError("Inputs are too long.")
-        else:  # drop begining
-            actual_maxlen = maxlen
-            inputs = [p[max(0, len(p) - maxlen) :] for p in inputs]
-    if pad_multiple > 0:
-        maxlen = (actual_maxlen + pad_multiple - 1) // pad_multiple
-        maxlen *= pad_multiple
-    assert PAD == 0
-    lr = np.array((1, 0) if pad_left else (0, 1))
-    return np.stack(
-        [np.pad(np.array(x, np.int32), lr * (maxlen - len(x))) for x in inputs],
+def _pad_inputs(inputs):
+    lengths = np.array([len(x) for x in inputs])
+    maxlen = lengths.max()
+    padded = np.stack(
+        [np.pad(x, (0, maxlen - len(x))) for x in inputs],
        axis=0,
    )
+    return mx.array(padded), mx.array(lengths)
+
+
+def chat_template_fn(**extra_kwargs):
+    def apply_chat_template(self, chat_history, add_generation_prompt=True) -> str:
+        return self.tokenizer.apply_chat_template(
+            chat_history,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            continue_final_message=not add_generation_prompt,
+            **extra_kwargs,
+        )
+
+    return apply_chat_template


@register_model("mlxlm")
 class MLXLM(LM):
+
+    tokenizer_name = lm_eval.models.huggingface.HFLM.tokenizer_name
+    apply_chat_template = chat_template_fn()
+
    def __init__(
        self,
        path_or_hf_repo: str,
-        batch_size: int = 16,
        max_tokens: Optional[int] = None,
+        use_chat_template: Optional[bool] = None,
    ) -> None:
        super().__init__()
-        self._batch_size = batch_size
-        self._model, self._tokenizer = load(path_or_hf_repo)
-        self._max_tokens = max_tokens or self._tokenizer.model_max_length
+        self._model, self.tokenizer = load(path_or_hf_repo)
+        self._max_tokens = max_tokens or self.tokenizer.model_max_length
+        self._batch_size = 8
+        self.use_chat_template = use_chat_template
+        if use_chat_template is None:
+            self.use_chat_template = self.tokenizer.chat_template is not None

-    def _score_fn(self, inputs, tokenize=True, step_size=32):
-        if tokenize:
-            inputs = self._tokenizer.encode(inputs)
-        inputs = _pad_inputs(inputs, self._max_tokens, truncate=False)
-        inputs = mx.array(inputs)
+    def _process_prompt(self, prompt, step_size: int = 2048):
+        prompt = mx.array(prompt)[None]
+        cache = make_prompt_cache(self._model)
+        for i in range(0, prompt.shape[1], step_size):
+            logits = self._model(prompt[:, i : i + step_size], cache=cache)
+            mx.eval([c.state for c in cache])
+            mx.clear_cache()
+        logprobs = nn.log_softmax(logits[:, -1, :].astype(mx.float32))
+        return logprobs, cache
+
+    def _score_fn(self, inputs, cache: Optional[Any] = None, step_size: int = 2048):
+        inputs, lengths = _pad_inputs(inputs)
        inputs, targets = inputs[..., :-1], inputs[..., 1:]

-        cache = make_prompt_cache(self._model)
-
-        mask = targets != PAD
+        cache = cache or make_prompt_cache(self._model)
+        lengths += cache[0].offset

        scores, is_greedy = [], []
        for i in range(0, inputs.shape[1], step_size):
-            logits = self._model(inputs[:, i : i + step_size], cache=cache)
+            inp = inputs[:, i : i + step_size]
+            T = inp.shape[1]

+            offset = cache[0].offset
+            mask = create_causal_mask(T, offset, lengths=lengths)
+
+            logits = self._model(inp, cache=cache, mask=mask)
            log_probs = nn.log_softmax(logits.astype(mx.float32))
+
            score = mx.take_along_axis(
                log_probs, targets[:, i : i + step_size, mx.newaxis], axis=-1
            )[..., 0]
-            ig = mask[:, i : i + step_size] * (
-                targets[:, i : i + step_size] == mx.argmax(logits, axis=-1)
-            )
+            ig = targets[:, i : i + step_size] == mx.argmax(logits, axis=-1)
+            ig = mx.where(mx.arange(T) + offset < lengths[:, None], ig, False)

            mx.eval(score, ig)
-            mx.metal.clear_cache()
+            mx.clear_cache()

            is_greedy.append(ig)
            scores.append(score)
@@ -111,41 +122,15 @@ class MLXLM(LM):
        scores = mx.concatenate(scores, axis=1)
        is_greedy = mx.concatenate(is_greedy, axis=1)

-        return scores, mask.sum(axis=-1), is_greedy
-
-    def _loglikelihood(self, texts, score_spans=None, tokenize=True):
-        # sort by length to get batches with little padding.
-        sorted_indices = sorted(range(len(texts)), key=lambda i: -len(texts[i]))
-        sorted_inputs = [texts[sorted_indices[i]] for i in range(len(texts))]
-        sorted_spans = None
-        if score_spans is not None:
-            sorted_spans = [score_spans[sorted_indices[i]] for i in range(len(texts))]
-
-        results = []
-        for i in tqdm(range(0, len(sorted_inputs), self._batch_size)):
-            batch = sorted_inputs[i : i + self._batch_size]
-            scores, length, is_greedy = self._score_fn(batch, tokenize=tokenize)
-            for j in range(len(batch)):
-                if sorted_spans is None:  # full sequence score
-                    mask = mx.arange(scores[j].shape[-1]) < length
-                    score = (scores[j].astype(mx.float32) * mask).sum(axis=-1)
-                    ig = (is_greedy[j].astype(mx.int32) * mask).sum(axis=-1)
-                else:  # subsequence score
-                    start, end = sorted_spans[i + j]
-                    score = scores[j][start:end].astype(mx.float32).sum()
-                    ig = is_greedy[j][start:end].astype(mx.int32).sum()
-                    length = end - start
-
-                results.append((score.item(), ig.item(), length))
-
-        # reorder the outputs
-        inv_sort = np.argsort(sorted_indices)
-        results = [results[inv_sort[i]] for i in range(len(results))]
-
-        return results
+        return scores, lengths, is_greedy

    def _tokenize(self, texts):
-        return [tuple(self._tokenizer.encode(t)) for t in texts]
+        return [
+            tuple(
+                self.tokenizer.encode(t, add_special_tokens=not self.use_chat_template)
+            )
+            for t in texts
+        ]

    def loglikelihood(self, requests) -> list[tuple[float, bool]]:
        """Compute log-likelihood of generating a continuation from a context.
@@ -169,39 +154,65 @@ class MLXLM(LM):
        """
        logging.info("Estimating loglikelihood for %d pairs." % len(requests))

-        # tokenize prefix and prefix + completion for all requests.
-        tokenized = self._tokenize(
-            [t for r in requests for t in [r.args[0], r.args[0] + r.args[1]]]
-        )
+        group = mx.distributed.init()

-        # max length (prefix + completion) and longest common prefix per question.
-        length_stats = {}
-        for prefix, completed in zip(tokenized[0::2], tokenized[1::2]):
-            max_completed_l, min_prefix_l = length_stats.get(prefix, (0, 1e8))
-            length_stats[prefix] = (
-                max(max_completed_l, len(completed)),
-                min(min_prefix_l, _len_longest_common_prefix(prefix, completed)),
-            )
+        # Group by common prefix
+        group_reqs = collections.defaultdict(list)
+        for idx, req in enumerate(requests):
+            group_reqs[req.args[0]].append((idx, req.args[1]))
+        questions = list(group_reqs.keys())
+        responses = []
+        indices = []
+        for v in group_reqs.values():
+            idx, resp = zip(*v)
+            indices.extend(idx)
+            responses.append(resp)
+
+        # split data accross ranks
+        questions = questions[group.rank() :: group.size()]
+        responses = responses[group.rank() :: group.size()]

-        # truncate requests for completed sequences longer than model context.
-        shortened = []
-        completion_spans = []
        long_completions = 0
-        for prefix, completed in zip(tokenized[0::2], tokenized[1::2]):
-            max_completed_l, prefix_l = length_stats[prefix]
+        scores, is_greedy = [], []
+        for q, rs in tqdm(zip(questions, responses), total=len(questions)):
+            prefix = self._tokenize([q])[0]
+            full_sequences = self._tokenize([q + r for r in rs])
+            max_completed_l = max(len(s) for s in full_sequences)
+
            # compute truncation length
            truncation = max(0, max_completed_l - self._max_tokens - 1)
-            prefix_l = prefix_l - truncation
-            if prefix_l <= 0:
-                # completion too long, prefix is eliminated for some requests.
+            orig_prefix_l = len(prefix)
+            prefix_l = max(len(prefix) - truncation, 0)
+            prefix = prefix[len(prefix) - prefix_l :]
+
+            # If the entire prompt got truncated ignore the question
+            if prefix_l == 0:
                long_completions += 1
-                truncation = max(0, len(completed) - self._max_tokens - 1)
-                prefix_l = 1
-            # truncate the completed sequence
-            completed = completed[truncation:]
-            shortened.append(completed)
-            # scores do not include initial bos, substract 1 to span bounds
-            completion_spans.append((prefix_l - 1, len(completed) - 1))
+                all_scores.extend([-float("inf")] * len(rs))
+                all_is_greedy.extend([False] * len(rs))
+                continue
+
+            # model scoring, returns num_requests x (logp, is_greedy, length).
+            logprobs, cache = self._process_prompt(prefix)
+            max_idx = mx.argmax(logprobs).item()
+
+            for s in full_sequences:
+                inputs = s[len(prefix) :]
+                # The logprobs from the last token of the prompt are
+                # for the first input token
+                scores.append(logprobs[0, inputs[0]].item())
+                is_greedy.append((inputs[0] == max_idx))
+
+                if len(inputs) == 1:
+                    continue
+                score, _, ig = self._score_fn(
+                    mx.array(inputs)[None, :], cache=copy.deepcopy(cache)
+                )
+                scores[-1] += mx.sum(score).item()
+                is_greedy[-1] &= mx.all(ig).item()
+
+        scores = mx.array(scores)
+        is_greedy = mx.array(is_greedy)

        if long_completions > 0:
            logging.info(
@@ -209,13 +220,23 @@ class MLXLM(LM):
                + "completion longer than context."
            )

-        # model scoring, returns num_requests x (logp, is_greedy, length).
-        results = self._loglikelihood(
-            shortened,
-            score_spans=completion_spans,
-            tokenize=False,
-        )
-        return [(r[0], r[1] == r[2]) for r in results]
+        num_results = len(requests)
+
+        # all gather the results across groups
+        if group.size() > 1:
+            per_group = int(np.ceil(num_results / group.size()))
+            scores = mx.pad(scores, ((0, per_group - len(scores)),))
+            is_greedy = mx.pad(is_greedy, ((0, per_group - len(is_greedy))))
+            scores = mx.distributed.all_gather(scores[mx.newaxis], stream=mx.cpu)
+            is_greedy = mx.distributed.all_gather(is_greedy[mx.newaxis], stream=mx.cpu)
+            mx.eval(scores, is_greedy)
+            scores = scores.T.reshape(-1)
+            is_greedy = is_greedy.T.reshape(-1)
+
+        inv_sort = mx.argsort(mx.array(indices))
+        scores = scores[:num_results][inv_sort]
+        is_greedy = is_greedy[:num_results][inv_sort]
+        return list(zip(scores.tolist(), is_greedy.tolist()))

    def loglikelihood_rolling(self, requests) -> list[float]:
        """Compute full log-likelihood of a string, with no truncation, for perplexity computation
@@ -252,8 +273,15 @@ class MLXLM(LM):
        logging.info(
            "Estimating loglikelihood rolling for %d sequences." % len(requests)
        )
-        inputs = [req.args[0] for req in requests]
-        return [t[0] for t in self._loglikelihood(inputs)]
+        inputs = self._tokenize([req.args[0] for req in requests])
+        all_scores = []
+        for i in tqdm(range(0, len(texts), self._batch_size)):
+            batch = texts[i : i + self._batch_size]
+            scores, lengths, _ = self._score_fn(batch)
+            mask = mx.arange(scores.shape[-1]) < lengths[:, None]
+            all_scores.extend((mask * scores).sum(axis=-1).tolist())
+
+        return all_scores

    def generate_until(self, requests) -> list[str]:
        """Generate greedily until a stopping sequence
@@ -273,27 +301,20 @@ class MLXLM(LM):
        contexts, options = zip(*[req.args for req in requests])
        # contrary to the doc the second element of the tuple contains
        # {'do_sample': False, 'until': ['\n\n'], 'temperature': 0}
-        keys = list(options[0].keys())
-        assert "until" in keys
-        untils = [x["until"] for x in options]
        completions = []
-        for context, until in tqdm(zip(contexts, untils), total=len(contexts)):
-            if (
-                hasattr(self._tokenizer, "apply_chat_template")
-                and self._tokenizer.chat_template is not None
-            ):
-                messages = [{"role": "user", "content": context}]
-                context = self._tokenizer.apply_chat_template(
-                    messages, tokenize=False, add_generation_prompt=True
-                )

+        for context, opt in tqdm(zip(contexts, options), total=len(contexts)):
+            until = opt["until"]
+            context = self.tokenizer.encode(
+                context, add_special_tokens=not self.use_chat_template
+            )
            max_tokens = min(
-                self._max_tokens,
-                self._tokenizer.model_max_length - len(self._tokenizer.encode(context)),
+                opt.get("max_gen_tokens", self._max_tokens),
+                self.tokenizer.model_max_length - len(context),
            )
            text = ""
            for response in stream_generate(
-                self._model, self._tokenizer, prompt=context, max_tokens=max_tokens
+                self._model, self.tokenizer, prompt=context, max_tokens=max_tokens
            ):
                text += response.text
                if any(u in text for u in until):
@@ -315,13 +336,42 @@ def main():
        "--output-dir", default=".", help="Output directory for result files."
    )
    parser.add_argument("--batch-size", type=int, default=16, help="Batch size")
-    parser.add_argument("--num-shots", type=int, default=0, help="Number of shots")
+    parser.add_argument("--num-shots", type=int, default=None, help="Number of shots")
    parser.add_argument(
        "--max-tokens",
        type=int,
        help="Maximum nunber of tokens to generate. Defaults to the model's max context length.",
    )
+    parser.add_argument(
+        "--limit",
+        default=None,
+        help="Limit the number of examples per task.",
+        type=int,
+    )
    parser.add_argument("--seed", type=int, default=123, help="Random seed.")
+    parser.add_argument(
+        "--fewshot-as-multiturn",
+        action="store_true",
+        help="Whether to provide the fewshot examples as a multiturn "
+        "conversation or a single user turn.",
+        default=False,
+    )
+    parser.add_argument(
+        "--apply-chat-template",
+        action=argparse.BooleanOptionalAction,
+        help="Specifies whether to apply a chat template to the prompt. If "
+        "the model has a chat template, this defaults to `True`, "
+        "otherwise `False`.",
+        default=None,
+    )
+    parser.add_argument(
+        "--chat-template-args",
+        type=json.loads,
+        help="""A JSON formatted string of arguments for the tokenizer's "
+        "apply_chat_template, e.g. '{"enable_thinking":false}'""",
+        default="{}",
+    )
+
    args = parser.parse_args()

    output_dir = Path(args.output_dir)
@@ -332,24 +382,34 @@ def main():

    mx.random.seed(args.seed)

-    lm = MLXLM(args.model, batch_size=args.batch_size, max_tokens=args.max_tokens)
+    lm = MLXLM(
+        args.model,
+        max_tokens=args.max_tokens,
+        use_chat_template=args.apply_chat_template,
+    )
+    MLXLM.apply_chat_template = chat_template_fn(**args.chat_template_args)

    results = lm_eval.simple_evaluate(
        model=lm,
        tasks=args.tasks,
+        fewshot_as_multiturn=args.fewshot_as_multiturn,
+        apply_chat_template=lm.use_chat_template,
        num_fewshot=args.num_shots,
+        limit=args.limit,
        random_seed=args.seed,
        numpy_random_seed=args.seed,
        torch_random_seed=args.seed,
        fewshot_random_seed=args.seed,
    )

-    model_name = args.model.replace("/", "_")
-    task_names = "_".join(args.tasks)
-    ver = version("lm_eval")
-    filename = f"eval_{model_name}_{task_names}_{args.num_shots:02d}_v_{ver}.json"
-    output_path = output_dir / filename
-    output_path.write_text(json.dumps(results["results"], indent=4))
-    print("Results:")
-    for result in results["results"].values():
-        print(json.dumps(result, indent=4))
+    file_keys = ["eval", args.model.replace("/", "_"), version("lm_eval")]
+    if args.num_shots is not None:
+        file_keys += [f"{args.num_shots:02d}"]
+    file_keys += args.tasks
+    filename = "_".join(file_keys)
+    if mx.distributed.init().rank() == 0:
+        output_path = output_dir / filename
+        output_path.write_text(json.dumps(results["results"], indent=4))
+        print("Results:")
+        for result in results["results"].values():
+            print(json.dumps(result, indent=4))
@@ -15,9 +15,7 @@ prompt_cache = make_prompt_cache(model)
 # User turn
 prompt = "Hi my name is <Name>."
 messages = [{"role": "user", "content": prompt}]
-prompt = tokenizer.apply_chat_template(
-    messages, tokenize=False, add_generation_prompt=True
-)
+prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

 # Assistant response
 response = generate(
@@ -25,16 +23,13 @@ response = generate(
    tokenizer,
    prompt=prompt,
    verbose=True,
-    temp=0.0,
    prompt_cache=prompt_cache,
 )

 # User turn
 prompt = "What's my name?"
 messages = [{"role": "user", "content": prompt}]
-prompt = tokenizer.apply_chat_template(
-    messages, tokenize=False, add_generation_prompt=True
-)
+prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

 # Assistant response
 response = generate(
@@ -14,7 +14,7 @@ conversation = [{"role": "user", "content": prompt}]

 # Transform the prompt into the chat template
 prompt = tokenizer.apply_chat_template(
-    conversation=conversation, tokenize=False, add_generation_prompt=True
+    conversation=conversation, add_generation_prompt=True
 )

 # Specify the maximum number of tokens
@@ -1,5 +1,5 @@
 # The path to the local model directory or Hugging Face repo.
-model: "mlx_model"
+model: "mlx-community/Llama-3.2-1B-Instruct"

 # Whether or not to train (boolean)
 train: true
@@ -7,8 +7,17 @@ train: true
 # The fine-tuning method: "lora", "dora", or "full".
 fine_tune_type: lora

+# The Optimizer with its possible inputs
+optimizer: adamw
+# optimizer_config:
+#   adamw:
+#     betas: [0.9, 0.98]
+#     eps: 1e-6
+#     weight_decay: 0.05
+#     bias_correction: true
+
 # Directory with {train, valid, test}.jsonl files
-data: "/path/to/training/data"
+data: "mlx-community/WikiSQL"

 # The PRNG seed
 seed: 0
@@ -28,6 +37,9 @@ val_batches: 25
 # Adam learning rate.
 learning_rate: 1e-5

+# Whether to report the logs to WandB
+# wand: "wandb-project"
+
 # Number of training steps between loss reporting.
 steps_per_report: 10

@@ -72,7 +84,7 @@ lora_parameters:
 #  arguments: [1e-5, 1000, 1e-7] # passed to scheduler

 #hf_dataset:
-#  name: "billsum"
+#  path: "billsum"
 #  train_split: "train[:1000]"
 #  valid_split: "train[-100:]"
 #  prompt_feature: "text"
@@ -0,0 +1,65 @@
+# Copyright © 2025 Apple Inc.
+"""
+This is an example of tool use with mlx_lm and the OpenAI client.
+
+To run, first start the server:
+
+>>> mlx_lm.server
+
+Then run this script.
+"""
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-needed")
+
+model = "mlx-community/qwen3-4b-4bit-DWQ"
+messages = [{"role": "user", "content": "What's the weather in Boston?"}]
+
+tools = [
+    {
+        "type": "function",
+        "function": {
+            "name": "get_current_weather",
+            "description": "Get the current weather in a given location",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "location": {
+                        "type": "string",
+                        "description": "The city and state, e.g. San Francisco, CA",
+                    },
+                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
+                },
+                "required": ["location"],
+            },
+        },
+    }
+]
+
+
+def get_current_weather(**kwargs):
+    return "51 Farenheit, clear skies"
+
+
+functions = {"get_current_weather": get_current_weather}
+
+# The first query generates a tool call
+response = client.chat.completions.create(
+    model=model,
+    messages=messages,
+    tools=tools,
+)
+
+# Call the function
+function = response.choices[0].message.tool_calls[0].function
+tool_result = functions[function.name](**function.arguments)
+
+# Put the result of the function in the messages and generate the final
+# response:
+messages.append({"role": "tool", "name": function.name, "content": tool_result})
+response = client.chat.completions.create(
+    model=model,
+    messages=messages,
+    tools=tools,
+)
+print(response.choices[0].message.content)
@@ -0,0 +1,131 @@
+# Copyright © 2024 Apple Inc.
+
+"""
+Run with:
+
+```
+mlx.launch \
+ --hostfile /path/to/hosts.json \
+ /path/to/pipeline_generate.py \
+ --prompt "hello world"
+```
+
+Make sure you can run MLX over MPI on two hosts. For more information see the
+documentation:
+
+https://ml-explore.github.io/mlx/build/html/usage/distributed.html).
+"""
+
+import argparse
+import json
+import resource
+from pathlib import Path
+
+import mlx.core as mx
+from huggingface_hub import snapshot_download
+from mlx.utils import tree_flatten
+
+from mlx_lm import load, stream_generate
+from mlx_lm.utils import load_model, load_tokenizer
+
+# Needed for 8 bit model
+resource.setrlimit(resource.RLIMIT_NOFILE, (2048, 4096))
+
+
+def download(repo: str, allow_patterns: list[str]) -> Path:
+    return Path(
+        snapshot_download(
+            repo,
+            allow_patterns=allow_patterns,
+        )
+    )
+
+
+def shard_and_load(repo):
+    # Get model path with everything but weight safetensors
+    model_path = download(
+        args.model,
+        allow_patterns=["*.json", "*.py", "tokenizer.model", "*.tiktoken", "*.txt"],
+    )
+
+    # Lazy load and shard model to figure out
+    # which weights we need
+    model, _ = load_model(model_path, lazy=True, strict=False)
+
+    group = mx.distributed.init()
+    rank = group.rank()
+    model.model.pipeline(group)
+
+    # Figure out which files we need for the local shard
+    with open(model_path / "model.safetensors.index.json", "r") as fid:
+        weight_index = json.load(fid)["weight_map"]
+
+    local_files = set()
+    for k, _ in tree_flatten(model.parameters()):
+        local_files.add(weight_index[k])
+
+    # Download weights for local shard
+    download(args.model, allow_patterns=local_files)
+
+    # Load and shard the model, and load the weights
+    tokenizer = load_tokenizer(model_path)
+    model, _ = load_model(model_path, lazy=True, strict=False)
+    model.model.pipeline(group)
+    mx.eval(model.parameters())
+
+    # Synchronize processes before generation to avoid timeout if downloading
+    # model for the first time.
+    mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu))
+    return model, tokenizer
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="LLM pipelined inference example")
+    parser.add_argument(
+        "--model",
+        default="mlx-community/DeepSeek-R1-3bit",
+        help="HF repo or path to local model.",
+    )
+    parser.add_argument(
+        "--prompt",
+        "-p",
+        default="Write a quicksort in C++.",
+        help="Message to be processed by the model ('-' reads from stdin)",
+    )
+    parser.add_argument(
+        "--max-tokens",
+        "-m",
+        type=int,
+        default=256,
+        help="Maximum number of tokens to generate",
+    )
+    args = parser.parse_args()
+
+    group = mx.distributed.init()
+    rank = group.rank()
+
+    def rprint(*args, **kwargs):
+        if rank == 0:
+            print(*args, **kwargs)
+
+    model, tokenizer = shard_and_load(args.model)
+
+    messages = [{"role": "user", "content": args.prompt}]
+    prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+
+    for response in stream_generate(
+        model, tokenizer, prompt, max_tokens=args.max_tokens
+    ):
+        rprint(response.text, end="", flush=True)
+
+    rprint()
+    rprint("=" * 10)
+    rprint(
+        f"Prompt: {response.prompt_tokens} tokens, "
+        f"{response.prompt_tps:.3f} tokens-per-sec"
+    )
+    rprint(
+        f"Generation: {response.generation_tokens} tokens, "
+        f"{response.generation_tps:.3f} tokens-per-sec"
+    )
+    rprint(f"Peak memory: {response.peak_memory:.3f} GB")
@@ -0,0 +1,73 @@
+# Copyright © 2025 Apple Inc.
+
+import json
+
+from mlx_lm import generate, load
+from mlx_lm.models.cache import make_prompt_cache
+
+# Specify the checkpoint
+checkpoint = "mlx-community/Qwen2.5-32B-Instruct-4bit"
+
+# Load the corresponding model and tokenizer
+model, tokenizer = load(path_or_hf_repo=checkpoint)
+
+
+# An example tool, make sure to include a docstring and type hints
+def multiply(a: float, b: float):
+    """
+    A function that multiplies two numbers
+
+    Args:
+        a: The first number to multiply
+        b: The second number to multiply
+    """
+    return a * b
+
+
+tools = {"multiply": multiply}
+
+# Specify the prompt and conversation history
+prompt = "Multiply 12234585 and 48838483920."
+messages = [{"role": "user", "content": prompt}]
+
+prompt = tokenizer.apply_chat_template(
+    messages, add_generation_prompt=True, tools=list(tools.values())
+)
+
+prompt_cache = make_prompt_cache(model)
+
+# Generate the initial tool call:
+response = generate(
+    model=model,
+    tokenizer=tokenizer,
+    prompt=prompt,
+    max_tokens=2048,
+    verbose=True,
+    prompt_cache=prompt_cache,
+)
+
+# Parse the tool call:
+# (Note, the tool call format is model specific)
+tool_open = "<tool_call>"
+tool_close = "</tool_call>"
+start_tool = response.find(tool_open) + len(tool_open)
+end_tool = response.find(tool_close)
+tool_call = json.loads(response[start_tool:end_tool].strip())
+tool_result = tools[tool_call["name"]](**tool_call["arguments"])
+
+# Put the tool result in the prompt
+messages = [{"role": "tool", "name": tool_call["name"], "content": tool_result}]
+prompt = tokenizer.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+)
+
+# Generate the final response:
+response = generate(
+    model=model,
+    tokenizer=tokenizer,
+    prompt=prompt,
+    max_tokens=2048,
+    verbose=True,
+    prompt_cache=prompt_cache,
+)
@@ -1,19 +1,14 @@
 import argparse
-import glob
-import shutil
 from pathlib import Path

 from mlx.utils import tree_flatten, tree_unflatten

 from .gguf import convert_to_gguf
-from .tuner.dora import DoRAEmbedding, DoRALinear
-from .tuner.lora import LoRAEmbedding, LoRALinear, LoRASwitchLinear
 from .tuner.utils import dequantize, load_adapters
 from .utils import (
    fetch_from_hub,
    get_model_path,
-    save_config,
-    save_weights,
+    save,
    upload_to_hub,
 )

@@ -38,12 +33,6 @@ def parse_arguments() -> argparse.Namespace:
        default="adapters",
        help="Path to the trained adapter weights and config.",
    )
-    parser.add_argument(
-        "--hf-path",
-        type=str,
-        default=None,
-        help="Path to the original Hugging Face model. Required for upload if --model is a local directory.",
-    )
    parser.add_argument(
        "--upload-repo",
        help="The Hugging Face repo to upload the model to.",
@@ -73,14 +62,16 @@ def main() -> None:
    print("Loading pretrained model")
    args = parse_arguments()

-    model_path = get_model_path(args.model)
+    model_path, hf_path = get_model_path(args.model)
    model, config, tokenizer = fetch_from_hub(model_path)

    model.freeze()
    model = load_adapters(model, args.adapter_path)

    fused_linears = [
-        (n, m.fuse()) for n, m in model.named_modules() if hasattr(m, "fuse")
+        (n, m.fuse(de_quantize=args.de_quantize))
+        for n, m in model.named_modules()
+        if hasattr(m, "fuse")
    ]

    if fused_linears:
@@ -89,23 +80,18 @@ def main() -> None:
    if args.de_quantize:
        print("De-quantizing model")
        model = dequantize(model)
-
-    weights = dict(tree_flatten(model.parameters()))
-
-    save_path = Path(args.save_path)
-
-    save_weights(save_path, weights)
-
-    py_files = glob.glob(str(model_path / "*.py"))
-    for file in py_files:
-        shutil.copy(file, save_path)
-
-    tokenizer.save_pretrained(save_path)
-
-    if args.de_quantize:
        config.pop("quantization", None)

-    save_config(config, config_path=save_path / "config.json")
+    save_path = Path(args.save_path)
+    save(
+        save_path,
+        model_path,
+        model,
+        tokenizer,
+        config,
+        hf_repo=hf_path,
+        donate_model=False,
+    )

    if args.export_gguf:
        model_type = config["model_type"]
@@ -113,18 +99,20 @@ def main() -> None:
            raise ValueError(
                f"Model type {model_type} not supported for GGUF conversion."
            )
+        weights = dict(tree_flatten(model.parameters()))
        convert_to_gguf(model_path, weights, config, str(save_path / args.gguf_path))

    if args.upload_repo is not None:
-        hf_path = args.hf_path or (
-            args.model if not Path(args.model).exists() else None
-        )
        if hf_path is None:
            raise ValueError(
                "Must provide original Hugging Face repo to upload local model."
            )
-        upload_to_hub(args.save_path, args.upload_repo, hf_path)
+        upload_to_hub(args.save_path, args.upload_repo)


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.fuse...` directly is deprecated."
+        " Use `mlx_lm.fuse...` or `python -m mlx_lm fuse ...` instead."
+    )
    main()
@@ -1,23 +1,46 @@
 # Copyright © 2023-2024 Apple Inc.

 import argparse
-import codecs
+import contextlib
+import functools
 import json
 import sys
+import time
+from dataclasses import dataclass
+from typing import (
+    Any,
+    Callable,
+    Generator,
+    List,
+    Optional,
+    Tuple,
+    Union,
+)

 import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_reduce
+from transformers import PreTrainedTokenizer

-from .models.cache import QuantizedKVCache, load_prompt_cache
+from .models import cache
+from .models.cache import (
+    QuantizedKVCache,
+    load_prompt_cache,
+)
 from .sample_utils import make_sampler
-from .utils import generate, load
+from .tokenizer_utils import TokenizerWrapper
+from .utils import does_model_support_input_embeddings, load

 DEFAULT_PROMPT = "hello"
 DEFAULT_MAX_TOKENS = 100
 DEFAULT_TEMP = 0.0
 DEFAULT_TOP_P = 1.0
 DEFAULT_MIN_P = 0.0
+DEFAULT_TOP_K = 0
+DEFAULT_XTC_PROBABILITY = 0.0
+DEFAULT_XTC_THRESHOLD = 0.0
 DEFAULT_MIN_TOKENS_TO_KEEP = 1
-DEFAULT_SEED = 0
+DEFAULT_SEED = None
 DEFAULT_MODEL = "mlx-community/Llama-3.2-3B-Instruct-4bit"
 DEFAULT_QUANTIZED_KV_START = 5000

@@ -44,10 +67,11 @@ def setup_arg_parser():
        help="Optional path for the trained adapter weights and config.",
    )
    parser.add_argument(
-        "--eos-token",
+        "--extra-eos-token",
        type=str,
-        default=None,
-        help="End of sequence token for tokenizer",
+        default=(),
+        nargs="+",
+        help="Add tokens in the list of eos tokens that stop generation.",
    )
    parser.add_argument(
        "--system-prompt",
@@ -60,6 +84,11 @@ def setup_arg_parser():
        default=DEFAULT_PROMPT,
        help="Message to be processed by the model ('-' reads from stdin)",
    )
+    parser.add_argument(
+        "--prefill-response",
+        default=None,
+        help="Prefill response to be used for the chat template",
+    )
    parser.add_argument(
        "--max-tokens",
        "-m",
@@ -76,13 +105,33 @@ def setup_arg_parser():
    parser.add_argument(
        "--min-p", type=float, default=DEFAULT_MIN_P, help="Sampling min-p"
    )
+    parser.add_argument(
+        "--top-k", type=int, default=DEFAULT_TOP_K, help="Sampling top-k"
+    )
+    parser.add_argument(
+        "--xtc-probability",
+        type=float,
+        default=DEFAULT_XTC_PROBABILITY,
+        help="Probability of XTC sampling to happen each next token",
+    )
+    parser.add_argument(
+        "--xtc-threshold",
+        type=float,
+        default=0.0,
+        help="Thresold the probs of each next token candidate to be sampled by XTC",
+    )
    parser.add_argument(
        "--min-tokens-to-keep",
        type=int,
        default=DEFAULT_MIN_TOKENS_TO_KEEP,
        help="Minimum tokens to keep for min-p sampling.",
    )
-    parser.add_argument("--seed", type=int, default=DEFAULT_SEED, help="PRNG seed")
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=DEFAULT_SEED,
+        help="PRNG seed",
+    )
    parser.add_argument(
        "--ignore-chat-template",
        action="store_true",
@@ -93,6 +142,12 @@ def setup_arg_parser():
        action="store_true",
        help="Use the default chat template",
    )
+    parser.add_argument(
+        "--chat-template-config",
+        help="Additional config for `apply_chat_template`. Should be a dictionary of"
+        " string keys to values represented as a JSON decodable string.",
+        default=None,
+    )
    parser.add_argument(
        "--verbose",
        type=str2bool,
@@ -131,14 +186,568 @@ def setup_arg_parser():
        type=int,
        default=DEFAULT_QUANTIZED_KV_START,
    )
+    parser.add_argument(
+        "--draft-model",
+        type=str,
+        help="A model to be used for speculative decoding.",
+        default=None,
+    )
+    parser.add_argument(
+        "--num-draft-tokens",
+        type=int,
+        help="Number of tokens to draft when using speculative decoding.",
+        default=3,
+    )
    return parser


+# A stream on the default device just for generation
+generation_stream = mx.new_stream(mx.default_device())
+
+
+@contextlib.contextmanager
+def wired_limit(model: nn.Module, streams: Optional[List[mx.Stream]] = None):
+    """
+    A context manager to temporarily change the wired limit.
+
+    Note, the wired limit should not be changed during an async eval.  If an
+    async eval could be running pass in the streams to synchronize with prior
+    to exiting the context manager.
+    """
+    if not mx.metal.is_available():
+        try:
+            yield
+        finally:
+            return
+
+    model_bytes = tree_reduce(
+        lambda acc, x: acc + x.nbytes if isinstance(x, mx.array) else acc, model, 0
+    )
+    max_rec_size = mx.metal.device_info()["max_recommended_working_set_size"]
+    if model_bytes > 0.9 * max_rec_size:
+        model_mb = model_bytes // 2**20
+        max_rec_mb = max_rec_size // 2**20
+        print(
+            f"[WARNING] Generating with a model that requires {model_mb} MB "
+            f"which is close to the maximum recommended size of {max_rec_mb} "
+            "MB. This can be slow. See the documentation for possible work-arounds: "
+            "https://github.com/ml-explore/mlx-lm/tree/main#large-models"
+        )
+    old_limit = mx.set_wired_limit(max_rec_size)
+    try:
+        yield
+    finally:
+        if streams is not None:
+            for s in streams:
+                mx.synchronize(s)
+        else:
+            mx.synchronize()
+        mx.set_wired_limit(old_limit)
+
+
+@dataclass
+class GenerationResponse:
+    """
+    The output of :func:`stream_generate`.
+
+    Args:
+        text (str): The next segment of decoded text. This can be an empty string.
+        token (int): The next token.
+        from_draft (bool): Whether the token was generated by the draft model.
+        logprobs (mx.array): A vector of log probabilities.
+        prompt_tokens (int): The number of tokens in the prompt.
+        prompt_tps (float): The prompt processing tokens-per-second.
+        generation_tokens (int): The number of generated tokens.
+        generation_tps (float): The tokens-per-second for generation.
+        peak_memory (float): The peak memory used so far in GB.
+        finish_reason (str): The reason the response is being sent: "length", "stop" or `None`
+    """
+
+    text: str
+    token: int
+    logprobs: mx.array
+    from_draft: bool
+    prompt_tokens: int
+    prompt_tps: float
+    generation_tokens: int
+    generation_tps: float
+    peak_memory: float
+    finish_reason: Optional[str] = None
+
+
+def maybe_quantize_kv_cache(prompt_cache, quantized_kv_start, kv_group_size, kv_bits):
+    if (
+        kv_bits is not None
+        and not isinstance(prompt_cache[0], cache.QuantizedKVCache)
+        and prompt_cache[0].offset > quantized_kv_start
+    ):
+        for i in range(len(prompt_cache)):
+            if isinstance(prompt_cache[i], cache.KVCache):
+                prompt_cache[i] = prompt_cache[i].to_quantized(
+                    group_size=kv_group_size, bits=kv_bits
+                )
+
+
+def generate_step(
+    prompt: mx.array,
+    model: nn.Module,
+    *,
+    max_tokens: int = 256,
+    sampler: Optional[Callable[mx.array, mx.array]] = None,
+    logits_processors: Optional[List[Callable[[mx.array, mx.array], mx.array]]] = None,
+    max_kv_size: Optional[int] = None,
+    prompt_cache: Optional[Any] = None,
+    prefill_step_size: int = 2048,
+    kv_bits: Optional[int] = None,
+    kv_group_size: int = 64,
+    quantized_kv_start: int = 0,
+    prompt_progress_callback: Optional[Callable[int, int]] = None,
+    input_embeddings: Optional[mx.array] = None,
+) -> Generator[Tuple[mx.array, mx.array], None, None]:
+    """
+    A generator producing token ids based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        max_tokens (int): The maximum number of tokens. Use``-1`` for an infinite
+          generator. Default: ``256``.
+        sampler (Callable[mx.array, mx.array], optional): A sampler for sampling a
+          token from a vector of log probabilities. Default: ``None``.
+        logits_processors (List[Callable[[mx.array, mx.array], mx.array]], optional):
+          A list of functions that take tokens and logits and return the processed
+          logits. Default: ``None``.
+        max_kv_size (int, optional): Maximum size of the key-value cache. Old
+          entries (except the first 4 tokens) will be overwritten.
+        prompt_cache (List[Any], optional): A pre-computed prompt cache. Note, if
+          provided, the cache will be updated in place.
+        prefill_step_size (int): Step size for processing the prompt.
+        kv_bits (int, optional): Number of bits to use for KV cache quantization.
+          None implies no cache quantization. Default: ``None``.
+        kv_group_size (int): Group size for KV cache quantization. Default: ``64``.
+        quantized_kv_start (int): Step to begin using a quantized KV cache.
+           when ``kv_bits`` is non-None. Default: ``0``.
+        prompt_progress_callback (Callable[int, int]): A call-back which takes the
+           prompt tokens processed so far and the total number of prompt tokens.
+        input_embeddings (mx.array, optional): Input embeddings to use in place of
+           prompt tokens. Default: ``None``.
+
+    Yields:
+        Tuple[mx.array, mx.array]: One token and a vector of log probabilities.
+    """
+    if input_embeddings is not None:
+        if not does_model_support_input_embeddings(model):
+            raise ValueError("Model does not support input embeddings.")
+        if len(prompt) != 0:
+            raise ValueError(
+                "If using input embeddings, prompt tokens must be an empty array."
+            )
+
+    tokens = None
+
+    # Create the KV cache for generation
+    if prompt_cache is None:
+        prompt_cache = cache.make_prompt_cache(
+            model,
+            max_kv_size=max_kv_size,
+        )
+
+    prompt_progress_callback = prompt_progress_callback or (lambda *_: None)
+
+    quantize_cache_fn = functools.partial(
+        maybe_quantize_kv_cache,
+        quantized_kv_start=quantized_kv_start,
+        kv_group_size=kv_group_size,
+        kv_bits=kv_bits,
+    )
+
+    sampler = sampler or (lambda x: mx.argmax(x, axis=-1))
+
+    def _model_call(y):
+        if y.ndim == 3:
+            return model(None, cache=prompt_cache, input_embeddings=y)
+        else:
+            return model(y, cache=prompt_cache)
+
+    def _step(y):
+        nonlocal tokens
+
+        with mx.stream(generation_stream):
+            logits = _model_call(y[None])
+
+            logits = logits[:, -1, :]
+
+            if logits_processors and input_embeddings is None:
+                tokens = mx.concat([tokens, y]) if tokens is not None else y
+                for processor in logits_processors:
+                    logits = processor(tokens, logits)
+
+            quantize_cache_fn(prompt_cache)
+
+            logprobs = logits - mx.logsumexp(logits, keepdims=True)
+            y = sampler(logprobs)
+            return y, logprobs.squeeze(0)
+
+    using_embeddings = input_embeddings is not None
+
+    y = input_embeddings if using_embeddings else prompt
+    with mx.stream(generation_stream):
+        total_prompt_tokens = y.shape[0]
+        prompt_processed_tokens = 0
+        while y.shape[0] > prefill_step_size:
+            _model_call(y[:prefill_step_size][None])
+            quantize_cache_fn(prompt_cache)
+            mx.eval([c.state for c in prompt_cache])
+            prompt_progress_callback(prompt_processed_tokens, total_prompt_tokens)
+            prompt_processed_tokens += prefill_step_size
+            y = y[prefill_step_size:]
+            mx.clear_cache()
+
+        y, logprobs = _step(y)
+
+    mx.async_eval(y, logprobs)
+    n = 0
+    while True:
+        if n != max_tokens:
+            next_y, next_logprobs = _step(y)
+            mx.async_eval(next_y, next_logprobs)
+        if n == 0:
+            mx.eval(y)
+            prompt_progress_callback(total_prompt_tokens, total_prompt_tokens)
+        if n == max_tokens:
+            break
+        yield y.item(), logprobs
+        if n % 256 == 0:
+            mx.clear_cache()
+        y, logprobs = next_y, next_logprobs
+        n += 1
+
+
+def speculative_generate_step(
+    prompt: mx.array,
+    model: nn.Module,
+    draft_model: nn.Module,
+    *,
+    num_draft_tokens=2,
+    max_tokens: int = 256,
+    sampler: Optional[Callable[mx.array, mx.array]] = None,
+    logits_processors: Optional[List[Callable[[mx.array, mx.array], mx.array]]] = None,
+    prompt_cache: Optional[Any] = None,
+    prefill_step_size: int = 512,
+    kv_bits: Optional[int] = None,
+    kv_group_size: int = 64,
+    quantized_kv_start: int = 0,
+) -> Generator[Tuple[mx.array, mx.array, bool], None, None]:
+    """
+    A generator producing token ids based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        draft_model (nn.Module): The draft model for speculative decoding.
+        num_draft_tokens (int, optional): The number of draft tokens for
+          speculative decoding. Default: ``2``.
+        max_tokens (int): The maximum number of tokens. Use``-1`` for an infinite
+          generator. Default: ``256``.
+        sampler (Callable[mx.array, mx.array], optional): A sampler for sampling a
+          token from a vector of log probabilities. Default: ``None``.
+        logits_processors (List[Callable[[mx.array, mx.array], mx.array]], optional):
+          A list of functions that take tokens and logits and return the processed
+          logits. Default: ``None``.
+        prompt_cache (List[Any], optional): A pre-computed prompt cache. Note, if
+          provided, the cache will be updated in place. The cache must be trimmable.
+        prefill_step_size (int): Step size for processing the prompt.
+        kv_bits (int, optional): Number of bits to use for KV cache quantization.
+          None implies no cache quantization. Default: ``None``.
+        kv_group_size (int): Group size for KV cache quantization. Default: ``64``.
+        quantized_kv_start (int): Step to begin using a quantized KV cache.
+           when ``kv_bits`` is non-None. Default: ``0``.
+
+    Yields:
+        Tuple[mx.array, mx.array, bool]: One token, a vector of log probabilities,
+          and a bool indicating if the token was generated by the draft model
+    """
+
+    y = prompt.astype(mx.uint32)
+    prev_tokens = None
+
+    # Create the KV cache for generation
+    if prompt_cache is None:
+        model_cache = cache.make_prompt_cache(model)
+        draft_cache = cache.make_prompt_cache(draft_model)
+    else:
+        model_cache = prompt_cache[: len(model.layers)]
+        draft_cache = prompt_cache[len(model.layers) :]
+
+    sampler = sampler or (lambda x: mx.argmax(x, axis=-1))
+
+    quantize_cache_fn = functools.partial(
+        maybe_quantize_kv_cache,
+        quantized_kv_start=quantized_kv_start,
+        kv_group_size=kv_group_size,
+        kv_bits=kv_bits,
+    )
+
+    def _process_and_sample(tokens, logits):
+        if logits_processors:
+            for processor in logits_processors:
+                logits = processor(tokens, logits)
+
+        logprobs = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
+        y = sampler(logprobs)
+        return y, logprobs
+
+    def _step(model, cache, y, n_predict=1):
+        with mx.stream(generation_stream):
+            logits = model(y[None], cache=cache)
+            logits = logits[:, -n_predict:, :]
+
+            quantize_cache_fn(cache)
+            if logits_processors:
+                nonlocal prev_tokens
+                out_y, out_logprobs = [], []
+                if n_predict > 1:
+                    y = y[: -(n_predict - 1)]
+                for i in range(n_predict):
+                    prev_tokens = (
+                        mx.concat([prev_tokens, y]) if prev_tokens is not None else y
+                    )
+                    y, logprobs = _process_and_sample(prev_tokens, logits[:, i, :])
+                    out_y.append(y)
+                    out_logprobs.append(logprobs)
+                return mx.concatenate(out_y, axis=0), mx.concatenate(
+                    out_logprobs, axis=0
+                )
+            else:
+                return _process_and_sample(None, logits.squeeze(0))
+
+    def _prefill(model, cache, y):
+        while y.size > prefill_step_size:
+            model(y[:prefill_step_size][None], cache=cache)
+            quantize_cache_fn(cache)
+            mx.eval([c.state for c in cache])
+            y = y[prefill_step_size:]
+            mx.clear_cache()
+        return y
+
+    def _rewind_cache(num_draft, num_accept):
+        cache.trim_prompt_cache(model_cache, num_draft - num_accept)
+        cache.trim_prompt_cache(draft_cache, max(num_draft - num_accept - 1, 0))
+
+    def _draft_generate(y, num_draft):
+        if num_draft == 0:
+            return mx.array([], mx.uint32)
+        ys = []
+        for _ in range(num_draft):
+            y, _ = _step(draft_model, draft_cache, y)
+            mx.async_eval(y)
+            ys.append(y)
+        return mx.concatenate(ys)
+
+    with mx.stream(generation_stream):
+        draft_y = _prefill(draft_model, draft_cache, y)
+        y = _prefill(model, model_cache, y)
+
+    ntoks = 0
+    # Set these so the finally block doesn't raise
+    num_draft = 0
+    n = 0
+    try:
+        while True:
+            num_draft = min(max_tokens - ntoks, num_draft_tokens)
+            draft_tokens = _draft_generate(draft_y, num_draft)
+            if prev_tokens is not None:
+                prev_tokens = prev_tokens[: prev_tokens.size - y.size - num_draft + 1]
+            y = mx.concatenate([y, draft_tokens])
+            tokens, logprobs = _step(model, model_cache, y, num_draft + 1)
+            mx.eval(tokens, draft_tokens)
+            draft_tokens = draft_tokens.tolist()
+            tokens = tokens.tolist()
+            n = 0
+            while n < num_draft:
+                tn, dtn, lpn = tokens[n], draft_tokens[n], logprobs[n]
+                if tn != dtn:
+                    break
+                n += 1
+                ntoks += 1
+                yield tn, lpn, True
+                if ntoks == max_tokens:
+                    break
+            if ntoks < max_tokens:
+                ntoks += 1
+                yield tokens[n], logprobs[n], False
+
+            if ntoks == max_tokens:
+                break
+
+            y = mx.array([tokens[n]], mx.uint32)
+            draft_y = y
+
+            # If we accepted all the draft tokens, include the last
+            # draft token in the next draft step since it hasn't been
+            # processed yet by the draft model
+            if n == num_draft:
+                draft_y = mx.concatenate(
+                    [mx.array(draft_tokens[-1:], mx.uint32), draft_y]
+                )
+
+            if prev_tokens is not None:
+                prev_tokens = prev_tokens[: -max(num_draft - n, 1)]
+            _rewind_cache(num_draft, n)
+    finally:
+        _rewind_cache(num_draft, n)
+
+
+def stream_generate(
+    model: nn.Module,
+    tokenizer: Union[PreTrainedTokenizer, TokenizerWrapper],
+    prompt: Union[str, mx.array, List[int]],
+    draft_model: Optional[nn.Module] = None,
+    **kwargs,
+) -> Generator[GenerationResponse, None, None]:
+    """
+    A generator producing text based on the given prompt from the model.
+
+    Args:
+        model (nn.Module): The model to use for generation.
+        tokenizer (PreTrainedTokenizer): The tokenizer.
+        prompt (Union[str, mx.array, List[int]]): The input prompt string or
+          integer tokens.
+        draft_model (Optional[nn.Module]): An optional draft model. If provided
+          then speculative decoding is used. The draft model must use the same
+          tokenizer as the main model. Default: ``None``.
+        kwargs: The remaining options get passed to :func:`generate_step`.
+          See :func:`generate_step` for more details.
+
+    Yields:
+        GenerationResponse: An instance containing the generated text segment and
+            associated metadata. See :class:`GenerationResponse` for details.
+    """
+    if not isinstance(tokenizer, TokenizerWrapper):
+        tokenizer = TokenizerWrapper(tokenizer)
+
+    if not isinstance(prompt, mx.array):
+        if isinstance(prompt, str):
+            # Try to infer if special tokens are needed
+            add_special_tokens = tokenizer.bos_token is None or not prompt.startswith(
+                tokenizer.bos_token
+            )
+            prompt = tokenizer.encode(prompt, add_special_tokens=add_special_tokens)
+        prompt = mx.array(prompt)
+
+    detokenizer = tokenizer.detokenizer
+
+    if draft_model is None:
+        kwargs.pop("num_draft_tokens", None)
+        token_generator = generate_step(prompt, model, **kwargs)
+        # from_draft always false for non-speculative generation
+        token_generator = (
+            (token, logprobs, False) for token, logprobs in token_generator
+        )
+    else:
+        kwargs.pop("max_kv_size", None)
+        token_generator = speculative_generate_step(
+            prompt, model, draft_model, **kwargs
+        )
+    with wired_limit(model, [generation_stream]):
+        detokenizer.reset()
+        tic = time.perf_counter()
+        for n, (token, logprobs, from_draft) in enumerate(token_generator):
+            if n == 0:
+                prompt_time = time.perf_counter() - tic
+                prompt_tps = prompt.size / prompt_time
+                tic = time.perf_counter()
+            if token in tokenizer.eos_token_ids:
+                break
+
+            detokenizer.add_token(token)
+
+            yield GenerationResponse(
+                text=detokenizer.last_segment,
+                token=token,
+                logprobs=logprobs,
+                from_draft=from_draft,
+                prompt_tokens=prompt.size,
+                prompt_tps=prompt_tps,
+                generation_tokens=n + 1,
+                generation_tps=(n + 1) / (time.perf_counter() - tic),
+                peak_memory=mx.get_peak_memory() / 1e9,
+                finish_reason=None,
+            )
+
+        detokenizer.finalize()
+        yield GenerationResponse(
+            text=detokenizer.last_segment,
+            token=token,
+            logprobs=logprobs,
+            from_draft=from_draft,
+            prompt_tokens=prompt.size,
+            prompt_tps=prompt_tps,
+            generation_tokens=n + 1,
+            generation_tps=(n + 1) / (time.perf_counter() - tic),
+            peak_memory=mx.get_peak_memory() / 1e9,
+            finish_reason="stop" if token in tokenizer.eos_token_ids else "length",
+        )
+
+
+def generate(
+    model: nn.Module,
+    tokenizer: Union[PreTrainedTokenizer, TokenizerWrapper],
+    prompt: Union[str, List[int]],
+    verbose: bool = False,
+    formatter: Optional[Callable] = None,
+    **kwargs,
+) -> str:
+    """
+    Generate a complete response from the model.
+
+    Args:
+       model (nn.Module): The language model.
+       tokenizer (PreTrainedTokenizer): The tokenizer.
+       prompt (Union[str, List[int]]): The input prompt string or integer tokens.
+       verbose (bool): If ``True``, print tokens and timing information.
+           Default: ``False``.
+       kwargs: The remaining options get passed to :func:`stream_generate`.
+          See :func:`stream_generate` for more details.
+    """
+    if formatter is not None:
+        print(
+            "[Warning] Text formatting is deprecated and no longer used. "
+            "The argument will be removed in a future version."
+        )
+    if verbose:
+        print("=" * 10)
+
+    text = ""
+    for response in stream_generate(model, tokenizer, prompt, **kwargs):
+        if verbose:
+            print(response.text, end="", flush=True)
+        text += response.text
+
+    if verbose:
+        print()
+        print("=" * 10)
+        if len(text) == 0:
+            print("No text generated for this prompt")
+            return
+        print(
+            f"Prompt: {response.prompt_tokens} tokens, "
+            f"{response.prompt_tps:.3f} tokens-per-sec"
+        )
+        print(
+            f"Generation: {response.generation_tokens} tokens, "
+            f"{response.generation_tps:.3f} tokens-per-sec"
+        )
+        print(f"Peak memory: {response.peak_memory:.3f} GB")
+    return text
+
+
 def main():
    parser = setup_arg_parser()
    args = parser.parse_args()

-    mx.random.seed(args.seed)
+    if args.seed is not None:
+        mx.random.seed(args.seed)

    # Load the prompt cache and metadata if a cache file is provided
    using_cache = args.prompt_cache_file is not None
@@ -162,8 +771,6 @@ def main():
        {} if not using_cache else json.loads(metadata["tokenizer_config"])
    )
    tokenizer_config["trust_remote_code"] = True
-    if args.eos_token is not None:
-        tokenizer_config["eos_token"] = args.eos_token

    model_path = args.model
    if using_cache:
@@ -182,31 +789,37 @@ def main():
        adapter_path=args.adapter_path,
        tokenizer_config=tokenizer_config,
    )
+    for eos_token in args.extra_eos_token:
+        tokenizer.add_eos_token(eos_token)
+
+    template_kwargs = {}
+    if args.chat_template_config is not None:
+        template_kwargs = json.loads(args.chat_template_config)

    if args.use_default_chat_template:
        if tokenizer.chat_template is None:
            tokenizer.chat_template = tokenizer.default_chat_template
    elif using_cache:
-        tokenizer.chat_template = metadata["chat_template"]
+        tokenizer.chat_template = json.loads(metadata["chat_template"])

-    prompt = codecs.decode(args.prompt, "unicode_escape")
-
-    if not args.ignore_chat_template and (
-        hasattr(tokenizer, "apply_chat_template")
-        and tokenizer.chat_template is not None
-    ):
+    prompt = args.prompt.replace("\\n", "\n").replace("\\t", "\t")
+    prompt = sys.stdin.read() if prompt == "-" else prompt
+    if not args.ignore_chat_template and tokenizer.chat_template is not None:
        if args.system_prompt is not None:
            messages = [{"role": "system", "content": args.system_prompt}]
        else:
            messages = []
-        messages.append(
-            {
-                "role": "user",
-                "content": sys.stdin.read() if prompt == "-" else prompt,
-            }
-        )
+        messages.append({"role": "user", "content": prompt})
+
+        has_prefill = args.prefill_response is not None
+        if has_prefill:
+            messages.append({"role": "assistant", "content": args.prefill_response})
        prompt = tokenizer.apply_chat_template(
-            messages, tokenize=False, add_generation_prompt=True
+            messages,
+            tokenize=False,
+            continue_final_message=has_prefill,
+            add_generation_prompt=not has_prefill,
+            **template_kwargs,
        )

        # Treat the prompt as a suffix assuming that the prefix is in the
@@ -216,11 +829,30 @@ def main():
            test_prompt = tokenizer.apply_chat_template(
                messages,
                tokenize=False,
-                add_generation_prompt=True,
+                continue_final_message=has_prefill,
+                add_generation_prompt=not has_prefill,
            )
            prompt = prompt[test_prompt.index("<query>") :]
+        prompt = tokenizer.encode(prompt, add_special_tokens=False)
+    else:
+        prompt = tokenizer.encode(prompt)

-    sampler = make_sampler(args.temp, args.top_p, args.min_p, args.min_tokens_to_keep)
+    if args.draft_model is not None:
+        draft_model, draft_tokenizer = load(args.draft_model)
+        if draft_tokenizer.vocab_size != tokenizer.vocab_size:
+            raise ValueError("Draft model tokenizer does not match model tokenizer.")
+    else:
+        draft_model = None
+    sampler = make_sampler(
+        args.temp,
+        args.top_p,
+        args.min_p,
+        args.min_tokens_to_keep,
+        top_k=args.top_k,
+        xtc_probability=args.xtc_probability,
+        xtc_threshold=args.xtc_threshold,
+        xtc_special_tokens=tokenizer.encode("\n") + list(tokenizer.eos_token_ids),
+    )
    response = generate(
        model,
        tokenizer,
@@ -233,10 +865,16 @@ def main():
        kv_bits=args.kv_bits,
        kv_group_size=args.kv_group_size,
        quantized_kv_start=args.quantized_kv_start,
+        draft_model=draft_model,
+        num_draft_tokens=args.num_draft_tokens,
    )
    if not args.verbose:
        print(response)


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.generate...` directly is deprecated."
+        " Use `mlx_lm.generate...` or `python -m mlx_lm generate ...` instead."
+    )
    main()
@@ -1,18 +1,18 @@
-# Copyright © 2024 Apple Inc.
-
 import argparse
 import math
+import os
 import re
 import types
 from pathlib import Path

+import mlx.core as mx
 import mlx.nn as nn
 import mlx.optimizers as optim
 import numpy as np
 import yaml

-from .tokenizer_utils import TokenizerWrapper
-from .tuner.datasets import load_dataset
+from .tuner.callbacks import WandBCallback
+from .tuner.datasets import CacheDataset, load_dataset
 from .tuner.trainer import TrainingArgs, TrainingCallback, evaluate, train
 from .tuner.utils import (
    build_schedule,
@@ -42,6 +42,11 @@ CONFIG_DEFAULTS = {
    "model": "mlx_model",
    "train": False,
    "fine_tune_type": "lora",
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+    },
    "data": "data/",
    "seed": 0,
    "num_layers": 16,
@@ -57,8 +62,12 @@ CONFIG_DEFAULTS = {
    "test": False,
    "test_batches": 500,
    "max_seq_length": 2048,
+    "config": None,
+    "grad_checkpoint": False,
    "lr_schedule": None,
-    "lora_parameters": {"rank": 8, "alpha": 16, "dropout": 0.0, "scale": 10.0},
+    "lora_parameters": {"rank": 8, "dropout": 0.0, "scale": 20.0},
+    "mask_prompt": False,
+    "wandb": None,
 }


@@ -66,6 +75,7 @@ def build_parser():
    parser = argparse.ArgumentParser(description="LoRA or QLoRA finetuning.")
    parser.add_argument(
        "--model",
+        type=str,
        help="The path to the local model directory or Hugging Face repo.",
    )

@@ -88,9 +98,21 @@ def build_parser():
        "--fine-tune-type",
        type=str,
        choices=["lora", "dora", "full"],
-        default="lora",
        help="Type of fine-tuning to perform: lora, dora, or full.",
    )
+    parser.add_argument(
+        "--optimizer",
+        type=str,
+        choices=["adam", "adamw"],
+        default=None,
+        help="Optimizer to use for training: adam or adamw",
+    )
+    parser.add_argument(
+        "--mask-prompt",
+        action="store_true",
+        help="Mask the prompt in the loss when training",
+        default=None,
+    )
    parser.add_argument(
        "--num-layers",
        type=int,
@@ -148,7 +170,7 @@ def build_parser():
    parser.add_argument(
        "-c",
        "--config",
-        default=None,
+        type=str,
        help="A YAML configuration file with the training options",
    )
    parser.add_argument(
@@ -157,22 +179,36 @@ def build_parser():
        help="Use gradient checkpointing to reduce memory use.",
        default=None,
    )
-    parser.add_argument("--seed", type=int, default=None, help="The PRNG seed")
+    parser.add_argument(
+        "--wandb",
+        type=str,
+        default=None,
+        help="WandB project name to report training metrics. Disabled if None.",
+    )
+    parser.add_argument("--seed", type=int, help="The PRNG seed")
    return parser


 def train_model(
    args,
    model: nn.Module,
-    tokenizer: TokenizerWrapper,
    train_set,
    valid_set,
    training_callback: TrainingCallback = None,
 ):
+    mx.random.seed(args.seed)
    model.freeze()
+    if args.num_layers > len(model.layers):
+        raise ValueError(
+            f"Requested to train {args.num_layers} layers "
+            f"but the model only has {len(model.layers)} layers."
+        )
+
    if args.fine_tune_type == "full":
-        for l in model.layers[-min(args.num_layers, 0) :]:
+        for l in model.layers[-max(args.num_layers, 0) :]:
            l.unfreeze()
+
+        args.lora_parameters = None
    elif args.fine_tune_type in ["lora", "dora"]:
        # Convert linear layers to lora/dora layers and unfreeze in the process
        linear_to_lora_layers(
@@ -210,31 +246,36 @@ def train_model(
        grad_checkpoint=args.grad_checkpoint,
    )

-    model.train()
-    opt = optim.Adam(
-        learning_rate=(
-            build_schedule(args.lr_schedule) if args.lr_schedule else args.learning_rate
-        )
-    )
+    # Initialize the selected optimizer
+    lr = build_schedule(args.lr_schedule) if args.lr_schedule else args.learning_rate
+
+    optimizer_name = args.optimizer.lower()
+    optimizer_config = args.optimizer_config.get(optimizer_name, {})
+
+    if optimizer_name == "adam":
+        opt_class = optim.Adam
+    elif optimizer_name == "adamw":
+        opt_class = optim.AdamW
+    else:
+        raise ValueError(f"Unsupported optimizer: {optimizer_name}")
+
+    opt = opt_class(learning_rate=lr, **optimizer_config)
+
    # Train model
    train(
        model=model,
-        tokenizer=tokenizer,
        args=training_args,
        optimizer=opt,
-        train_dataset=train_set,
-        val_dataset=valid_set,
+        train_dataset=CacheDataset(train_set),
+        val_dataset=CacheDataset(valid_set),
        training_callback=training_callback,
    )


-def evaluate_model(args, model: nn.Module, tokenizer: TokenizerWrapper, test_set):
-    model.eval()
-
+def evaluate_model(args, model: nn.Module, test_set):
    test_loss = evaluate(
        model=model,
-        dataset=test_set,
-        tokenizer=tokenizer,
+        dataset=CacheDataset(test_set),
        batch_size=args.batch_size,
        num_batches=args.test_batches,
        max_seq_length=args.max_seq_length,
@@ -248,6 +289,14 @@ def evaluate_model(args, model: nn.Module, tokenizer: TokenizerWrapper, test_set
 def run(args, training_callback: TrainingCallback = None):
    np.random.seed(args.seed)

+    if args.wandb is not None:
+        training_callback = WandBCallback(
+            project_name=args.wandb,
+            log_dir=args.adapter_path,
+            config=vars(args),
+            wrapped_callback=training_callback,
+        )
+
    print("Loading pretrained model")
    model, tokenizer = load(args.model)

@@ -261,16 +310,17 @@ def run(args, training_callback: TrainingCallback = None):

    elif args.train:
        print("Training")
-        train_model(args, model, tokenizer, train_set, valid_set, training_callback)
+        train_model(args, model, train_set, valid_set, training_callback)
    else:
        raise ValueError("Must provide at least one of --train or --test")

    if args.test:
        print("Testing")
-        evaluate_model(args, model, tokenizer, test_set)
+        evaluate_model(args, model, test_set)


 def main():
+    os.environ["TOKENIZERS_PARALLELISM"] = "true"
    parser = build_parser()
    args = parser.parse_args()
    config = args.config
@@ -292,4 +342,8 @@ def main():


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.lora...` directly is deprecated."
+        " Use `mlx_lm.lora...` or `python -m mlx_lm lora ...` instead."
+    )
    main()
@@ -2,23 +2,37 @@ import argparse
 from typing import List, Union

 from huggingface_hub import scan_cache_dir
-from transformers.commands.user import tabulate
+
+
+def tabulate(rows: List[List[Union[str, int]]], headers: List[str]) -> str:
+    """
+    Inspired by:
+    - stackoverflow.com/a/8356620/593036
+    - stackoverflow.com/questions/9535954/printing-lists-as-tabular-data
+    """
+    col_widths = [max(len(str(x)) for x in col) for col in zip(*rows, headers)]
+    row_format = ("{{:{}}} " * len(headers)).format(*col_widths)
+    lines = []
+    lines.append(row_format.format(*headers))
+    lines.append(row_format.format(*["-" * w for w in col_widths]))
+    for row in rows:
+        lines.append(row_format.format(*row))
+    return "\n".join(lines)


 def ask_for_confirmation(message: str) -> bool:
+    """Ask user for confirmation with Y/N prompt.
+    Returns True for Y/yes, False for N/no/empty."""
    y = ("y", "yes", "1")
-    n = ("n", "no", "0")
-    all_values = y + n + ("",)
-    full_message = f"{message} (Y/n) "
+    n = ("n", "no", "0", "")
+    full_message = f"{message} (y/n) "
    while True:
        answer = input(full_message).lower()
-        if answer == "":
-            return False
        if answer in y:
            return True
        if answer in n:
            return False
-        print(f"Invalid input. Must be one of {all_values}")
+        print(f"Invalid input. Must be one of: yes/no/y/n or empty for no")


 def main():
@@ -43,9 +57,7 @@ def main():
    args = parser.parse_args()

    if args.scan:
-        print(
-            "Scanning Hugging Face cache for models with" f'pattern "{args.pattern}".'
-        )
+        print(f'Scanning Hugging Face cache for models with pattern "{args.pattern}".')
        hf_cache_info = scan_cache_dir()
        print(
            tabulate(
@@ -86,36 +98,46 @@ def main():
            if args.pattern in repo.repo_id
        ]
        if repos:
+            print("\nFound the following models:")
            print(
                tabulate(
                    rows=[
                        [
                            repo.repo_id,
+                            repo.size_on_disk_str,  # Added size information
                            str(repo.repo_path),
                        ]
                        for repo in repos
                    ],
                    headers=[
                        "REPO ID",
+                        "SIZE",  # Added size header
                        "LOCAL PATH",
                    ],
                )
            )

-            confirmed = ask_for_confirmation(f"Confirm deletion ?")
+            confirmed = ask_for_confirmation(
+                "\nAre you sure you want to delete these models?"
+            )
            if confirmed:
                for model_info in repos:
+                    print(f"\nDeleting {model_info.repo_id}...")
                    for revision in sorted(
                        model_info.revisions, key=lambda revision: revision.commit_hash
                    ):
                        strategy = hf_cache_info.delete_revisions(revision.commit_hash)
                        strategy.execute()
-                print("Model(s) deleted.")
+                print("\nModel(s) deleted successfully.")
            else:
-                print("Deletion is cancelled. Do nothing.")
+                print("\nDeletion cancelled - no changes made.")
        else:
-            print(f"No models found.")
+            print(f'No models found matching pattern "{args.pattern}"')


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.manage...` directly is deprecated."
+        " Use `mlx_lm.manage...` or `python -m mlx_lm manage ...` instead."
+    )
    main()
@@ -1,172 +0,0 @@
-# Copyright © 2023-2024 Apple Inc.
-
-import argparse
-import glob
-import shutil
-from pathlib import Path
-from typing import Optional
-
-import mlx.core as mx
-import mlx.nn as nn
-import numpy as np
-import yaml
-from mlx.utils import tree_flatten, tree_map
-
-from .utils import (
-    fetch_from_hub,
-    get_model_path,
-    save_config,
-    save_weights,
-    upload_to_hub,
-)
-
-
-def configure_parser() -> argparse.ArgumentParser:
-    """
-    Configures and returns the argument parser for the script.
-
-    Returns:
-        argparse.ArgumentParser: Configured argument parser.
-    """
-    parser = argparse.ArgumentParser(description="Merge multiple models.")
-
-    parser.add_argument("--config", type=str, help="Path to the YAML config.")
-    parser.add_argument(
-        "--mlx-path",
-        type=str,
-        default="mlx_merged_model",
-        help="Path to save the MLX model.",
-    )
-    parser.add_argument(
-        "--upload-repo",
-        help="The Hugging Face repo to upload the model to.",
-        type=str,
-        default=None,
-    )
-    return parser
-
-
-def slerp(t, w1, w2, eps=1e-5):
-    """
-    Spherical linear interpolation
-
-    Args:
-        t (float): Interpolation weight in [0.0, 1.0]
-        w1 (mx.array): First input
-        w2 (mx.array): Second input
-        eps (float): Constant for numerical stability
-    Returns:
-        mx.array: Interpolated result
-    """
-    t = float(t)
-    if t == 0:
-        return w1
-    elif t == 1:
-        return w2
-    # Normalize
-    v1 = w1 / mx.linalg.norm(w1)
-    v2 = w2 / mx.linalg.norm(w2)
-    # Angle
-    dot = mx.clip((v1 * v2).sum(), 0.0, 1.0)
-    theta = mx.arccos(dot)
-    sin_theta = mx.sin(theta + eps)
-    s1 = mx.sin(theta * (1 - t)) / sin_theta
-    s2 = mx.sin(theta * t) / sin_theta
-    return s1 * w1 + s2 * w2
-
-
-def merge_models(base_model: nn.Module, model: nn.Module, config: dict):
-    method = config.get("method", None)
-    if method != "slerp":
-        raise ValueError(f"Merge method {method} not supported")
-
-    num_layers = len(model.layers)
-
-    def unpack_values(vals):
-        if isinstance(vals, (int, float)):
-            return np.full(num_layers, vals)
-        bins = len(vals) - 1
-        sizes = [num_layers // bins] * bins
-        sizes[-1] = num_layers - sum(sizes[:-1])
-        return np.concatenate(
-            [np.linspace(v1, v2, s) for v1, v2, s in zip(vals[:-1], vals[1:], sizes)]
-        )
-
-    param_list = config["parameters"]["t"]
-    params = {}
-    filter_keys = set()
-    for pl in param_list[:-1]:
-        params[pl["filter"]] = unpack_values(pl["value"])
-        filter_keys.add(pl["filter"])
-    default = unpack_values(param_list[-1]["value"])
-
-    for e in range(num_layers):
-        bl = base_model.layers[e]
-        l = model.layers[e]
-        base_weights = bl.parameters()
-        weights = l.parameters()
-        for k, w1 in base_weights.items():
-            w2 = weights[k]
-            t = params.get(k, default)[e]
-            base_weights[k] = tree_map(lambda x, y: slerp(t, x, y), w1, w2)
-        base_model.update(base_weights)
-
-
-def merge(
-    config: str,
-    mlx_path: str = "mlx_model",
-    upload_repo: Optional[str] = None,
-):
-    with open(config, "r") as fid:
-        merge_conf = yaml.safe_load(fid)
-    print("[INFO] Loading")
-
-    model_paths = merge_conf.get("models", [])
-    if len(model_paths) < 2:
-        raise ValueError(f"Expected at least 2 models, got {len(model_paths)}.")
-
-    # Load all models
-    base_hf_path = model_paths[0]
-    base_path = get_model_path(base_hf_path)
-    base_model, base_config, tokenizer = fetch_from_hub(base_path, lazy=True)
-    models = []
-    for mp in model_paths[1:]:
-        model, model_config, _ = fetch_from_hub(get_model_path(mp), lazy=True)
-        base_type = base_config["model_type"]
-        model_type = model_config["model_type"]
-        if base_type != model_type:
-            raise ValueError(
-                f"Can only merge models of the same type,"
-                f" but got {base_type} and {model_type}."
-            )
-        models.append(model)
-
-    # Merge models into base model
-    for m in models:
-        merge_models(base_model, m, merge_conf)
-
-    # Save base model
-    mlx_path = Path(mlx_path)
-    weights = dict(tree_flatten(base_model.parameters()))
-    del models, base_model
-    save_weights(mlx_path, weights, donate_weights=True)
-    py_files = glob.glob(str(base_path / "*.py"))
-    for file in py_files:
-        shutil.copy(file, mlx_path)
-
-    tokenizer.save_pretrained(mlx_path)
-
-    save_config(config, config_path=mlx_path / "config.json")
-
-    if upload_repo is not None:
-        upload_to_hub(mlx_path, upload_repo, base_hf_path)
-
-
-def main():
-    parser = configure_parser()
-    args = parser.parse_args()
-    merge(**vars(args))
-
-
-if __name__ == "__main__":
-    main()
@@ -0,0 +1,397 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from functools import partial
+from itertools import accumulate
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import ConcatenateKVCache, KVCache
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_dim: int
+    num_layers: int
+    num_kv_reuse_layers: int
+    num_heads: int
+    num_kv_heads: int
+    hidden_dim_scale_factor: float = 3.25
+    rope_theta: float = 50000
+    rms_norm_eps: float = 1e-5
+
+
+class FusedLoRALinear(nn.Module):
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: list[int],
+        r: int = 8,
+        dropout: float = 0.0,
+        scale: float = 20.0,
+    ):
+        super().__init__()
+
+        self.linear = FusedLinear(input_dims, output_dims)
+        self.dropout = nn.Dropout(p=dropout)
+        self.scale = scale
+
+        scale = 1 / math.sqrt(input_dims)
+        self.lora_a = [
+            mx.random.uniform(low=-scale, high=scale, shape=(input_dims, r))
+            for _ in output_dims
+        ]
+        self.lora_b = [mx.zeros((r, od)) for od in output_dims]
+
+    def fuse(self, de_quantize: bool = False):
+        linear = self.linear
+        weight = linear.weight
+        is_quantized = isinstance(linear, FusedQuantizedLinear)
+
+        # Use the same type as the linear weight if not quantized
+        dtype = weight.dtype
+
+        if is_quantized:
+            dtype = linear.scales.dtype
+            weight = mx.dequantize(
+                weight,
+                linear.scales,
+                linear.biases,
+                linear.group_size,
+                linear.bits,
+            )
+
+        input_dims = weight.shape[-1]
+        output_dims = linear.output_dims
+        fused_linear = FusedLinear(input_dims, output_dims)
+        fused_linear.weight = weight
+        deltas = [
+            ((self.scale * b.T) @ a.T).astype(dtype)
+            for a, b in zip(self.lora_a, self.lora_b)
+        ]
+        delta = mx.concatenate(deltas, axis=0)
+        fused_linear.weight = weight + delta
+
+        if is_quantized and not de_quantize:
+            fused_linear = fused_linear.to_quantized(linear.group_size, linear.bits)
+
+        return fused_linear
+
+    def __call__(self, x):
+        dt = x.dtype
+        y = self.linear(x)
+        x = self.dropout(x)
+        z = [(x @ a) @ b for a, b in zip(self.lora_a, self.lora_b)]
+        return tuple(yi + (self.scale * zi).astype(dt) for yi, zi in zip(y, z))
+
+
+class FusedQuantizedLinear(nn.QuantizedLinear):
+    def __init__(self, input_dims, output_dims, group_size: int = 64, bits: int = 4):
+        *indices, output_dims = accumulate(output_dims)
+        self.indices = indices
+        super().__init__(
+            input_dims, output_dims, bias=False, group_size=group_size, bits=bits
+        )
+
+    @property
+    def input_dims(self):
+        return self.scales.shape[-1] * self.group_size
+
+    @property
+    def output_dims(self):
+        indices = [0] + self.indices + [self.weight.shape[0]]
+        return [indices[i] - indices[i - 1] for i in range(1, len(indices))]
+
+    def __call__(self, x):
+        x = super().__call__(x)
+        return x.split(self.indices, axis=-1)
+
+    def to_lora(self, r: int = 8, dropout: float = 0.0, scale: float = 20.0):
+        lora_lin = FusedLoRALinear(self.input_dims, self.output_dims, r, dropout, scale)
+        lora_lin.linear = self
+        return lora_lin
+
+
+class FusedLinear(nn.Linear):
+    def __init__(self, input_dims, output_dims):
+        *indices, output_dims = accumulate(output_dims)
+        self.indices = indices
+        super().__init__(input_dims, output_dims, bias=False)
+
+    @property
+    def input_dims(self):
+        return self.weight.shape[-1]
+
+    @property
+    def output_dims(self):
+        indices = [0] + self.indices + [self.weight.shape[0]]
+        return [indices[i] - indices[i - 1] for i in range(1, len(indices))]
+
+    def __call__(self, x):
+        x = super().__call__(x)
+        return x.split(self.indices, axis=-1)
+
+    def to_quantized(self, group_size: int = 64, bits: int = 4):
+        input_dims = self.input_dims
+        output_dims = self.output_dims
+        ql = FusedQuantizedLinear(input_dims, output_dims, group_size, bits)
+        ql.weight, ql.scales, ql.biases = mx.quantize(self.weight, group_size, bits)
+
+        return ql
+
+    def to_lora(self, r: int = 8, dropout: float = 0.0, scale: float = 20.0):
+        lora_lin = FusedLoRALinear(self.input_dims, self.output_dims, r, dropout, scale)
+        lora_lin.linear = self
+        return lora_lin
+
+
+@partial(mx.compile, shapeless=True)
+def fake_8bit_quant(x, scale):
+    dt = x.dtype
+    x = x.astype(mx.float32)
+    x = (x / scale).round()
+    x = mx.clip(x, -128, 127)
+    return (x * scale).astype(dt)
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_dim
+        self.n_heads = n_heads = args.num_heads
+        self.n_kv_heads = n_kv_heads = args.num_kv_heads
+        self.head_dim = head_dim = args.hidden_dim // n_heads
+        self.scale = head_dim**-0.5
+
+        qkv_dim = (n_heads + 2 * n_kv_heads) * head_dim
+        self.qkv_proj = FusedLinear(
+            dim, [n_heads * head_dim] + 2 * [n_kv_heads * head_dim]
+        )
+        self.out_proj = nn.Linear(dim, dim, bias=False)
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            True,
+        )
+        self.q_norm = nn.RMSNorm(head_dim)
+        self.k_norm = nn.RMSNorm(head_dim)
+        self.quant_key_scale = mx.array(1.0)
+        self.quant_value_scale = mx.array(1.0)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        # Get the queries, keys and values
+        queries, keys, values = self.qkv_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.q_norm(self.rope(queries, offset=cache.offset))
+            keys = self.k_norm(self.rope(keys, offset=cache.offset))
+            keys = fake_8bit_quant(keys, self.quant_key_scale)
+            values = fake_8bit_quant(values, self.quant_value_scale)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.q_norm(self.rope(queries))
+            keys = self.k_norm(self.rope(keys))
+            keys = fake_8bit_quant(keys, self.quant_key_scale)
+            values = fake_8bit_quant(values, self.quant_value_scale)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.out_proj(output)
+
+
+class KVReuseAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_dim
+        self.n_heads = n_heads = args.num_heads
+        self.head_dim = head_dim = args.hidden_dim // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, dim, bias=False)
+        self.out_proj = nn.Linear(dim, dim, bias=False)
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            True,
+        )
+        self.q_norm = nn.RMSNorm(head_dim)
+
+    def __call__(
+        self,
+        x: mx.array,
+        keys: mx.array,
+        values: mx.array,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+        _, _, S, _ = keys.shape
+
+        queries = self.q_proj(x)
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        queries = self.q_norm(self.rope(queries, offset=S - L))
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=None, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.out_proj(output)
+
+
+@partial(mx.compile, shapeless=True)
+def _swiglu(g, x):
+    return nn.silu(g) * x
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_dim
+        hidden_dim = int(dim * args.hidden_dim_scale_factor)
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        g = self.gate_proj(x)
+        x = self.up_proj(x)
+        return self.down_proj(_swiglu(g, x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_dim, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_dim, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class KVReuseTransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = KVReuseAttention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_dim, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_dim, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        keys: mx.array,
+        values: mx.array,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), keys, values, mask)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class AFMModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+
+        self.embedding = nn.Embedding(args.vocab_size, args.hidden_dim)
+        self.layers = [
+            TransformerBlock(args)
+            for _ in range(args.num_layers - args.num_kv_reuse_layers)
+        ]
+        self.kv_reuse_layers = [
+            KVReuseTransformerBlock(args) for _ in range(args.num_kv_reuse_layers)
+        ]
+        self.output_norm = nn.RMSNorm(args.hidden_dim, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embedding(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+            cache[-1] = ConcatenateKVCache()
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        keys, values = cache[-1].state
+        for layer in self.kv_reuse_layers:
+            h = layer(h, keys, values, mask)
+
+        return self.output_norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = AFMModel(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        out = self.model.embedding.as_linear(out)
+        return out
+
+    def make_cache(self):
+        return [KVCache() for _ in range(len(self.model.layers))]
+
+    @property
+    def layers(self):
+        return self.model.layers + self.model.kv_reuse_layers
@@ -0,0 +1,226 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import CacheList, KVCache, MambaCache, RotatingKVCache
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    rope_theta: float
+    sliding_window: int
+    sliding_window_layers: List[int]
+    conv_window: int
+    rms_norm_eps: float
+    model_type: str = "baichuan_m1"
+    num_swa_attention_heads: Optional[int] = None
+    num_swa_key_value_heads: Optional[int] = None
+    tie_word_embeddings: bool = False
+
+
+class Attention(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            raise ValueError("Layer index must be provided to Attention module.")
+
+        self.is_swa = layer_idx in config.sliding_window_layers
+        self.num_heads = (
+            config.num_swa_attention_heads
+            if self.is_swa and config.num_swa_attention_heads
+            else config.num_attention_heads
+        )
+        self.num_kv_heads = (
+            config.num_swa_key_value_heads
+            if self.is_swa and config.num_swa_key_value_heads
+            else config.num_key_value_heads
+        )
+
+        self.hidden_size = config.hidden_size
+        self.head_dim = self.hidden_size // self.num_heads
+        assert self.head_dim * self.num_heads == self.hidden_size
+
+        self.scale = self.head_dim**-0.5
+
+        self.W_pack = nn.Linear(
+            config.hidden_size,
+            self.hidden_size + 2 * self.num_kv_heads * self.head_dim,
+            bias=False,
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, config.hidden_size, bias=False
+        )
+
+        self.rope = nn.RoPE(self.head_dim, traditional=False, base=config.rope_theta)
+
+        self.conv_window = config.conv_window
+        assert self.conv_window == 2
+        self.conv_k = mx.zeros((1, 1, self.num_kv_heads, 1, self.conv_window))
+        self.conv_v = mx.zeros((1, 1, self.num_kv_heads, 1, self.conv_window))
+
+    def _custom_convolution(self, u, weights, state=None):
+        B, H, L, D = u.shape
+        weights = weights.reshape((1, H, self.conv_window, 1, 1))
+        w0 = weights[:, :, 0]
+        w1 = weights[:, :, 1]
+        if state is None:
+            state = mx.zeros((B, H, 1, D), u.dtype)
+        if L > 1:
+            u_prev = mx.concatenate([state, u[:, :, :-1]], axis=2)
+        else:
+            u_prev = state
+        return u_prev * w0 + u * w1
+
+    def __call__(
+        self, x: mx.array, mask: mx.array = None, cache: Any = None
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        proj = self.W_pack(x)
+        q, k, v = mx.split(proj, (D, D + self.num_kv_heads * self.head_dim), axis=-1)
+
+        q = q.reshape(B, L, self.num_heads, self.head_dim).transpose(0, 2, 1, 3)
+        k = k.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+        v = v.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            offset = cache[1].offset
+            last_k, last_v = cache[0][0], cache[0][1]
+        else:
+            offset = 0
+            last_k, last_v = None, None
+
+        k_init = k
+        v_init = v
+        k = self._custom_convolution(k, self.conv_k, state=last_k)
+        v = self._custom_convolution(v, self.conv_v, state=last_v)
+        q = self.rope(q, offset=offset)
+        k = self.rope(k, offset=offset)
+
+        if cache is not None:
+            k, v = cache[1].update_and_fetch(k, v)
+            if L > 0:
+                cache[0][0] = k_init[:, :, -1:, :]
+                cache[0][1] = v_init[:, :, -1:, :]
+
+        out = scaled_dot_product_attention(
+            q, k, v, cache=cache[1], scale=self.scale, mask=mask
+        )
+        out = out.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(out)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.gate_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.up_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.down_proj = nn.Linear(
+            config.intermediate_size, config.hidden_size, bias=False
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = Attention(config, layer_idx)
+        self.mlp = MLP(config)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def __call__(
+        self, x: mx.array, mask: mx.array = None, cache: Any = None
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        x = x + r
+        r = self.mlp(self.post_attention_layernorm(x))
+        return x + r
+
+
+class BaichuanModel(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [DecoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self, inputs: mx.array, mask: mx.array = None, cache: Any = None
+    ) -> mx.array:
+        x = self.embed_tokens(inputs)
+        if mask is None:
+            if cache is not None:
+                c = [cache[0][1]]
+            mask = create_attention_mask(x, c)
+        if cache is None:
+            cache = [None] * len(self.layers)
+        for layer, c in zip(self.layers, cache):
+            x = layer(x, mask, c)
+        return self.norm(x)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        self.model = BaichuanModel(config)
+        self.tie_word_embeddings = config.tie_word_embeddings
+        if not config.tie_word_embeddings:
+            self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def make_cache(self) -> List[Any]:
+        caches = []
+        for i, layer in enumerate(self.model.layers):
+            is_swa = i in self.config.sliding_window_layers
+            conv_cache = MambaCache()
+            if is_swa:
+                kv_cache = RotatingKVCache(max_size=self.config.sliding_window)
+            else:
+                kv_cache = KVCache()
+            caches.append(CacheList(conv_cache, kv_cache))
+        return caches
+
+    def sanitize(self, weights: dict) -> dict:
+        is_quantized = "lm_head.scales" in weights
+        if not is_quantized and "lm_head.weight" in weights:
+            w = weights["lm_head.weight"]
+            dtype = w.dtype
+            w = w.astype(mx.float32)
+            norm = mx.linalg.norm(w, axis=-1, keepdims=True)
+            w = (w / (norm + 1e-7)).astype(dtype)
+            weights["lm_head.weight"] = w
+        return weights
+
+    def __call__(
+        self, inputs: mx.array, mask: mx.array = None, cache: Any = None
+    ) -> mx.array:
+        outputs = self.model(inputs, mask, cache)
+        return self.lm_head(outputs)
+
+    @property
+    def layers(self) -> List[nn.Module]:
+        return self.model.layers
@@ -23,31 +23,43 @@ class BaseModelArgs:
        )


-def create_causal_mask(N: int, offset: int = 0, window_size: Optional[int] = None):
+def create_causal_mask(
+    N: int,
+    offset: int = 0,
+    window_size: Optional[int] = None,
+    lengths: Optional[mx.array] = None,
+):
    rinds = mx.arange(offset + N)
    linds = mx.arange(offset, offset + N) if offset else rinds
    linds = linds[:, None]
    rinds = rinds[None]
-    mask = linds < rinds
+    mask = linds >= rinds
    if window_size is not None:
-        mask = mask | (linds > rinds + window_size)
-    return mask * -1e9
+        mask = mask & (linds <= rinds + window_size)
+    if lengths is not None:
+        lengths = lengths[:, None, None, None]
+        mask = mask & (rinds < lengths)
+    return mask


-def create_attention_mask(h: mx.array, cache: Optional[Any] = None):
+def create_attention_mask(
+    h: mx.array, cache: Optional[Any] = None, return_array: bool = False
+):
    T = h.shape[1]
    if T > 1:
-        window_size = None
        offset = 0
+        window_size = None
        if cache is not None and cache[0] is not None:
            c = cache[0]
+            offset = c.offset
            if hasattr(c, "max_size"):
-                offset = min(c.max_size, c.offset)
                window_size = c.max_size
-            else:
-                offset = c.offset
-        mask = create_causal_mask(T, offset, window_size=window_size)
-        mask = mask.astype(h.dtype)
+                offset = min(window_size, offset)
+                return_array = return_array or offset + T > window_size
+        if return_array:
+            return create_causal_mask(T, offset, window_size=window_size)
+        else:
+            return "causal"
    else:
        mask = None
    return mask
@@ -77,7 +89,15 @@ def quantized_scaled_dot_product_attention(
        queries, *q_keys, transpose=True, group_size=group_size, bits=bits
    )
    if mask is not None:
-        scores += mask
+        if isinstance(mask, str):
+            qL, kL = scores.shape[-2:]
+            q_indices = mx.arange(kL - qL, kL)
+            k_indices = mx.arange(kL)
+            mask = q_indices[:, None] >= k_indices[None]
+        if mask.dtype == mx.bool_:
+            scores = mx.where(mask, scores, mx.finfo(scores.dtype).min)
+        else:
+            scores += mask
    scores = mx.softmax(scores, axis=-1, precise=True)
    out = mx.quantized_matmul(
        scores, *q_values, transpose=False, group_size=group_size, bits=bits
@@ -0,0 +1,131 @@
+# Copyright © 2025 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.quantized import QuantizedLinear
+
+
+def make_bitlinear_kernel():
+    """
+    Custom Metal kernel that performs matrix multiplication directly on
+    packed weights and scales the output. This eliminates the need to
+    store unpacked weights in memory.
+    """
+    source = """
+    constexpr int M = 4;
+    constexpr int BLOCK = 32;
+
+    uint tid = thread_position_in_grid.y;
+    uint in_offset = thread_position_in_grid.x;
+
+    uint batch_idx = tid / (out_features / 4);
+    uint row_idx = tid % (out_features / 4);
+
+    float sum[4] = {0.0};
+
+    for (uint i = in_offset * M; i < in_features; i += BLOCK * M) {
+        float v[M];
+        for (int j=0; j<M; j++) {
+            v[j] = x[batch_idx * in_features + i + j];
+        }
+
+        for (int j=0; j<M; j++) {
+            uint8_t w = packed_weights[row_idx * in_features + i + j];
+            sum[0] += v[j] * ((w & 3) - 1);
+            sum[1] += v[j] * (((w >> 2) & 3) - 1);
+            sum[2] += v[j] * (((w >> 4) & 3) - 1);
+            sum[3] += v[j] * (((w >> 6) & 3) - 1);
+        }
+    }
+
+    for (int j=0; j<4; j++) {
+        sum[j] = simd_sum(sum[j]);
+    }
+
+    // Apply weight scaling by diving them or multiplying them
+    if (in_offset == 0) {
+        float scale = invert_weight_scales ? 1 / weight_scale[0] : weight_scale[0];
+        for (int i=0; i<4; i++) {
+            out[batch_idx * out_features + row_idx + i * (out_features/4)] = static_cast<T>(sum[i] * scale);
+        }
+    }
+    """
+
+    return mx.fast.metal_kernel(
+        name="bitlinear_matmul",
+        input_names=["x", "packed_weights", "weight_scale"],
+        output_names=["out"],
+        source=source,
+    )
+
+
+_bitlinear_kernel = make_bitlinear_kernel()
+
+
+class BitLinear(nn.Module):
+    """
+    BitLinear module with memory-efficient weight handling.
+    """
+
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        bias=True,
+        invert_weight_scales=False,
+    ):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        # Calculate packed dimensions - the first dimension gets packed 4:1
+        # The weights are ternary so can be represented with 2 bits, and they
+        # are packed in uint8 tensors, hence the number of values per item is 4
+        packed_out_features = (out_features + 3) // 4
+        self.weight = mx.zeros((packed_out_features, in_features), dtype=mx.uint8)
+
+        self.invert_weight_scales = invert_weight_scales
+        self.weight_scale = mx.array([1.0])
+
+        if bias:
+            self.bias = mx.zeros((out_features,))
+        else:
+            self.bias = None
+
+    def execute_matmul_kernel(self, x, packed_weights):
+        original_shape = x.shape
+        if len(original_shape) > 2:
+            x = x.reshape(-1, original_shape[-1])
+        total_batch_elements, in_features = x.shape
+
+        out_features = self.out_features
+
+        dtype = self.weight_scale.dtype
+        assert x.dtype == dtype, "Wrong type for input."
+        out = _bitlinear_kernel(
+            inputs=[
+                x,
+                packed_weights,
+                self.weight_scale,
+            ],
+            template=[
+                ("T", dtype),
+                ("invert_weight_scales", self.invert_weight_scales),
+                ("in_features", in_features),
+                ("out_features", out_features),
+            ],
+            grid=(32, total_batch_elements * out_features // 4, 1),
+            threadgroup=(32, 1, 1),  # SIMD width is 32 threads
+            output_shapes=[(total_batch_elements, out_features)],
+            output_dtypes=[dtype],
+        )[0]
+
+        if len(original_shape) > 2:
+            out = out.reshape(*original_shape[:-1], out_features)
+        return out
+
+    def __call__(self, x):
+        y = self.execute_matmul_kernel(x, self.weight)
+
+        if self.bias is not None:
+            y = mx.add(y, self.bias)
+        return y
@@ -0,0 +1,215 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .bitlinear_layers import BitLinear
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    head_dim: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
+    attention_bias: bool = False
+    mlp_bias: bool = False
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.head_dim or args.hidden_size // n_heads
+
+        self.scale = head_dim**-0.5
+        attention_bias = args.attention_bias
+
+        self.q_proj = BitLinear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = BitLinear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = BitLinear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = BitLinear(n_heads * head_dim, dim, bias=attention_bias)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            args.rope_traditional,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+        self.attn_sub_norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        output = self.attn_sub_norm(output)
+        output = self.o_proj(output)
+
+        return output
+
+
+@partial(mx.compile, shapeless=True)
+def relu2(x):
+    return mx.square(nn.relu(x))
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        hidden_dim = args.intermediate_size
+        if hasattr(args, "mlp_bias"):
+            mlp_bias = args.mlp_bias
+        else:
+            mlp_bias = False
+
+        self.gate_proj = BitLinear(dim, hidden_dim, bias=mlp_bias)
+        self.down_proj = BitLinear(hidden_dim, dim, bias=mlp_bias)
+        self.up_proj = BitLinear(dim, hidden_dim, bias=mlp_bias)
+        self.ffn_sub_norm = nn.RMSNorm(args.intermediate_size, eps=args.rms_norm_eps)
+
+    def __call__(self, x) -> mx.array:
+        x = relu2(self.gate_proj(x)) * self.up_proj(x)
+        x = self.ffn_sub_norm(x)
+        x = self.down_proj(x)
+        return x
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class LlamaModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = LlamaModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        # Remove unused precomputed rotary freqs
+        weights = {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -12,7 +12,7 @@ def make_prompt_cache(
    max_kv_size: Optional[int] = None,
 ) -> List[Any]:
    """
-    Construct the model's cache for use when cgeneration.
+    Construct the model's cache for use in generation.

    This function will defer the cache construction to the model if it has a
    ``make_cache`` method, otherwise it will make a default KV cache.
@@ -129,6 +129,40 @@ class _BaseCache:
        return False


+class ConcatenateKVCache(_BaseCache):
+    """ConcatenateKVCache the simplest KV cache implementation.
+
+    Can be used as a mock KV cache or when large blocks are being processed at
+    a time in which case KVCache isn't necessarily faster. Consider using the
+    KVCache with a larger step size before using this cache.
+    """
+
+    def __init__(self):
+        self.keys = None
+        self.values = None
+        self.offset = 0
+
+    def update_and_fetch(self, keys, values):
+        if self.keys is None:
+            self.keys = keys
+            self.values = values
+        else:
+            self.keys = mx.concatenate([self.keys, keys], axis=-2)
+            self.values = mx.concatenate([self.values, values], axis=-2)
+        self.offset = self.keys.shape[-2]
+
+        return self.keys, self.values
+
+    @property
+    def state(self):
+        return self.keys, self.values
+
+    @state.setter
+    def state(self, v):
+        self.keys, self.values = v
+        self.offset = self.keys.shape[-2]
+
+
 class QuantizedKVCache(_BaseCache):
    def __init__(self, group_size: int = 64, bits: int = 8):
        self.keys = None
@@ -436,3 +470,76 @@ class MambaCache(_BaseCache):
    @state.setter
    def state(self, v):
        self.cache = v
+
+
+class ChunkedKVCache(KVCache):
+    def __init__(self, chunk_size=None):
+        super().__init__()
+        self.chunk_size = chunk_size
+        self.start_position = 0
+
+    def maybe_trim_front(self):
+        # Maintain the cache below the chunk size
+        if self.keys is not None and self.keys.shape[2] >= self.chunk_size:
+            self.start_position += self.keys.shape[2] - self.chunk_size
+            self.keys = self.keys[..., -self.chunk_size :, :]
+            self.values = self.values[..., -self.chunk_size :, :]
+
+    def update_and_fetch(self, keys, values):
+        prev = self.offset - self.start_position
+        if self.keys is None or (prev + keys.shape[2]) > self.keys.shape[2]:
+            B, n_kv_heads, _, k_head_dim = keys.shape
+            v_head_dim = values.shape[3]
+            n_steps = (self.step + keys.shape[2] - 1) // self.step
+            k_shape = (B, n_kv_heads, n_steps * self.step, k_head_dim)
+            v_shape = (B, n_kv_heads, n_steps * self.step, v_head_dim)
+            new_k = mx.zeros(k_shape, keys.dtype)
+            new_v = mx.zeros(v_shape, values.dtype)
+            if self.keys is not None:
+                if prev % self.step != 0:
+                    self.keys = self.keys[..., :prev, :]
+                    self.values = self.values[..., :prev, :]
+                self.keys = mx.concatenate([self.keys, new_k], axis=2)
+                self.values = mx.concatenate([self.values, new_v], axis=2)
+            else:
+                self.keys, self.values = new_k, new_v
+
+        self.offset += keys.shape[2]
+        end = self.offset - self.start_position
+        self.keys[..., prev:end, :] = keys
+        self.values[..., prev:end, :] = values
+        return self.keys[..., :end, :], self.values[..., :end, :]
+
+    def trim(self, n):
+        n = min(self.offset - self.start_position, n)
+        self.offset -= n
+        return n
+
+    @property
+    def meta_state(self):
+        return tuple(map(str, (self.chunk_size, self.start_position)))
+
+    @meta_state.setter
+    def meta_state(self, v):
+        self.chunk_size, self.start_position = map(int, v)
+
+
+class CacheList(KVCache):
+    def __init__(self, *caches):
+        self.caches = caches
+
+    def __getitem__(self, idx):
+        return self.caches[idx]
+
+    @property
+    def state(self):
+        return [s for c in self.caches for s in c.state]
+
+    @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
@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Optional, Tuple
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -155,11 +155,13 @@ class CohereModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -180,9 +182,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.embed_tokens.as_linear(out)
        out = out * self.model.args.logit_scale
        return out
@@ -6,7 +6,7 @@ from typing import Optional, Tuple
 import mlx.core as mx
 import mlx.nn as nn

-from .base import BaseModelArgs, create_causal_mask, scaled_dot_product_attention
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
 from .cache import KVCache, RotatingKVCache


@@ -83,15 +83,22 @@ class Attention(nn.Module):
        if cache is not None:
            keys, values = cache.update_and_fetch(keys, values)

-        if self.use_sliding_window and mask is not None:
+        if self.use_sliding_window and isinstance(mask, mx.array):
            key_len = keys.shape[-2]
            if mask.shape[-1] != key_len:
                mask = mask[..., -key_len:]

+        # TODO: maybe remove cast once fused mask is supported since attention
+        # may be in higher precision
+        sdpa_type = mx.float32 if queries.dtype == mx.float16 else queries.dtype
        output = scaled_dot_product_attention(
-            queries, keys, values, cache=cache, scale=self.scale, mask=mask
-        )
-
+            queries.astype(sdpa_type),
+            keys,
+            values,
+            cache=cache,
+            scale=self.scale,
+            mask=mask,
+        ).astype(queries.dtype)
        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
        return self.o_proj(output)

@@ -126,9 +133,11 @@ class TransformerBlock(nn.Module):
        mask: Optional[mx.array] = None,
        cache: Optional[Tuple[mx.array, mx.array]] = None,
    ) -> mx.array:
+
        h = self.input_layernorm(x)
        attn_h = self.self_attn(h, mask, cache)
        ff_h = self.mlp(h)
+
        return attn_h + ff_h + x


@@ -151,22 +160,32 @@ class CohereModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        T = h.shape[1]
-        if T > 1:
-            offset = cache[0].offset if cache else 0
-            mask = create_causal_mask(T, offset).astype(h.dtype)
-        else:
-            mask = None
-
        if cache is None:
            cache = [None] * len(self.layers)

-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, c)
+        if mask is None:
+            j = self.args.sliding_window_pattern
+            full_mask = create_attention_mask(h, cache[j - 1 : j])
+            sliding_window_mask = create_attention_mask(h, cache)
+
+        for i, (layer, c) in enumerate(zip(self.layers, cache)):
+            is_global = (
+                i % self.args.sliding_window_pattern
+                == self.args.sliding_window_pattern - 1
+            )
+
+            local_mask = mask
+            if mask is None and is_global:
+                local_mask = full_mask
+            elif mask is None:
+                local_mask = sliding_window_mask
+
+            h = layer(h, local_mask, c)

        return self.norm(h)

@@ -181,9 +200,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.embed_tokens.as_linear(out)
        out = out * self.model.args.logit_scale
        return out
@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Optional, Tuple
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -105,10 +105,9 @@ class MLP(nn.Module):
        self.v1 = nn.Linear(d_model, ffn_dim, bias=False)
        self.w1 = nn.Linear(d_model, ffn_dim, bias=False)
        self.w2 = nn.Linear(ffn_dim, d_model, bias=False)
-        self.act_fn = nn.silu

    def __call__(self, x: mx.array) -> mx.array:
-        current_hidden_states = self.act_fn(self.w1(x)) * self.v1(x)
+        current_hidden_states = nn.silu(self.w1(x)) * self.v1(x)
        current_hidden_states = self.w2(current_hidden_states)
        return current_hidden_states

@@ -197,11 +196,13 @@ class DBRX(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.wte(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.blocks)
@@ -223,9 +224,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.transformer(inputs, cache)
+        out = self.transformer(inputs, mask, cache)
        return self.lm_head(out)

    @property
@@ -118,10 +118,9 @@ class DeepseekMLP(nn.Module):
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
-        self.act_fn = nn.silu

    def __call__(self, x: mx.array) -> mx.array:
-        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))


 class MoEGate(nn.Module):
@@ -211,9 +210,11 @@ class DeepseekModel(nn.Module):
        self,
        x: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ) -> mx.array:
        h = self.embed_tokens(x)
-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -236,8 +237,9 @@ class Model(nn.Module):
        self,
        inputs: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, cache, mask)
        return self.lm_head(out)

    def sanitize(self, weights):
@@ -2,7 +2,7 @@

 import math
 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Dict, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -148,7 +148,7 @@ class DeepseekV2Attention(nn.Module):
            self.q_a_proj = nn.Linear(
                self.hidden_size, self.q_lora_rank, bias=config.attention_bias
            )
-            self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank)
+            self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank, eps=1e-6)
            self.q_b_proj = nn.Linear(
                self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
            )
@@ -158,7 +158,7 @@ class DeepseekV2Attention(nn.Module):
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=config.attention_bias,
        )
-        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank)
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=1e-6)
        self.kv_b_proj = nn.Linear(
            self.kv_lora_rank,
            self.num_heads
@@ -282,12 +282,12 @@ class MoEGate(nn.Module):
        if self.topk_method == "group_limited_greedy":
            bsz, seq_len = x.shape[:2]
            scores = scores.reshape(bsz, seq_len, self.n_group, -1)
-            group_scores = scores.max(axis=-1)
+            group_scores = scores.max(axis=-1, keepdims=True)
            k = self.n_group - self.topk_group
-            group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-1)[..., :k]
-            batch_idx = mx.expand_dims(mx.arange(bsz), (1, 2))
-            seq_idx = mx.expand_dims(mx.arange(seq_len), (0, 2))
-            scores[batch_idx, seq_idx, group_idx] = 0.0
+            group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+            scores = mx.put_along_axis(
+                scores, group_idx, mx.array(0.0, scores.dtype), axis=-2
+            )
            scores = scores.reshape(bsz, seq_len, -1)

        k = self.top_k
@@ -364,21 +364,61 @@ class DeepseekV2Model(nn.Module):
            DeepseekV2DecoderLayer(config, idx)
            for idx in range(config.num_hidden_layers)
        ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

+        self.pipeline_rank = 0
+        self.pipeline_size = 1
+
+    def pipeline(self, group):
+        # Split layers in reverse so rank=0 gets the last layers and
+        # rank=pipeline_size-1 gets the first
+        self.pipeline_rank = group.rank()
+        self.pipeline_size = group.size()
+        layers_per_rank = len(self.layers) // self.pipeline_size
+        extra = len(self.layers) - layers_per_rank * self.pipeline_size
+        if self.pipeline_rank < extra:
+            layers_per_rank += 1
+
+        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
+        self.end_idx = self.start_idx + layers_per_rank
+        self.num_layers = layers_per_rank
+        self.layers = self.layers[: self.end_idx]
+        self.layers[: self.start_idx] = [None] * self.start_idx
+        self.num_layers = len(self.layers) - self.start_idx
+
    def __call__(
        self,
        x: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ) -> mx.array:
        h = self.embed_tokens(x)
-        mask = create_attention_mask(h, cache)
+
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
-            cache = [None] * len(self.layers)
+            cache = [None] * self.num_layers

-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, c)
+        # Receive from the previous process in the pipeline
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1))
+
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i])
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+
+        # Broadcast h while keeping it in the graph
+        h = mx.distributed.all_gather(h)[: h.shape[0]]

        return self.norm(h)

@@ -395,8 +435,9 @@ class Model(nn.Module):
        self,
        inputs: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, cache, mask)
        return self.lm_head(out)

    def sanitize(self, weights):
@@ -414,4 +455,4 @@ class Model(nn.Module):

    @property
    def layers(self):
-        return self.model.layers
+        return self.model.layers[self.model.start_idx : self.model.end_idx]
@@ -0,0 +1,545 @@
+# Copyright © 2024 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "deepseek_v3"
+    vocab_size: int = 102400
+    hidden_size: int = 4096
+    intermediate_size: int = 11008
+    moe_intermediate_size: int = 1407
+    num_hidden_layers: int = 30
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 32
+    n_shared_experts: Optional[int] = None
+    n_routed_experts: Optional[int] = None
+    routed_scaling_factor: float = 1.0
+    kv_lora_rank: int = 512
+    q_lora_rank: int = 1536
+    qk_rope_head_dim: int = 64
+    v_head_dim: int = 128
+    qk_nope_head_dim: int = 128
+    topk_method: str = "noaux_tc"
+    scoring_func: str = "sigmoid"
+    norm_topk_prob: bool = True
+    n_group: Optional[int] = None
+    topk_group: Optional[int] = None
+    num_experts_per_tok: Optional[int] = None
+    moe_layer_freq: int = 1
+    first_k_dense_replace: int = 0
+    max_position_embeddings: int = 2048
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 10000.0
+    rope_scaling: Dict = None
+    attention_bias: bool = False
+
+
+def yarn_find_correction_dim(
+    num_rotations, dim, base=10000, max_position_embeddings=2048
+):
+    return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (
+        2 * math.log(base)
+    )
+
+
+def yarn_find_correction_range(
+    low_rot, high_rot, dim, base=10000, max_position_embeddings=2048
+):
+    low = math.floor(
+        yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings)
+    )
+    high = math.ceil(
+        yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings)
+    )
+    return max(low, 0), min(high, dim - 1)
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+def yarn_linear_ramp_mask(min_val, max_val, dim):
+    if min_val == max_val:
+        max_val += 0.001  # Prevent singularity
+
+    linear_func = (mx.arange(dim, dtype=mx.float32) - min_val) / (max_val - min_val)
+    return mx.clip(linear_func, 0, 1)
+
+
+class DeepseekV3YarnRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim,
+        max_position_embeddings=2048,
+        base=10000,
+        scaling_factor=1.0,
+        original_max_position_embeddings=4096,
+        beta_fast=32,
+        beta_slow=1,
+        mscale=1,
+        mscale_all_dim=0,
+    ):
+        super().__init__()
+        self.mscale = yarn_get_mscale(scaling_factor, mscale) / yarn_get_mscale(
+            scaling_factor, mscale_all_dim
+        )
+        freq_extra = base ** (mx.arange(0, dim, 2, dtype=mx.float32) / dim)
+        freq_inter = scaling_factor * freq_extra
+        low, high = yarn_find_correction_range(
+            beta_fast,
+            beta_slow,
+            dim,
+            base,
+            original_max_position_embeddings,
+        )
+        freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2)
+        self._freqs = (freq_inter * freq_extra) / (
+            freq_inter * freq_mask + freq_extra * (1 - freq_mask)
+        )
+
+    def __call__(self, x, offset=0):
+        if self.mscale != 1.0:
+            x = self.mscale * x
+        return mx.fast.rope(
+            x,
+            x.shape[-1],
+            traditional=True,
+            base=None,
+            scale=1.0,
+            offset=offset,
+            freqs=self._freqs,
+        )
+
+
+# A clipped silu to prevent fp16 from overflowing
+@partial(mx.compile, shapeless=True)
+def clipped_silu(x):
+    return mx.clip(x * mx.sigmoid(x), a_min=-100, a_max=100)
+
+
+class ClippedSilu(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def __call__(self, x):
+        return clipped_silu(x)
+
+
+class DeepseekV3Attention(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
+
+        self.scale = self.q_head_dim**-0.5
+
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(
+                self.hidden_size, self.num_heads * self.q_head_dim, bias=False
+            )
+        else:
+            self.q_a_proj = nn.Linear(
+                self.hidden_size, self.q_lora_rank, bias=config.attention_bias
+            )
+            self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank, eps=1e-6)
+            self.q_b_proj = nn.Linear(
+                self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
+            )
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=1e-6)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads
+            * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        if self.config.rope_scaling is not None:
+            mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_scaling["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scale = self.scale * mscale * mscale
+
+            rope_kwargs = {
+                key: self.config.rope_scaling[key]
+                for key in [
+                    "original_max_position_embeddings",
+                    "beta_fast",
+                    "beta_slow",
+                    "mscale",
+                    "mscale_all_dim",
+                ]
+                if key in self.config.rope_scaling
+            }
+            self.rope = DeepseekV3YarnRotaryEmbedding(
+                dim=self.qk_rope_head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                scaling_factor=scaling_factor,
+                base=self.rope_theta,
+                **rope_kwargs,
+            )
+        else:
+            self.rope = nn.RoPE(
+                dims=self.qk_rope_head_dim,
+                base=self.rope_theta,
+                traditional=True,
+            )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(x)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
+
+        q = q.reshape(B, L, self.num_heads, self.q_head_dim).transpose(0, 2, 1, 3)
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+        kv = self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
+        kv = kv.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+
+        k_nope, values = mx.split(kv, [self.qk_nope_head_dim], axis=-1)
+
+        if cache is not None:
+            q_pe = self.rope(q_pe, cache.offset)
+            k_pe = self.rope(k_pe, cache.offset)
+            k_pe = mx.repeat(k_pe, self.num_heads, axis=1)
+            keys, values = cache.update_and_fetch(
+                mx.concatenate([k_nope, k_pe], axis=-1), values
+            )
+        else:
+            q_pe = self.rope(q_pe)
+            k_pe = self.rope(k_pe)
+            k_pe = mx.repeat(k_pe, self.num_heads, axis=1)
+            keys = mx.concatenate([k_nope, k_pe], axis=-1)
+
+        queries = mx.concatenate([q_nope, q_pe], axis=-1)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class DeepseekV3MLP(nn.Module):
+    def __init__(
+        self, config: ModelArgs, hidden_size: int = None, intermediate_size: int = None
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x):
+        down_proj = self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+
+    k = top_k
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+    scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+    group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+    k = n_group - topk_group
+    group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+    scores = mx.put_along_axis(scores, group_idx, mx.array(0.0), axis=-2)
+    scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
+class MoEGate(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, config.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert config.topk_method == "noaux_tc", "Unsupported topk method."
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class DeepseekV3MoE(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.switch_mlp = SwitchGLU(
+            config.hidden_size,
+            config.moe_intermediate_size,
+            config.n_routed_experts,
+            activation=ClippedSilu(),
+        )
+
+        self.gate = MoEGate(config)
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = DeepseekV3MLP(
+                config=config, intermediate_size=intermediate_size
+            )
+
+    def __call__(self, x):
+        inds, scores = self.gate(x)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.config.n_shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        return y
+
+
+class DeepseekV3DecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = DeepseekV3Attention(config)
+        self.mlp = (
+            DeepseekV3MoE(config)
+            if (
+                config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+                and layer_idx % config.moe_layer_freq == 0
+            )
+            else DeepseekV3MLP(config)
+        )
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class DeepseekV3Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            DeepseekV3DecoderLayer(config, idx)
+            for idx in range(config.num_hidden_layers)
+        ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.pipeline_rank = 0
+        self.pipeline_size = 1
+
+    def pipeline(self, group):
+        # Split layers in reverse so rank=0 gets the last layers and
+        # rank=pipeline_size-1 gets the first
+        self.pipeline_rank = group.rank()
+        self.pipeline_size = group.size()
+        layers_per_rank = len(self.layers) // self.pipeline_size
+        extra = len(self.layers) - layers_per_rank * self.pipeline_size
+        if self.pipeline_rank < extra:
+            layers_per_rank += 1
+        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
+        self.end_idx = self.start_idx + layers_per_rank
+        self.layers = self.layers[: self.end_idx]
+        self.layers[: self.start_idx] = [None] * self.start_idx
+        self.num_layers = len(self.layers) - self.start_idx
+
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        h = self.embed_tokens(x)
+
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * self.num_layers
+
+        # Receive from the previous process in the pipeline
+
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1))
+
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i])
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size)
+
+        # Broadcast h while keeping it in the graph
+        h = mx.distributed.all_gather(h)[: h.shape[0]]
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.model = DeepseekV3Model(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
+    ):
+        out = self.model(inputs, cache, mask)
+        return self.lm_head(out)
+
+    def sanitize(self, weights):
+        def dequant(weight, scale_inv):
+            dtype = weight.dtype
+            bs = 128  # block size
+            m, n = weight.shape
+            pad_bottom = (-m) % bs
+            pad_side = (-n) % bs
+            weight = mx.pad(weight, ((0, pad_bottom), (0, pad_side)))
+            weight = weight.reshape(
+                ((m + pad_bottom) // bs, bs, (n + pad_side) // bs, bs)
+            )
+            weight = (weight * scale_inv[:, None, :, None]).reshape(
+                m + pad_bottom, n + pad_side
+            )
+            return weight[:m, :n].astype(dtype)
+
+        # Dequantize
+        new_weights = {}
+        for k, v in weights.items():
+            if "weight_scale_inv" in k:
+                scale_inv = v
+                wk = k.replace("_scale_inv", "")
+                weight = weights[wk]
+                weight = dequant(weight, scale_inv)
+                new_weights[wk] = weight
+            elif k not in new_weights:
+                new_weights[k] = v
+        weights = new_weights
+
+        # Stack experts
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+
+        # 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
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers[self.model.start_idx : self.model.end_idx]
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
@@ -0,0 +1,320 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    max_position_embeddings: Optional[int]
+    num_key_value_heads: Optional[int]
+    first_k_dense_replace: int
+    moe_intermediate_size: int
+    moe_layer_freq: int
+    n_routed_experts: int
+    n_shared_experts: int
+    norm_topk_prob: bool
+    num_experts_per_tok: int
+    rope_theta: float
+    routed_scaling_factor: float
+    head_dim: Optional[int] = None
+    scoring_func: str = ("noaux_tc",)
+    n_group: Optional[int] = 1
+    topk_group: Optional[int] = 1
+    attention_bias: bool = False
+    mlp_bias: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+
+
+class Dots1Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        assert args.num_key_value_heads is not None
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.head_dim or args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.q_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+        self.rope = initialize_rope(
+            head_dim,
+            base=args.rope_theta,
+            traditional=False,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = self.q_norm(queries.reshape(B, L, self.n_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        keys = self.k_norm(keys.reshape(B, L, self.n_kv_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+
+    k = top_k
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+    k = n_group - topk_group
+    if k != 0:
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(scores, group_idx, mx.array(0.0), axis=-2)
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
+class Dots1TopkRouter(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+        self.n_routed_experts = args.n_routed_experts
+        self.routed_scaling_factor = args.routed_scaling_factor
+        self.n_group = args.n_group
+        self.topk_group = args.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, args.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class Dots1MLP(nn.Module):
+    def __init__(
+        self, args: ModelArgs, hidden_size: int = None, intermediate_size: int = None
+    ):
+        super().__init__()
+
+        self.hidden_size = args.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            args.intermediate_size if intermediate_size is None else intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=args.mlp_bias
+        )
+        self.up_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=args.mlp_bias
+        )
+        self.down_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=args.mlp_bias
+        )
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class Dots1MoE(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_experts_per_tok = args.num_experts_per_tok
+        self.n_shared_experts = args.n_shared_experts
+
+        self.experts = SwitchGLU(
+            args.hidden_size,
+            args.moe_intermediate_size,
+            args.n_routed_experts,
+        )
+
+        self.gate = Dots1TopkRouter(args)
+
+        self.shared_experts = Dots1MLP(
+            args=args,
+            intermediate_size=args.moe_intermediate_size * args.n_shared_experts,
+        )
+
+    def __call__(self, x):
+        inds, scores = self.gate(x)
+        y = self.experts(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.n_shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        return y
+
+
+class Dots1DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = Dots1Attention(args)
+
+        if layer_idx >= args.first_k_dense_replace:
+            self.mlp = Dots1MoE(args)
+        else:
+            self.mlp = Dots1MLP(args)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class Dots1Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            Dots1DecoderLayer(args, layer_idx)
+            for layer_idx in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Dots1Model(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        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)
+
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            if l >= self.args.first_k_dense_replace:
+                for n, m in [
+                    ("w1", "gate_proj"),
+                    ("w2", "down_proj"),
+                    ("w3", "up_proj"),
+                ]:
+                    for k in ["weight", "scales", "biases"]:
+                        if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                            to_join = [
+                                weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                                for e in range(self.args.n_routed_experts)
+                            ]
+                            weights[f"{prefix}.mlp.experts.{m}.{k}"] = mx.stack(to_join)
+
+        return {k: v for k, v in weights.items() if "rotary_emb.inv_freq" not in k}
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -0,0 +1,167 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    hidden_size: int
+    intermediate_size: int
+    model_type: str
+    max_position_embeddings: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: Optional[int]
+    num_hidden_layers: int
+    rms_norm_eps: float
+    vocab_size: int
+    rope_theta: float
+    use_bias: bool
+    tie_word_embeddings: bool
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.head_dim or dim // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.use_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.use_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.use_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.use_bias)
+
+        self.rope = initialize_rope(
+            head_dim,
+            base=args.rope_theta,
+            traditional=True,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim, use_bias=False):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=use_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=use_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=use_bias)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size, args.use_bias)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class Ernie45Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [DecoderLayer(args) for _ in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Ernie45Model(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        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
@@ -123,10 +123,12 @@ class ExaoneModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.wte(inputs)
-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.h)
@@ -149,9 +151,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.transformer(inputs, cache)
+        out = self.transformer(inputs, mask, cache)
        if self.args.tie_word_embeddings:
            out = self.transformer.wte.as_linear(out)
        else:
@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Optional, Tuple
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -138,12 +138,14 @@ class GemmaModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)
        h = h * (self.args.hidden_size**0.5)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -164,9 +166,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.embed_tokens.as_linear(out)
        return out

@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Optional, Tuple
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -94,7 +94,12 @@ class Attention(nn.Module):
        scores *= self.attn_logit_softcapping

        if mask is not None:
-            scores = scores + mask
+            if mask.dtype == mx.bool_:
+                scores = mx.where(
+                    mask, scores, mx.array(mx.finfo(scores.dtype).min, scores.dtype)
+                )
+            else:
+                scores = scores + mask
        scores = mx.softmax(scores, precise=True, axis=-1)
        output = scores @ values
        if self.repeats > 1:
@@ -160,12 +165,14 @@ class GemmaModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)
        h = h * (self.args.hidden_size**0.5)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache, return_array=True)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -187,9 +194,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.embed_tokens.as_linear(out)
        out = mx.tanh(out / self.final_logit_softcapping)
        out = out * self.final_logit_softcapping
@@ -0,0 +1,64 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from . import gemma3_text
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+    vocab_size: int = 262208
+
+    def __post_init__(self):
+        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", 4
+        )
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = gemma3_text.Model(
+            gemma3_text.ModelArgs.from_dict(args.text_config)
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+        )
+
+    def sanitize(self, weights):
+        weights = tree_unflatten(list(weights.items()))
+        weights.pop("vision_tower", None)
+        weights.pop("multi_modal_projector", None)
+        lm_weights = dict(tree_flatten(weights["language_model"]))
+        lm_weights = self.language_model.sanitize(lm_weights)
+        weights["language_model"] = tree_unflatten(list(lm_weights.items()))
+        return dict(tree_flatten(weights))
+
+    @property
+    def layers(self):
+        return self.language_model.layers
+
+    def make_cache(self):
+        return self.language_model.make_cache()
@@ -0,0 +1,252 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int = 1152
+    num_hidden_layers: int = 26
+    intermediate_size: int = 6912
+    num_attention_heads: int = 4
+    head_dim: int = 256
+    rms_norm_eps: float = 1.0e-6
+    vocab_size: int = 262144
+    num_key_value_heads: int = 1
+    rope_global_base_freq: float = 1_000_000.0
+    rope_local_base_freq: float = 10_000.0
+    rope_traditional: bool = False
+    query_pre_attn_scalar: float = 256
+    sliding_window: int = 512
+    sliding_window_pattern: int = 6
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+        self.repeats = n_heads // n_kv_heads
+        self.head_dim = head_dim = args.head_dim
+        self.layer_idx = layer_idx
+
+        self.scale = args.query_pre_attn_scalar**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.q_norm = RMSNorm(dims=head_dim, eps=args.rms_norm_eps)
+        self.k_norm = RMSNorm(dims=head_dim, eps=args.rms_norm_eps)
+        self.is_sliding = (layer_idx + 1) % args.sliding_window_pattern != 0
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=args.rope_traditional,
+            base=(
+                args.rope_local_base_freq
+                if self.is_sliding
+                else args.rope_global_base_freq
+            ),
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        queries = self.q_norm(queries)
+        keys = self.k_norm(keys)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        # Sliding window
+        if isinstance(mask, mx.array) and mask.shape[-1] != keys.shape[-2]:
+            mask = mask[..., -keys.shape[-2] :]
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = mx.ones((dims,))
+        self.eps = eps
+
+    def __call__(self, x):
+        return mx.fast.rms_norm(x, 1.0 + self.weight, self.eps)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.gelu_approx(self.gate_proj(x)) * self.up_proj(x))
+
+
+@partial(mx.compile, shapeless=True)
+def clip_residual(x, y):
+    if x.dtype != mx.float16:
+        return x + y
+    bound = mx.finfo(mx.float16).max
+    return mx.clip(x.astype(mx.float32) + y.astype(mx.float32), -bound, bound).astype(
+        mx.float16
+    )
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args, layer_idx)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        self.input_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.pre_feedforward_layernorm = RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.post_feedforward_layernorm = RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = clip_residual(x, self.post_attention_layernorm(r))
+        r = self.mlp(self.pre_feedforward_layernorm(h))
+        out = clip_residual(h, self.post_feedforward_layernorm(r))
+        return out
+
+
+class Gemma3Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args, layer_idx=layer_idx)
+            for layer_idx in range(args.num_hidden_layers)
+        ]
+        self.norm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        if input_embeddings is not None:
+            h = input_embeddings
+        else:
+            h = self.embed_tokens(inputs)
+        h *= mx.array(self.args.hidden_size**0.5, mx.bfloat16).astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            j = self.args.sliding_window_pattern
+            full_mask = create_attention_mask(h, cache[j - 1 : j])
+            sliding_window_mask = create_attention_mask(h, cache)
+
+        for i, (layer, c) in enumerate(zip(self.layers, cache)):
+            is_global = (
+                i % self.args.sliding_window_pattern
+                == self.args.sliding_window_pattern - 1
+            )
+
+            local_mask = mask
+            if mask is None and is_global:
+                local_mask = full_mask
+            elif mask is None:
+                local_mask = sliding_window_mask
+
+            h = layer(h, local_mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Gemma3Model(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        out = self.model(inputs, mask, cache, input_embeddings)
+        out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        weights = dict(weights)
+        if "lm_head.weight" not in weights:
+            weights["lm_head.weight"] = weights["model.embed_tokens.weight"]
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    def make_cache(self):
+        caches = []
+        for i in range(self.args.num_hidden_layers):
+            if (
+                i % self.args.sliding_window_pattern
+                == self.args.sliding_window_pattern - 1
+            ):
+                caches.append(KVCache())
+            else:
+                caches.append(
+                    RotatingKVCache(max_size=self.args.sliding_window, keep=0)
+                )
+        return caches
@@ -0,0 +1,621 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+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.utils import tree_flatten, tree_unflatten
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+
+
+@dataclass
+class TextConfig(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    head_dim: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    num_kv_shared_layers: int
+    query_pre_attn_scalar: float
+    vocab_size_per_layer_input: int
+    sliding_window: int
+    max_position_embeddings: int
+    rope_local_base_freq: float
+    rope_theta: float
+    final_logit_softcapping: float
+    layer_types: List[str]
+    activation_sparsity_pattern: List[float]
+    hidden_size_per_layer_input: int
+    altup_num_inputs: int
+    altup_coef_clip: float
+    altup_correct_scale: bool
+    altup_active_idx: int
+    laurel_rank: int
+    rope_scaling: Optional[Dict] = None
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+
+
+class RMSNoScale(nn.Module):
+    def __init__(self, eps: float = 1e-5):
+        super().__init__()
+        self.eps = eps
+
+    def __call__(self, x):
+        return mx.fast.rms_norm(x, None, self.eps)
+
+
+class Gemma3nLaurelBlock(nn.Module):
+    """Learned Augmented Residual Layer"""
+
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+
+        self.linear_left = nn.Linear(
+            self.config.hidden_size, self.config.laurel_rank, bias=False
+        )
+        self.linear_right = nn.Linear(
+            self.config.laurel_rank, self.config.hidden_size, bias=False
+        )
+        self.post_laurel_norm = nn.RMSNorm(
+            dims=self.config.hidden_size,
+            eps=self.config.rms_norm_eps,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        laurel_x = self.linear_left(x)
+        laurel_x = self.linear_right(laurel_x)
+        normed_laurel_x = self.post_laurel_norm(laurel_x)
+        return x + normed_laurel_x
+
+
+class Gemma3nAttention(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int, is_kv_shared_layer: bool):
+        super().__init__()
+        self.is_sliding = config.layer_types[layer_idx] == "sliding_attention"
+
+        dim = config.hidden_size
+        self.n_heads = n_heads = config.num_attention_heads
+        self.n_kv_heads = n_kv_heads = config.num_key_value_heads
+        self.repeats = n_heads // n_kv_heads
+        self.head_dim = head_dim = config.head_dim
+        self.layer_idx = layer_idx
+
+        self.scale = 1.0
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.q_norm = nn.RMSNorm(dims=config.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(dims=config.head_dim, eps=config.rms_norm_eps)
+        self.v_norm = RMSNoScale(eps=config.rms_norm_eps)
+
+        self.is_kv_shared_layer = is_kv_shared_layer
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=False,
+            base=(
+                config.rope_local_base_freq if self.is_sliding else config.rope_theta
+            ),
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        queries = self.q_proj(x)
+        queries = queries.reshape(B, L, -1, self.head_dim)
+        queries = self.q_norm(queries)
+
+        offset = 0
+        if self.is_kv_shared_layer and cache is not None:
+            # For shared layers, retrieve KV from the designated cache layer
+            keys, values = cache.state
+            offset = cache.offset
+
+        else:
+            if cache is not None:
+                offset = cache.offset
+            keys = self.k_proj(x).reshape(B, L, -1, self.head_dim)
+            keys = self.k_norm(keys)
+            keys = keys.transpose(0, 2, 1, 3)
+            keys = self.rope(keys, offset=offset)
+
+            values = self.v_proj(x).reshape(B, L, -1, self.head_dim)
+            values = self.v_norm(values)
+            values = values.transpose(0, 2, 1, 3)
+
+            if cache is not None:
+                keys, values = cache.update_and_fetch(keys, values)
+
+        queries = queries.transpose(0, 2, 1, 3)
+        queries = self.rope(queries, offset=offset)
+
+        if isinstance(mask, mx.array) and mask.shape[-1] != keys.shape[-2]:
+            mask = mask[:, : keys.shape[-2]]
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.o_proj(output)
+
+
+@partial(mx.compile, shapeless=True)
+def gelu_topk(inputs, std_multiplier):
+    inputs_mean = mx.mean(inputs, axis=-1, keepdims=True)
+    inputs_std = mx.std(inputs, axis=-1, keepdims=True)
+    cutoff_x = inputs_mean + inputs_std * std_multiplier.astype(inputs_std.dtype)
+    return nn.gelu_approx(mx.maximum(0, inputs - cutoff_x))
+
+
+class MLP(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        if config.activation_sparsity_pattern is not None:
+            self.activation_sparsity = config.activation_sparsity_pattern[layer_idx]
+        else:
+            self.activation_sparsity = 0.0
+        if self.activation_sparsity > 0:
+            self._std_multiplier = math.sqrt(2.0) * mx.erfinv(
+                2 * self.activation_sparsity - 1
+            )
+
+    def __call__(self, x: mx.array):
+        gate_proj = self.gate_proj(x)
+        if self.activation_sparsity > 0.0:
+            activations = gelu_topk(gate_proj, self._std_multiplier)
+        else:
+            activations = nn.gelu_approx(gate_proj)
+        up_proj = self.up_proj(x)
+        down_proj = self.down_proj(activations * up_proj)
+        return down_proj
+
+
+class Gemma3nAltUp(nn.Module):
+    """Alternating Updates (AltUp)"""
+
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+
+        self.correct_output_scale = mx.zeros((self.config.hidden_size,))
+        self.correction_coefs = nn.Linear(
+            self.config.altup_num_inputs, self.config.altup_num_inputs, bias=False
+        )
+        self.prediction_coefs = nn.Linear(
+            self.config.altup_num_inputs, self.config.altup_num_inputs**2, bias=False
+        )
+        self.modality_router = nn.Linear(
+            self.config.hidden_size, self.config.altup_num_inputs, bias=False
+        )
+        self.router_norm = nn.RMSNorm(
+            dims=self.config.hidden_size,
+            eps=self.config.rms_norm_eps,
+        )
+
+    def compute_router_modalities(self, x: mx.array) -> mx.array:
+        router_inputs = self.router_norm(x) * (self.config.hidden_size**-1.0)
+        routed = self.modality_router(router_inputs).astype(mx.float32)
+        return mx.tanh(routed)
+
+    def predict(self, x: mx.array) -> mx.array:
+        modalities = self.compute_router_modalities(x[self.config.altup_active_idx])
+
+        self.prediction_coefs.weight = self.prediction_coefs.weight.astype(mx.float32)
+
+        if self.config.altup_coef_clip is not None:
+            self.prediction_coefs.weight = mx.clip(
+                self.prediction_coefs.weight,
+                -self.config.altup_coef_clip,
+                self.config.altup_coef_clip,
+            )
+
+        all_coefs = (
+            self.prediction_coefs(modalities)
+            .reshape(
+                *modalities.shape[:-1],
+                self.config.altup_num_inputs,
+                self.config.altup_num_inputs,
+            )
+            .transpose(0, 1, 3, 2)
+        )
+
+        x_up = x.astype(mx.float32)
+        x_permuted = x_up.transpose(1, 2, 3, 0)
+        predictions = mx.matmul(x_permuted, all_coefs)
+        predictions = predictions.transpose(3, 0, 1, 2)
+        predictions += x_up
+        return predictions.astype(x.dtype)
+
+    def correct(self, predictions: mx.array, activated: mx.array):
+        modalities = self.compute_router_modalities(activated)
+
+        self.correction_coefs.weight = self.correction_coefs.weight.astype(mx.float32)
+
+        if self.config.altup_coef_clip is not None:
+            self.correction_coefs.weight = mx.clip(
+                self.correction_coefs.weight,
+                -self.config.altup_coef_clip,
+                self.config.altup_coef_clip,
+            )
+
+        all_coefs = self.correction_coefs(modalities) + 1.0
+
+        active_x = predictions[self.config.altup_active_idx]
+        innovation = activated - active_x
+
+        all_coefs = all_coefs.transpose(2, 1, 0)
+        corrected = innovation[None] * all_coefs[:, None]
+        corrected += predictions
+
+        return corrected.astype(activated.dtype)
+
+
+class Gemma3nDecoderLayer(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int, is_kv_shared_layer: bool):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+        self.self_attn = Gemma3nAttention(config, layer_idx, is_kv_shared_layer)
+        self.mlp = MLP(config, layer_idx=layer_idx)
+        self.input_layernorm = nn.RMSNorm(
+            self.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+        self.post_attention_layernorm = nn.RMSNorm(
+            self.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+        self.pre_feedforward_layernorm = nn.RMSNorm(
+            self.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+        self.post_feedforward_layernorm = nn.RMSNorm(
+            self.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+        self.is_sliding = self.self_attn.is_sliding
+        self.sliding_window = config.sliding_window
+
+        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
+
+        self.altup = Gemma3nAltUp(config)
+        self.laurel = Gemma3nLaurelBlock(config)
+        self.per_layer_input_gate = nn.Linear(
+            self.hidden_size, self.hidden_size_per_layer_input, bias=False
+        )
+        self.per_layer_projection = nn.Linear(
+            self.hidden_size_per_layer_input, self.hidden_size, bias=False
+        )
+        self.post_per_layer_input_norm = nn.RMSNorm(
+            self.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        per_layer_input: Optional[mx.array] = None,
+    ):
+        predictions = self.altup.predict(x)
+        active_prediction = predictions[self.config.altup_active_idx]
+
+        active_prediction_normed = self.input_layernorm(active_prediction)
+        laurel_output = self.laurel(active_prediction_normed)
+
+        attn = self.self_attn(
+            active_prediction_normed,
+            mask,
+            cache,
+        )
+
+        attn = self.post_attention_layernorm(attn)
+
+        attn_gated = active_prediction + attn
+        attn_laurel = (attn_gated + laurel_output) * (2.0**-0.5)
+
+        attn_norm = self.pre_feedforward_layernorm(attn_laurel)
+        attn_ffw = self.mlp(attn_norm)
+        attn_ffw_norm = self.post_feedforward_layernorm(attn_ffw)
+        attn_ffw_laurel_gated = attn_laurel + attn_ffw_norm
+
+        corrected_predictions = self.altup.correct(predictions, attn_ffw_laurel_gated)
+
+        first_prediction = corrected_predictions[self.config.altup_active_idx]
+        if self.config.altup_correct_scale:
+            first_prediction = first_prediction * self.altup.correct_output_scale
+
+        first_prediction = self.per_layer_input_gate(first_prediction)
+        first_prediction = nn.gelu_approx(first_prediction)
+
+        first_prediction = mx.multiply(first_prediction, per_layer_input)
+
+        first_prediction = self.per_layer_projection(first_prediction)
+        first_prediction = self.post_per_layer_input_norm(first_prediction)
+
+        corrected_predictions[1:] = corrected_predictions[1:] + first_prediction
+
+        return corrected_predictions
+
+
+@partial(mx.compile, shapeless=True)
+def logit_softcap(softcap, x):
+    out = mx.tanh(x / softcap)
+    out = out * softcap
+    return out
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.hidden_size_per_layer_input = config.hidden_size_per_layer_input
+        self.vocab_size = config.vocab_size
+        self.vocab_size_per_layer_input = config.vocab_size_per_layer_input
+        self.num_hidden_layers = config.num_hidden_layers
+        self.final_logit_softcapping = config.final_logit_softcapping
+        self.first_kv_shared_layer_idx = (
+            config.num_hidden_layers - config.num_kv_shared_layers
+        )
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            Gemma3nDecoderLayer(
+                config=config,
+                layer_idx=layer_idx,
+                is_kv_shared_layer=layer_idx >= self.first_kv_shared_layer_idx,
+            )
+            for layer_idx in range(config.num_hidden_layers)
+        ]
+
+        self.embed_tokens_per_layer = nn.Embedding(
+            config.vocab_size_per_layer_input,
+            config.num_hidden_layers * config.hidden_size_per_layer_input,
+        )
+
+        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(
+            dims=config.hidden_size_per_layer_input,
+            eps=config.rms_norm_eps,
+        )
+
+        self.altup_projections = [
+            nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+            for _ in range(1, self.config.altup_num_inputs)
+        ]
+
+        self.altup_unembed_projections = [
+            nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+            for _ in range(1, self.config.altup_num_inputs)
+        ]
+
+        self.norm = nn.RMSNorm(
+            config.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+        self.first_sliding_idx = self.config.layer_types.index("sliding_attention")
+        self.first_full_idx = self.config.layer_types.index("full_attention")
+
+        concrete_layers = self.config.layer_types[: self.first_kv_shared_layer_idx]
+        shared_full_idx = (
+            len(concrete_layers) - 1 - concrete_layers[::-1].index("full_attention")
+        )
+        shared_sliding_idx = (
+            len(concrete_layers) - 1 - concrete_layers[::-1].index("sliding_attention")
+        )
+
+        self.layer_idx_to_cache_idx = []
+        for i, layer_type in enumerate(self.config.layer_types):
+            if i < self.first_kv_shared_layer_idx:
+                self.layer_idx_to_cache_idx.append(i)
+            else:
+                if layer_type == "full_attention":
+                    self.layer_idx_to_cache_idx.append(shared_full_idx)
+                elif layer_type == "sliding_attention":
+                    self.layer_idx_to_cache_idx.append(shared_sliding_idx)
+                else:
+                    raise NotImplementedError(f"Unknown layer type: {layer_type}")
+
+    def __call__(
+        self,
+        inputs: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+        input_embeddings: mx.array = None,
+    ):
+        if input_embeddings is None:
+            h = self.embed_tokens(inputs) * (self.hidden_size**0.5)
+        else:
+            h = input_embeddings
+
+        per_layer_inputs = self.get_per_layer_inputs(inputs)
+        per_layer_inputs = self.project_per_layer_inputs(h, per_layer_inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            full_mask = create_attention_mask(
+                h,
+                cache[self.first_full_idx :],
+            )
+            sliding_window_mask = create_attention_mask(
+                h,
+                cache[self.first_sliding_idx :],
+            )
+        h0 = h
+
+        # Expand hidden_states to support per-layer inputs
+        target_magnitude = mx.mean(h0**2, axis=-1, keepdims=True) ** 0.5
+
+        h_list = [h0]
+        h_list.extend([proj(h0) for proj in self.altup_projections])
+        h = mx.stack(h_list, axis=0)
+        mags = mx.mean(h[1:] ** 2, axis=-1, keepdims=True) ** 0.5
+        h[1:] = h[1:] * (target_magnitude / mx.maximum(mags, mx.finfo(h0.dtype).min))
+
+        for i, layer in enumerate(self.layers):
+            per_layer_input = per_layer_inputs[:, :, i, :]
+
+            is_global = self.config.layer_types[i] == "full_attention"
+
+            local_mask = mask
+            if mask is None and is_global:
+                local_mask = full_mask
+            elif mask is None:
+                local_mask = sliding_window_mask
+
+            h = layer(
+                h,
+                local_mask,
+                cache[self.layer_idx_to_cache_idx[i]],
+                per_layer_input,
+            )
+
+        # Per-layer inputs to single output
+        target_magnitude = mx.mean(h[0] ** 2, axis=-1, keepdims=True) ** 0.5
+        for i, proj in enumerate(self.altup_unembed_projections):
+            h[i + 1] = proj(h[i + 1])
+        mags = mx.mean(h[1:] ** 2, axis=-1, keepdims=True) ** 0.5
+        h[1:] = h[1:] * (target_magnitude / mx.maximum(mags, mx.finfo(h0.dtype).min))
+
+        h = mx.mean(h, axis=0)
+
+        out = self.norm(h)
+        out = self.embed_tokens.as_linear(out)
+        if self.final_logit_softcapping is not None:
+            out = logit_softcap(self.final_logit_softcapping, out)
+        return out
+
+    def get_per_layer_inputs(self, input_ids: mx.array) -> mx.array:
+        per_layer_inputs_mask = input_ids < self.vocab_size_per_layer_input
+        tokens = mx.where(per_layer_inputs_mask, input_ids, mx.zeros_like(input_ids))
+        result = self.embed_tokens_per_layer(tokens) * (
+            self.hidden_size_per_layer_input**0.5
+        )
+        return result.reshape(
+            *input_ids.shape,
+            self.num_hidden_layers,
+            self.hidden_size_per_layer_input,
+        )
+
+    def project_per_layer_inputs(
+        self,
+        inputs_embeds: mx.array,
+        per_layer_inputs: mx.array,
+    ) -> mx.array:
+        per_layer_projection = self.per_layer_model_projection(inputs_embeds) * (
+            self.hidden_size**-0.5
+        )
+        per_layer_projection = per_layer_projection.reshape(
+            *inputs_embeds.shape[:-1],
+            self.config.num_hidden_layers,
+            self.config.hidden_size_per_layer_input,
+        )
+        per_layer_projection = self.per_layer_projection_norm(per_layer_projection)
+        return (per_layer_projection + per_layer_inputs) * (2.0**-0.5)
+
+    def make_cache(self):
+        caches = []
+        for layer_type in self.config.layer_types[: self.first_kv_shared_layer_idx]:
+            if layer_type == "full_attention":
+                caches.append(KVCache())
+            elif layer_type == "sliding_attention":
+                caches.append(
+                    RotatingKVCache(max_size=self.config.sliding_window, keep=0)
+                )
+            else:
+                raise NotImplementedError(f"Unknown layer type: {layer_type}")
+        return caches
+
+
+class Gemma3n(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.language_model = LanguageModel(TextConfig.from_dict(args.text_config))
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+        )
+
+    def make_cache(self):
+        return self.language_model.make_cache()
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model = Gemma3n(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.model(
+            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+        )
+
+    def sanitize(self, weights):
+        weights = tree_unflatten(list(weights.items()))
+        for k in ["vision_tower", "audio_tower", "embed_audio", "embed_vision"]:
+            weights["model"].pop(k, None)
+        return dict(tree_flatten(weights))
+
+    @property
+    def layers(self):
+        return self.model.language_model.layers
+
+    def make_cache(self):
+        return self.model.make_cache()
@@ -0,0 +1,183 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    attention_bias: bool
+    head_dim: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    partial_rotary_factor: float
+    rope_theta: float
+    rope_traditional: bool = True
+    max_position_embeddings: int = 32768
+
+
+class Glm4MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.gate_up_proj = nn.Linear(
+            args.hidden_size, 2 * args.intermediate_size, bias=False
+        )
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        x = self.gate_up_proj(x)
+        gate, up_states = mx.split(x, 2, axis=-1)
+        return self.down_proj(nn.silu(gate) * up_states)
+
+
+class Glm4Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.head_dim = getattr(
+            args, "head_dim", args.hidden_size // args.num_attention_heads
+        )
+        self.n_heads = args.num_attention_heads
+        self.n_kv_heads = args.num_key_value_heads
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size,
+            args.num_attention_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size,
+            args.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size,
+            args.num_key_value_heads * self.head_dim,
+            bias=args.attention_bias,
+        )
+        self.o_proj = nn.Linear(
+            args.num_attention_heads * self.head_dim, args.hidden_size, bias=False
+        )
+
+        self.rope = nn.RoPE(
+            dims=int(self.head_dim * args.partial_rotary_factor),
+            base=args.rope_theta,
+            traditional=args.rope_traditional,
+        )
+
+    def __call__(
+        self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class Glm4DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Glm4Attention(args=args)
+
+        self.mlp = Glm4MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.post_self_attn_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.post_mlp_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None
+    ) -> mx.array:
+        x = x + self.post_self_attn_layernorm(
+            self.self_attn(self.input_layernorm(x), mask, cache)
+        )
+        residual = x
+        x = (
+            self.post_mlp_layernorm(self.mlp(self.post_attention_layernorm(x)))
+            + residual
+        )
+        return x
+
+
+class Glm4Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            Glm4DecoderLayer(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Glm4Model(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ):
+        out = self.model(inputs, mask, cache)
+        return self.lm_head(out)
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -1,11 +1,10 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
-import numpy as np

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention

@@ -126,18 +125,21 @@ class GPT2Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        _, L = inputs.shape

        hidden_states = self.wte(inputs)

-        mask = None
-        if hidden_states.shape[1] > 1:
+        offset = 0
+        if cache is not None and len(cache) > 0 and cache[0] is not None:
+            offset = cache[0].offset

-            position_ids = mx.array(np.arange(L))
-            hidden_states += self.wpe(position_ids)
+        position_ids = mx.arange(offset, offset + L)
+        hidden_states += self.wpe(position_ids)

+        if mask is None:
            mask = create_attention_mask(hidden_states, cache)

        if cache is None:
@@ -159,9 +161,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.wte.as_linear(out)
        return out

@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -137,6 +137,7 @@ class GPTBigCodeModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        B, L = inputs.shape
@@ -144,15 +145,16 @@ class GPTBigCodeModel(nn.Module):
        hidden_states = self.wte(inputs)

        mask = None
-        if hidden_states.shape[1] > 1:
-
-            position_ids = mx.array(np.arange(L))
-            hidden_states += self.wpe(position_ids)
-
+        if mask is not None and hidden_states.shape[1] > 1:
            mask = create_attention_mask(hidden_states, cache)

        if cache is None:
            cache = [None] * len(self.h)
+            position_ids = mx.array(np.arange(L))
+        else:
+            position_ids = mx.array(np.arange(cache[0].offset, cache[0].offset + L))
+
+        hidden_states += self.wpe(position_ids)

        for layer, c in zip(self.h, cache):
            hidden_states = layer(hidden_states, mask, cache=c)
@@ -172,9 +174,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.transformer(inputs, cache)
+        out = self.transformer(inputs, mask, cache)
        if self.args.tie_word_embeddings:
            out = self.transformer.wte.as_linear(out)
        else:
@@ -1,11 +1,10 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
-import numpy as np

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention

@@ -146,13 +145,15 @@ class GPTNeoXModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        _, L = inputs.shape

        hidden_states = self.embed_in(inputs)

-        mask = create_attention_mask(hidden_states, cache)
+        if mask is None:
+            mask = create_attention_mask(hidden_states, cache)

        if cache is None:
            cache = [None] * len(self.h)
@@ -176,9 +177,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        return out

    def sanitize(self, weights):
@@ -0,0 +1,195 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    logits_scaling: float
+    attention_multiplier: float
+    embedding_multiplier: float
+    residual_multiplier: float
+    max_position_embeddings: int
+    num_key_value_heads: int
+    attention_bias: bool
+    mlp_bias: bool
+    rope_theta: float
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.scale = args.attention_multiplier
+        attention_bias = args.attention_bias
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attention_bias)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            False,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        hidden_dim = args.intermediate_size
+        if hasattr(args, "mlp_bias"):
+            mlp_bias = args.mlp_bias
+        else:
+            mlp_bias = False
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=mlp_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.residual_multiplier = args.residual_multiplier
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r * self.residual_multiplier
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r * self.residual_multiplier
+        return out
+
+
+class GraniteModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.embedding_multiplier = args.embedding_multiplier
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs) * self.embedding_multiplier
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = GraniteModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self.logits_scaling = args.logits_scaling
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out / self.logits_scaling
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -0,0 +1,185 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    attention_bias: bool
+    head_dim: int
+    max_position_embeddings: int
+    mlp_bias: bool
+    model_type: str
+    rope_theta: float
+    tie_word_embeddings: bool
+
+
+class HeliumAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        assert args.num_key_value_heads is not None
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+        self.rope = nn.RoPE(head_dim, traditional=True, base=args.rope_theta)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class HeliumMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.intermediate_size = args.intermediate_size
+
+        self.gate_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=args.mlp_bias
+        )
+        self.up_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=args.mlp_bias
+        )
+        self.down_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=args.mlp_bias
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class HeliumDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+
+        self.self_attn = HeliumAttention(args)
+        self.mlp = HeliumMLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class HeliumModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_hidden_layers = args.num_hidden_layers
+        self.vocab_size = args.vocab_size
+
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+
+        self.layers = [HeliumDecoderLayer(args) for _ in range(args.num_hidden_layers)]
+
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+
+        self.model = HeliumModel(args)
+
+        self.vocab_size = args.vocab_size
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ) -> mx.array:
+        out = self.model(inputs, mask, cache)
+        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
@@ -1,6 +1,5 @@
 # Copyright © 2023-2024 Apple Inc.

-import math
 from dataclasses import dataclass
 from typing import Any, Dict, Optional, Tuple, Union

@@ -76,7 +75,6 @@ class Attention(nn.Module):

        head_dim = args.hidden_size // n_heads
        self.scale = head_dim**-0.5
-
        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
        if kv_proj:
            self.k_proj = nn.Linear(
@@ -107,7 +105,6 @@ class Attention(nn.Module):
        B, L, D = x.shape

        queries = self.q_proj(x)
-
        if kv_states is None:
            keys, values = self.k_proj(x), self.v_proj(x)
            kv_states = keys, values
@@ -198,7 +195,10 @@ class DecoderLayer(nn.Module):
        super().__init__()
        self.hidden_size = args.hidden_size
        self.self_attn = Attention(kv_proj, args)
-        self.mlp = MoeBlock(args)
+        if args.num_experts == 1:
+            self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        else:
+            self.mlp = MoeBlock(args)

        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
        self.post_attention_layernorm = nn.RMSNorm(
@@ -231,7 +231,10 @@ class HunYuanModel(nn.Module):
        assert self.vocab_size > 0
        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
        self.layers = [
-            DecoderLayer(args=args, kv_proj=(i % args.cla_share_factor) == 0)
+            DecoderLayer(
+                args=args,
+                kv_proj=(not args.use_cla) or (i % args.cla_share_factor) == 0,
+            )
            for i in range(args.num_hidden_layers)
        ]
        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
@@ -239,17 +242,19 @@ class HunYuanModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)

        for i, (layer, c) in enumerate(zip(self.layers, cache)):
-            if i % self.args.cla_share_factor == 0:
+            if (not self.args.use_cla) or i % self.args.cla_share_factor == 0:
                shared_kv_states = None
            h, shared_kv_states = layer(h, mask, c, shared_kv_states)

@@ -266,12 +271,36 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        return self.model.embed_tokens.as_linear(out)

    def sanitize(self, weights):
+
+        if "model.layers.0.mlp.gate_and_up_proj.weight" in weights:
+            new_weights = {}
+            D = self.args.hidden_size
+            n_kv_heads = self.args.num_key_value_heads
+            n_kv_groups = self.args.num_attention_heads // n_kv_heads
+            head_dim = D // self.args.num_attention_heads
+            for k, v in weights.items():
+                if "qkv_proj" in k:
+                    v = v.reshape(n_kv_heads, n_kv_groups + 2, head_dim, -1)
+                    splits = v.split([n_kv_groups, n_kv_groups + 1], axis=1)
+                    for k_up, v_new in zip(["q_proj", "k_proj", "v_proj"], splits):
+                        k_new = k.replace("qkv_proj", k_up)
+                        new_weights[k_new] = mx.flatten(v_new, 0, 2)
+                elif "gate_and_up_proj" in k:
+                    splits = v.split(2, axis=0)
+                    for k_up, v_new in zip(["up_proj", "gate_proj"], splits):
+                        k_new = k.replace("gate_and_up_proj", k_up)
+                        new_weights[k_new] = v_new
+                else:
+                    new_weights[k] = v
+            weights = new_weights
+
        if "model.layers.0.mlp.experts.0.up_proj.weight" not in weights:
            return weights
        for l in range(self.args.num_hidden_layers):
@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Dict, Optional, Union

 import mlx.core as mx
 import mlx.nn as nn
@@ -193,11 +193,13 @@ class InternLM2Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.tok_embeddings(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -220,9 +222,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        if self.args.tie_word_embeddings:
            out = self.model.tok_embeddings.as_linear(out)
        else:
@@ -0,0 +1,241 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    bias: bool = False
+    qkv_bias: bool = False
+    max_position_embeddings: int = 32768
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    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:
+            self.num_key_value_heads = self.num_attention_heads
+
+        if self.rope_scaling:
+            required_keys = {"factor", "rope_type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if self.rope_scaling["rope_type"] not in ["linear", "dynamic"]:
+                raise ValueError(
+                    "rope_scaling 'rope_type' currently only supports 'linear' or 'dynamic"
+                )
+
+
+class DynamicNTKScalingRoPE(nn.Module):
+    """Implements the rotary positional encoding with Dynamic NTK scaling."""
+
+    def __init__(
+        self,
+        dims: int,
+        max_position_embeddings: int = 2048,
+        traditional: bool = False,
+        base: float = 10000,
+        scale: float = 1.0,
+    ):
+        super().__init__()
+        self.max_position_embeddings = max_position_embeddings
+        self.original_base = base
+        self.dims = dims
+        self.traditional = traditional
+        self.scale = scale
+
+    def extra_repr(self):
+        return f"{self.dims}, traditional={self.traditional}, max_position_embeddings={self.max_position_embeddings}, scaling_factor={self.scaling_factor}"
+
+    def __call__(self, x, offset: int = 0):
+        seq_len = x.shape[1] + offset
+        if seq_len > self.max_position_embeddings:
+            base = self.original_base * (
+                (self.scale * seq_len / self.max_position_embeddings) - (self.scale - 1)
+            ) ** (self.dims / (self.dims - 2))
+        else:
+            base = self.original_base
+
+        return mx.fast.rope(
+            x,
+            self.dims,
+            traditional=self.traditional,
+            base=base,
+            scale=self.scale,
+            offset=offset,
+        )
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        qkv_bias = args.qkv_bias
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+        self.n_kv_groups = n_heads // args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=qkv_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=qkv_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=qkv_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=qkv_bias)
+
+        rope_scale = (
+            1 / args.rope_scaling["factor"]
+            if args.rope_scaling is not None
+            and args.rope_scaling["rope_type"] == "linear"
+            else 2.0
+        )
+
+        self.rope = DynamicNTKScalingRoPE(
+            head_dim,
+            max_position_embeddings=args.max_position_embeddings,
+            traditional=args.rope_traditional,
+            base=args.rope_theta,
+            scale=rope_scale,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim, bias):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=bias)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size, args.bias)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class InternLM2Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        assert args.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = InternLM2Model(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        # Remove unused precomputed rotary freqs
+        return {k: v for k, v in weights.items() if "attention.rope.inv_freq" not in k}
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -0,0 +1,118 @@
+# Copyright © 2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .deepseek_v3 import DeepseekV3Model
+
+
+@dataclass
+class TextArgs(BaseModelArgs):
+    vocab_size: int = 102400
+    hidden_size: int = 4096
+    intermediate_size: int = 11008
+    moe_intermediate_size: int = 1407
+    num_hidden_layers: int = 30
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 32
+    n_shared_experts: Optional[int] = None
+    n_routed_experts: Optional[int] = None
+    routed_scaling_factor: float = 1.0
+    kv_lora_rank: int = 512
+    q_lora_rank: int = 1536
+    qk_rope_head_dim: int = 64
+    v_head_dim: int = 128
+    qk_nope_head_dim: int = 128
+    topk_method: str = "noaux_tc"
+    scoring_func: str = "sigmoid"
+    norm_topk_prob: bool = True
+    n_group: Optional[int] = None
+    topk_group: Optional[int] = None
+    num_experts_per_tok: Optional[int] = None
+    moe_layer_freq: int = 1
+    first_k_dense_replace: int = 0
+    max_position_embeddings: int = 2048
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 10000.0
+    rope_scaling: Dict = None
+    attention_bias: bool = False
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    text_config: Union[TextArgs, dict]
+    model_type: str
+
+    def __post_init__(self):
+        self.text_config = TextArgs.from_dict(self.text_config)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextArgs):
+        super().__init__()
+        self.args = config
+        self.model = DeepseekV3Model(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
+    ):
+        out = self.model(inputs, cache, mask)
+        return self.lm_head(out)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.language_model = LanguageModel(config.text_config)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
+    ):
+        return self.language_model(inputs, cache, mask)
+
+    def sanitize(self, weights):
+        def keep(key):
+            return (
+                "vision_tower" not in key
+                and "rotary_emb" not in key
+                and "multi_modal_projector" not in key
+            )
+
+        weights = {k: v for k, v in weights.items() if keep(k)}
+        # Stack experts
+        for l in range(self.args.text_config.num_hidden_layers):
+            prefix = f"language_model.model.layers.{l}"
+            for m in [("gate_proj"), ("down_proj"), ("up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.text_config.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+
+        return weights
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
@@ -155,11 +155,17 @@ class LlamaModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
+        input_embeddings: Optional[mx.array] = None,
    ):
-        h = self.embed_tokens(inputs)
+        if input_embeddings is not None:
+            h = input_embeddings
+        else:
+            h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -182,9 +188,11 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
+        input_embeddings: Optional[mx.array] = None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache, input_embeddings)
        if self.args.tie_word_embeddings:
            out = self.model.embed_tokens.as_linear(out)
        else:
@@ -193,9 +201,12 @@ class Model(nn.Module):

    def sanitize(self, weights):
        # Remove unused precomputed rotary freqs
-        return {
+        weights = {
            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
        }
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights

    @property
    def layers(self):
@@ -0,0 +1,333 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import ChunkedKVCache, KVCache
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class TextArgs(BaseModelArgs):
+    attention_bias: bool
+    attention_chunk_size: int
+    head_dim: int
+    hidden_act: str
+    hidden_size: int
+    interleave_moe_layer_step: int
+    intermediate_size: int
+    intermediate_size_mlp: int
+    max_position_embeddings: int
+    model_type: str
+    num_attention_heads: int
+    num_experts_per_tok: int
+    num_hidden_layers: int
+    num_key_value_heads: int
+    num_local_experts: int
+    rms_norm_eps: float
+    rope_scaling: Any
+    rope_theta: float
+    use_qk_norm: bool
+    vocab_size: int
+    attn_temperature_tuning: int = 4
+    floor_scale: int = 8192
+    attn_scale: float = 0.1
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    text_config: Union[TextArgs, dict]
+    model_type: str
+
+    def __post_init__(self):
+        self.text_config = TextArgs.from_dict(self.text_config)
+
+
+class Attention(nn.Module):
+    def __init__(self, args: TextArgs, layer_idx: int):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.use_rope = int((layer_idx + 1) % 4 != 0)  # rope unused for dense layers
+        self.attn_temperature_tuning = args.attn_temperature_tuning
+        self.floor_scale = args.floor_scale
+        self.attn_scale = args.attn_scale
+
+        self.head_dim = head_dim = args.head_dim or args.hidden_size // n_heads
+
+        self.scale = head_dim**-0.5
+        if hasattr(args, "attention_bias"):
+            attention_bias = args.attention_bias
+        else:
+            attention_bias = False
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attention_bias)
+
+        self.use_qk_norm = args.use_qk_norm and self.use_rope
+
+        if self.use_rope:
+            self.rope = initialize_rope(
+                head_dim,
+                args.rope_theta,
+                traditional=True,
+                scaling_config=args.rope_scaling,
+                max_position_embeddings=args.max_position_embeddings,
+            )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            offset = cache.offset
+        else:
+            offset = 0
+
+        if self.use_rope:
+            queries = self.rope(queries, offset=offset)
+            keys = self.rope(keys, offset=offset)
+
+        if self.use_qk_norm:
+            queries = mx.fast.rms_norm(queries, weight=None, eps=1e-6)
+            keys = mx.fast.rms_norm(keys, weight=None, eps=1e-6)
+
+        if self.attn_temperature_tuning and not self.use_rope:
+            attn_scales = (
+                mx.log(
+                    mx.floor(mx.arange(offset + 1, offset + L + 1) / self.floor_scale)
+                    + 1.0
+                )
+                * self.attn_scale
+                + 1.0
+            )
+            attn_scales = attn_scales[:, None]
+            queries = (queries * attn_scales).astype(queries.dtype)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs, intermediate_size: int = None):
+        super().__init__()
+
+        dim = args.hidden_size
+        hidden_dim = intermediate_size or args.intermediate_size
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class MoE(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.top_k = args.num_experts_per_tok
+        self.num_experts = args.num_local_experts
+        self.experts = SwitchGLU(
+            args.hidden_size, args.intermediate_size, self.num_experts
+        )
+        self.router = nn.Linear(args.hidden_size, args.num_local_experts, bias=False)
+        self.shared_expert = MLP(args)
+
+    def __call__(self, x) -> mx.array:
+        logits = self.router(x)
+        k = self.top_k
+        indices = mx.argpartition(-logits, kth=k - 1, axis=-1)[..., :k]
+        scores = mx.take_along_axis(logits, indices, axis=-1)
+        scores = mx.sigmoid(scores.astype(mx.float32)).astype(x.dtype)
+
+        out = self.experts(x * scores, indices).squeeze(2)
+        return out + self.shared_expert(x)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: TextArgs, layer_idx: int):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args, layer_idx)
+        self.is_moe_layer = (layer_idx % args.interleave_moe_layer_step) == (
+            args.interleave_moe_layer_step - 1
+        )
+        if self.is_moe_layer:
+            self.feed_forward = MoE(args)
+        else:
+            self.feed_forward = MLP(args, args.intermediate_size_mlp)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.feed_forward(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class LlamaModel(nn.Module):
+    def __init__(self, args: TextArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [TransformerBlock(args, i) for i in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.attention_chunk_size = args.attention_chunk_size
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if cache is not None:
+            for idx, c in enumerate(cache):
+                if (idx + 1) % 4 != 0:
+                    c.maybe_trim_front()
+            start = cache[0].start_position
+            offset = cache[0].offset
+        else:
+            start = 0
+            offset = 0
+        end = offset + h.shape[1]
+        linds = mx.arange(start, end)
+        rinds = mx.arange(offset, end)[:, None]
+        block_pos = mx.abs(
+            (linds // self.attention_chunk_size) - (rinds // self.attention_chunk_size)
+        )
+        token_pos = linds <= rinds
+        chunk_mask = (block_pos == 0) & token_pos
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+        else:
+            chunk_mask &= mask
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for idx, (layer, c) in enumerate(zip(self.layers, cache)):
+            use_chunked_attention = (idx + 1) % 4 != 0
+            if use_chunked_attention:
+                local_mask = chunk_mask
+            else:
+                local_mask = mask
+            h = layer(h, local_mask, cache=c)
+
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, args: TextArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = LlamaModel(self.args)
+        self.lm_head = nn.Linear(
+            self.args.hidden_size, self.args.vocab_size, bias=False
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        return self.lm_head(out)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = LanguageModel(args.text_config)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        return self.language_model(inputs, mask, cache)
+
+    def sanitize(self, weights):
+        def to_remove(k):
+            return "vision_model" in k or "multi_modal_projector" in k
+
+        # Remove vision weights
+        weights = {k: v for k, v in weights.items() if not to_remove(k)}
+
+        # Rename expert weights for SwitchGLU
+        for l in range(self.args.text_config.num_hidden_layers):
+            prefix = f"language_model.model.layers.{l}.feed_forward.experts"
+            if f"{prefix}.gate_up_proj" in weights:
+                v = weights.pop(f"{prefix}.gate_up_proj")
+                gate_k = f"{prefix}.gate_proj.weight"
+                up_k = f"{prefix}.up_proj.weight"
+                gate_proj, up_proj = mx.split(v, 2, axis=-1)
+                weights[gate_k] = mx.swapaxes(gate_proj, 1, 2)
+                weights[up_k] = mx.swapaxes(up_proj, 1, 2)
+            if f"{prefix}.down_proj" in weights:
+                down_proj = weights.pop(f"{prefix}.down_proj")
+                weights[f"{prefix}.down_proj.weight"] = mx.swapaxes(down_proj, 1, 2)
+        return weights
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def make_cache(self):
+        chunk_size = self.args.text_config.attention_chunk_size
+        caches = []
+        for i in range(len(self.layers)):
+            if (i + 1) % 4 != 0:
+                caches.append(ChunkedKVCache(chunk_size))
+            else:
+                caches.append(KVCache())
+        return caches
@@ -1,4 +1,4 @@
-# Copyright © 2024 Apple Inc.
+# Copyright © 2024-2025 Apple Inc.

 import math
 from dataclasses import dataclass
@@ -123,17 +123,16 @@ class MambaBlock(nn.Module):
            self.intermediate_size, self.hidden_size, bias=args.use_bias
        )

-    def ssm_step(self, x, state=None):
-        A = -mx.exp(self.A_log)
+    def ssm_step(self, x, A, state=None):
        D = self.D
        deltaBC = self.x_proj(x)
-        delta, B, C = mx.split(
-            deltaBC,
-            indices_or_sections=[
-                self.time_step_rank,
-                self.time_step_rank + self.ssm_state_size,
-            ],
-            axis=-1,
+        delta, B, C = map(
+            self.mixer_norm if self.use_bcdt_rms else lambda x: x,
+            mx.split(
+                deltaBC,
+                [self.time_step_rank, self.time_step_rank + self.ssm_state_size],
+                axis=-1,
+            ),
        )
        if self.use_bcdt_rms:
            delta, B, C = map(self.mixer_norm, (delta, B, C))
@@ -145,25 +144,40 @@ class MambaBlock(nn.Module):
        y = y + D * x
        return y, new_state

-    def __call__(self, x, cache):
+    def _process_sequence(self, x, conv_cache, state_cache):
        B, T, D = x.shape
-        if cache is None:
-            cache = [None, None]
+        xz = self.in_proj(x)
+        x, z = xz.split(indices_or_sections=2, axis=-1)
+
+        conv_out, new_conv_cache = self.conv1d(x, conv_cache)
+        x = nn.silu(conv_out)
+
+        A = -mx.exp(self.A_log)

        outputs = []
+        current_state = state_cache
+        y = []
        for t in range(T):
-            xt = x[:, t, :]
-            xz = self.in_proj(xt)
-            x_t, z_t = xz.split(indices_or_sections=2, axis=1)
-            conv_out, cache[0] = self.conv1d(mx.expand_dims(x_t, 1), cache[0])
-            x_t = conv_out.squeeze(1)
-            x_t = nn.silu(x_t)
-            y_t, cache[1] = self.ssm_step(x_t, cache[1])
-            z_t = nn.silu(z_t)
-            output_t = y_t * z_t
-            output_t = self.out_proj(output_t)
-            outputs.append(output_t)
-        output = mx.stack(outputs, axis=1)
+            y_t, current_state = self.ssm_step(x[:, t], A, current_state)
+            y.append(y_t)
+        y = mx.stack(y, axis=1)
+        z = self.out_proj(nn.silu(z) * y)
+        return z, (new_conv_cache, current_state)
+
+    def __call__(self, x, cache):
+        if cache is None:
+            conv_cache, state_cache = None, None
+        else:
+            conv_cache, state_cache = cache[0], cache[1]
+
+        output, (new_conv_cache, new_state_cache) = self._process_sequence(
+            x, conv_cache, state_cache
+        )
+
+        if isinstance(cache, MambaCache):
+            cache[0] = new_conv_cache
+            cache[1] = new_state_cache
+
        return output


@@ -0,0 +1,196 @@
+# Copyright © 2023-2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    max_position_embeddings: int = 32768
+    rope_theta: float = 10000.0
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+    num_nextn_predict_layers: int = 2
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        assert args.num_key_value_heads is not None
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=True)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.rope = initialize_rope(
+            head_dim,
+            base=args.rope_theta,
+            traditional=args.rope_traditional,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class MiMoModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.num_nextn_predict_layers = args.num_nextn_predict_layers
+
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        h = self.norm(h)
+
+        return h
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = MiMoModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+
+        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)
+
+        return {
+            k: v
+            for k, v in weights.items()
+            if "self_attn.rotary_emb.inv_freq" not in k
+            and not k.startswith("model.mtp_layers.")
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -1,13 +1,13 @@
-# Copyright © 2023-2024 Apple Inc.
+# Copyright © 2023-2025 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Dict, Optional, Union

 import mlx.core as mx
 import mlx.nn as nn
-import numpy as np

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope


@dataclass
@@ -23,6 +23,7 @@ class ModelArgs(BaseModelArgs):
    num_key_value_heads: int
    scale_depth: float
    scale_emb: float
+    max_position_embeddings: Optional[int] = None
    rope_theta: float = 1000000.0
    rope_traditional: bool = False
    rope_scaling: Optional[Dict[str, Union[str, float]]] = None
@@ -68,17 +69,12 @@ class Attention(nn.Module):
            self.num_heads * self.head_dim, self.hidden_size, bias=False
        )

-        rope_scale = (
-            1 / args.rope_scaling["factor"]
-            if args.rope_scaling is not None and args.rope_scaling["type"] == "linear"
-            else 1
-        )
-
-        self.rope = nn.RoPE(
-            dims=self.head_dim,
-            traditional=args.rope_traditional,
-            base=self.rope_theta,
-            scale=rope_scale,
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            args.rope_traditional,
+            args.rope_scaling,
+            args.max_position_embeddings,
        )

    def __call__(
@@ -138,9 +134,9 @@ class DecoderLayer(nn.Module):
        cache: Optional[Any] = None,
    ) -> mx.array:
        r = self.self_attn(self.input_layernorm(x), mask, cache)
-        h = x + r * (self.scale_depth / np.sqrt(self.num_hidden_layers))
+        h = x + r * (self.scale_depth / self.num_hidden_layers**0.5)
        r = self.mlp(self.post_attention_layernorm(h))
-        out = h + r * (self.scale_depth / np.sqrt(self.num_hidden_layers))
+        out = h + r * (self.scale_depth / self.num_hidden_layers**0.5)
        return out


@@ -158,11 +154,13 @@ class MiniCPMModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs) * self.args.scale_emb

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -186,9 +184,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)

        if not self.args.tie_word_embeddings:
            out = self.lm_head(out / (self.args.hidden_size / self.args.dim_model_base))
@@ -0,0 +1,250 @@
+# Copyright © 2023-2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import SuScaledRoPE
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    dim_model_base: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    q_lora_rank: int
+    qk_nope_head_dim: int
+    qk_rope_head_dim: int
+    kv_lora_rank: int
+    scale_depth: float
+    scale_emb: float
+    max_position_embeddings: int
+    attention_bias: bool = False
+    rope_theta: float = 1000000.0
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[str, float]]] = None
+    tie_word_embeddings: bool = False
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+
+        self.qk_rope_head_dim = self.args.qk_rope_head_dim
+        self.qk_nope_head_dim = self.args.qk_nope_head_dim
+        self.attention_bias = self.args.attention_bias
+        self.kv_lora_rank = self.args.kv_lora_rank
+        self.num_heads = self.args.num_attention_heads
+        self.q_lora_rank = self.args.q_lora_rank
+        self.hidden_size = self.args.hidden_size
+
+        self.v_head_dim = self.hidden_size // self.args.num_attention_heads
+        self.q_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+        self.softmax_scale = self.q_head_dim ** (-0.5)
+
+        self.q_a_proj = nn.Linear(
+            self.hidden_size, self.q_lora_rank, bias=self.attention_bias
+        )
+        self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank)
+
+        self.q_b_proj = nn.Linear(
+            self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
+        )
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=self.attention_bias,
+        )
+
+        self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank)
+
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads
+            * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=self.attention_bias,
+        )
+
+        self.rope = SuScaledRoPE(
+            dims=args.qk_rope_head_dim,
+            base=args.rope_theta,
+            max_position_embeddings=args.max_position_embeddings,
+            original_max_position_embeddings=args.rope_scaling.get(
+                "original_max_position_embeddings", 4096
+            ),
+            short_factor=args.rope_scaling.get("short_factor", 1.0),
+            long_factor=args.rope_scaling.get("long_factor", 1.0),
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Dict[str, mx.array]] = None,
+    ):
+        B, L, _ = x.shape
+
+        # Project query
+        q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x)))
+        q = q.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+        q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1)
+
+        # Project key and value
+        compressed_kv = self.kv_a_proj_with_mqa(x)
+        compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1)
+        k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
+
+        kv = self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
+        kv = kv.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+
+        k_nope, values = mx.split(kv, [self.qk_nope_head_dim], axis=-1)
+
+        # Apply RoPE to the query and key parts that need position embedding
+        if cache is not None:
+            q_pe = self.rope(q_pe, offset=cache.offset)
+            k_pe = self.rope(k_pe, offset=cache.offset)
+        else:
+            q_pe = self.rope(q_pe)
+            k_pe = self.rope(k_pe)
+
+        # Create the full query and key tensors by combining the parts
+        # Broadcast k_pe to all heads
+        k_pe_broadcasted = mx.broadcast_to(
+            k_pe, (B, self.num_heads, L, self.qk_rope_head_dim)
+        )
+
+        # Use concatenate for queries
+        queries = mx.concatenate([q_nope, q_pe], axis=-1)
+
+        # Use concatenate for keys
+        keys = mx.concatenate([k_nope, k_pe_broadcasted], axis=-1)
+
+        # Update cache if needed
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        # Perform attention
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.softmax_scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.gate_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.hidden_size = args.hidden_size
+        self.num_hidden_layers = args.num_hidden_layers
+
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+        self.scale_depth = args.scale_depth
+        self.num_hidden_layers = args.num_hidden_layers
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r * (self.scale_depth / (self.num_hidden_layers**0.5))
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r * (self.scale_depth / (self.num_hidden_layers**0.5))
+        return out
+
+
+class MiniCPM3Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        assert self.vocab_size > 0
+
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [DecoderLayer(args) for _ in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs) * self.args.scale_emb
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = MiniCPM3Model(args)
+
+        if not self.args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+
+        if not self.args.tie_word_embeddings:
+            out = self.lm_head(out / (self.args.hidden_size / self.args.dim_model_base))
+        else:
+            out = self.model.embed_tokens.as_linear(out)
+
+        return out
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -0,0 +1,49 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from . import llama
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+
+    def __post_init__(self):
+        self.text_config["tie_word_embeddings"] = False
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = llama.Model(llama.ModelArgs.from_dict(args.text_config))
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+        )
+
+    def sanitize(self, weights):
+        weights = tree_unflatten(list(weights.items()))
+        weights.pop("vision_tower", None)
+        weights.pop("multi_modal_projector", None)
+        return dict(tree_flatten(weights))
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
@@ -1,8 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

-import math
 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Dict, Optional, Union

 import mlx.core as mx
 import mlx.nn as nn
@@ -162,11 +161,13 @@ class MixtralModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -188,9 +189,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        return self.lm_head(out)

    def sanitize(self, weights):
@@ -0,0 +1,385 @@
+# Copyright © 2025 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 .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass(frozen=True)
+class AttentionConfig:
+    no_op: bool = False
+    replace_with_linear: bool = False
+    sparsify: Optional[list[str]] = None
+    n_heads_in_group: Optional[int] = None  # GQA group size
+    window_length: Optional[int] = None  # Not directly used here, placeholder
+    num_sink_tokens: Optional[int] = None  # Not directly used here, placeholder
+    use_prefill_window_in_sink_attention: bool = (
+        False  # Not directly used here, placeholder
+    )
+    unshifted_sink: bool = False  # Not directly used here, placeholder
+
+    def __post_init__(self):
+        # Ensure consistency: If no-op or linear, other attn params are irrelevant
+        if self.no_op or self.replace_with_linear:
+            # Use object.__setattr__ because the dataclass is frozen
+            object.__setattr__(self, "n_heads_in_group", None)
+            object.__setattr__(self, "window_length", None)
+            object.__setattr__(self, "num_sink_tokens", None)
+        # If it's a standard attention block, n_heads_in_group must be provided
+        elif not self.no_op:
+            if self.n_heads_in_group is None:
+                raise ValueError(
+                    "n_heads_in_group must be specified for active attention blocks"
+                )
+            if self.n_heads_in_group <= 0:
+                raise ValueError(
+                    f"n_heads_in_group must be positive, got {self.n_heads_in_group}"
+                )
+
+
+@dataclass(frozen=True)
+class FFNConfig:
+    no_op: bool = False
+    replace_with_linear: bool = False
+    sparsify: Optional[list[str]] = None
+    ffn_mult: Optional[float] = None
+
+    def __post_init__(self):
+        # Ensure consistency: If no-op or linear, ffn_mult is irrelevant
+        if self.no_op or self.replace_with_linear:
+            object.__setattr__(self, "ffn_mult", None)
+        # If it's a standard FFN block, ffn_mult must be provided
+        elif not self.no_op:
+            if self.ffn_mult is None:
+                raise ValueError("ffn_mult must be specified for active FFN blocks")
+            # Round to prevent potential floating point inconsistencies if needed
+            object.__setattr__(self, "ffn_mult", round(self.ffn_mult, 6))
+
+
+@dataclass(frozen=True)
+class BlockConfig:
+    attention: AttentionConfig
+    ffn: FFNConfig
+
+    @classmethod
+    def from_dict(cls, data: dict):
+        # Helper to create BlockConfig from a dictionary (e.g., loaded from JSON)
+        attn_conf = AttentionConfig(**data.get("attention", {}))
+        ffn_conf = FFNConfig(**data.get("ffn", {}))
+        return cls(attention=attn_conf, ffn=ffn_conf)
+
+
+def _find_multiple(n: int, k: int) -> int:
+    """Finds the smallest multiple of k greater than or equal to n."""
+    if n % k == 0:
+        return n
+    return n + k - (n % k)
+
+
+def _ffn_mult_to_intermediate_size(ffn_mult: float, n_embd: int) -> int:
+    """Calculates intermediate size based on multiplier, rounding up to multiple of 256."""
+    intermediate_size = int(2 * ffn_mult * n_embd / 3)
+    return _find_multiple(intermediate_size, 256)
+
+
+# Activation function mapping
+_ACT2FN = {
+    "silu": nn.silu,
+    "relu": nn.relu,
+    "gelu": nn.gelu,
+    "gelu_new": nn.gelu_approx,
+    "gelu_fast": nn.gelu_approx,
+}
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "nemotron-nas"
+    hidden_size: int = 8192
+    num_hidden_layers: int = 80
+    num_attention_heads: int = 64
+    rms_norm_eps: float = 1e-5
+    vocab_size: int = 128256
+    block_configs: list = field(default_factory=list)  # List of BlockConfig or dicts
+    hidden_act: str = "silu"
+    attention_bias: bool = False
+    mlp_bias: bool = False
+    rope_theta: float = 500000.0
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    max_position_embeddings: int = 131072
+    tie_word_embeddings: bool = False
+
+    def __post_init__(self):
+        # Automatically parse block_configs if they are loaded as dicts
+        if self.block_configs and isinstance(self.block_configs[0], dict):
+            self.block_configs = [
+                BlockConfig.from_dict(conf) for conf in self.block_configs
+            ]
+
+        if len(self.block_configs) != self.num_hidden_layers:
+            raise ValueError(
+                f"Number of block_configs ({len(self.block_configs)}) must match "
+                f"num_hidden_layers ({self.num_hidden_layers})"
+            )
+
+        # Basic validation for RoPE scaling if provided
+        if self.rope_scaling:
+            if "factor" not in self.rope_scaling:
+                raise ValueError("rope_scaling must contain 'factor'")
+            rope_type = self.rope_scaling.get("rope_type")
+            if rope_type is None:
+                raise ValueError("rope_scaling must contain 'rope_type'")
+
+        # Validate individual block configs (post_init in dataclasses already does some)
+        for i, block_conf in enumerate(self.block_configs):
+            attn_conf = block_conf.attention
+            if not attn_conf.no_op and not attn_conf.replace_with_linear:
+                if self.num_attention_heads % attn_conf.n_heads_in_group != 0:
+                    raise ValueError(
+                        f"Layer {i}: num_attention_heads ({self.num_attention_heads}) "
+                        f"must be divisible by n_heads_in_group ({attn_conf.n_heads_in_group})"
+                    )
+
+
+class Attention(nn.Module):
+    """Standard GQA Attention mechanism for layers that use it."""
+
+    def __init__(self, args: ModelArgs, attention_config: AttentionConfig):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = n_heads // attention_config.n_heads_in_group
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+        if (self.head_dim * n_heads) != dim:
+            raise ValueError(
+                f"hidden_size ({dim}) must be divisible by num_attention_heads ({n_heads})"
+            )
+
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.attention_bias)
+
+        # Initialize RoPE based on global config
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            False,  # Llama uses traditional=False
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        keys = keys.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    """Standard Feed-Forward Network for layers that use it."""
+
+    def __init__(self, args: ModelArgs, ffn_config: FFNConfig):
+        super().__init__()
+
+        dim = args.hidden_size
+        # Calculate intermediate dim based on layer's specific config
+        hidden_dim = _ffn_mult_to_intermediate_size(ffn_config.ffn_mult, dim)
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=args.mlp_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=args.mlp_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=args.mlp_bias)
+
+        try:
+            self.act_fn = _ACT2FN[args.hidden_act]
+        except KeyError:
+            raise ValueError(f"Unknown activation function: {args.hidden_act}")
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LinearSubblockReplacement(nn.Module):
+    """A simple linear layer used to replace Attention or MLP blocks."""
+
+    def __init__(self, hidden_size: int, bias: bool):
+        super().__init__()
+        self.linear = nn.Linear(hidden_size, hidden_size, bias=bias)
+
+    def __call__(self, x: mx.array, *args, **kwargs) -> mx.array:
+        # Accepts potential extra args (like mask, cache) but ignores them
+        return self.linear(x)
+
+
+class TransformerBlock(nn.Module):
+    """A single transformer block, potentially heterogeneous based on config."""
+
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        # Get the specific configuration for this layer
+        block_config = args.block_configs[layer_idx]
+        self.attention_config = block_config.attention
+        self.ffn_config = block_config.ffn
+
+        # Conditionally initialize Input LayerNorm (needed unless Attention is no-op)
+        if not self.attention_config.no_op:
+            self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        else:
+            self.input_layernorm = None
+
+        # Conditionally initialize Attention block
+        if self.attention_config.no_op:
+            self.self_attn = None
+        elif self.attention_config.replace_with_linear:
+            self.self_attn = LinearSubblockReplacement(
+                args.hidden_size, args.attention_bias
+            )
+        else:
+            # Standard attention for this layer
+            self.self_attn = Attention(args, self.attention_config)
+
+        # Conditionally initialize Post-Attention LayerNorm (needed unless FFN is no-op)
+        if not self.ffn_config.no_op:
+            self.post_attention_layernorm = nn.RMSNorm(
+                args.hidden_size, eps=args.rms_norm_eps
+            )
+        else:
+            self.post_attention_layernorm = None
+
+        # Conditionally initialize MLP block
+        if self.ffn_config.no_op:
+            self.mlp = None
+        elif self.ffn_config.replace_with_linear:
+            self.mlp = LinearSubblockReplacement(args.hidden_size, args.mlp_bias)
+        else:
+            # Standard MLP for this layer
+            self.mlp = MLP(args, self.ffn_config)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+
+        # Attention part (Input Norm -> Attention -> Residual)
+        if self.self_attn is not None:
+            residual = x
+            h = self.input_layernorm(x)
+            attn_out = self.self_attn(h, mask=mask, cache=cache)
+            x = residual + attn_out
+
+        # MLP part (Post-Attention Norm -> MLP -> Residual)
+        if self.mlp is not None:
+            residual = x
+            h = self.post_attention_layernorm(x)
+            mlp_out = self.mlp(h)
+            x = residual + mlp_out
+
+        return x
+
+
+class NemotronNASModel(nn.Module):
+    """The core Nemotron-NAS style transformer model."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args, layer_idx=i)
+            for i in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[List[Any]] = None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for i, layer in enumerate(self.layers):
+            h = layer(h, mask, cache=cache[i])
+
+        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 = NemotronNASModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        else:
+            self.lm_head = None
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask=None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask=mask, cache=cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+    def sanitize(self, weights):
+        if self.args.tie_word_embeddings:
+            weights.pop("lm_head.weight", None)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -176,11 +176,13 @@ class NemotronModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -203,9 +205,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        if self.args.tie_word_embeddings:
            out = self.model.embed_tokens.as_linear(out)
        else:
@@ -2,7 +2,7 @@

 import sys
 from dataclasses import dataclass
-from typing import Any, Optional, Tuple
+from typing import Any, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -124,11 +124,13 @@ class Transformer(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.wte(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.blocks)
@@ -152,9 +154,10 @@ class OlmoModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        return self.transformer(inputs, cache)
+        return self.transformer(inputs, mask, cache)


 class Model(nn.Module):
@@ -167,9 +170,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        return self.model(inputs, cache)
+        return self.model(inputs, mask, cache)

    @property
    def layers(self):
@@ -163,10 +163,12 @@ class LlamaModel(nn.Module):
        self,
        inputs: mx.array,
        cache=None,
+        mask=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -190,8 +192,9 @@ class Model(nn.Module):
        self,
        inputs: mx.array,
        cache=None,
+        mask=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, cache, mask)
        if self.args.tie_word_embeddings:
            out = self.model.embed_tokens.as_linear(out)
        else:
@@ -0,0 +1,217 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_experts: int
+    num_experts_per_tok: int
+    norm_topk_prob: bool = False
+    head_dim: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
+    attention_bias: bool = False
+    mlp_bias: bool = False
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.head_dim or args.hidden_size // n_heads
+
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.attention_bias)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            args.rope_traditional,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+        self.q_norm = nn.RMSNorm(n_heads * head_dim, args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(n_kv_heads * head_dim, args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+        queries = self.q_norm(queries)
+        keys = self.k_norm(keys)
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class OlmoeSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+
+        self.gate = nn.Linear(args.hidden_size, self.num_experts, bias=False)
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size,
+            args.intermediate_size,
+            self.num_experts,
+            bias=args.mlp_bias,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        B, L, D = x.shape
+        x_flat = x.reshape(-1, D)
+        router_logits = self.gate(x_flat)
+        routing_weights = mx.softmax(router_logits, axis=1, precise=True)
+        k = self.top_k
+        indices = mx.stop_gradient(
+            mx.argpartition(-routing_weights, kth=k - 1, axis=-1)[..., :k]
+        )
+        scores = mx.take_along_axis(routing_weights, indices, axis=-1)
+        if self.norm_topk_prob:
+            scores = scores / scores.sum(axis=-1, keepdims=True)
+        y = self.switch_mlp(x_flat, indices)
+        y = (y * scores[..., None]).sum(axis=-2)
+        return y.reshape(B, L, D)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = OlmoeSparseMoeBlock(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        x = x + self.self_attn(self.input_layernorm(x), mask, cache)
+        x = x + self.mlp(self.post_attention_layernorm(x))
+        return x
+
+
+class OlmoeModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask=None,
+    ):
+        h = self.embed_tokens(inputs)
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+        if cache is None:
+            cache = [None] * len(self.layers)
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = OlmoeModel(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,
+        mask=None,
+    ):
+        out = self.model(inputs, cache, mask)
+        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 "model.layers.0.mlp.experts.0.up_proj.weight" not in weights:
+            return weights
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            for n in ["up_proj", "down_proj", "gate_proj"]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{n}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{n}.{k}")
+                            for e in range(self.args.num_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{n}.{k}"] = mx.stack(to_join)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -1,7 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.

 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Union

 import mlx.core as mx
 import mlx.nn as nn
@@ -178,11 +178,13 @@ class OpenELMModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.token_embeddings(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -205,9 +207,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.transformer(inputs, cache)
+        out = self.transformer(inputs, mask, cache)
        if self.args.share_input_output_layers:
            out = self.transformer.token_embeddings.as_linear(out)
        else:
@@ -2,7 +2,6 @@

 import math
 from dataclasses import dataclass
-from typing import Tuple

 import mlx.core as mx
 import mlx.nn as nn
@@ -112,10 +111,9 @@ class PhiMLP(nn.Module):
        super().__init__()
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
-        self.act = nn.GELU(approx="precise")

    def __call__(self, x) -> mx.array:
-        return self.fc2(self.act(self.fc1(x)))
+        return self.fc2(nn.gelu_approx(self.fc1(x)))


 class PhiDecoderLayer(nn.Module):
@@ -143,10 +141,11 @@ class PhiModel(nn.Module):
            config.hidden_size, eps=config.layer_norm_eps
        )

-    def __call__(self, x, cache):
+    def __call__(self, x, mask, cache):
        x = self.embed_tokens(x)

-        mask = create_attention_mask(x, cache)
+        if mask is None:
+            mask = create_attention_mask(x, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -167,9 +166,10 @@ class Model(nn.Module):
    def __call__(
        self,
        x: mx.array,
+        mask: mx.array = None,
        cache=None,
    ) -> mx.array:
-        y = self.model(x, cache)
+        y = self.model(x, mask, cache)
        return self.lm_head(y)

    @property
@@ -7,7 +7,7 @@ import mlx.core as mx
 import mlx.nn as nn

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
-from .su_rope import SuScaledRotaryEmbedding
+from .rope_utils import SuScaledRoPE


@dataclass
@@ -23,8 +23,10 @@ class ModelArgs(BaseModelArgs):
    rope_theta: float = 10000
    rope_traditional: bool = False
    rope_scaling: Optional[Dict[str, Union[float, List[float]]]] = None
+    partial_rotary_factor: float = 1.0
    max_position_embeddings: int = 131072
    original_max_position_embeddings: int = 4096
+    tie_word_embeddings: bool = False

    def __post_init__(self):
        if self.num_key_value_heads is None:
@@ -59,9 +61,10 @@ class Attention(nn.Module):
        self.qkv_proj = nn.Linear(dim, op_size, bias=False)
        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)

+        rope_dim = int(head_dim * args.partial_rotary_factor)
        if args.rope_scaling and args.rope_scaling["type"] in ["longrope", "su"]:
-            self.rope = SuScaledRotaryEmbedding(
-                head_dim,
+            self.rope = SuScaledRoPE(
+                rope_dim,
                base=args.rope_theta,
                max_position_embeddings=args.max_position_embeddings,
                original_max_position_embeddings=args.original_max_position_embeddings,
@@ -74,7 +77,7 @@ class Attention(nn.Module):
                assert isinstance(args.rope_scaling["factor"], float)
                rope_scale = 1 / args.rope_scaling["factor"]
            self.rope = nn.RoPE(
-                head_dim,
+                rope_dim,
                traditional=args.rope_traditional,
                base=args.rope_theta,
                scale=rope_scale,
@@ -168,11 +171,13 @@ class Phi3Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -188,16 +193,22 @@ class Model(nn.Module):
        super().__init__()
        self.model_type = args.model_type
        self.model = Phi3Model(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)
        self.args = args

    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
-        return self.lm_head(out)
+        out = self.model(inputs, mask, cache)
+        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):
@@ -258,13 +258,15 @@ class Phi3Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)
        if self.mup_embedding_multiplier:
            h = self.mup_embedding_multiplier * h

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache, return_array=True)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -290,9 +292,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        out = self.model.embed_tokens.as_linear(out)
        if self.mup_width_multiplier:
            out = out / self.mup_width_multiplier
@@ -7,7 +7,7 @@ import mlx.core as mx
 import mlx.nn as nn

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
-from .su_rope import SuScaledRotaryEmbedding
+from .rope_utils import SuScaledRoPE
 from .switch_layers import SwitchGLU


@@ -45,7 +45,7 @@ class Attention(nn.Module):
        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=True)

-        self.rope = SuScaledRotaryEmbedding(
+        self.rope = SuScaledRoPE(
            head_dim,
            base=args.rope_theta,
            max_position_embeddings=args.max_position_embeddings,
@@ -155,11 +155,13 @@ class PhiMoEModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ) -> mx.array:
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -181,9 +183,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        return self.lm_head(out)

    def sanitize(self, weights):
@@ -175,7 +175,9 @@ class Model(nn.Module):
        mask: mx.array = None,
        cache=None,
    ) -> mx.array:
-        mask = create_attention_mask(x, cache)
+
+        if mask is None:
+            mask = create_attention_mask(x, cache)

        y = self.transformer(x, mask, cache)
        return self.lm_head(y)
@@ -0,0 +1,52 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from . import llama
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+
+    def __post_init__(self):
+        self.text_config["tie_word_embeddings"] = False
+        self.text_config["num_attention_heads"] = self.text_config.get(
+            "num_attention_heads", 32
+        )
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.language_model = llama.Model(llama.ModelArgs.from_dict(args.text_config))
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        mask: Optional[mx.array] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ):
+        return self.language_model(
+            inputs, cache=cache, mask=mask, input_embeddings=input_embeddings
+        )
+
+    def sanitize(self, weights):
+        weights = tree_unflatten(list(weights.items()))
+        weights.pop("vision_tower", None)
+        weights.pop("multi_modal_projector", None)
+        return dict(tree_flatten(weights))
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
@@ -174,10 +174,12 @@ class PlamoModel(nn.Module):
        self,
        inputs: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ) -> mx.array:
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None for _ in range(len(self.layers.layers))]
@@ -202,8 +204,9 @@ class Model(nn.Module):
        self,
        inputs: mx.array,
        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
    ) -> mx.array:
-        out = self.model(inputs, cache)
+        out = self.model(inputs, cache, mask)
        return self.lm_head(out)

    @property
@@ -0,0 +1,599 @@
+# Copyright © 2025 Apple Inc.
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from mlx_lm.models.base import BaseModelArgs, create_attention_mask
+
+from .cache import KVCache, MambaCache
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str = "plamo2"
+    hidden_size: int = 4096
+    num_hidden_layers: int = 32
+    rms_norm_eps: float = 1e-6
+    tie_word_embeddings: bool = True
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 4
+    hidden_size_per_head: int = 128
+    max_position_embeddings: int = 2048
+    attention_window_size: int = 2048
+    full_attention_idx: Optional[list[int]] = None
+    mamba_d_state: int = 64
+    mamba_d_conv: int = 4
+    mamba_num_heads: int = 64
+    mamba_step: int = 2
+    mamba_chunk_size: int = 256
+    mamba_enabled: bool = True
+    intermediate_size: int = 13312
+    vocab_size: int = 32000
+
+
+class RMSNorm(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-6,
+        offset: float = 1.0,
+    ) -> None:
+        super().__init__()
+        self.weight = mx.zeros(hidden_size)
+        self.variance_epsilon = eps
+        self.offset = offset
+
+    def __call__(self, hidden_states: mx.array) -> mx.array:
+        return mx.fast.rms_norm(
+            hidden_states, self.weight + self.offset, self.variance_epsilon
+        )
+
+
+def get_initial_dt_bias(num_heads: int) -> mx.array:
+    dt_min = 0.001
+    dt_max = 0.1
+    dt = mx.exp(
+        mx.random.uniform(shape=(num_heads,)) * (math.log(dt_max) - math.log(dt_min))
+        + math.log(dt_min)
+    )
+    dt = mx.clip(dt, a_min=1e-4, a_max=None)
+    inv_dt = dt + mx.log(-mx.expm1(-dt))
+    return inv_dt
+
+
+def get_initial_A(num_heads: int) -> mx.array:
+    A = mx.arange(1, num_heads + 1, dtype=mx.float32)
+    return mx.log(A)
+
+
+# From: https://github.com/state-spaces/mamba/blob/0cce0fa645f100f00620ddf2333c2b7712abfdec/mamba_ssm/ops/triton/selective_state_update.py#L219
+def selective_state_update_ref(
+    state, x, dt, A, B, C, D=None, z=None, dt_bias=None, dt_softplus=False
+) -> tuple[mx.array, mx.array]:
+    """
+    Argument:
+        state: (batch, dim, dstate) or (batch, nheads, dim, dstate)
+        x: (batch, dim) or (batch, nheads, dim)
+        dt: (batch, dim) or (batch, nheads, dim)
+        A: (dim, dstate) or (nheads, dim, dstate)
+        B: (batch, dstate) or (batch, ngroups, dstate)
+        C: (batch, dstate) or (batch, ngroups, dstate)
+        D: (dim,) or (nheads, dim)
+        z: (batch, dim) or (batch, nheads, dim)
+        dt_bias: (dim,) or (nheads, dim)
+    Return:
+        out: (batch, dim) or (batch, nheads, dim)
+    """
+    has_heads = state.ndim > 3
+    if state.ndim == 3:
+        state = mx.expand_dims(state, 1)
+    if x.ndim == 2:
+        x = mx.expand_dims(x, 1)
+    if dt.ndim == 2:
+        dt = mx.expand_dims(dt, 1)
+    if A.ndim == 2:
+        A = mx.expand_dims(A, 0)
+    if B.ndim == 2:
+        B = mx.expand_dims(B, 1)
+    if C.ndim == 2:
+        C = mx.expand_dims(C, 1)
+    if D is not None and D.ndim == 1:
+        D = mx.expand_dims(D, 0)
+    if z is not None and z.ndim == 2:
+        z = mx.expand_dims(z, 1)
+    if dt_bias is not None and dt_bias.ndim == 1:
+        dt_bias = mx.expand_dims(dt_bias, 0)
+    batch, nheads, dim, dstate = state.shape
+    assert x.shape == (batch, nheads, dim)
+    assert dt.shape == x.shape
+    assert A.shape == (nheads, dim, dstate)
+    ngroups = B.shape[1]
+    assert nheads % ngroups == 0, "nheads must be divisible by ngroups"
+    assert B.shape == (batch, ngroups, dstate)
+    assert C.shape == B.shape
+    if D is not None:
+        assert D.shape == (nheads, dim)
+    if z is not None:
+        assert z.shape == x.shape
+    if dt_bias is not None:
+        assert dt_bias.shape == (nheads, dim)
+        dt = dt + dt_bias
+    dt = nn.softplus(dt) if dt_softplus else dt
+    dA = mx.exp(mx.expand_dims(dt, axis=-1) * A)  # (batch, nheads, dim, dstate)
+    B = mx.reshape(
+        mx.repeat(mx.expand_dims(B, axis=2), nheads // ngroups, 2),
+        (batch, nheads, dstate),
+    )  # (batch, nheads, dstate)
+    C = mx.reshape(
+        mx.repeat(mx.expand_dims(C, axis=2), nheads // ngroups, 2),
+        (batch, nheads, dstate),
+    )  # (batch, nheads, dstate)
+    dB = mx.expand_dims(dt, axis=-1) * mx.expand_dims(
+        B, axis=-2
+    )  # (batch, nheads, dim, dstate)
+    state = state * dA + dB * mx.expand_dims(x, axis=-1)  # (batch, dim, dstate)
+    out = mx.einsum("bhdn,bhn->bhd", state.astype(C.dtype), C)
+    if D is not None:
+        out += (x * D).astype(out.dtype)
+    out = (out if z is None else out * nn.silu(z)).astype(x.dtype)
+    if not has_heads:
+        out = out.squeeze(1)
+    return out, state
+
+
+def ssd_update_state(
+    ssm_state: mx.array,
+    x: mx.array,
+    dt: mx.array,
+    A: mx.array,
+    B: mx.array,
+    C: mx.array,
+    D: mx.array,
+    z: mx.array,
+    dt_bias: mx.array,
+    dt_softplus: bool,
+) -> tuple[mx.array, mx.array]:
+    assert ssm_state.dtype == mx.float32
+    dtype = x.dtype
+
+    hidden_size_per_head = x.shape[-1]
+    d_state = B.shape[-1]
+    A = mx.broadcast_to(
+        A[:, None, None], (A.shape[0], hidden_size_per_head, d_state)
+    ).astype(mx.float32)
+    dt = mx.broadcast_to(
+        dt[..., None], (dt.shape[0], dt.shape[1], hidden_size_per_head)
+    )
+    dt_bias = mx.broadcast_to(
+        dt_bias[:, None], (dt_bias.shape[0], hidden_size_per_head)
+    )
+    D = mx.broadcast_to(D[:, None], (D.shape[0], hidden_size_per_head))
+    out, ssm_state = selective_state_update_ref(
+        ssm_state,
+        x.astype(dtype),
+        dt.astype(dtype),
+        A.astype(mx.float32),
+        B.astype(dtype),
+        C.astype(dtype),
+        D.astype(mx.float32),
+        z.astype(dtype),
+        dt_bias.astype(mx.float32),
+        dt_softplus=dt_softplus,
+    )
+    return out[:, None], ssm_state
+
+
+def ssd_chunk_scan_combined(
+    x: mx.array,
+    dt: mx.array,
+    A: mx.array,
+    B: mx.array,
+    C: mx.array,
+    D: mx.array,
+    z: mx.array,
+    dt_bias: mx.array,
+    dt_softplus: bool,
+    ssm_state: mx.array,
+) -> tuple[mx.array, mx.array]:
+    assert ssm_state.dtype == mx.float32
+    length = x.shape[1]
+    ys = []
+    for i in range(length):
+        y, ssm_state = ssd_update_state(
+            ssm_state,
+            x[:, i],
+            dt[:, i],
+            A,
+            B[:, i],
+            C[:, i],
+            D if D.ndim == 1 else D[:, i],
+            z=z[:, i],
+            dt_bias=dt_bias,
+            dt_softplus=dt_softplus,
+        )
+        ys.append(y)
+    return mx.concatenate(ys, axis=1), ssm_state
+
+
+def causal_conv1d_update(conv_state, x, weight) -> tuple[mx.array, mx.array]:
+    _, seqlen, dim = x.shape
+    state_len = conv_state.shape[-2]
+    x = mx.concatenate([conv_state, x], axis=-2)
+    conv_state = x[:, -state_len:]
+    out = mx.conv1d(
+        x,
+        weight,
+        padding=0,
+        groups=dim,
+    )[:, -seqlen:]
+    return nn.silu(out), conv_state
+
+
+class Mamba(nn.Module):
+    def __init__(self, config: ModelArgs) -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.d_state = config.mamba_d_state
+        self.d_conv = config.mamba_d_conv
+        self.chunk_size = config.mamba_chunk_size
+        self.num_heads = config.mamba_num_heads
+        self.hidden_size_per_head = config.hidden_size_per_head
+
+        self.intermediate_size = self.num_heads * self.hidden_size_per_head
+
+        self.in_proj = nn.Linear(
+            self.hidden_size, 2 * self.intermediate_size, bias=False
+        )
+        self.conv1d = nn.Conv1d(
+            in_channels=self.intermediate_size,
+            out_channels=self.intermediate_size,
+            bias=False,
+            kernel_size=self.d_conv,
+            groups=self.intermediate_size,
+            padding=0,
+        )
+        self.dt_dim = max(64, self.hidden_size // 16)
+        self.bcdt_proj = nn.Linear(
+            self.intermediate_size,
+            self.dt_dim + 2 * self.d_state,
+            bias=False,
+        )
+        self.dt_proj = nn.Linear(self.dt_dim, self.num_heads, bias=False)
+
+        self.dt_bias = get_initial_dt_bias(self.num_heads)
+        self.A_log = get_initial_A(self.num_heads)
+        self.D = mx.ones(self.num_heads, dtype=mx.float32)
+
+        self.dt_norm_weight = mx.ones(self.dt_dim)
+        self.B_norm_weight = mx.ones(self.d_state)
+        self.C_norm_weight = mx.ones(self.d_state)
+
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        bsize, length, _ = hidden_states.shape
+
+        if cache is not None and cache[0] is not None:
+            conv_state = cache[0]
+            ssm_state = cache[1]
+        else:
+            conv_state = mx.zeros(
+                (bsize, self.d_conv - 1, self.intermediate_size),
+                dtype=hidden_states.dtype,
+            )
+            ssm_state = mx.zeros(
+                (bsize, self.num_heads, self.hidden_size_per_head, self.d_state),
+                dtype=mx.float32,
+            )
+
+        zx = self.in_proj(hidden_states)
+        zx = zx.reshape(bsize, length, self.num_heads, -1)
+        # z: (bsize, length, num_heads, hidden_size_per_head)
+        # x: (bsize, length, num_heads, hidden_size_per_head)
+        z, x = mx.split(
+            zx,
+            [
+                self.hidden_size_per_head,
+            ],
+            axis=-1,
+        )
+
+        x = x.reshape(bsize, -1, self.num_heads * self.hidden_size_per_head)
+        x, conv_state = causal_conv1d_update(conv_state, x, self.conv1d.weight)
+        BCdt = self.bcdt_proj(x)
+        x = x.reshape(bsize, length, self.num_heads, -1)
+        B, C, dt = mx.split(BCdt, [self.d_state, self.d_state * 2], axis=-1)
+
+        A = -mx.exp(self.A_log.astype(mx.float32))  # (num_heads,)
+        dt = mx.fast.rms_norm(dt, self.dt_norm_weight, self.config.rms_norm_eps)
+        B = mx.fast.rms_norm(B, self.B_norm_weight, self.config.rms_norm_eps)
+        C = mx.fast.rms_norm(C, self.C_norm_weight, self.config.rms_norm_eps)
+
+        # (bsize, length, num_heads, 1)
+        dt = self.dt_proj(dt)[..., None]
+
+        out, ssm_state = ssd_chunk_scan_combined(
+            x,
+            dt.reshape(bsize, length, -1),
+            A,
+            B,
+            C,
+            D=self.D,
+            z=z,
+            dt_bias=self.dt_bias,
+            dt_softplus=True,
+            ssm_state=ssm_state,
+        )
+
+        if cache is not None:
+            cache[0] = conv_state
+            cache[1] = ssm_state
+        y = self.out_proj(out.reshape(bsize, length, -1))
+
+        return y
+
+
+class Attention(nn.Module):
+    def __init__(self, config: ModelArgs) -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        head_dim = config.hidden_size_per_head
+        self.max_position_embeddings = config.max_position_embeddings
+        self.scale = head_dim**-0.5
+
+        self.q_num_heads = config.num_attention_heads
+        self.qk_dim = self.v_dim = head_dim
+        self.k_num_heads = self.v_num_heads = config.num_key_value_heads
+        assert self.q_num_heads % self.k_num_heads == 0
+        self.n_group = self.q_num_heads // self.k_num_heads
+
+        self.q_proj_dim = self.q_num_heads * self.qk_dim
+        self.k_proj_dim = self.k_num_heads * self.qk_dim
+        self.v_proj_dim = self.k_num_heads * self.v_dim
+        self.qkv_proj = nn.Linear(
+            self.hidden_size,
+            self.q_proj_dim + self.k_proj_dim + self.v_proj_dim,
+            bias=False,
+        )
+        self.o_proj = nn.Linear(
+            self.q_num_heads * self.v_dim, self.hidden_size, bias=False
+        )
+
+        self.q_weight = mx.ones((self.q_num_heads, self.qk_dim))
+        self.k_weight = mx.ones((self.k_num_heads, self.qk_dim))
+
+        self.rope = nn.RoPE(self.qk_dim)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        B, T, _ = hidden_states.shape
+
+        qkv = self.qkv_proj(hidden_states)
+        q, k, v = mx.split(
+            qkv, [self.q_proj_dim, self.q_proj_dim + self.k_proj_dim], axis=-1
+        )
+        q = q.reshape(B, T, self.q_num_heads, self.qk_dim).transpose(0, 2, 1, 3)
+        k = k.reshape(B, T, self.k_num_heads, self.qk_dim).transpose(0, 2, 1, 3)
+        v = v.reshape(B, T, self.v_num_heads, self.v_dim).transpose(0, 2, 1, 3)
+
+        q = mx.fast.rms_norm(q, weight=None, eps=1e-6) * self.q_weight[:, None]
+        k = mx.fast.rms_norm(k, weight=None, eps=1e-6) * self.k_weight[:, None]
+
+        if cache is not None:
+            q = self.rope(q, offset=cache.offset)
+            k = self.rope(k, offset=cache.offset)
+            k, v = cache.update_and_fetch(k, v)
+        else:
+            q = self.rope(q)
+            k = self.rope(k)
+
+        output = mx.fast.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            scale=self.scale,
+            mask=mask,
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(
+            B, T, self.q_num_heads * self.v_dim
+        )
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: ModelArgs) -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_up_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size * 2, bias=False
+        )
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        h = self.gate_up_proj(x)
+        hs = mx.split(h, 2, axis=-1)
+        return self.down_proj(nn.silu(hs[0]) * hs[1])
+
+
+class PlamoDecoderLayer(nn.Module):
+    def __init__(self, config: ModelArgs, is_mamba: bool) -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.is_mamba = is_mamba
+        self.mixer: nn.Module
+        if is_mamba:
+            self.mixer = Mamba(config)
+        else:
+            self.mixer = Attention(config)
+        self.mlp = MLP(config)
+        self.pre_mixer_norm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps, offset=1.0
+        )
+        self.post_mixer_norm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps, offset=1.0 / 5
+        )
+        self.pre_mlp_norm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps, offset=1.0
+        )
+        self.post_mlp_norm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps, offset=1.0 / (5**1.5)
+        )
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        residual = hidden_states
+        hidden_states = self.pre_mixer_norm(hidden_states)
+
+        hidden_states_sa = self.mixer(
+            hidden_states=hidden_states,
+            mask=mask,
+            cache=cache,
+        )
+
+        hidden_states_sa = self.post_mixer_norm(hidden_states_sa)
+        hidden_states = residual + hidden_states_sa
+
+        residual = hidden_states
+        hidden_states = self.pre_mlp_norm(hidden_states)
+
+        # Fully Connected
+        hidden_states_mlp = self.mlp(hidden_states)
+
+        # Residual
+        hidden_states_mlp = self.post_mlp_norm(hidden_states_mlp)
+        return residual + hidden_states_mlp
+
+
+def is_mamba(config: ModelArgs, i: int) -> bool:
+    if not config.mamba_enabled:
+        return False
+    assert config.mamba_step > 1
+    assert i < config.num_hidden_layers
+
+    if config.num_hidden_layers <= (config.mamba_step // 2):
+        # use attention in last layer
+        return i != config.num_hidden_layers - 1
+    return (i % config.mamba_step) != (config.mamba_step // 2)
+
+
+class PlamoDecoder(nn.Module):
+    def __init__(self, config: ModelArgs) -> None:
+        super().__init__()
+
+        self.layers = [
+            PlamoDecoderLayer(config, is_mamba=is_mamba(config, i))
+            for i in range(config.num_hidden_layers)
+        ]
+
+    def __call__(self, x: mx.array, mask: mx.array, cache):
+        for i, decoder_layer in enumerate(self.layers):
+            x = decoder_layer(
+                x,
+                mask=mask,
+                cache=cache[i],
+            )
+        return x
+
+
+class PlamoModel(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+
+        self.config = config
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = PlamoDecoder(config)  # type: ignore
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        batch_size, seq_length = inputs.shape
+
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, [cache[1]] if cache is not None else None)
+
+        if cache is None:
+            cache = [None] * len(self.layers.layers)
+
+        # decoder layers
+        out = self.layers(
+            h,
+            mask,
+            cache,
+        )
+
+        return self.norm(out)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs) -> None:
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        self.model = PlamoModel(config)
+
+        self.vocab_size = config.vocab_size
+
+        if not config.tie_word_embeddings:
+            self.lm_head: nn.Module = nn.Linear(
+                config.hidden_size, self.vocab_size, bias=False
+            )
+
+    def sanitize(self, weights: dict[Any, Any]) -> dict[Any, Any]:
+        for k, v in weights.items():
+            if "conv1d.weight" in k and v.shape[-1] != 1:
+                weights[k] = v.moveaxis(2, 1)
+        return weights
+
+    def make_cache(self):
+        # TODO use RotatingKVCache is not full_attn
+        # full_attn = self.layer_idx in self.config.full_attention_idx
+        return [MambaCache() if l.is_mamba else KVCache() for l in self.layers]
+
+    def __call__(
+        self, inputs: mx.array, mask: Optional[mx.array] = None, cache=None
+    ) -> mx.array:
+        outputs = self.model(
+            inputs=inputs,
+            mask=None,
+            cache=cache,
+        )
+        if self.config.tie_word_embeddings:
+            logits = self.model.embed_tokens.as_linear(outputs)
+        else:
+            logits = self.lm_head(outputs)
+
+        return logits
+
+    @property
+    def layers(self):
+        return self.model.layers.layers
@@ -123,7 +123,8 @@ class QwenModel(nn.Module):
    def __call__(self, inputs, mask=None, cache=None):
        x = self.wte(inputs)

-        mask = create_attention_mask(x, cache)
+        if mask is None:
+            mask = create_attention_mask(x, cache)

        if cache is None:
            cache = [None] * len(self.h)
@@ -7,6 +7,7 @@ import mlx.core as mx
 import mlx.nn as nn

 from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope


@dataclass
@@ -18,24 +19,13 @@ class ModelArgs(BaseModelArgs):
    num_attention_heads: int
    rms_norm_eps: float
    vocab_size: int
-    num_key_value_heads: Optional[int] = None
+    num_key_value_heads: int
+    max_position_embeddings: int = 32768
    rope_theta: float = 1000000
    rope_traditional: bool = False
    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
    tie_word_embeddings: bool = True

-    def __post_init__(self):
-        if self.num_key_value_heads is None:
-            self.num_key_value_heads = self.num_attention_heads
-
-        if self.rope_scaling:
-            required_keys = {"factor", "type"}
-            if not all(key in self.rope_scaling for key in required_keys):
-                raise ValueError(f"rope_scaling must contain keys {required_keys}")
-
-            if self.rope_scaling["type"] != "linear":
-                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
-

 class Attention(nn.Module):
    def __init__(self, args: ModelArgs):
@@ -54,16 +44,12 @@ class Attention(nn.Module):
        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)

-        rope_scale = (
-            1 / args.rope_scaling["factor"]
-            if args.rope_scaling is not None and args.rope_scaling["type"] == "linear"
-            else 1
-        )
-        self.rope = nn.RoPE(
+        self.rope = initialize_rope(
            head_dim,
-            traditional=args.rope_traditional,
            base=args.rope_theta,
-            scale=rope_scale,
+            traditional=args.rope_traditional,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
        )

    def __call__(
@@ -149,11 +135,17 @@ class Qwen2Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
+        input_embeddings: Optional[mx.array] = None,
    ):
-        h = self.embed_tokens(inputs)
+        if input_embeddings is not None:
+            h = input_embeddings
+        else:
+            h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -176,9 +168,11 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
+        input_embeddings: Optional[mx.array] = None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache, input_embeddings)
        if self.args.tie_word_embeddings:
            out = self.model.embed_tokens.as_linear(out)
        else:
@@ -1,6 +1,5 @@
 # Copyright © 2023-2024 Apple Inc.

-import math
 from dataclasses import dataclass
 from typing import Any, Dict, Optional, Union

@@ -187,11 +186,13 @@ class Qwen2MoeModel(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -213,9 +214,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        return self.lm_head(out)

    def sanitize(self, weights):
@@ -0,0 +1,189 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    max_position_embeddings: int
+    rope_theta: float
+    head_dim: int
+    tie_word_embeddings: bool
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        assert args.num_key_value_heads is not None
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.head_dim
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.q_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+        self.rope = initialize_rope(
+            head_dim,
+            base=args.rope_theta,
+            traditional=False,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = self.q_norm(queries.reshape(B, L, self.n_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        keys = self.k_norm(keys.reshape(B, L, self.n_kv_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class Qwen3Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Qwen3Model(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        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)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -0,0 +1,240 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_experts: int
+    num_experts_per_tok: int
+    decoder_sparse_step: int
+    mlp_only_layers: List[int]
+    moe_intermediate_size: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    head_dim: int
+    rope_theta: float
+    tie_word_embeddings: bool
+    max_position_embeddings: int
+    norm_topk_prob: bool
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        assert args.num_key_value_heads is not None
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = getattr(
+            args, "head_dim", args.hidden_size // args.num_attention_heads
+        )
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.q_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(head_dim, eps=args.rms_norm_eps)
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=False,
+            base=args.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = self.q_norm(queries.reshape(B, L, self.n_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        keys = self.k_norm(keys.reshape(B, L, self.n_kv_heads, -1)).transpose(
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class Qwen3MoeSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        intermediate_size = args.moe_intermediate_size
+
+        self.num_experts = num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+
+        self.gate = nn.Linear(dim, num_experts, bias=False)
+        self.switch_mlp = SwitchGLU(dim, intermediate_size, num_experts)
+
+    def __call__(
+        self,
+        x: mx.array,
+    ):
+        gates = self.gate(x)
+        gates = mx.softmax(gates, axis=-1, precise=True)
+
+        k = self.top_k
+        inds = mx.stop_gradient(mx.argpartition(-gates, kth=k - 1, axis=-1)[..., :k])
+        scores = mx.take_along_axis(gates, inds, axis=-1)
+        if self.norm_topk_prob:
+            scores /= mx.sum(scores, axis=-1, keepdims=True)
+
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2)
+
+        return y
+
+
+class Qwen3MoeDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args, layer_idx)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.args = args
+
+        if (layer_idx not in args.mlp_only_layers) and (
+            args.num_experts > 0 and (layer_idx + 1) % args.decoder_sparse_step == 0
+        ):
+            self.mlp = Qwen3MoeSparseMoeBlock(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:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class Qwen3MoeModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            Qwen3MoeDecoderLayer(args=args, layer_idx=i)
+            for i in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Qwen3MoeModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        return self.lm_head(out)
+
+    def sanitize(self, weights):
+        if "model.layers.0.mlp.experts.0.up_proj.weight" not in weights:
+            return weights
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"model.layers.{l}"
+            for n in ["up_proj", "down_proj", "gate_proj"]:
+                if f"{prefix}.mlp.experts.0.{n}.weight" in weights:
+                    to_join = [
+                        weights.pop(f"{prefix}.mlp.experts.{e}.{n}.weight")
+                        for e in range(self.args.num_experts)
+                    ]
+                    weights[f"{prefix}.mlp.switch_mlp.{n}.weight"] = mx.stack(to_join)
+        return weights
+
+    @property
+    def layers(self):
+        return self.model.layers
@@ -2,7 +2,7 @@

 import math
 from dataclasses import dataclass
-from typing import List, Literal, Optional
+from typing import List, Optional

 import mlx.core as mx
 import mlx.nn as nn
@@ -389,6 +389,7 @@ class Griffin(nn.Module):
    def __call__(
        self,
        tokens,
+        mask: mx.array = None,
        cache=None,
    ):
        x = self.embed_tokens(tokens)
@@ -402,7 +403,8 @@ class Griffin(nn.Module):
            if block.temporal_block_type != "recurrent":
                mask_cache = [cache[i]]

-        mask = create_attention_mask(x, mask_cache)
+        if mask is None:
+            mask = create_attention_mask(x, mask_cache)

        for i, block in enumerate(self.layers):
            x = block(x, mask=mask, cache=cache[i])
@@ -418,12 +420,12 @@ class Model(nn.Module):
        self.model_type = config.model_type
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

-    def __call__(self, tokens: mx.array, cache=None) -> mx.array:
+    def __call__(self, tokens: mx.array, mask: mx.array = None, cache=None) -> mx.array:
        """
        Args:
          tokens: Sequence of input tokens.
        """
-        logits = self.model(tokens, cache=cache)
+        logits = self.model(tokens, mask=mask, cache=cache)
        if "lm_head" in self:
            logits = self.lm_head(logits)
        else:
@@ -1,11 +1,71 @@
 # Copyright © 2023-2024 Apple Inc.

-from typing import Optional
+import math
+from typing import List, Optional, Union

 import mlx.core as mx
 import mlx.nn as nn


+class SuScaledRoPE(nn.Module):
+    def __init__(
+        self,
+        dims: int,
+        base: float = 10000.0,
+        max_position_embeddings: int = 131072,
+        original_max_position_embeddings: int = 4096,
+        short_factor: Union[List[float], float] = 1.0,
+        long_factor: Union[List[float], float] = 1.0,
+        short_mscale: float = None,
+        long_mscale: float = None,
+    ):
+        """
+        Su Scaled Rotary Embedding layer.
+
+        Args:
+            dims (int): The feature dimensions to be rotated.
+            base (int, optional): Base for the exponential scaling.
+            max_position_embeddings (int, optional): The maximum sequence
+              length that this model was trained with. This is used to determine
+              the size of the original RoPE embeddings when using long scaling.
+              Default: ``131072``.
+            original_max_position_embeddings (int, optional): The maximum
+              sequence length that this model was trained with. This is used to
+              determine the size of the original RoPE embeddings when using long
+              scaling. Default: ``4096``.
+            short_factor (float or list[float], optional): List of scaling
+              factors for sequences of length lesser than
+              ``original_max_position_embeddings``. Default: ``1.0``.
+            long_factor (float or list[float], optional): List of scaling
+              factors for sequences of length greater than
+              ``original_max_position_embeddings``.  Default: ``1.0``.
+            short_mscale (float, optional): Scale the input prior to embedding.
+            long_mscale (float, optional): Scale the input prior to embedding.
+        """
+        super().__init__()
+        freqs = base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        self._freqs = mx.array(long_factor, dtype=mx.float32) * freqs
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.scale = long_mscale or math.sqrt(
+            1
+            + math.log(max_position_embeddings / original_max_position_embeddings)
+            / math.log(original_max_position_embeddings)
+        )
+        self.dim = dims
+
+    def __call__(self, x, offset: int = 0):
+        x[..., : self.dim] = self.scale * x[..., : self.dim]
+        return mx.fast.rope(
+            x,
+            self.dim,
+            traditional=False,
+            base=None,
+            scale=1.0,
+            offset=offset,
+            freqs=self._freqs,
+        )
+
+
 class Llama3RoPE(nn.Module):

    def __init__(
@@ -61,6 +121,78 @@ class Llama3RoPE(nn.Module):
        )


+class YarnRoPE(nn.Module):
+    def __init__(
+        self,
+        dims,
+        traditional=False,
+        max_position_embeddings=2048,
+        base=10000,
+        scaling_factor=1.0,
+        original_max_position_embeddings=4096,
+        beta_fast=32,
+        beta_slow=1,
+        mscale=1,
+        mscale_all_dim=0,
+    ):
+        super().__init__()
+
+        def yarn_find_correction_dim(num_rotations):
+            return (
+                dims
+                * math.log(
+                    original_max_position_embeddings / (num_rotations * 2 * math.pi)
+                )
+            ) / (2 * math.log(base))
+
+        def yarn_find_correction_range():
+            low = math.floor(yarn_find_correction_dim(beta_fast))
+            high = math.ceil(yarn_find_correction_dim(beta_slow))
+            return max(low, 0), min(high, dims - 1)
+
+        def yarn_get_mscale(scale=1, mscale=1):
+            if scale <= 1:
+                return 1.0
+            return 0.1 * mscale * math.log(scale) + 1.0
+
+        def yarn_linear_ramp_mask(min_val, max_val, dim):
+            if min_val == max_val:
+                max_val += 0.001  # Prevent singularity
+
+            linear_func = (mx.arange(dim, dtype=mx.float32) - min_val) / (
+                max_val - min_val
+            )
+            return mx.clip(linear_func, 0, 1)
+
+        self.mscale = yarn_get_mscale(scaling_factor, mscale) / yarn_get_mscale(
+            scaling_factor, mscale_all_dim
+        )
+        freq_extra = base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        freq_inter = scaling_factor * base ** (
+            mx.arange(0, dims, 2, dtype=mx.float32) / dims
+        )
+        low, high = yarn_find_correction_range()
+        freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dims // 2)
+        self._freqs = (freq_inter * freq_extra) / (
+            freq_inter * freq_mask + freq_extra * (1 - freq_mask)
+        )
+        self.dims = dims
+        self.traditional = traditional
+
+    def __call__(self, x, offset=0):
+        if self.mscale != 1.0:
+            x[..., : self.dims] = self.mscale * x[..., : self.dims]
+        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,
@@ -87,5 +219,37 @@ def initialize_rope(
            base=base,
            scaling_config=scaling_config,
        )
+    elif rope_type == "yarn":
+        scaling_factor = scaling_config["factor"]
+        rope_kwargs = {
+            key: scaling_config[key]
+            for key in [
+                "original_max_position_embeddings",
+                "beta_fast",
+                "beta_slow",
+                "mscale",
+                "mscale_all_dim",
+            ]
+            if key in scaling_config
+        }
+        return YarnRoPE(
+            dims=dims,
+            max_position_embeddings=max_position_embeddings,
+            traditional=traditional,
+            base=base,
+            **rope_kwargs,
+        )
+    elif rope_type == "longrope":
+        return SuScaledRoPE(
+            dims=dims,
+            base=base,
+            max_position_embeddings=max_position_embeddings,
+            original_max_position_embeddings=scaling_config[
+                "original_max_position_embeddings"
+            ],
+            short_factor=scaling_config["short_factor"],
+            long_factor=scaling_config["long_factor"],
+        )
+
    else:
        raise ValueError(f"Unsupported RoPE type {rope_type}")
@@ -199,7 +199,10 @@ class Model(nn.Module):
        mask: mx.array = None,
        cache=None,
    ) -> mx.array:
-        mask = create_attention_mask(x, cache)
+
+        if mask is None:
+            mask = create_attention_mask(x, cache)
+
        y = self.model(x, mask, cache)
        return self.lm_head(y)

@@ -125,11 +125,13 @@ class Starcoder2Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
        h = self.embed_tokens(inputs)

-        mask = create_attention_mask(h, cache)
+        if mask is None:
+            mask = create_attention_mask(h, cache)

        if cache is None:
            cache = [None] * len(self.layers)
@@ -152,9 +154,10 @@ class Model(nn.Module):
    def __call__(
        self,
        inputs: mx.array,
+        mask: mx.array = None,
        cache=None,
    ):
-        out = self.model(inputs, cache)
+        out = self.model(inputs, mask, cache)
        if self.args.tie_word_embeddings:
            out = self.model.embed_tokens.as_linear(out)
        else:
@@ -1,64 +0,0 @@
-# Copyright © 2023-2024 Apple Inc.
-
-import math
-from typing import List, Union
-
-import mlx.core as mx
-import mlx.nn as nn
-
-
-class SuScaledRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        dims: int,
-        base: float = 10000.0,
-        max_position_embeddings: int = 131072,
-        original_max_position_embeddings: int = 4096,
-        short_factor: Union[List[float], float] = 1.0,
-        long_factor: Union[List[float], float] = 1.0,
-        short_mscale: float = None,
-        long_mscale: float = None,
-    ):
-        """
-        Phi3Su Scaled Rotary Embedding layer for Phi-3 models.
-
-        Args:
-            dims (int): The feature dimensions to be rotated.
-            base (int, optional): Base for the exponential scaling.
-            max_position_embeddings (int, optional): The maximum sequence
-              length that this model was trained with. This is used to determine
-              the size of the original RoPE embeddings when using long scaling.
-              Default: ``131072``.
-            original_max_position_embeddings (int, optional): The maximum
-              sequence length that this model was trained with. This is used to
-              determine the size of the original RoPE embeddings when using long
-              scaling. Default: ``4096``.
-            short_factor (float or list[float], optional): List of scaling
-              factors for sequences of length lesser than
-              ``original_max_position_embeddings``. Default: ``1.0``.
-            long_factor (float or list[float], optional): List of scaling
-              factors for sequences of length greater than
-              ``original_max_position_embeddings``.  Default: ``1.0``.
-            short_mscale (float, optional): Scale the input prior to embedding.
-            long_mscale (float, optional): Scale the input prior to embedding.
-        """
-        super().__init__()
-        freqs = base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
-        self._freqs = mx.array(long_factor, dtype=mx.float32) * freqs
-        self.original_max_position_embeddings = original_max_position_embeddings
-        self.scale = long_mscale or math.sqrt(
-            1
-            + math.log(max_position_embeddings / original_max_position_embeddings)
-            / math.log(original_max_position_embeddings)
-        )
-
-    def __call__(self, x, offset: int = 0):
-        return mx.fast.rope(
-            self.scale * x,
-            x.shape[-1],
-            traditional=False,
-            base=None,
-            scale=1.0,
-            offset=offset,
-            freqs=self._freqs,
-        )
@@ -6,6 +6,21 @@ import mlx.core as mx
 import mlx.nn as nn


+def _gather_sort(x, indices):
+    *_, M = indices.shape
+    indices = indices.flatten()
+    order = mx.argsort(indices)
+    inv_order = mx.argsort(order)
+    return x.flatten(0, -3)[order // M], indices[order], inv_order
+
+
+def _scatter_unsort(x, inv_order, shape=None):
+    x = x[inv_order]
+    if shape is not None:
+        x = mx.unflatten(x, 0, shape)
+    return x
+
+
 class QuantizedSwitchLinear(nn.Module):
    def __init__(
        self,
@@ -38,12 +53,6 @@ class QuantizedSwitchLinear(nn.Module):
        # Freeze this model's parameters
        self.freeze()

-    def unfreeze(self, *args, **kwargs):
-        """Wrap unfreeze so that we unfreeze any layers we might contain but
-        our parameters will remain frozen."""
-        super().unfreeze(*args, **kwargs)
-        self.freeze(recurse=False)
-
    @property
    def input_dims(self):
        return self.scales.shape[2] * self.group_size
@@ -56,7 +65,7 @@ class QuantizedSwitchLinear(nn.Module):
    def num_experts(self):
        return self.weight.shape[0]

-    def __call__(self, x, indices):
+    def __call__(self, x, indices, sorted_indices=False):
        x = mx.gather_qmm(
            x,
            self["weight"],
@@ -66,6 +75,7 @@ class QuantizedSwitchLinear(nn.Module):
            transpose=True,
            group_size=self.group_size,
            bits=self.bits,
+            sorted_indices=sorted_indices,
        )
        if "bias" in self:
            x = x + mx.expand_dims(self["bias"][indices], -2)
@@ -99,8 +109,13 @@ class SwitchLinear(nn.Module):
    def num_experts(self):
        return self.weight.shape[0]

-    def __call__(self, x, indices):
-        x = mx.gather_mm(x, self["weight"].swapaxes(-1, -2), rhs_indices=indices)
+    def __call__(self, x, indices, sorted_indices=False):
+        x = mx.gather_mm(
+            x,
+            self["weight"].swapaxes(-1, -2),
+            rhs_indices=indices,
+            sorted_indices=sorted_indices,
+        )
        if "bias" in self:
            x = x + mx.expand_dims(self["bias"][indices], -2)
        return x
@@ -122,7 +137,7 @@ class SwitchGLU(nn.Module):
        input_dims: int,
        hidden_dims: int,
        num_experts: int,
-        activation=nn.silu,
+        activation=nn.SiLU(),
        bias: bool = False,
    ):
        super().__init__()
@@ -135,9 +150,24 @@ class SwitchGLU(nn.Module):
    def __call__(self, x, indices) -> mx.array:
        x = mx.expand_dims(x, (-2, -3))

-        x_up = self.up_proj(x, indices)
-        x_gate = self.gate_proj(x, indices)
-        x = self.down_proj(self.activation(x_gate) * x_up, indices)
+        # When we have many tokens, then sort them to make sure that the access
+        # of different experts is in order.
+        do_sort = indices.size >= 64
+        idx = indices
+        inv_order = None
+        if do_sort:
+            x, idx, inv_order = _gather_sort(x, indices)
+
+        x_up = self.up_proj(x, idx, sorted_indices=do_sort)
+        x_gate = self.gate_proj(x, idx, sorted_indices=do_sort)
+        x = self.down_proj(
+            self.activation(x_gate) * x_up,
+            idx,
+            sorted_indices=do_sort,
+        )
+
+        if do_sort:
+            x = _scatter_unsort(x, inv_order, indices.shape)

        return x.squeeze(-2)

@@ -148,7 +178,7 @@ class SwitchMLP(nn.Module):
        input_dims: int,
        hidden_dims: int,
        num_experts: int,
-        activation=nn.gelu_approx,
+        activation=nn.GELU(approx="precise"),
        bias: bool = False,
    ):
        super().__init__()
@@ -160,8 +190,19 @@ class SwitchMLP(nn.Module):
    def __call__(self, x, indices) -> mx.array:
        x = mx.expand_dims(x, (-2, -3))

-        x = self.fc1(x, indices)
+        # When we have many tokens, then sort them to make sure that the access
+        # of different experts is in order.
+        do_sort = indices.size >= 64
+        idx = indices
+        inv_order = None
+        if do_sort:
+            x, idx, inv_order = _gather_sort(x, indices)
+
+        x = self.fc1(x, idx, sorted_indices=do_sort)
        x = self.activation(x)
-        x = self.fc2(x, indices)
+        x = self.fc2(x, idx, sorted_indices=do_sort)
+
+        if do_sort:
+            x = _scatter_unsort(x, inv_order, indices.shape)

        return x.squeeze(-2)
@@ -0,0 +1,588 @@
+# Copyright © 2025 Apple Inc.
+
+import argparse
+import copy
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Callable, Dict
+from urllib import request
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_map, tree_map_with_path
+from tqdm import tqdm
+
+from mlx_lm.models.base import create_attention_mask
+from mlx_lm.models.switch_layers import SwitchLinear
+from mlx_lm.quant.utils import load_data
+from mlx_lm.utils import (
+    fetch_from_hub,
+    get_model_path,
+    save,
+)
+
+
+@dataclass
+class ScaleConfig:
+    prev: nn.Module
+    layers: list[nn.Module]
+    block: nn.Module | None = None
+    kwargs: list = field(default_factory=list)
+    use_config: Callable[[nn.Module], bool] | None = None
+
+
+@dataclass
+class AWQConfig:
+    embed: str
+    lm_head: str
+    no_clip: list[str]
+    scale_configs: list[ScaleConfig]
+    lm_key: str | None = None
+
+
+def update(cfg, **kwargs):
+    cfg = copy.deepcopy(cfg)
+    for k, v in kwargs.items():
+        setattr(cfg, k, v)
+    return cfg
+
+
+llama_awq = AWQConfig(
+    embed="embed_tokens",
+    lm_head="lm_head",
+    no_clip=["q_proj", "k_proj"],
+    scale_configs=[
+        ScaleConfig(
+            block="self_attn",
+            prev="input_layernorm",
+            layers=["q_proj", "k_proj", "v_proj"],
+            kwargs=["mask"],
+        ),
+        ScaleConfig(prev="mlp.up_proj", layers=["mlp.down_proj"]),
+        ScaleConfig(
+            block="mlp",
+            prev="post_attention_layernorm",
+            layers=["gate_proj", "up_proj"],
+        ),
+    ],
+)
+
+gemma3_text_awq = AWQConfig(
+    embed="embed_tokens",
+    lm_head="lm_head",
+    no_clip=["q_proj", "k_proj"],
+    scale_configs=[
+        ScaleConfig(
+            block="self_attn",
+            prev="input_layernorm",
+            layers=["q_proj", "k_proj", "v_proj"],
+            kwargs=["mask"],
+        ),
+        ScaleConfig(prev="mlp.up_proj", layers=["mlp.down_proj"]),
+        ScaleConfig(
+            block="mlp",
+            prev="pre_feedforward_layernorm",
+            layers=["gate_proj", "up_proj"],
+        ),
+    ],
+)
+
+gemma3_awq = update(gemma3_text_awq, lm_key="language_model")
+
+deepseek_v2_awq = AWQConfig(
+    embed="embed_tokens",
+    lm_head="lm_head",
+    no_clip=["q_proj", "q_a_proj", "q_b_proj", "kv_a_proj_with_mqa", "kv_b_proj"],
+    scale_configs=[
+        ScaleConfig(
+            block="self_attn",
+            prev="input_layernorm",
+            layers=["q_proj", "kv_a_proj_with_mqa"],
+            kwargs=["mask"],
+        ),
+        ScaleConfig(
+            prev="self_attn.kv_a_layernorm",
+            layers=["self_attn.kv_b_proj"],
+        ),
+        ScaleConfig(
+            prev="mlp.up_proj",
+            layers=["mlp.down_proj"],
+            use_config=lambda block: not "switch_mlp" in block.mlp,
+        ),
+        ScaleConfig(
+            prev="mlp.shared_experts.up_proj",
+            layers=["mlp.shared_experts.down_proj"],
+            use_config=lambda block: "switch_mlp" in block.mlp,
+        ),
+        ScaleConfig(
+            prev="mlp.switch_mlp.up_proj",
+            layers=["mlp.switch_mlp.down_proj"],
+            use_config=lambda block: "switch_mlp" in block.mlp,
+            kwargs=["indices"],
+        ),
+        ScaleConfig(
+            block="mlp",
+            prev="post_attention_layernorm",
+            layers=["gate_proj", "up_proj"],
+            use_config=lambda block: not "switch_mlp" in block.mlp,
+        ),
+        ScaleConfig(
+            block="mlp",
+            prev="post_attention_layernorm",
+            layers=[
+                "switch_mlp.gate_proj",
+                "switch_mlp.up_proj",
+                "shared_experts.gate_proj",
+                "shared_experts.up_proj",
+                "gate",  # not quantized, just scaled
+            ],
+            use_config=lambda block: "switch_mlp" in block.mlp,
+        ),
+    ],
+)
+
+AWQ_MODEL_CONFIGS = {
+    "llama": llama_awq,
+    "mistral": llama_awq,
+    "qwen2": llama_awq,
+    "qwen3": llama_awq,
+    "gemma3_text": gemma3_text_awq,
+    "gemma3": update(gemma3_text_awq, lm_key="language_model"),
+    "deepseek_v2": deepseek_v2_awq,
+}
+
+
+def mse(x, y):
+    return ((x - y).astype(mx.float32)) ** 2
+
+
+def submodule_from_key(module, key):
+    keys = key.split(".")
+    for k in keys:
+        module = module[k]
+    return module
+
+
+def run_layer(
+    layer: nn.Module,
+    x: mx.array,
+    indices: mx.array | None = None,
+    batch_size: int = 32,
+    **kwargs,
+):
+    y = []
+    for i in range(0, x.shape[0], batch_size):
+        if indices is not None:
+            y.append(
+                layer(x[i : i + batch_size], indices[i : i + batch_size], **kwargs)
+            )
+        else:
+            y.append(layer(x[i : i + batch_size], **kwargs))
+        mx.eval(y)
+    y = mx.concatenate(y, axis=0)
+    return y
+
+
+def dist_split(x: mx.array, group: mx.distributed.Group):
+    N = group.size()
+    if N == 1:
+        return x
+    B = x.shape[0]
+    assert B % N == 0
+    r = group.rank()
+    local_B = (B + N - 1) // N
+    return x[r * local_B : (r + 1) * local_B]
+
+
+def search_best_scale(
+    layers: list[nn.Module],
+    quantize_func: Callable,
+    block: nn.Module | None,
+    layer_kwargs: dict,
+    n_grid: int,
+):
+    group = mx.distributed.init()
+
+    layer_kwargs = layer_kwargs or {}
+
+    x = layers[0].input_feat
+
+    block = block or layers[0]
+    out = block(x, **layer_kwargs)
+
+    x_max = x.abs().mean(axis=(0, 1))
+
+    best_error = float("inf")
+    best_scales = None
+
+    weights = tree_flatten(block.parameters())
+
+    # Search across different scaling ratios
+    # and take the best loss.
+    for ratio in range(n_grid):
+        ratio = ratio / n_grid
+        scales = mx.maximum(x_max**ratio, 1e-4).reshape(-1)
+        scales = scales / (scales.max() * scales.min()).sqrt()
+        for layer in layers:
+            if isinstance(layer, (nn.Linear, SwitchLinear)):
+                layer.weight = quantize_func(layer.weight * scales) / scales
+
+        out_q = run_layer(block, x, **layer_kwargs)
+        loss = mse(out, out_q).sum()
+        if group is not None:
+            loss = mx.distributed.all_sum(loss) / group.size()
+        loss /= out.size
+        mx.eval(loss)
+        if loss.item() < best_error:
+            best_error = loss.item()
+            best_scales = scales
+
+        # reload the original weights
+        block.load_weights(weights)
+
+    best_scales = best_scales.reshape(-1)
+    mx.eval(best_scales)
+    return best_scales
+
+
+def apply_scale(prev_op, layers, scales):
+    # Fuse the scales into the previous op
+    if isinstance(prev_op, (nn.Linear, SwitchLinear)):
+        assert len(layers) == 1
+        prev_op.weight = prev_op.weight / scales[:, mx.newaxis]
+        if hasattr(prev_op, "bias"):
+            prev_op.bias = prev_op.bias / scales
+        layers[0].weight = layers[0].weight * scales
+    elif isinstance(prev_op, (nn.LayerNorm, nn.RMSNorm)):
+        prev_op.weight = prev_op.weight / scales
+        if hasattr(prev_op, "bias"):
+            prev_op.bias = prev_op.bias / scales
+        for layer in layers:
+            layer.weight = layer.weight * scales
+    elif prev_op.__class__.__name__ == "RMSNorm":  # For gemma models
+        dt = prev_op.weight.dtype
+        prev_op.weight = (
+            (1.0 + prev_op.weight.astype(mx.float32)) / scales - 1.0
+        ).astype(dt)
+        for layer in layers:
+            layer.weight = layer.weight * scales
+    else:
+        raise NotImplementedError(f"Could not apply scale to prev_op: {prev_op}")
+
+    for layer in layers:
+        if hasattr(layer, "input_feat"):
+            layer.input_feat = layer.input_feat / scales
+
+
+def scale_block(
+    block: nn.Module,
+    configs: list[ScaleConfig],
+    quantize_func: Callable,
+    layer_kwargs: dict,
+    n_grid: int,
+):
+    for conf in configs:
+        if conf.use_config is not None and not conf.use_config(block):
+            continue
+        if conf.block is not None:
+            local_block = block[conf.block]
+            layers = [submodule_from_key(local_block, l) for l in conf.layers]
+        else:
+            local_block = None
+            layers = [submodule_from_key(block, l) for l in conf.layers]
+        local_kwargs = {k: layer_kwargs[k] for k in conf.kwargs if k in layer_kwargs}
+        for k in conf.kwargs:
+            if hasattr(layers[0], k):
+                local_kwargs[k] = getattr(layers[0], k)
+
+        scales = search_best_scale(
+            layers=layers,
+            block=local_block,
+            layer_kwargs=local_kwargs,
+            quantize_func=quantize_func,
+            n_grid=n_grid,
+        )
+        apply_scale(submodule_from_key(block, conf.prev), layers, scales)
+
+
+def search_best_clip(
+    module: nn.Module,
+    quantize_func: Callable,
+    group_size: int,
+    n_grid: int,
+    max_shrink: float = 0.5,
+    batch_size: int = 64,
+    n_frames: int = 512,
+):
+    group = mx.distributed.init()
+
+    # subsample the input features
+    x = module.input_feat.flatten(0, 1)
+    stride = (x.shape[0] + n_frames - 1) // n_frames
+    x = x[::stride]
+
+    w = module.weight
+    x = x.reshape(x.shape[0], -1, group_size)
+
+    w_init_shape = w.shape
+    w_all = mx.flatten(w, 0, w.ndim - 2)
+    w_max_all = []
+
+    # batch across W to save memory
+    for b in range(0, w_all.shape[0], batch_size):
+        w = w_all[b : b + batch_size]
+
+        group_shape = (w.shape[0], w.shape[-1] // group_size)
+        best_error = mx.full(group_shape, float("inf"))
+        best_w_max = mx.zeros((*group_shape, 1), dtype=x.dtype)
+
+        w_shape = w.shape
+
+        w = w.reshape(*w.shape[:-1], -1, group_size)
+        out = mx.einsum("bdg,odg->bod", x, w)
+        init_max = w.abs().max(axis=-1, keepdims=True)
+
+        # try a range of clips and pick the one with the smallest loss
+        for i in range(int(max_shrink * n_grid)):
+            p = 1 - i / n_grid
+            w_max = p * init_max
+            w_m = mx.clip(w, -w_max, w_max).reshape(w_shape)
+
+            w_q = quantize_func(w_m)
+
+            w_q = w_q.reshape(*w_q.shape[:-1], -1, group_size)
+            out_q = mx.einsum("bdg,odg->bod", x, w_q)
+
+            # Take the mean across the input batch
+            loss = mse(out, out_q).sum(axis=0)
+            if group is not None:
+                loss = mx.distributed.all_sum(loss) / group.size()
+            loss /= out.shape[0]
+            best_indices = loss < best_error
+            best_error = mx.where(best_indices, loss, best_error)
+            best_w_max = mx.where(best_indices[..., mx.newaxis], w_max, best_w_max)
+            mx.eval(best_w_max, best_error)
+
+        w_max_all.append(best_w_max)
+
+    best_w_max = mx.concatenate(w_max_all, axis=0)
+
+    w_r = w_all.reshape(*w_all.shape[:-1], -1, group_size)
+    best_w = mx.clip(w_r, -best_w_max, best_w_max)
+    best_w = best_w.reshape(w_init_shape)
+
+    mx.eval(best_w)
+    return best_w
+
+
+def clip_block(
+    block: nn.Module,
+    no_clip_keys: list[str],
+    quantize_func: Callable,
+    group_size: int,
+    n_grid: int = 20,
+):
+    def apply_clip(path, module):
+        if isinstance(module, (nn.Linear, SwitchLinear)) and all(
+            k not in path for k in no_clip_keys
+        ):
+            best_weight = search_best_clip(
+                module,
+                quantize_func=quantize_func,
+                group_size=group_size,
+                n_grid=n_grid,
+            )
+            module.weight = best_weight
+
+    tree_map_with_path(apply_clip, block.leaf_modules(), is_leaf=nn.Module.is_module)
+
+
+def awq_quantize(
+    model,
+    inputs: mx.array,
+    awq_config: AWQConfig,
+    group_size: int = 64,
+    bits: int = 3,
+    embed_group_size: int = 32,
+    embed_bits: int = 4,
+    n_grid: int = 20,
+):
+    if awq_config.lm_key is not None:
+        model = model[awq_config.lm_key]
+
+    group = mx.distributed.init()
+
+    def quantize_func(w):
+        wq = mx.quantize(w, bits=bits, group_size=group_size)
+        return mx.dequantize(*wq, bits=bits, group_size=group_size)
+
+    mask = create_attention_mask(inputs)
+
+    embed_key = awq_config.embed
+    model.model[embed_key] = model.model[embed_key].to_quantized(
+        group_size=embed_group_size, bits=embed_bits
+    )
+    inputs = model.model[embed_key](inputs)
+
+    def capture(module):
+        if not isinstance(module, (nn.Linear, SwitchLinear)):
+            return module
+
+        class Catcher(nn.Module):
+            def __call__(self, x: mx.array, *args, **kwargs):
+                # Store the input features on the original modules.
+                if hasattr(module, "input_feat"):
+                    module.input_feat = mx.concatenate([module.input_feat, x], axis=0)
+                else:
+                    module.input_feat = x
+
+                # Also store the MOE indices if applicabale
+                if isinstance(module, SwitchLinear):
+                    indices = args[0]
+                    if hasattr(module, "indices"):
+                        module.indices = mx.concatenate(
+                            [module.indices, indices], axis=0
+                        )
+                    else:
+                        module.indices = indices
+
+                return module(x, *args, **kwargs)
+
+        return Catcher()
+
+    for e, block in enumerate(tqdm(model.layers)):
+        # Capture the input features for each of the layers in the transformer block
+        orig_leaves = block.leaf_modules()
+        capture_leaves = tree_map(capture, orig_leaves, is_leaf=nn.Module.is_module)
+        block.update_modules(capture_leaves)
+        outputs = run_layer(block, inputs, mask=mask)
+        block.update_modules(orig_leaves)
+        del capture_leaves
+
+        # Quantize the block without AWQ to obtain a reference loss
+        nn.quantize(block, group_size=group_size, bits=bits)
+        outputs_q = run_layer(block, inputs, mask=mask)
+        before_loss = mse(outputs, outputs_q).sum()
+        if group is not None:
+            before_loss = mx.distributed.all_sum(before_loss) / group.size()
+        before_loss /= outputs.size
+        block.update_modules(orig_leaves)
+        orig_params = block.parameters()
+
+        scale_block(
+            block=block,
+            configs=awq_config.scale_configs,
+            quantize_func=quantize_func,
+            n_grid=n_grid,
+            layer_kwargs={"mask": mask},
+        )
+
+        clip_block(
+            block=block,
+            no_clip_keys=awq_config.no_clip,
+            quantize_func=quantize_func,
+            group_size=group_size,
+            n_grid=n_grid,
+        )
+
+        # Quantize the scaled and clipped block
+        nn.quantize(block, group_size=group_size, bits=bits)
+        outputs_q = run_layer(block, inputs, mask=mask)
+        after_loss = mse(outputs, outputs_q).sum()
+        if group is not None:
+            after_loss = mx.distributed.all_sum(after_loss) / group.size()
+        after_loss /= outputs.size
+        tqdm.write(f"Loss reduction: {after_loss / before_loss}")
+        if after_loss > before_loss:
+            # Reload original weights and quantize
+            block.update_modules(orig_leaves)
+            block.update(orig_params)
+            nn.quantize(block, group_size=group_size, bits=bits)
+            tqdm.write("Loss is not reduced, falling back to original weights.")
+
+        inputs = outputs
+
+        mx.eval(block)
+        mx.clear_cache()
+
+    if (lm_head := awq_config.lm_head) in model:
+        model[lm_head] = model[lm_head].to_quantized(
+            group_size=embed_group_size, bits=embed_bits
+        )
+
+
+def update_config(
+    model: nn.Module,
+    config: Dict[str, Any],
+):
+    # dummy
+    config["quantization"] = {"group_size": 64, "bits": 4}
+
+    def update_config(path, module):
+        if hasattr(module, "bits"):
+            config["quantization"][path] = {
+                "group_size": module.group_size,
+                "bits": module.bits,
+            }
+        else:
+            config["quantization"][path] = False
+
+    tree_map_with_path(update_config, model.leaf_modules(), is_leaf=nn.Module.is_module)
+    return config
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model", "-m", default="mlx-community/Qwen2.5-7B-Instruct-bf16"
+    )
+    parser.add_argument("--mlx-path", default="mlx_model")
+    parser.add_argument("--bits", type=int, default=4)
+    parser.add_argument("--group-size", type=int, default=64)
+    parser.add_argument("--embed-bits", type=int, default=4)
+    parser.add_argument("--embed-group-size", type=int, default=32)
+    parser.add_argument("--num-samples", type=int, default=128)
+    parser.add_argument("--sequence-length", type=int, default=512)
+    parser.add_argument("--n-grid", type=int, default=20)
+    parser.add_argument("--seed", type=int, default=123)
+    args = parser.parse_args()
+
+    group = mx.distributed.init()
+
+    num_samples = args.num_samples
+    if group is not None and num_samples % group.size() > 0:
+        num_samples += group.size() - num_samples % group.size()
+
+    mx.random.seed(args.seed)
+
+    model_path, hf_repo = get_model_path(args.model, revision=None)
+    model, config, tokenizer = fetch_from_hub(model_path, lazy=True)
+
+    model_type = config["model_type"]
+    if (awq_config := AWQ_MODEL_CONFIGS.get(model_type, None)) is None:
+        raise NotImplementedError(f"AWQ support for {model_type} models NYI.")
+
+    calibration_data = load_data(tokenizer, args.num_samples, args.sequence_length)
+
+    calibration_data = dist_split(calibration_data, group)
+
+    awq_quantize(
+        model,
+        calibration_data,
+        awq_config,
+        bits=args.bits,
+        group_size=args.group_size,
+        embed_bits=args.embed_bits,
+        embed_group_size=args.embed_group_size,
+        n_grid=args.n_grid,
+    )
+
+    config = update_config(model, config)
+    save(
+        args.mlx_path,
+        model_path,
+        model,
+        tokenizer,
+        config,
+        hf_repo=hf_repo,
+    )
@@ -0,0 +1,251 @@
+# Copyright © 2025 Apple Inc.
+
+import argparse
+import copy
+import time
+import types
+
+import mlx.core as mx
+import mlx.nn as nn
+import mlx.optimizers as optimizers
+import numpy as np
+from mlx.utils import tree_flatten, tree_map
+from tqdm import tqdm
+
+from mlx_lm.tuner.datasets import load_dataset
+from mlx_lm.tuner.losses import kl_div_loss
+from mlx_lm.tuner.trainer import grad_checkpoint, iterate_batches
+from mlx_lm.tuner.utils import print_trainable_parameters
+from mlx_lm.utils import (
+    fetch_from_hub,
+    get_model_path,
+    quantize_model,
+    save,
+)
+
+
+class Catcher(nn.Module):
+    def __init__(self, module):
+        super().__init__()
+        self.module = module
+
+    def __call__(self, *args, **kwargs):
+        self.outputs = self.module(*args, **kwargs)
+        return self.outputs
+
+
+def dwq_quantize(
+    model,
+    q_model,
+    opt,
+    data,
+    batch_size: int = 2,
+    max_seq_length: int = 2048,
+    activation_layer_step: float = 0.25,
+    activation_loss_weight: float = 1.0,
+    dtype: mx.Dtype = mx.bfloat16,
+    gradient_checkpoint: bool = False,
+):
+    group = mx.distributed.init()
+    world_size = group.size()
+    rank = group.rank()
+
+    def unfreeze(_, m):
+        if hasattr(m, "bits") and hasattr(m, "group_size"):
+            m.unfreeze(keys=["scales", "biases"], recurse=False)
+
+    q_model.apply_to_modules(unfreeze)
+    print_trainable_parameters(q_model)
+
+    layer_id_step = max(int(activation_layer_step * len(model.layers)), 1)
+    layer_ids = list(range(len(model.layers)))[layer_id_step::layer_id_step]
+
+    for lid in layer_ids:
+        model.layers[lid] = Catcher(model.layers[lid])
+        q_model.layers[lid] = Catcher(q_model.layers[lid])
+
+    if gradient_checkpoint:
+        grad_checkpoint(q_model.layers[0])
+
+    def forward(model, inputs):
+        logits = model(inputs)
+        extra_targets = [
+            model.layers[lid].outputs.astype(mx.float32) for lid in layer_ids
+        ]
+        for lid in layer_ids:
+            model.layers[lid].outputs = None
+        return logits, extra_targets
+
+    def loss_fn(params, x, targets, extra_targets, lengths):
+        q_model.update(tree_map(lambda x: x.astype(dtype), params))
+        logits, q_extra_targets = forward(q_model, x)
+        losses = kl_div_loss(logits, targets)
+        mask = mx.arange(1, 1 + targets.shape[1]) < lengths[:, 1:]
+        ntoks = mask.sum()
+        kl_loss = (mask * losses).sum() / ntoks
+        act_loss = mx.stack(
+            [
+                (mask * (qe - e).abs().mean(axis=-1)).sum() / ntoks
+                for qe, e in zip(q_extra_targets, extra_targets)
+            ]
+        )
+        loss = kl_loss + activation_loss_weight * act_loss.mean()
+        return loss, ntoks
+
+    def step(inputs, targets, extra_targets, lengths, params):
+        (loss, ntoks), grads = mx.value_and_grad(loss_fn)(
+            params, inputs, targets, extra_targets, lengths
+        )
+        grads = nn.average_gradients(grads)
+        params = opt.apply_gradients(grads, params)
+        return loss, ntoks, params
+
+    # Accumulate learned weights in higher precision
+    params = tree_map(
+        lambda x: x.astype(mx.float32),
+        q_model.trainable_parameters(),
+    )
+
+    total_loss = 0.0
+    total_tokens = 0
+    tokens = 0
+    tic = time.time()
+    for it, (batch, lengths) in (
+        pbar := tqdm(
+            enumerate(iterate_batches(data, batch_size, max_seq_length)),
+            total=len(data) // batch_size,
+        )
+    ):
+        batch = batch[:, :-1]
+        targets, extra_targets = forward(model, batch)
+        mx.eval(targets, extra_targets)
+        loss, ntoks, params = step(batch, targets, extra_targets, lengths, params)
+        mx.eval(loss, params)
+        loss = mx.distributed.all_sum(loss, stream=mx.cpu).item() / world_size
+        ntoks = mx.distributed.all_sum(ntoks, stream=mx.cpu).item()
+        tokens += ntoks
+        total_loss += loss * ntoks
+        if rank == 0:
+            pbar.set_description(desc=f"{loss=:.4f}")
+            if (it + 1) % 20 == 0:
+                toks_per_sec = tokens / (time.time() - tic)
+                peak_memory_gb = mx.get_peak_memory() / 1e9
+                avg_loss = total_loss / tokens
+                total_tokens += tokens
+                tqdm.write(
+                    f"{it=}, {avg_loss=:.4f}, {total_tokens=},"
+                    f" {toks_per_sec=:.3f}, {peak_memory_gb=:.3f}",
+                )
+                tic = time.time()
+                tokens = 0
+                total_loss = 0
+    q_model.update(tree_map(lambda x: x.astype(dtype), params))
+    for lid in layer_ids:
+        q_model.layers[lid] = q_model.layers[lid].module
+
+
+def load_data(tokenizer, data_path: str, num_samples: int, max_seq_length: int):
+    args = types.SimpleNamespace(
+        hf_dataset={
+            "path": data_path,
+            "train_split": f"train",
+            "valid_split": "train[:1]",
+        },
+        train=True,
+        test=False,
+    )
+    dataset = load_dataset(args, tokenizer)[0]
+    perm = np.random.permutation(len(dataset))[:num_samples].tolist()
+
+    def process(idx):
+        tokens, offset = dataset.process(dataset[idx])
+        return (tokens[:max_seq_length], offset)
+
+    return [process(i) for i in perm]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", "-m", default="Qwen/Qwen3-4B")
+    parser.add_argument("--quantized-model", default=None)
+    parser.add_argument(
+        "--mlx-path", default="mlx_model", help="Path to save the quantized model."
+    )
+    parser.add_argument(
+        "--bits",
+        type=int,
+        default=4,
+        help="Bits per weight for quantization.",
+    )
+    parser.add_argument(
+        "--group-size", type=int, default=64, help="Group size for quantization."
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=2048,
+        help="Number of samples to use for training.",
+    )
+    parser.add_argument("--max-seq-length", type=int, default=2049)
+    parser.add_argument("--seed", type=int, default=123)
+    parser.add_argument("--learning-rate", type=float, default=1e-6)
+    parser.add_argument("--batch-size", type=int, default=4)
+    parser.add_argument(
+        "--data-path",
+        type=str,
+        default="allenai/tulu-3-sft-mixture",
+        help="A Hugging Face dataset which is compatible with an mlx-lm dataset format.",
+    )
+    parser.add_argument(
+        "--grad-checkpoint",
+        action="store_true",
+        help="Use gradient checkpointing to reduce memory use.",
+    )
+    args = parser.parse_args()
+
+    group = mx.distributed.init()
+
+    num_samples = args.num_samples
+    if num_samples % group.size() > 0:
+        num_samples += group.size() - num_samples % group.size()
+
+    np.random.seed(args.seed)
+    mx.random.seed(args.seed)
+
+    model_path, hf_repo = get_model_path(args.model, revision=None)
+    model, config, tokenizer = fetch_from_hub(model_path, lazy=True)
+
+    calibration_data = load_data(
+        tokenizer, args.data_path, args.num_samples, args.max_seq_length
+    )
+
+    if args.quantized_model is not None:
+        q_model_path = get_model_path(args.quantized_model, revision=None)
+        q_model, config, _ = fetch_from_hub(q_model_path, lazy=True)
+    else:
+        q_model = copy.deepcopy(model)
+        _, config = quantize_model(
+            q_model,
+            config,
+            q_group_size=args.group_size,
+            q_bits=args.bits,
+        )
+
+    opt = optimizers.Adam(learning_rate=args.learning_rate, bias_correction=True)
+    dwq_quantize(
+        model,
+        q_model,
+        opt,
+        calibration_data,
+        batch_size=args.batch_size,
+        max_seq_length=args.max_seq_length,
+        gradient_checkpoint=args.grad_checkpoint,
+    )
+    save(
+        args.mlx_path,
+        model_path,
+        q_model,
+        tokenizer,
+        config,
+        hf_repo=hf_repo,
+    )
@@ -0,0 +1,349 @@
+# Copyright © 2025 Apple Inc.
+
+import argparse
+import copy
+import json
+import math
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx.utils import tree_flatten, tree_map, tree_reduce, tree_unflatten
+from tqdm import tqdm
+
+from mlx_lm.quant.utils import load_data
+from mlx_lm.tuner.losses import kl_div_loss
+from mlx_lm.tuner.trainer import grad_checkpoint
+from mlx_lm.tuner.utils import get_total_parameters
+from mlx_lm.utils import (
+    compute_bits_per_weight,
+    fetch_from_hub,
+    get_model_path,
+    load,
+    quantize_model,
+    save,
+)
+
+
+def make_quant_predicate(config):
+    def quant_predicate(p, m, _):
+        if not hasattr(m, "to_quantized"):
+            return False
+        return config.get(p, True)
+
+    return quant_predicate
+
+
+def eval_ppl(model, data, batch_size=8):
+    all_loss = 0.0
+    ntoks = 0
+    for s in range(0, len(data), batch_size):
+        batch = data[s : s + batch_size]
+        logits = model(batch[:, :-1]).astype(mx.float32)
+        losses = nn.losses.cross_entropy(logits, batch[:, 1:])
+        all_loss += losses.sum().item()
+        ntoks += losses.size
+    ppl = math.exp(all_loss / ntoks)
+    return ppl
+
+
+def make_options(
+    low_bits, low_group_size, high_bits, high_group_size, include_bpw=True
+):
+    options = []
+    min_bpw = low_bits + 32 / low_group_size
+    max_bpw = high_bits + 32 / high_group_size
+    for b in range(low_bits, high_bits + 1):
+        for g in [32, 64, 128]:
+            cbpw = b + 32 / g
+            if b == 7 or not (min_bpw <= cbpw <= max_bpw):
+                continue
+            options.append({"bits": b, "group_size": g, "bpw": cbpw})
+    options.sort(key=lambda x: x["bpw"])
+    if not include_bpw:
+        for o in options:
+            o.pop("bpw")
+
+    return options
+
+
+def estimate_sensitivities(
+    model,
+    data,
+    low_bits,
+    low_group_size,
+    high_bits,
+    high_group_size,
+    batch_size: int = 4,
+    gradient_accum_dtype: mx.Dtype = mx.float32,
+    gradient_checkpoint: bool = False,
+):
+    def qdq(w, bits, group_size):
+        w, s, b = mx.quantize(w, bits=bits, group_size=group_size)
+        return mx.dequantize(w, scales=s, biases=b, bits=bits, group_size=group_size)
+
+    layers = tree_flatten(model.leaf_modules(), is_leaf=nn.Module.is_module)
+    layers = {k: l for k, l in layers if hasattr(l, "to_quantized")}
+    q_model = copy.deepcopy(model)
+    q_layers = copy.deepcopy(layers)
+    for l in q_layers.values():
+        l.weight = qdq(l.weight, low_bits, low_group_size)
+        # Freeze everything but the quantizable weight
+        l.freeze()
+        l.unfreeze(keys=["weight"])
+    q_model.freeze()
+    q_model.update_modules(tree_unflatten(list(q_layers.items())))
+
+    def loss_fn(batch, targets):
+        return kl_div_loss(q_model(batch), targets).mean()
+
+    if gradient_checkpoint:
+        grad_checkpoint(q_model.layers[0])
+
+    grad_accum = tree_map(
+        lambda x: mx.zeros(x.shape, dtype=gradient_accum_dtype),
+        q_model.trainable_parameters(),
+    )
+    for e, s in tqdm(
+        enumerate(range(0, len(data), batch_size)),
+        total=len(data) // batch_size,
+        desc="Estimating sensitivities",
+    ):
+        batch = data[s : s + batch_size]
+        targets = model(batch)
+        mx.eval(targets)
+        _, grads = nn.value_and_grad(q_model, loss_fn)(batch, targets)
+        grad_accum = tree_map(lambda x, y: x + y, grad_accum, grads)
+        del grads
+        mx.eval(grad_accum)
+
+    options = make_options(low_bits, low_group_size, high_bits, high_group_size)
+    current_bpw = options[0]["bpw"]
+
+    def compute_sensitivity(gradient, low_q_weight, original_weight):
+        n_batches = (len(data) + batch_size - 1) // batch_size
+        gradient = gradient / n_batches
+        scores = [{"loss_change": 0, "extra_bits": 0}]
+        for opt in options[1:]:
+            extra_bits = (opt["bpw"] - current_bpw) * original_weight.size
+            other_weight = qdq(original_weight, opt["bits"], opt["group_size"])
+            loss_change = (gradient * (low_q_weight - other_weight)).sum()
+            scores.append({"loss_change": loss_change, "extra_bits": extra_bits})
+        return scores
+
+    sensitivities = tree_map(
+        compute_sensitivity,
+        grad_accum,
+        q_model.parameters(),
+        model.parameters(),
+    )
+    mx.eval(sensitivities)
+
+    sensitivities = [
+        (k.replace(".weight", ""), s.item() if isinstance(s, mx.array) else s)
+        for k, s in tree_flatten(sensitivities)
+    ]
+
+    return sensitivities
+
+
+def compute_bit_budget(model, target_bpw):
+    model_bytes = tree_reduce(
+        lambda acc, x: acc + x.nbytes if isinstance(x, mx.array) else acc, model, 0
+    )
+    model_params = get_total_parameters(model)
+
+    return model_params * target_bpw - model_bytes * 8
+
+
+def estimate_threshold(
+    model,
+    sensitivities,
+    target_bpw,
+    low_bits,
+    low_group_size,
+    high_bits,
+    high_group_size,
+):
+    options = make_options(
+        low_bits, low_group_size, high_bits, high_group_size, include_bpw=False
+    )
+    sensitivities = tree_flatten(
+        tree_unflatten(list(sensitivities.items())),
+        is_leaf=lambda x: isinstance(x, list) and "loss_change" in x[0],
+    )
+
+    q_model = copy.deepcopy(model)
+    nn.quantize(q_model, group_size=low_group_size, bits=low_bits)
+    budget = int(compute_bit_budget(q_model, target_bpw))
+    benefit_map = {}
+
+    def benefit(layer, option, budget):
+        if (layer, option, budget) in benefit_map:
+            return benefit_map[layer, option, budget]
+
+        stack = [(layer, option, budget)]
+        while stack:
+            layer, option, budget = stack[-1]
+
+            if budget <= 0 or layer < 0 or option < 0:
+                benefit_map[layer, option, budget] = 0
+                stack.pop()
+                continue
+
+            # We either not use this option
+            prev_layer = layer if option > 0 else layer - 1
+            prev_option = (option if option > 0 else len(options)) - 1
+            if (prev_layer, prev_option, budget) not in benefit_map:
+                stack.append((prev_layer, prev_option, budget))
+                continue
+            a = benefit_map[prev_layer, prev_option, budget]
+
+            # Or we use it so we have less budget for before
+            b = float("-inf")
+            info = sensitivities[layer][1][option]
+            prev_layer = layer - 1
+            prev_option = len(options) - 1
+            prev_budget = budget - info["extra_bits"]
+            if (
+                prev_layer,
+                prev_option,
+                prev_budget,
+            ) not in benefit_map and prev_budget >= 0:
+                stack.append((prev_layer, prev_option, prev_budget))
+                continue
+            if prev_budget >= 0:
+                b = benefit_map[prev_layer, prev_option, prev_budget]
+                b += info["loss_change"]
+
+            benefit_map[layer, option, budget] = max(a, b)
+            stack.pop()
+
+        return benefit_map[layer, option, budget]
+
+    def backtrack(layer, budget):
+        selected = []
+        while layer >= 0:
+            prev_benefit = benefit(layer - 1, len(options) - 1, budget)
+            option_benefits = [benefit(layer, i, budget) for i in range(len(options))]
+            idx, v = max(enumerate(option_benefits), key=lambda x: x[1] - prev_benefit)
+            info = sensitivities[layer][1][idx]
+            if v != 0:
+                budget -= info["extra_bits"]
+                selected.append((layer, idx))
+            layer -= 1
+        return selected[::-1]
+
+    selected = backtrack(len(sensitivities) - 1, budget)
+    config = {sensitivities[l][0]: options[i] for l, i in selected}
+
+    return config
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", "-m", default="Qwen/Qwen3-0.6B-base")
+    parser.add_argument(
+        "--mlx-path", default="mlx_model", help="Path to save the model"
+    )
+    parser.add_argument("--seed", type=int, default=123)
+    parser.add_argument(
+        "--sensitivities",
+        type=str,
+        default=None,
+        help="Path to a pre-computed sensitivity JSON file.",
+    )
+    parser.add_argument(
+        "--target-bpw", type=float, default=5.0, help="Target bits per weight."
+    )
+    parser.add_argument("--low-bits", type=int, default=4)
+    parser.add_argument("--low-group-size", type=int, default=128)
+    parser.add_argument("--high-bits", type=int, default=5)
+    parser.add_argument("--high-group-size", type=int, default=32)
+    parser.add_argument(
+        "--report-ppl",
+        action="store_true",
+        help="Compute the perplexity of the base and quantized models.",
+    )
+    parser.add_argument(
+        "--grad-checkpoint",
+        action="store_true",
+        help="Use gradient checkpointing to reduce memory use.",
+    )
+    parser.add_argument(
+        "--accumulation-dtype",
+        default="float32",
+        choices=["float32", "bfloat16"],
+        help="What type to use to accumulate the gradients for the sensitivities",
+    )
+    args = parser.parse_args()
+
+    group = mx.distributed.init()
+
+    if args.sensitivities is None:
+        model, tokenizer = load(args.model)
+        mx.random.seed(args.seed)
+        data = load_data(tokenizer, num_samples=-1, sequence_length=512)
+
+        sensitivities = estimate_sensitivities(
+            model,
+            data,
+            args.low_bits,
+            args.low_group_size,
+            args.high_bits,
+            args.high_group_size,
+            gradient_accum_dtype=getattr(mx, args.accumulation_dtype),
+            gradient_checkpoint=args.grad_checkpoint,
+        )
+        model_name = args.model.replace("/", "_")
+        with open(f"{model_name}_sensitivities.json", "w") as fid:
+            json.dump(sensitivities, fid)
+    else:
+        with open(args.sensitivities, "r") as fid:
+            sensitivities = json.load(fid)
+
+    sensitivities = dict(sensitivities)
+    model_path, hf_repo = get_model_path(args.model, revision=None)
+    model, config, tokenizer = fetch_from_hub(model_path, lazy=True)
+    mx.random.seed(args.seed)
+    data = load_data(tokenizer, num_samples=-1, sequence_length=512)
+
+    if args.report_ppl:
+        ppl = eval_ppl(model, data)
+        print(f"Original PPL: {ppl:.3f}")
+
+    quant_config = estimate_threshold(
+        model,
+        sensitivities,
+        target_bpw=args.target_bpw,
+        low_bits=args.low_bits,
+        low_group_size=args.low_group_size,
+        high_bits=args.high_bits,
+        high_group_size=args.high_group_size,
+    )
+
+    model, config = quantize_model(
+        model,
+        config,
+        q_group_size=args.low_group_size,
+        q_bits=args.low_bits,
+        quant_predicate=make_quant_predicate(quant_config),
+    )
+
+    if args.report_ppl:
+        ppl = eval_ppl(model, data)
+        print(f"Quantized PPL: {ppl:.3f}")
+
+    save(
+        args.mlx_path,
+        model_path,
+        model,
+        tokenizer,
+        config,
+        hf_repo=hf_repo,
+    )
+    print(f"Peak memory used: {mx.get_peak_memory() / 1000**3:.3f}GB")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,26 @@
+# Copyright © 2025 Apple Inc.
+
+from pathlib import Path
+
+import mlx.core as mx
+
+
+def load_data(tokenizer, num_samples: int, sequence_length: int) -> mx.array:
+    save_dir = Path.home() / ".cache/mlx-lm/calibration_v5.txt"
+    if not save_dir.exists():
+        from urllib import request
+
+        save_dir.parent.mkdir(parents=True, exist_ok=True)
+        url = "https://gist.githubusercontent.com/tristandruyen/9e207a95c7d75ddf37525d353e00659c/raw/571fda718462de863e5a0171078c175420c7649a/calibration_data_v5_rc.txt"
+        request.urlretrieve(url, save_dir)
+    with open(save_dir) as fid:
+        texts = fid.read()
+    tokens = tokenizer.encode(texts, return_tensors="mlx")[0]
+
+    # select random non-overlapping chunks
+    tokens = tokens[: (tokens.size // sequence_length) * sequence_length]
+    tokens = tokens.reshape(-1, sequence_length)
+    segments = mx.random.permutation(tokens.shape[0])
+    if num_samples > 0:
+        segments = segments[:num_samples]
+    return tokens[segments]
@@ -2,7 +2,7 @@

 import math
 from functools import partial
-from typing import Callable, Dict, Optional
+from typing import Callable, Dict, List, Optional

 import mlx.core as mx

@@ -12,6 +12,10 @@ def make_sampler(
    top_p: float = 0.0,
    min_p: float = 0.0,
    min_tokens_to_keep: int = 1,
+    top_k: int = 0,
+    xtc_probability: float = 0.0,
+    xtc_threshold: float = 0.0,
+    xtc_special_tokens: List[int] = [],
 ) -> Callable[mx.array, mx.array]:
    """
    Make a sampler function for use with ``generate_step``.
@@ -25,6 +29,15 @@ def make_sampler(
          probability) that a token probability must have to be considered.
        min_tokens_to_keep (int, optional): Minimum number of tokens that cannot
          be filtered by min_p sampling.
+        top_k (int, optional): The top k tokens ranked by probability to constrain
+          the sampling to.
+        xtc_probability (float, optional): The probability of applying XTC
+            sampling.
+        xtc_threshold (float, optional): The threshold the probs need to reach
+            for being sampled.
+        xtc_special_tokens (list(int), optional): List of special tokens IDs to
+            be excluded from XTC sampling.
+

    Returns:
        Callable[mx.array, mx.array]:
@@ -32,12 +45,29 @@ def make_sampler(
    """
    if temp == 0:
        return lambda x: mx.argmax(x, axis=-1)
-    elif top_p > 0 and top_p < 1.0:
-        return lambda x: top_p_sampling(x, top_p, temp)
-    elif min_p != 0.0:
-        return lambda x: min_p_sampling(x, min_p, min_tokens_to_keep, temp)
-    else:
-        return lambda x: categorical_sampling(x, temp)
+
+    # Create sampler chain
+    sampling_methods = []
+    if top_k > 0:
+        sampling_methods.append(lambda x: apply_top_k(x, top_k))
+    if top_p > 0 and top_p < 1.0:
+        sampling_methods.append(lambda x: apply_top_p(x, top_p))
+    if min_p != 0.0:
+        sampling_methods.append(lambda x: apply_min_p(x, min_p, min_tokens_to_keep))
+    if xtc_probability > 0.0:
+        sampling_methods.append(
+            lambda x: apply_xtc(x, xtc_probability, xtc_threshold, xtc_special_tokens)
+        )
+
+    # Apply the sampling methods
+    def sampler(logits):
+        for method in sampling_methods:
+            logits = method(logits)
+
+        # Return the sampled token
+        return categorical_sampling(logits, temp)
+
+    return sampler


 def make_logits_processors(
@@ -80,14 +110,38 @@ def make_logits_processors(


@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
-def min_p_sampling(
+def apply_top_k(
+    logprobs: mx.array,
+    top_k: int,
+) -> mx.array:
+    """
+    Sample from only the top K tokens ranked by probability.
+
+    Args:
+        logprobs: A vector of log probabilities.
+        top_k (int): Top k tokens to sample from.
+    """
+    vocab_size = logprobs.shape[-1]
+    if not isinstance(top_k, int) or not (0 < top_k < vocab_size):
+        raise ValueError(
+            f"`top_k` has to be an integer in the (0, {vocab_size}] interval,"
+            f" but is {top_k}."
+        )
+    mask_idx = mx.argpartition(-logprobs, kth=top_k - 1, axis=-1)[..., top_k:]
+    masked_logprobs = mx.put_along_axis(
+        logprobs, mask_idx, mx.array(-float("inf"), logprobs.dtype), axis=-1
+    )
+    return masked_logprobs
+
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_min_p(
    logprobs: mx.array,
    min_p: float,
    min_tokens_to_keep: int = 1,
-    temperature=1.0,
 ) -> mx.array:
    """
-    Apply min-p sampling to the logits.
+    Apply min-p sampling to the logprobs.

    Min-p keeps all tokens that are above a minimum probability, scaled by the
    probability of the most likely token. As a result, the filter is more
@@ -112,14 +166,12 @@ def min_p_sampling(
        )
    # reference implementation: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L531-L605

-    logprobs = logprobs * (1 / temperature)
-
    # Indices sorted in decreasing order
-    sorted_indices = mx.argsort(-logprobs).squeeze(0)
-    sorted_logprobs = logprobs[..., sorted_indices]
+    sorted_indices = mx.argsort(-logprobs, axis=-1)
+    sorted_logprobs = mx.take_along_axis(logprobs, sorted_indices, axis=-1)

    # Top probability
-    top_logprobs = logprobs[..., sorted_indices[0]]
+    top_logprobs = sorted_logprobs[:, 0:1]

    # Calculate the min_p threshold
    scaled_min_p = top_logprobs + math.log(min_p)
@@ -131,43 +183,93 @@ def min_p_sampling(
    # Create pool of tokens with probability less than scaled min_p
    selected_logprobs = mx.where(tokens_to_remove, -float("inf"), sorted_logprobs)

-    # Return sampled token
-    sorted_token = mx.random.categorical(selected_logprobs)
-    return sorted_indices[sorted_token]
+    # 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


@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
-def top_p_sampling(logits: mx.array, top_p: float, temperature: float) -> mx.array:
+def apply_top_p(logprobs: mx.array, top_p: float) -> mx.array:
    """
    Apply top-p (nucleus) sampling to logits.

    Args:
-        logits: The logits from the model's output.
+        logprobs: A vector of log probabilities.
        top_p: The cumulative probability threshold for top-p filtering.
-        temperature: Temperature parameter for softmax distribution reshaping.
    Returns:
        token selected based on the top-p criterion.
    """
    # referenced implementation from https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L449-L460
-    probs = mx.softmax(logits * (1 / temperature), axis=-1)
-
-    # sort probs in ascending order
-    sorted_indices = mx.argsort(probs, axis=-1)
-    sorted_probs = probs[..., sorted_indices.squeeze(0)]
+    probs = mx.exp(logprobs)
+    # sort in ascending order
+    sorted_indices = mx.argsort(logprobs, axis=-1)
+    sorted_probs = mx.take_along_axis(probs, sorted_indices, axis=-1)

    cumulative_probs = mx.cumsum(sorted_probs, axis=-1)

+    # Rearrange cumulative probs back to original order
+    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,
+    )
+    cumulative_probs = mx.take_along_axis(cumulative_probs, inverse_indices, axis=-1)
+
    # select tokens with cumulative probs below threshold
-    top_probs = mx.where(
+    return mx.where(
        cumulative_probs > 1 - top_p,
-        sorted_probs,
-        0,
+        logprobs,
+        -float("inf"),
    )

-    sorted_token = mx.random.categorical(mx.log(top_probs))
-    token = sorted_indices.squeeze(0)[sorted_token]

-    return token
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_xtc(
+    logits: mx.array,
+    xtc_probability: float,
+    xtc_threshold: float,
+    xtc_special_tokens: List[int],
+) -> mx.array:
+    """
+    Apply XTC sampling to the logits.
+
+    Args:
+        logits: The logits from the model's output.
+        xtc_probability (float): Probability of XTC sampling to happen for each token
+        xtc_threshold (float): The threshold the probs need to reach for being sampled.
+        special_tokens_ids (list(int)): List of special tokens IDs to be excluded from XTC sampling.
+    """
+    if not (0 <= xtc_threshold <= 0.5):
+        raise ValueError(
+            f"`threshold` has to be a float in the [0, 0.5] interval, but is {xtc_threshold}"
+        )
+    if not (0 <= xtc_probability <= 1.0):
+        raise ValueError(
+            f"`probability` has to be a float in the [0, 1] interval, but is {xtc_probability}"
+        )
+
+    probs = mx.softmax(logits, -1)
+    mask = probs > mx.where(probs > xtc_threshold, probs, mx.inf).min()
+    if xtc_special_tokens:
+        mask[..., xtc_special_tokens] = False
+
+    return mx.where(
+        mx.random.uniform(0, 1) > xtc_probability,
+        logits,
+        mx.where(mask, -mx.inf, logits),
+    )


@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
@@ -4,6 +4,7 @@ import argparse
 import json
 import logging
 import platform
+import socket
 import time
 import uuid
 import warnings
@@ -26,9 +27,10 @@ import mlx.core as mx
 from huggingface_hub import scan_cache_dir

 from ._version import __version__
-from .models.cache import make_prompt_cache
+from .generate import stream_generate
+from .models.cache import can_trim_prompt_cache, make_prompt_cache, trim_prompt_cache
 from .sample_utils import make_logits_processors, make_sampler
-from .utils import load, stream_generate
+from .utils import common_prefix_len, load


 def get_system_fingerprint():
@@ -114,10 +116,39 @@ def convert_chat(messages: List[dict], role_mapping: Optional[dict] = None):
    return prompt.rstrip()


+def process_message_content(messages):
+    """
+    Convert message content to a format suitable for `apply_chat_template`.
+
+    The function operates on messages in place. It converts the 'content' field
+    to a string instead of a list of text fragments.
+
+    Args:
+        message_list (list): A list of dictionaries, where each dictionary may
+          have a 'content' key containing a list of dictionaries with 'type' and
+          'text' keys.
+
+    Raises:
+        ValueError: If the 'content' type is not supported or if 'text' is missing.
+
+    """
+    for message in messages:
+        content = message["content"]
+        if isinstance(content, list):
+            text_fragments = [
+                fragment["text"] for fragment in content if fragment["type"] == "text"
+            ]
+            if len(text_fragments) != len(content):
+                raise ValueError("Only 'text' content type is supported.")
+            message["content"] = "".join(text_fragments)
+        elif content is None:
+            message["content"] = ""
+
+
@dataclass
 class PromptCache:
    cache: List[Any] = field(default_factory=list)
-    model_key: Tuple[str, Optional[str]] = ("", None)
+    model_key: Tuple[str, Optional[str]] = ("", None, None)
    tokens: List[int] = field(default_factory=list)


@@ -128,10 +159,11 @@ class ModelProvider:
        self.model_key = None
        self.model = None
        self.tokenizer = None
+        self.draft_model = None

        # Preload the default model if it is provided
        if self.cli_args.model is not None:
-            self.load("default_model")
+            self.load("default_model", draft_model_path="default_model")

    def _validate_model_path(self, model_path: str):
        model_path = Path(model_path)
@@ -141,14 +173,15 @@ class ModelProvider:
            )

    # Added in adapter_path to load dynamically
-    def load(self, model_path, adapter_path=None):
-        if self.model_key == (model_path, adapter_path):
+    def load(self, model_path, adapter_path=None, draft_model_path=None):
+        if self.model_key == (model_path, adapter_path, draft_model_path):
            return self.model, self.tokenizer

        # Remove the old model if it exists.
        self.model = None
        self.tokenizer = None
        self.model_key = None
+        self.draft_model = None

        # Building tokenizer_config
        tokenizer_config = {
@@ -157,7 +190,12 @@ class ModelProvider:
        if self.cli_args.chat_template:
            tokenizer_config["chat_template"] = self.cli_args.chat_template

-        if model_path == "default_model" and self.cli_args.model is not None:
+        if model_path == "default_model":
+            if self.cli_args.model is None:
+                raise ValueError(
+                    "A model path has to be given as a CLI "
+                    "argument or in the HTTP request"
+                )
            model, tokenizer = load(
                self.cli_args.model,
                adapter_path=(
@@ -175,10 +213,30 @@ class ModelProvider:
            if tokenizer.chat_template is None:
                tokenizer.chat_template = tokenizer.default_chat_template

-        self.model_key = (model_path, adapter_path)
+        self.model_key = (model_path, adapter_path, draft_model_path)
        self.model = model
        self.tokenizer = tokenizer

+        def validate_draft_tokenizer(draft_tokenizer):
+            # Check if tokenizers are compatible
+            if draft_tokenizer.vocab_size != tokenizer.vocab_size:
+                logging.warning(
+                    "Draft model tokenizer does not match model tokenizer. "
+                    "Speculative decoding may not work as expected."
+                )
+
+        # Load draft model if specified
+        if (
+            draft_model_path == "default_model"
+            and self.cli_args.draft_model is not None
+        ):
+            self.draft_model, draft_tokenizer = load(self.cli_args.draft_model)
+            validate_draft_tokenizer(draft_tokenizer)
+
+        elif draft_model_path is not None and draft_model_path != "default_model":
+            self._validate_model_path(draft_model_path)
+            self.draft_model, draft_tokenizer = load(draft_model_path)
+            validate_draft_tokenizer(draft_tokenizer)
        return self.model, self.tokenizer


@@ -250,22 +308,35 @@ class APIHandler(BaseHTTPRequestHandler):
        self.stream = self.body.get("stream", False)
        self.stream_options = self.body.get("stream_options", None)
        self.requested_model = self.body.get("model", "default_model")
+        self.requested_draft_model = self.body.get("draft_model", "default_model")
+        self.num_draft_tokens = self.body.get(
+            "num_draft_tokens", self.model_provider.cli_args.num_draft_tokens
+        )
        self.adapter = self.body.get("adapters", None)
        self.max_tokens = self.body.get("max_completion_tokens", None)
        if self.max_tokens is None:
-            self.max_tokens = self.body.get("max_tokens", 512)
-        self.temperature = self.body.get("temperature", 0.0)
-        self.top_p = self.body.get("top_p", 1.0)
+            self.max_tokens = self.body.get(
+                "max_tokens", self.model_provider.cli_args.max_tokens
+            )
+        self.temperature = self.body.get(
+            "temperature", self.model_provider.cli_args.temp
+        )
+        self.top_p = self.body.get("top_p", self.model_provider.cli_args.top_p)
+        self.top_k = self.body.get("top_k", self.model_provider.cli_args.top_k)
+        self.min_p = self.body.get("min_p", self.model_provider.cli_args.min_p)
        self.repetition_penalty = self.body.get("repetition_penalty", 1.0)
        self.repetition_context_size = self.body.get("repetition_context_size", 20)
+        self.xtc_probability = self.body.get("xtc_probability", 0.0)
+        self.xtc_threshold = self.body.get("xtc_threshold", 0.0)
        self.logit_bias = self.body.get("logit_bias", None)
        self.logprobs = self.body.get("logprobs", -1)
        self.validate_model_parameters()
-
        # Load the model if needed
        try:
            self.model, self.tokenizer = self.model_provider.load(
-                self.requested_model, self.adapter
+                self.requested_model,
+                self.adapter,
+                self.requested_draft_model,
            )
        except:
            self._set_completion_headers(404)
@@ -309,6 +380,15 @@ class APIHandler(BaseHTTPRequestHandler):
        if not isinstance(self.top_p, (float, int)) or self.top_p < 0 or self.top_p > 1:
            raise ValueError("top_p must be a float between 0 and 1")

+        if not isinstance(self.top_k, int) or self.top_k < 0:
+            raise ValueError("top_k must be a non-negative integer")
+
+        if not isinstance(self.min_p, (float, int)) or self.min_p < 0 or self.min_p > 1:
+            raise ValueError("min_p must be a float between 0 and 1")
+
+        if not isinstance(self.num_draft_tokens, int) or self.num_draft_tokens < 0:
+            raise ValueError("num_draft_tokens must be a non-negative integer")
+
        if (
            not isinstance(self.repetition_penalty, (float, int))
            or self.repetition_penalty < 0
@@ -334,7 +414,15 @@ class APIHandler(BaseHTTPRequestHandler):
                self.logit_bias = {int(k): v for k, v in self.logit_bias.items()}
            except ValueError:
                raise ValueError("logit_bias must be a dict of int to float")
-
+        if not (
+            isinstance(self.xtc_probability, float)
+            and 0.00 <= self.xtc_probability <= 1.00
+        ):
+            raise ValueError(f"xtc_probability must be a float between 0.00 and 1.00")
+        if not (
+            isinstance(self.xtc_threshold, float) and 0.00 <= self.xtc_threshold <= 0.50
+        ):
+            raise ValueError(f"xtc_threshold must be a float between 0.00 and 0.5")
        if not isinstance(self.requested_model, str):
            raise ValueError("model must be a string")
        if self.adapter is not None and not isinstance(self.adapter, str):
@@ -349,6 +437,7 @@ class APIHandler(BaseHTTPRequestHandler):
        token_logprobs: Optional[List[float]] = None,
        top_tokens: Optional[List[Dict[int, float]]] = None,
        tokens: Optional[List[int]] = None,
+        tool_calls: Optional[List[str]] = None,
    ) -> dict:
        """
        Generate a single response packet based on response type (stream or
@@ -367,13 +456,26 @@ class APIHandler(BaseHTTPRequestHandler):
            top_tokens (Optional[List[Dict[int, float]]]): List of dictionaries mapping
              tokens to logprobs for the top N tokens at each token position.
            tokens (Optional[List[int]]): List of tokens to return with logprobs structure
+            tool_calls (Optional[List[str]]): List of tool calls.

        Returns:
            dict: A dictionary containing the response, in the same format as
              OpenAI's API.
        """
-        token_logprobs = token_logprobs if token_logprobs else []
-        top_logprobs = top_tokens if top_tokens else []
+        token_logprobs = token_logprobs or []
+        top_logprobs = top_tokens or []
+        tool_calls = tool_calls or []
+
+        def parse_function(tool_text):
+            tool_call = json.loads(tool_text.strip())
+            return {
+                "function": {
+                    "name": tool_call.get("name", None),
+                    "arguments": json.dumps(tool_call.get("arguments", "")),
+                },
+                "type": "function",
+                "id": None,
+            }

        # Static response
        response = {
@@ -391,7 +493,7 @@ class APIHandler(BaseHTTPRequestHandler):
                        "tokens": tokens,
                    },
                    "finish_reason": finish_reason,
-                }
+                },
            ],
        }

@@ -415,7 +517,11 @@ class APIHandler(BaseHTTPRequestHandler):
        # Add dynamic response
        if self.object_type.startswith("chat.completion"):
            key_name = "delta" if self.stream else "message"
-            choice[key_name] = {"role": "assistant", "content": text}
+            choice[key_name] = {
+                "role": "assistant",
+                "content": text,
+                "tool_calls": [parse_function(tool_text) for tool_text in tool_calls],
+            }
        elif self.object_type == "text_completion":
            choice.update(text=text)
        else:
@@ -423,18 +529,87 @@ class APIHandler(BaseHTTPRequestHandler):

        return response

+    def reset_prompt_cache(self, prompt):
+        """Resets the prompt cache and associated state.
+
+        Args:
+            prompt (List[int]): The tokenized new prompt which will populate the
+                reset cache.
+        """
+        logging.debug(f"*** Resetting cache. ***")
+        self.prompt_cache.model_key = self.model_provider.model_key
+        self.prompt_cache.cache = make_prompt_cache(self.model_provider.model)
+        if self.model_provider.draft_model is not None:
+            self.prompt_cache.cache += make_prompt_cache(
+                self.model_provider.draft_model
+            )
+        self.prompt_cache.tokens = list(prompt)  # Cache the new prompt fully
+
    def get_prompt_cache(self, prompt):
+        """
+        Determines the portion of the prompt that needs processing by comparing
+        it to the cached prompt and attempting to reuse the common prefix.
+
+        This function updates the internal prompt cache state (tokens and model cache)
+        based on the comparison. If a common prefix exists, it attempts to trim
+        the model cache (if supported) to match the common prefix length, avoiding
+        recomputation.
+
+        Args:
+            prompt (List[int]): The tokenized new prompt.
+
+        Returns:
+            List[int]: The suffix of the prompt that actually needs to be processed
+                       by the model. This will be the full prompt if the cache is
+                       reset or cannot be effectively used.
+        """
        cache_len = len(self.prompt_cache.tokens)
+        prompt_len = len(prompt)
+        com_prefix_len = common_prefix_len(self.prompt_cache.tokens, prompt)
+
+        # Leave at least one token in the prompt
+        com_prefix_len = min(com_prefix_len, len(prompt) - 1)
+
+        # Condition 1: Model changed or no common prefix at all. Reset cache.
        if (
            self.prompt_cache.model_key != self.model_provider.model_key
-            or cache_len >= len(prompt)
-            or self.prompt_cache.tokens != prompt[:cache_len]
+            or com_prefix_len == 0
        ):
-            self.prompt_cache.model_key = self.model_provider.model_key
-            self.prompt_cache.cache = make_prompt_cache(self.model_provider.model)
+            self.reset_prompt_cache(prompt)
+
+        # Condition 2: Common prefix exists and matches cache length. Process suffix.
+        elif com_prefix_len == cache_len:
+            logging.debug(
+                f"*** Cache is prefix of prompt (cache_len: {cache_len}, prompt_len: {prompt_len}). Processing suffix. ***"
+            )
+            prompt = prompt[com_prefix_len:]
+            self.prompt_cache.tokens.extend(prompt)
+
+        # Condition 3: Common prefix exists but is shorter than cache length. Attempt trim.
+        elif com_prefix_len < cache_len:
+            logging.debug(
+                f"*** Common prefix ({com_prefix_len}) shorter than cache ({cache_len}). Attempting trim. ***"
+            )
+
+            if can_trim_prompt_cache(self.prompt_cache.cache):
+                num_to_trim = cache_len - com_prefix_len
+                logging.debug(f"    Trimming {num_to_trim} tokens from cache.")
+                trim_prompt_cache(self.prompt_cache.cache, num_to_trim)
+                self.prompt_cache.tokens = self.prompt_cache.tokens[:com_prefix_len]
+                prompt = prompt[com_prefix_len:]
+                self.prompt_cache.tokens.extend(prompt)
+            else:
+                logging.debug(f"    Cache cannot be trimmed. Resetting cache.")
+                self.reset_prompt_cache(prompt)
+
+        # This case should logically not be reached if com_prefix_len <= cache_len
        else:
-            prompt = prompt[cache_len:]
-        self.prompt_cache.tokens.extend(prompt)
+            logging.error(
+                f"Unexpected cache state: com_prefix_len ({com_prefix_len}) > cache_len ({cache_len}). Resetting cache."
+            )
+            self.reset_prompt_cache(prompt)
+
+        logging.debug(f"Returning {len(prompt)} tokens for processing.")
        return prompt

    def handle_completion(
@@ -465,10 +640,28 @@ class APIHandler(BaseHTTPRequestHandler):

        text = ""
        tic = time.perf_counter()
-        sampler = make_sampler(self.temperature, top_p=self.top_p)
-        logits_processors = make_logits_processors(
-            self.logit_bias, self.repetition_penalty, self.repetition_context_size
+        sampler = make_sampler(
+            self.temperature,
+            top_p=self.top_p,
+            top_k=self.top_k,
+            min_p=self.min_p,
+            xtc_probability=self.xtc_probability,
+            xtc_threshold=self.xtc_threshold,
+            xtc_special_tokens=[
+                self.tokenizer.eos_token_id,
+                self.tokenizer.encode("\n"),
+            ],
        )
+        logits_processors = make_logits_processors(
+            self.logit_bias,
+            self.repetition_penalty,
+            self.repetition_context_size,
+        )
+
+        tool_calls = []
+        tool_text = ""
+        in_tool_call = False
+        segment = ""
        for gen_response in stream_generate(
            model=self.model,
            tokenizer=self.tokenizer,
@@ -477,10 +670,26 @@ class APIHandler(BaseHTTPRequestHandler):
            sampler=sampler,
            logits_processors=logits_processors,
            prompt_cache=self.prompt_cache.cache,
+            draft_model=self.model_provider.draft_model,
+            num_draft_tokens=self.num_draft_tokens,
        ):
-            segment = gen_response.text
-            text += segment
-            logging.debug(text)
+            logging.debug(gen_response.text)
+
+            if (
+                self.tokenizer.has_tool_calling
+                and gen_response.text == self.tokenizer.tool_call_start
+            ):
+                in_tool_call = True
+            elif in_tool_call:
+                if gen_response.text == self.tokenizer.tool_call_end:
+                    tool_calls.append(tool_text)
+                    tool_text = ""
+                    in_tool_call = False
+                else:
+                    tool_text += gen_response.text
+            else:
+                text += gen_response.text
+                segment += gen_response.text
            token = gen_response.token
            logprobs = gen_response.logprobs
            tokens.append(token)
@@ -504,9 +713,10 @@ class APIHandler(BaseHTTPRequestHandler):
                        tokens[-stop_condition.trim_length :]
                    )
                    text = text[: -len(stop_sequence_suffix)]
+                segment = ""
                break

-            if self.stream:
+            if self.stream and not in_tool_call:
                # If the end of tokens overlaps with a stop sequence, generate new
                # tokens until we know if the stop sequence is hit or not
                if any(
@@ -516,10 +726,14 @@ class APIHandler(BaseHTTPRequestHandler):
                    )
                ):
                    continue
-                elif segment:
-                    response = self.generate_response(segment, None)
+                elif segment or tool_calls:
+                    response = self.generate_response(
+                        segment, None, tool_calls=tool_calls
+                    )
                    self.wfile.write(f"data: {json.dumps(response)}\n\n".encode())
                    self.wfile.flush()
+                    segment = ""
+                    tool_calls = []

        self.prompt_cache.tokens.extend(tokens)

@@ -528,7 +742,9 @@ class APIHandler(BaseHTTPRequestHandler):
        logging.debug(f"Peak memory: {gen_response.peak_memory:.3f} GB")

        if self.stream:
-            response = self.generate_response(segment, finish_reason)
+            response = self.generate_response(
+                segment, finish_reason, tool_calls=tool_calls
+            )
            self.wfile.write(f"data: {json.dumps(response)}\n\n".encode())
            self.wfile.flush()
            if self.stream_options is not None and self.stream_options["include_usage"]:
@@ -546,6 +762,7 @@ class APIHandler(BaseHTTPRequestHandler):
                token_logprobs=token_logprobs,
                top_tokens=top_tokens,
                tokens=tokens,
+                tool_calls=tool_calls,
            )
            response_json = json.dumps(response).encode()
            indent = "\t"  # Backslashes can't be inside of f-strings
@@ -590,15 +807,14 @@ class APIHandler(BaseHTTPRequestHandler):
        # Determine response type
        self.request_id = f"chatcmpl-{uuid.uuid4()}"
        self.object_type = "chat.completion.chunk" if self.stream else "chat.completion"
-        if (
-            hasattr(self.tokenizer, "apply_chat_template")
-            and self.tokenizer.chat_template
-        ):
+        if self.tokenizer.chat_template:
+            messages = body["messages"]
+            process_message_content(messages)
            prompt = self.tokenizer.apply_chat_template(
-                body["messages"],
-                body.get("tools", None),
-                tokenize=True,
+                messages,
+                body.get("tools") or None,
                add_generation_prompt=True,
+                **self.model_provider.cli_args.chat_template_args,
            )
        else:
            prompt = convert_chat(body["messages"], body.get("role_mapping"))
@@ -625,11 +841,23 @@ class APIHandler(BaseHTTPRequestHandler):
        """
        if self.path == "/v1/models":
            self.handle_models_request()
+        elif self.path == "/health":
+            self.handle_health_check()
        else:
            self._set_completion_headers(404)
            self.end_headers()
            self.wfile.write(b"Not Found")

+    def handle_health_check(self):
+        """
+        Handle a GET request for the /health endpoint.
+        """
+        self._set_completion_headers(200)
+        self.end_headers()
+
+        self.wfile.write('{"status": "ok"}'.encode())
+        self.wfile.flush()
+
    def handle_models_request(self):
        """
        Handle a GET request for the /v1/models endpoint.
@@ -637,10 +865,20 @@ class APIHandler(BaseHTTPRequestHandler):
        self._set_completion_headers(200)
        self.end_headers()

+        files = ["config.json", "model.safetensors.index.json", "tokenizer_config.json"]
+
+        def probably_mlx_lm(repo):
+            if repo.repo_type != "model":
+                return False
+            if "main" not in repo.refs:
+                return False
+            file_names = {f.file_path.name for f in repo.refs["main"].files}
+            return all(f in file_names for f in files)
+
        # Scan the cache directory for downloaded mlx models
        hf_cache_info = scan_cache_dir()
        downloaded_models = [
-            repo for repo in hf_cache_info.repos if "mlx" in repo.repo_id
+            repo for repo in hf_cache_info.repos if probably_mlx_lm(repo)
        ]

        # Create a list of available models
@@ -669,6 +907,10 @@ def run(
 ):
    server_address = (host, port)
    prompt_cache = PromptCache()
+    infos = socket.getaddrinfo(
+        *server_address, type=socket.SOCK_STREAM, flags=socket.AI_PASSIVE
+    )
+    server_class.address_family, _, _, _, server_address = next(iter(infos))
    httpd = server_class(
        server_address,
        lambda *args, **kwargs: handler_class(
@@ -711,6 +953,18 @@ def main():
        default=8080,
        help="Port for the HTTP server (default: 8080)",
    )
+    parser.add_argument(
+        "--draft-model",
+        type=str,
+        help="A model to be used for speculative decoding.",
+        default=None,
+    )
+    parser.add_argument(
+        "--num-draft-tokens",
+        type=int,
+        help="Number of tokens to draft when using speculative decoding.",
+        default=3,
+    )
    parser.add_argument(
        "--trust-remote-code",
        action="store_true",
@@ -723,13 +977,6 @@ def main():
        choices=["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"],
        help="Set the logging level (default: INFO)",
    )
-    parser.add_argument(
-        "--cache-limit-gb",
-        type=int,
-        default=None,
-        help="Set the MLX cache limit in GB",
-        required=False,
-    )
    parser.add_argument(
        "--chat-template",
        type=str,
@@ -742,19 +989,54 @@ def main():
        action="store_true",
        help="Use the default chat template",
    )
+    parser.add_argument(
+        "--temp",
+        type=float,
+        default=0.0,
+        help="Default sampling temperature (default: 0.0)",
+    )
+    parser.add_argument(
+        "--top-p",
+        type=float,
+        default=1.0,
+        help="Default nucleus sampling top-p (default: 1.0)",
+    )
+    parser.add_argument(
+        "--top-k",
+        type=int,
+        default=0,
+        help="Default top-k sampling (default: 0, disables top-k)",
+    )
+    parser.add_argument(
+        "--min-p",
+        type=float,
+        default=0.0,
+        help="Default min-p sampling (default: 0.0, disables min-p)",
+    )
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=512,
+        help="Default maximum number of tokens to generate (default: 512)",
+    )
+    parser.add_argument(
+        "--chat-template-args",
+        type=json.loads,
+        help="""A JSON formatted string of arguments for the tokenizer's apply_chat_template, e.g. '{"enable_thinking":false}'""",
+        default="{}",
+    )
    args = parser.parse_args()

    logging.basicConfig(
        level=getattr(logging, args.log_level.upper(), None),
        format="%(asctime)s - %(levelname)s - %(message)s",
    )
-
-    if args.cache_limit_gb is not None:
-        logging.debug(f"Setting cache limit to {args.cache_limit_gb} GB")
-        mx.metal.set_cache_limit(args.cache_limit_gb * 1024 * 1024 * 1024)
-
    run(args.host, args.port, ModelProvider(args))


 if __name__ == "__main__":
+    print(
+        "Calling `python -m mlx_lm.server...` directly is deprecated."
+        " Use `mlx_lm.server...` or `python -m mlx_lm server ...` instead."
+    )
    main()
@@ -1,7 +1,9 @@
 import json
 from functools import partial
+from json import JSONDecodeError
+from typing import List

-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, PreTrainedTokenizerFast


 class StreamingDetokenizer:
@@ -89,6 +91,7 @@ class NaiveStreamingDetokenizer(StreamingDetokenizer):
            self._current_text = self._tokenizer.decode(self._current_tokens)
            if (
                self._tokenizer.clean_up_tokenization_spaces
+                and len(self._current_text) > 0
                and self._current_text[-1] == " "
            ):
                self._current_text = self._current_text[:-1]
@@ -265,6 +268,64 @@ class TokenizerWrapper:
            if eos_token_ids is not None
            else {tokenizer.eos_token_id}
        )
+        self._think_start = None
+        self._think_end = None
+        self._tool_call_start = None
+        self._tool_call_end = None
+
+        THINK_TOKENS = [("<think>", "</think>")]
+        TOOL_CALL_TOKENS = [("<tool_call>", "</tool_call>")]
+
+        vocab = tokenizer.get_vocab()
+        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
+                break
+        if tokenizer.chat_template and '"tool"' in tokenizer.chat_template:
+            self._tool_call_start = ""
+            self._tool_call_end = ""
+            for tool_call_start, tool_call_end in TOOL_CALL_TOKENS:
+                if tool_call_start in vocab and tool_call_end in vocab:
+                    self._tool_call_start = tool_call_start
+                    self._tool_call_end = tool_call_end
+                    break
+
+    def add_eos_token(self, token: str):
+        token_id = None
+        try:
+            token_id = int(token)
+        except ValueError:
+            token_id = self._tokenizer.convert_tokens_to_ids(token)
+
+        if token_id is None:
+            raise ValueError(f"'{token}' is not a token for this tokenizer")
+
+        self._eos_token_ids.add(token_id)
+
+    @property
+    def has_thinking(self):
+        return self._think_start is not None
+
+    @property
+    def think_start(self):
+        return self._think_start
+
+    @property
+    def think_end(self):
+        return self._think_end
+
+    @property
+    def has_tool_calling(self):
+        return self._tool_call_start is not None
+
+    @property
+    def tool_call_start(self):
+        return self._tool_call_start
+
+    @property
+    def tool_call_end(self):
+        return self._tool_call_end

    def __getattr__(self, attr):
        if attr == "detokenizer":
@@ -288,6 +349,35 @@ class TokenizerWrapper:
            setattr(self._tokenizer, attr, value)


+class NewlineTokenizer(PreTrainedTokenizerFast):
+    """A tokenizer that replaces newlines with <n> and <n> with new line."""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def _preprocess_text(self, text):
+        return text.replace("\n", "<n>")
+
+    def _postprocess_text(self, text):
+        return text.replace("<n>", "\n")
+
+    def encode(self, text, **kwargs):
+        return super().encode(self._preprocess_text(text), **kwargs)
+
+    def encode_batch(self, texts, **kwargs):
+        return super().encode_batch([self._preprocess_text(t) for t in texts], **kwargs)
+
+    def decode(self, *args, **kwargs):
+        return self._postprocess_text(super().decode(*args, **kwargs))
+
+    def batch_decode(self, *args, **kwargs):
+        decoded = super().batch_decode(*args, **kwargs)
+        return [self._postprocess_text(d) for d in decoded]
+
+
+AutoTokenizer.register("NewlineTokenizer", fast_tokenizer_class=NewlineTokenizer)
+
+
 def _match(a, b):
    if type(a) != type(b):
        return False
@@ -328,7 +418,9 @@ def _is_bpe_decoder(decoder):
    return isinstance(decoder, dict) and decoder.get("type", None) == "ByteLevel"


-def load_tokenizer(model_path, tokenizer_config_extra={}, eos_token_ids=None):
+def load_tokenizer(
+    model_path, tokenizer_config_extra={}, return_tokenizer=True, eos_token_ids=None
+):
    """Load a huggingface tokenizer and try to infer the type of streaming
    detokenizer to use.

@@ -339,8 +431,12 @@ def load_tokenizer(model_path, tokenizer_config_extra={}, eos_token_ids=None):

    tokenizer_file = model_path / "tokenizer.json"
    if tokenizer_file.exists():
-        with open(tokenizer_file, "r") as fid:
-            tokenizer_content = json.load(fid)
+        with open(tokenizer_file, "r", encoding="utf-8") as fid:
+            try:
+                tokenizer_content = json.load(fid)
+            except JSONDecodeError as e:
+                raise JSONDecodeError("Failed to parse tokenizer.json", e.doc, e.pos)
+
        if "decoder" in tokenizer_content:
            if _is_spm_decoder(tokenizer_content["decoder"]):
                detokenizer_class = SPMStreamingDetokenizer
@@ -351,8 +447,17 @@ def load_tokenizer(model_path, tokenizer_config_extra={}, eos_token_ids=None):

    if isinstance(eos_token_ids, int):
        eos_token_ids = [eos_token_ids]
-    return TokenizerWrapper(
-        AutoTokenizer.from_pretrained(model_path, **tokenizer_config_extra),
-        detokenizer_class,
-        eos_token_ids=eos_token_ids,
-    )
+
+    if return_tokenizer:
+        return TokenizerWrapper(
+            AutoTokenizer.from_pretrained(model_path, **tokenizer_config_extra),
+            detokenizer_class,
+            eos_token_ids=eos_token_ids,
+        )
+    else:
+        return detokenizer_class
+
+
+def no_bos_or_eos(sequence: List, bos: int, eos: int) -> List:
+    removed_bos = sequence if sequence[0] != bos else sequence[1:]
+    return removed_bos[:-1] if removed_bos[-1] == eos else removed_bos
@@ -0,0 +1,43 @@
+# Copyright © 2024 Apple Inc.
+
+try:
+    import wandb
+except ImportError:
+    wandb = None
+
+
+class TrainingCallback:
+
+    def on_train_loss_report(self, train_info: dict):
+        """Called to report training loss at specified intervals."""
+        pass
+
+    def on_val_loss_report(self, val_info: dict):
+        """Called to report validation loss at specified intervals or the beginning."""
+        pass
+
+
+class WandBCallback(TrainingCallback):
+    def __init__(
+        self,
+        project_name: str,
+        log_dir: str,
+        config: dict,
+        wrapped_callback: TrainingCallback = None,
+    ):
+        if wandb is None:
+            raise ImportError(
+                "wandb is not installed. Please install it to use WandBCallback."
+            )
+        self.wrapped_callback = wrapped_callback
+        wandb.init(project=project_name, dir=log_dir, config=config)
+
+    def on_train_loss_report(self, train_info: dict):
+        wandb.log(train_info, step=train_info.get("iteration"))
+        if self.wrapped_callback:
+            self.wrapped_callback.on_train_loss_report(train_info)
+
+    def on_val_loss_report(self, val_info: dict):
+        wandb.log(val_info, step=val_info.get("iteration"))
+        if self.wrapped_callback:
+            self.wrapped_callback.on_val_loss_report(val_info)
--- a/Show More
+++ b/Show More