Add MLX-VLM runtime, configuration, and model card

.gitattributes

@@ -1,35 +1,4 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/loophard_four_settings.png filter=lfs diff=lfs merge=lfs -text
+assets/looping_vs_gpqa.png filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,275 @@
+---
+library_name: mlx
+license: apache-2.0
+license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/995ad96eacd98c81ed38be0c5b274b04031597b0/LICENSE
+base_model:
+  - nvidia/Qwen3.6-35B-A3B-NVFP4
+  - N8Programs/Qwen3.6-35B-A3B-AntiLoop
+pipeline_tag: image-text-to-text
+tags:
+  - mlx
+  - mlx-vlm
+  - image-text-to-text
+  - qwen3.6
+  - moe
+  - modelopt
+  - quantized
+  - nvfp4
+  - fp4
+  - fp8
+  - lora
+  - merged
+  - antidoom
+---
+
+# Qwen3.6-35B-A3B-AntiLoop-NVFP4 for MLX-VLM
+
+![AntiLoop looping rate versus GPQA capability](assets/looping_vs_gpqa.png)
+
+![LoopHard judged-loop rates across four inference settings](assets/loophard_four_settings.png)
+
+This is the Apple-silicon MLX-VLM conversion of
+[`N8Programs/Qwen3.6-35B-A3B-AntiLoop-NVFP4`](https://huggingface.co/N8Programs/Qwen3.6-35B-A3B-AntiLoop-NVFP4).
+It supports both text and vision inputs, including Qwen thinking controls.
+
+The conversion does not re-quantize the mixed NVIDIA ModelOpt language
+checkpoint. Its FP8 and NVFP4 payloads and scales are re-expressed for MLX's
+native MXFP8/NVFP4 kernels, while the source tensor-level scales are applied by
+the included runtime. Activations remain in the model dtype. The vision tower
+is the original BF16 data, byte-for-byte rather than quantized.
+
+- 1,808 tensors across 42 safetensors shards
+- 130 scaled MXFP8 dense modules
+- 121 scaled NVFP4 dense modules
+- 120 scaled NVFP4 expert projections
+- 333 BF16 vision tensors (893,142,496 tensor-data bytes)
+
+The checkpoint was converted and smoke-tested with `mlx==0.31.2`,
+`mlx-lm==0.31.3`, and `mlx-vlm==0.6.4`.
+
+## Use with MLX-VLM
+
+Install MLX-VLM and download the repository:
+
+```bash
+pip install -U "mlx-vlm==0.6.4"
+hf download mlx-community/Qwen3.6-35B-A3B-AntiLoop-NVFP4 \
+  --local-dir Qwen3.6-35B-A3B-AntiLoop-NVFP4
+cd Qwen3.6-35B-A3B-AntiLoop-NVFP4
+```
+
+Image + text generation:
+
+```bash
+python run_mlx_vlm.py \
+  --model . \
+  --trust-remote-code \
+  --image /path/to/image.png \
+  --prompt "Describe this image." \
+  --max-tokens 256
+```
+
+Text-only generation with thinking enabled:
+
+```bash
+python run_mlx_vlm.py \
+  --model . \
+  --trust-remote-code \
+  --enable-thinking \
+  --prompt "Solve: 27 * 43" \
+  --max-tokens 512
+```
+
+For programmatic loading from the downloaded repository:
+
+```python
+from mlx_vlm_model_file_loader import load
+
+model, processor = load(".")
+```
+
+MLX-VLM 0.6.4 does not yet natively consult a model-local
+`vlm_model_file`. `run_mlx_vlm.py` installs that single lookup inside the
+current process without modifying the installed package. The local runtime is
+executed only when `--trust-remote-code` (or `trust_remote_code=True`) is
+explicitly enabled. Review the included Python files before trusting them.
+
+An HTTP server can be started similarly:
+
+```bash
+python run_mlx_vlm_server.py \
+  --model . \
+  --trust-remote-code \
+  --enable-thinking
+```
+
+## Original model card
+
+This is a mixed-precision NVIDIA ModelOpt deployment checkpoint for
+[`Qwen3.6-35B-A3B-AntiLoop`](https://huggingface.co/N8Programs/Qwen3.6-35B-A3B-AntiLoop),
+a narrow fine-tune intended to recover from pathological self-verification and
+enumeration loops while preserving ordinary long-form reasoning.
+
+No PEFT adapter is required at inference time. The MLX conversion retains the
+upstream multimodal architecture, tokenizer, chat template, and 262,144-token
+native context configuration. It does not include the source checkpoint's MTP
+draft weights.
+
+## Training data
+
+The exact 178 masked supervised targets used for the final AntiLoop training
+round are published in the
+[`Qwen3.6-35B-A3B-AntiLoop-SFT` dataset](https://huggingface.co/datasets/N8Programs/Qwen3.6-35B-A3B-AntiLoop-SFT).
+The dataset preserves each `loss_start_char` boundary so the pathological loop
+prefix remains conditioning context rather than a supervised target. It
+intentionally excludes the separately generated KL-regularization anchors.
+
+## Training procedure
+
+The AntiLoop adapter was trained on the 178 masked supervised targets using a
+standard supervised fine-tuning procedure, but regularized via KL-loss on separately generated non-loop anchors from the base model on everyday prompts.
+
+## Benchmark results
+
+### LoopHard
+
+**LoopHard** is our held-out set of 285 enumeration prompts designed to elicit
+futile recall, recounting, and self-verification loops. The primary metric is
+**judged loops**: whether the model's reasoning trace remains stuck in a futile
+cycle when generation ends.
+
+| Model | Judged loops | Loop rate |
+|---|---:|---:|
+| NVIDIA NVFP4 | 72 / 285 | 25.26% |
+| **AntiLoop NVFP4** | **10 / 285** | **3.51%** |
+| NVIDIA NVFP4 + `presence_penalty=1.5` | 30 / 285 | 10.53% |
+| **AntiLoop NVFP4 + `presence_penalty=1.5`** | **1 / 285** | **0.35%** |
+
+The matched `presence_penalty=1.5` comparison converted all 30 control loops to
+clean completions while introducing one different loop. Exact two-sided
+McNemar `p = 2.98e-8`.
+
+Generation used thinking mode, `temperature=0.7`, `top_p=0.95`, `top_k=20`, a
+6,144-token completion limit, and concurrency 24. The two
+`presence_penalty=1.5` arms used the exact original and AntiLoop NVFP4
+checkpoints with the same vLLM build and serving configuration: TP1, FP8 KV
+cache, FlashInfer attention, Marlin NVFP4 MoE, and MTP speculative decoding
+with three draft tokens.
+
+LoopHard is judged by GLM-5.2 using a convergence-aware rubric: systematic
+reasoning and verification that reaches a conclusion are not loops, and a trace
+that notices its own circling and exits is classified as recovered. The
+calibration set contained 42 manually labeled traces. Across three judge runs,
+accuracy was 88.1%, 92.9%, and 95.2%; all three runs identified all 17 labeled
+loops, with 2–5 false positives among the 25 non-loop traces.
+
+The 285 prompts, metadata, and GLM-5.2 evaluation code are published in the
+[LoopHard dataset](https://huggingface.co/datasets/N8Programs/LoopHard) on
+Hugging Face.
+
+### Capability preservation
+
+The capability checks below compare the same round-2 AntiLoop adapter against
+its FP8 reference model under a matched runtime-LoRA setup. These runs used the
+default presence penalty and should not be interpreted as evaluations of the
+exact mixed-precision artifact at `presence_penalty=1.5`.
+
+#### GPQA Diamond
+
+| Model | Accuracy |
+|---|---:|
+| Qwen3.6-35B-A3B official model card | 86.0% |
+| FP8 reference, our matched harness | 167 / 198 (84.34%) |
+| **AntiLoop FP8, our matched harness** | **166 / 198 (83.84%)** |
+
+The official-model-card number is included for context and was not produced by
+our harness. 
+
+Our GPQA run used thinking mode, paired per-question seeds,
+`temperature=0.7`, `top_p=0.95`, `top_k=20`, MTP3 speculative decoding, a
+65,536-token reasoning budget, and 4,096 tokens of answer headroom. The difference was not significant.
+
+Source for the published 86.0% result:
+[`Qwen/Qwen3.6-35B-A3B` model card](https://huggingface.co/Qwen/Qwen3.6-35B-A3B).
+
+#### GSM8K
+
+| Model | Accuracy |
+|---|---:|
+| FP8 reference | 1273 / 1319 (96.51%) |
+| **AntiLoop FP8** | **1270 / 1319 (96.29%)** |
+
+The GSM8K run used the exact 1,319-example `openai/gsm8k` `main` test split,
+thinking mode, paired seeds, `temperature=0.7`, `top_p=0.95`, `top_k=20`, MTP3,
+an 8,192-token reasoning budget and 1,024 tokens of answer headroom. The difference was not significant.
+
+Taken together, the matched GPQA and GSM8K results show no material or
+statistically detectable capability loss at these sample sizes. They do not
+establish equivalence across other tasks, modalities, or sampling settings.
+
+## Original NVIDIA ModelOpt usage
+
+Use a recent vLLM build with ModelOpt mixed-precision support:
+
+```bash
+vllm serve N8Programs/Qwen3.6-35B-A3B-AntiLoop-NVFP4 \
+  --quantization modelopt \
+  --trust-remote-code \
+  --max-model-len 262144 \
+  --kv-cache-dtype fp8 \
+  --reasoning-parser qwen3
+```
+
+MTP speculative decoding can be enabled on a compatible build with:
+
+```bash
+--speculative-config '{"method":"mtp","num_speculative_tokens":3}'
+```
+
+For the measured LoopHard setting, send the following sampling parameters:
+
+```json
+{
+  "temperature": 0.7,
+  "top_p": 0.95,
+  "top_k": 20,
+  "presence_penalty": 1.5
+}
+```
+
+The capability-preservation results above used the default presence penalty;
+`presence_penalty=1.5` has not yet been evaluated on GPQA or GSM8K.
+
+Follow the
+[`nvidia/Qwen3.6-35B-A3B-NVFP4` model card](https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4)
+for deployment requirements and the
+[`Qwen/Qwen3.6-35B-A3B` model card](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
+for chat templating, thinking controls, multimodal inputs, and base-model
+limitations.
+
+## Limitations
+
+- This is a narrow behavioral fine-tune, not a general alignment or safety model.
+- The MLX conversion does not include an MTP speculative drafter.
+- LoopHard is a task-specific, judge-based benchmark; its loop rate should not
+  be interpreted as a general safety, truthfulness, or factuality score.
+- The GPQA and GSM8K checks used runtime LoRA on an FP8 base, not this exact
+  mixed-precision artifact.
+- Capability preservation has not been tested at `presence_penalty=1.5`.
+- Fixed-scale FP8 re-quantization approximates the exact BF16 LoRA merge; small
+  adapter updates can round away or clip at the original E4M3 range.
+- Runtime validation used a 65,536-token configured context, not the full native
+  262,144-token context.
+- Multimodal generation quality has not been evaluated on this artifact.
+- Outputs may still be incorrect, overconfident, repetitive, biased, toxic, or
+  unsafe.
+
+## License
+
+Apache 2.0, following both the underlying Qwen checkpoint and NVIDIA's
+quantized derivative. See the
+[pinned Qwen license](https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/995ad96eacd98c81ed38be0c5b274b04031597b0/LICENSE),
+the [Qwen model card](https://huggingface.co/Qwen/Qwen3.6-35B-A3B), and the
+[NVIDIA ModelOpt checkpoint card](https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4).
+
+(co-written with GPT-5.6-Sol)

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if (preserve_thinking is defined and preserve_thinking is true) or (loop.index0 > ns.last_query_index) %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,1880 @@
+{
+  "architectures": [
+    "Qwen3_5MoeForConditionalGeneration"
+  ],
+  "dtype": "bfloat16",
+  "image_token_id": 248056,
+  "mlx_hybrid_format": {
+    "activations": "model_dtype_weight_only_quantized_matmul",
+    "format": "modelopt_fp8_nvfp4_v1",
+    "fp8_storage": "mxfp8_carrier_with_unit_e8m0_scales",
+    "nvfp4_storage": "native_e2m1_e4m3_with_output_tensor_scale",
+    "source_activation_scales_retained": false,
+    "source_quantization_config": {
+      "config_groups": {
+        "group_0": {
+          "input_activations": {
+            "dynamic": false,
+            "num_bits": 8,
+            "type": "float"
+          },
+          "targets": [
+            "model.language_model.layers.0.linear_attn.in_proj_qkv",
+            "model.language_model.layers.0.linear_attn.in_proj_z",
+            "model.language_model.layers.0.linear_attn.out_proj",
+            "model.language_model.layers.1.linear_attn.in_proj_qkv",
+            "model.language_model.layers.1.linear_attn.in_proj_z",
+            "model.language_model.layers.1.linear_attn.out_proj",
+            "model.language_model.layers.10.linear_attn.in_proj_qkv",
+            "model.language_model.layers.10.linear_attn.in_proj_z",
+            "model.language_model.layers.10.linear_attn.out_proj",
+            "model.language_model.layers.11.self_attn.k_proj",
+            "model.language_model.layers.11.self_attn.o_proj",
+            "model.language_model.layers.11.self_attn.q_proj",
+            "model.language_model.layers.11.self_attn.v_proj",
+            "model.language_model.layers.12.linear_attn.in_proj_qkv",
+            "model.language_model.layers.12.linear_attn.in_proj_z",
+            "model.language_model.layers.12.linear_attn.out_proj",
+            "model.language_model.layers.13.linear_attn.in_proj_qkv",
+            "model.language_model.layers.13.linear_attn.in_proj_z",
+            "model.language_model.layers.13.linear_attn.out_proj",
+            "model.language_model.layers.14.linear_attn.in_proj_qkv",
+            "model.language_model.layers.14.linear_attn.in_proj_z",
+            "model.language_model.layers.14.linear_attn.out_proj",
+            "model.language_model.layers.15.self_attn.k_proj",
+            "model.language_model.layers.15.self_attn.o_proj",
+            "model.language_model.layers.15.self_attn.q_proj",
+            "model.language_model.layers.15.self_attn.v_proj",
+            "model.language_model.layers.16.linear_attn.in_proj_qkv",
+            "model.language_model.layers.16.linear_attn.in_proj_z",
+            "model.language_model.layers.16.linear_attn.out_proj",
+            "model.language_model.layers.17.linear_attn.in_proj_qkv",
+            "model.language_model.layers.17.linear_attn.in_proj_z",
+            "model.language_model.layers.17.linear_attn.out_proj",
+            "model.language_model.layers.18.linear_attn.in_proj_qkv",
+            "model.language_model.layers.18.linear_attn.in_proj_z",
+            "model.language_model.layers.18.linear_attn.out_proj",
+            "model.language_model.layers.19.self_attn.k_proj",
+            "model.language_model.layers.19.self_attn.o_proj",
+            "model.language_model.layers.19.self_attn.q_proj",
+            "model.language_model.layers.19.self_attn.v_proj",
+            "model.language_model.layers.2.linear_attn.in_proj_qkv",
+            "model.language_model.layers.2.linear_attn.in_proj_z",
+            "model.language_model.layers.2.linear_attn.out_proj",
+            "model.language_model.layers.20.linear_attn.in_proj_qkv",
+            "model.language_model.layers.20.linear_attn.in_proj_z",
+            "model.language_model.layers.20.linear_attn.out_proj",
+            "model.language_model.layers.21.linear_attn.in_proj_qkv",
+            "model.language_model.layers.21.linear_attn.in_proj_z",
+            "model.language_model.layers.21.linear_attn.out_proj",
+            "model.language_model.layers.22.linear_attn.in_proj_qkv",
+            "model.language_model.layers.22.linear_attn.in_proj_z",
+            "model.language_model.layers.22.linear_attn.out_proj",
+            "model.language_model.layers.23.self_attn.k_proj",
+            "model.language_model.layers.23.self_attn.o_proj",
+            "model.language_model.layers.23.self_attn.q_proj",
+            "model.language_model.layers.23.self_attn.v_proj",
+            "model.language_model.layers.24.linear_attn.in_proj_qkv",
+            "model.language_model.layers.24.linear_attn.in_proj_z",
+            "model.language_model.layers.24.linear_attn.out_proj",
+            "model.language_model.layers.25.linear_attn.in_proj_qkv",
+            "model.language_model.layers.25.linear_attn.in_proj_z",
+            "model.language_model.layers.25.linear_attn.out_proj",
+            "model.language_model.layers.26.linear_attn.in_proj_qkv",
+            "model.language_model.layers.26.linear_attn.in_proj_z",
+            "model.language_model.layers.26.linear_attn.out_proj",
+            "model.language_model.layers.27.self_attn.k_proj",
+            "model.language_model.layers.27.self_attn.o_proj",
+            "model.language_model.layers.27.self_attn.q_proj",
+            "model.language_model.layers.27.self_attn.v_proj",
+            "model.language_model.layers.28.linear_attn.in_proj_qkv",
+            "model.language_model.layers.28.linear_attn.in_proj_z",
+            "model.language_model.layers.28.linear_attn.out_proj",
+            "model.language_model.layers.29.linear_attn.in_proj_qkv",
+            "model.language_model.layers.29.linear_attn.in_proj_z",
+            "model.language_model.layers.29.linear_attn.out_proj",
+            "model.language_model.layers.3.self_attn.k_proj",
+            "model.language_model.layers.3.self_attn.o_proj",
+            "model.language_model.layers.3.self_attn.q_proj",
+            "model.language_model.layers.3.self_attn.v_proj",
+            "model.language_model.layers.30.linear_attn.in_proj_qkv",
+            "model.language_model.layers.30.linear_attn.in_proj_z",
+            "model.language_model.layers.30.linear_attn.out_proj",
+            "model.language_model.layers.31.self_attn.k_proj",
+            "model.language_model.layers.31.self_attn.o_proj",
+            "model.language_model.layers.31.self_attn.q_proj",
+            "model.language_model.layers.31.self_attn.v_proj",
+            "model.language_model.layers.32.linear_attn.in_proj_qkv",
+            "model.language_model.layers.32.linear_attn.in_proj_z",
+            "model.language_model.layers.32.linear_attn.out_proj",
+            "model.language_model.layers.33.linear_attn.in_proj_qkv",
+            "model.language_model.layers.33.linear_attn.in_proj_z",
+            "model.language_model.layers.33.linear_attn.out_proj",
+            "model.language_model.layers.34.linear_attn.in_proj_qkv",
+            "model.language_model.layers.34.linear_attn.in_proj_z",
+            "model.language_model.layers.34.linear_attn.out_proj",
+            "model.language_model.layers.35.self_attn.k_proj",
+            "model.language_model.layers.35.self_attn.o_proj",
+            "model.language_model.layers.35.self_attn.q_proj",
+            "model.language_model.layers.35.self_attn.v_proj",
+            "model.language_model.layers.36.linear_attn.in_proj_qkv",
+            "model.language_model.layers.36.linear_attn.in_proj_z",
+            "model.language_model.layers.36.linear_attn.out_proj",
+            "model.language_model.layers.37.linear_attn.in_proj_qkv",
+            "model.language_model.layers.37.linear_attn.in_proj_z",
+            "model.language_model.layers.37.linear_attn.out_proj",
+            "model.language_model.layers.38.linear_attn.in_proj_qkv",
+            "model.language_model.layers.38.linear_attn.in_proj_z",
+            "model.language_model.layers.38.linear_attn.out_proj",
+            "model.language_model.layers.39.self_attn.k_proj",
+            "model.language_model.layers.39.self_attn.o_proj",
+            "model.language_model.layers.39.self_attn.q_proj",
+            "model.language_model.layers.39.self_attn.v_proj",
+            "model.language_model.layers.4.linear_attn.in_proj_qkv",
+            "model.language_model.layers.4.linear_attn.in_proj_z",
+            "model.language_model.layers.4.linear_attn.out_proj",
+            "model.language_model.layers.5.linear_attn.in_proj_qkv",
+            "model.language_model.layers.5.linear_attn.in_proj_z",
+            "model.language_model.layers.5.linear_attn.out_proj",
+            "model.language_model.layers.6.linear_attn.in_proj_qkv",
+            "model.language_model.layers.6.linear_attn.in_proj_z",
+            "model.language_model.layers.6.linear_attn.out_proj",
+            "model.language_model.layers.7.self_attn.k_proj",
+            "model.language_model.layers.7.self_attn.o_proj",
+            "model.language_model.layers.7.self_attn.q_proj",
+            "model.language_model.layers.7.self_attn.v_proj",
+            "model.language_model.layers.8.linear_attn.in_proj_qkv",
+            "model.language_model.layers.8.linear_attn.in_proj_z",
+            "model.language_model.layers.8.linear_attn.out_proj",
+            "model.language_model.layers.9.linear_attn.in_proj_qkv",
+            "model.language_model.layers.9.linear_attn.in_proj_z",
+            "model.language_model.layers.9.linear_attn.out_proj"
+          ],
+          "weights": {
+            "dynamic": false,
+            "num_bits": 8,
+            "type": "float"
+          }
+        },
+        "group_1": {
+          "input_activations": {
+            "dynamic": false,
+            "group_size": 16,
+            "num_bits": 4,
+            "type": "float"
+          },
+          "targets": [
+            "lm_head",
+            "model.language_model.layers.0.mlp.experts",
+            "model.language_model.layers.0.mlp.shared_expert.down_proj",
+            "model.language_model.layers.0.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.0.mlp.shared_expert.up_proj",
+            "model.language_model.layers.1.mlp.experts",
+            "model.language_model.layers.1.mlp.shared_expert.down_proj",
+            "model.language_model.layers.1.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.1.mlp.shared_expert.up_proj",
+            "model.language_model.layers.10.mlp.experts",
+            "model.language_model.layers.10.mlp.shared_expert.down_proj",
+            "model.language_model.layers.10.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.10.mlp.shared_expert.up_proj",
+            "model.language_model.layers.11.mlp.experts",
+            "model.language_model.layers.11.mlp.shared_expert.down_proj",
+            "model.language_model.layers.11.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.11.mlp.shared_expert.up_proj",
+            "model.language_model.layers.12.mlp.experts",
+            "model.language_model.layers.12.mlp.shared_expert.down_proj",
+            "model.language_model.layers.12.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.12.mlp.shared_expert.up_proj",
+            "model.language_model.layers.13.mlp.experts",
+            "model.language_model.layers.13.mlp.shared_expert.down_proj",
+            "model.language_model.layers.13.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.13.mlp.shared_expert.up_proj",
+            "model.language_model.layers.14.mlp.experts",
+            "model.language_model.layers.14.mlp.shared_expert.down_proj",
+            "model.language_model.layers.14.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.14.mlp.shared_expert.up_proj",
+            "model.language_model.layers.15.mlp.experts",
+            "model.language_model.layers.15.mlp.shared_expert.down_proj",
+            "model.language_model.layers.15.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.15.mlp.shared_expert.up_proj",
+            "model.language_model.layers.16.mlp.experts",
+            "model.language_model.layers.16.mlp.shared_expert.down_proj",
+            "model.language_model.layers.16.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.16.mlp.shared_expert.up_proj",
+            "model.language_model.layers.17.mlp.experts",
+            "model.language_model.layers.17.mlp.shared_expert.down_proj",
+            "model.language_model.layers.17.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.17.mlp.shared_expert.up_proj",
+            "model.language_model.layers.18.mlp.experts",
+            "model.language_model.layers.18.mlp.shared_expert.down_proj",
+            "model.language_model.layers.18.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.18.mlp.shared_expert.up_proj",
+            "model.language_model.layers.19.mlp.experts",
+            "model.language_model.layers.19.mlp.shared_expert.down_proj",
+            "model.language_model.layers.19.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.19.mlp.shared_expert.up_proj",
+            "model.language_model.layers.2.mlp.experts",
+            "model.language_model.layers.2.mlp.shared_expert.down_proj",
+            "model.language_model.layers.2.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.2.mlp.shared_expert.up_proj",
+            "model.language_model.layers.20.mlp.experts",
+            "model.language_model.layers.20.mlp.shared_expert.down_proj",
+            "model.language_model.layers.20.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.20.mlp.shared_expert.up_proj",
+            "model.language_model.layers.21.mlp.experts",
+            "model.language_model.layers.21.mlp.shared_expert.down_proj",
+            "model.language_model.layers.21.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.21.mlp.shared_expert.up_proj",
+            "model.language_model.layers.22.mlp.experts",
+            "model.language_model.layers.22.mlp.shared_expert.down_proj",
+            "model.language_model.layers.22.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.22.mlp.shared_expert.up_proj",
+            "model.language_model.layers.23.mlp.experts",
+            "model.language_model.layers.23.mlp.shared_expert.down_proj",
+            "model.language_model.layers.23.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.23.mlp.shared_expert.up_proj",
+            "model.language_model.layers.24.mlp.experts",
+            "model.language_model.layers.24.mlp.shared_expert.down_proj",
+            "model.language_model.layers.24.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.24.mlp.shared_expert.up_proj",
+            "model.language_model.layers.25.mlp.experts",
+            "model.language_model.layers.25.mlp.shared_expert.down_proj",
+            "model.language_model.layers.25.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.25.mlp.shared_expert.up_proj",
+            "model.language_model.layers.26.mlp.experts",
+            "model.language_model.layers.26.mlp.shared_expert.down_proj",
+            "model.language_model.layers.26.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.26.mlp.shared_expert.up_proj",
+            "model.language_model.layers.27.mlp.experts",
+            "model.language_model.layers.27.mlp.shared_expert.down_proj",
+            "model.language_model.layers.27.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.27.mlp.shared_expert.up_proj",
+            "model.language_model.layers.28.mlp.experts",
+            "model.language_model.layers.28.mlp.shared_expert.down_proj",
+            "model.language_model.layers.28.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.28.mlp.shared_expert.up_proj",
+            "model.language_model.layers.29.mlp.experts",
+            "model.language_model.layers.29.mlp.shared_expert.down_proj",
+            "model.language_model.layers.29.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.29.mlp.shared_expert.up_proj",
+            "model.language_model.layers.3.mlp.experts",
+            "model.language_model.layers.3.mlp.shared_expert.down_proj",
+            "model.language_model.layers.3.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.3.mlp.shared_expert.up_proj",
+            "model.language_model.layers.30.mlp.experts",
+            "model.language_model.layers.30.mlp.shared_expert.down_proj",
+            "model.language_model.layers.30.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.30.mlp.shared_expert.up_proj",
+            "model.language_model.layers.31.mlp.experts",
+            "model.language_model.layers.31.mlp.shared_expert.down_proj",
+            "model.language_model.layers.31.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.31.mlp.shared_expert.up_proj",
+            "model.language_model.layers.32.mlp.experts",
+            "model.language_model.layers.32.mlp.shared_expert.down_proj",
+            "model.language_model.layers.32.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.32.mlp.shared_expert.up_proj",
+            "model.language_model.layers.33.mlp.experts",
+            "model.language_model.layers.33.mlp.shared_expert.down_proj",
+            "model.language_model.layers.33.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.33.mlp.shared_expert.up_proj",
+            "model.language_model.layers.34.mlp.experts",
+            "model.language_model.layers.34.mlp.shared_expert.down_proj",
+            "model.language_model.layers.34.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.34.mlp.shared_expert.up_proj",
+            "model.language_model.layers.35.mlp.experts",
+            "model.language_model.layers.35.mlp.shared_expert.down_proj",
+            "model.language_model.layers.35.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.35.mlp.shared_expert.up_proj",
+            "model.language_model.layers.36.mlp.experts",
+            "model.language_model.layers.36.mlp.shared_expert.down_proj",
+            "model.language_model.layers.36.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.36.mlp.shared_expert.up_proj",
+            "model.language_model.layers.37.mlp.experts",
+            "model.language_model.layers.37.mlp.shared_expert.down_proj",
+            "model.language_model.layers.37.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.37.mlp.shared_expert.up_proj",
+            "model.language_model.layers.38.mlp.experts",
+            "model.language_model.layers.38.mlp.shared_expert.down_proj",
+            "model.language_model.layers.38.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.38.mlp.shared_expert.up_proj",
+            "model.language_model.layers.39.mlp.experts",
+            "model.language_model.layers.39.mlp.shared_expert.down_proj",
+            "model.language_model.layers.39.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.39.mlp.shared_expert.up_proj",
+            "model.language_model.layers.4.mlp.experts",
+            "model.language_model.layers.4.mlp.shared_expert.down_proj",
+            "model.language_model.layers.4.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.4.mlp.shared_expert.up_proj",
+            "model.language_model.layers.5.mlp.experts",
+            "model.language_model.layers.5.mlp.shared_expert.down_proj",
+            "model.language_model.layers.5.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.5.mlp.shared_expert.up_proj",
+            "model.language_model.layers.6.mlp.experts",
+            "model.language_model.layers.6.mlp.shared_expert.down_proj",
+            "model.language_model.layers.6.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.6.mlp.shared_expert.up_proj",
+            "model.language_model.layers.7.mlp.experts",
+            "model.language_model.layers.7.mlp.shared_expert.down_proj",
+            "model.language_model.layers.7.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.7.mlp.shared_expert.up_proj",
+            "model.language_model.layers.8.mlp.experts",
+            "model.language_model.layers.8.mlp.shared_expert.down_proj",
+            "model.language_model.layers.8.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.8.mlp.shared_expert.up_proj",
+            "model.language_model.layers.9.mlp.experts",
+            "model.language_model.layers.9.mlp.shared_expert.down_proj",
+            "model.language_model.layers.9.mlp.shared_expert.gate_proj",
+            "model.language_model.layers.9.mlp.shared_expert.up_proj"
+          ],
+          "weights": {
+            "dynamic": false,
+            "group_size": 16,
+            "num_bits": 4,
+            "type": "float"
+          }
+        }
+      },
+      "ignore": [
+        "mtp.layers.0*",
+        "mtp*"
+      ],
+      "producer": {
+        "name": "modelopt",
+        "version": "0.37.0"
+      },
+      "quant_algo": "MIXED_PRECISION",
+      "quant_method": "modelopt",
+      "quantized_layers": {
+        "lm_head": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.0.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.0.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.0.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.0.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.0.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.0.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.0.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.1.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.1.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.1.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.1.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.1.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.1.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.1.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.10.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.10.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.10.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.10.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.10.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.10.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.10.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.11.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.11.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.11.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.11.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.11.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.11.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.11.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.11.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.12.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.12.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.12.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.12.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.12.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.12.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.12.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.13.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.13.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.13.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.13.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.13.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.13.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.13.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.14.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.14.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.14.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.14.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.14.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.14.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.14.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.15.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.15.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.15.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.15.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.15.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.15.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.15.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.15.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.16.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.16.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.16.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.16.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.16.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.16.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.16.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.17.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.17.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.17.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.17.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.17.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.17.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.17.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.18.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.18.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.18.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.18.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.18.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.18.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.18.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.19.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.19.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.19.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.19.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.19.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.19.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.19.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.19.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.2.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.2.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.2.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.2.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.2.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.2.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.2.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.20.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.20.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.20.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.20.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.20.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.20.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.20.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.21.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.21.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.21.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.21.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.21.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.21.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.21.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.22.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.22.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.22.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.22.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.22.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.22.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.22.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.23.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.23.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.23.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.23.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.23.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.23.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.23.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.23.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.24.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.24.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.24.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.24.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.24.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.24.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.24.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.25.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.25.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.25.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.25.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.25.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.25.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.25.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.26.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.26.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.26.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.26.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.26.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.26.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.26.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.27.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.27.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.27.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.27.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.27.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.27.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.27.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.27.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.28.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.28.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.28.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.28.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.28.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.28.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.28.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.29.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.29.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.29.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.29.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.29.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.29.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.29.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.3.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.3.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.3.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.3.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.3.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.3.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.3.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.3.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.30.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.30.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.30.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.30.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.30.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.30.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.30.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.31.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.31.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.31.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.31.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.31.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.31.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.31.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.31.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.32.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.32.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.32.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.32.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.32.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.32.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.32.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.33.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.33.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.33.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.33.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.33.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.33.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.33.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.34.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.34.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.34.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.34.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.34.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.34.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.34.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.35.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.35.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.35.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.35.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.35.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.35.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.35.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.35.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.36.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.36.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.36.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.36.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.36.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.36.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.36.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.37.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.37.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.37.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.37.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.37.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.37.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.37.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.38.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.38.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.38.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.38.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.38.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.38.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.38.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.39.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.39.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.39.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.39.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.39.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.39.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.39.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.39.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.4.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.4.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.4.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.4.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.4.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.4.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.4.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.5.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.5.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.5.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.5.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.5.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.5.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.5.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.6.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.6.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.6.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.6.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.6.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.6.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.6.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.7.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.7.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.7.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.7.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.7.self_attn.k_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.7.self_attn.o_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.7.self_attn.q_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.7.self_attn.v_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.8.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.8.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.8.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.8.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.8.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.8.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.8.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.9.linear_attn.in_proj_qkv": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.9.linear_attn.in_proj_z": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.9.linear_attn.out_proj": {
+          "quant_algo": "FP8"
+        },
+        "model.language_model.layers.9.mlp.experts": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.9.mlp.shared_expert.down_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.9.mlp.shared_expert.gate_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        },
+        "model.language_model.layers.9.mlp.shared_expert.up_proj": {
+          "group_size": 16,
+          "quant_algo": "W4A16_NVFP4"
+        }
+      }
+    },
+    "vision": {
+      "storage": "bf16_exact_source_weights",
+      "tensor_count": 333,
+      "tensor_data_bytes": 893142496,
+      "weight_bytes_requantized": false
+    }
+  },
+  "mlx_modelopt_quantization": {
+    "language_model.lm_head": "scaled_nvfp4",
+    "language_model.model.layers.0.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.0.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.0.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.0.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.0.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.0.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.0.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.0.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.0.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.1.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.1.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.1.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.1.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.1.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.1.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.1.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.1.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.1.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.10.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.10.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.10.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.10.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.10.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.10.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.10.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.10.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.10.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.11.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.11.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.11.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.11.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.11.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.11.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.11.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.11.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.11.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.11.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.12.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.12.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.12.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.12.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.12.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.12.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.12.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.12.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.12.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.13.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.13.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.13.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.13.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.13.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.13.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.13.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.13.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.13.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.14.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.14.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.14.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.14.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.14.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.14.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.14.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.14.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.14.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.15.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.15.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.15.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.15.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.15.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.15.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.15.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.15.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.15.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.15.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.16.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.16.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.16.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.16.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.16.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.16.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.16.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.16.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.16.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.17.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.17.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.17.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.17.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.17.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.17.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.17.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.17.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.17.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.18.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.18.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.18.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.18.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.18.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.18.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.18.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.18.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.18.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.19.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.19.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.19.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.19.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.19.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.19.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.19.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.19.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.19.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.19.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.2.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.2.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.2.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.2.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.2.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.2.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.2.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.2.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.2.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.20.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.20.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.20.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.20.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.20.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.20.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.20.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.20.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.20.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.21.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.21.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.21.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.21.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.21.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.21.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.21.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.21.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.21.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.22.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.22.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.22.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.22.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.22.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.22.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.22.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.22.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.22.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.23.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.23.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.23.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.23.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.23.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.23.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.23.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.23.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.23.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.23.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.24.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.24.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.24.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.24.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.24.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.24.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.24.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.24.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.24.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.25.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.25.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.25.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.25.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.25.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.25.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.25.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.25.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.25.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.26.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.26.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.26.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.26.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.26.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.26.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.26.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.26.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.26.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.27.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.27.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.27.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.27.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.27.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.27.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.27.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.27.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.27.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.27.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.28.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.28.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.28.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.28.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.28.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.28.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.28.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.28.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.28.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.29.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.29.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.29.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.29.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.29.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.29.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.29.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.29.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.29.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.3.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.3.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.3.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.3.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.3.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.3.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.3.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.3.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.3.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.3.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.30.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.30.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.30.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.30.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.30.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.30.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.30.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.30.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.30.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.31.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.31.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.31.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.31.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.31.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.31.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.31.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.31.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.31.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.31.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.32.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.32.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.32.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.32.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.32.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.32.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.32.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.32.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.32.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.33.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.33.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.33.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.33.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.33.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.33.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.33.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.33.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.33.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.34.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.34.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.34.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.34.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.34.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.34.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.34.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.34.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.34.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.35.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.35.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.35.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.35.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.35.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.35.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.35.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.35.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.35.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.35.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.36.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.36.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.36.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.36.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.36.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.36.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.36.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.36.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.36.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.37.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.37.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.37.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.37.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.37.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.37.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.37.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.37.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.37.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.38.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.38.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.38.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.38.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.38.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.38.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.38.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.38.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.38.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.39.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.39.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.39.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.39.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.39.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.39.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.39.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.39.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.39.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.39.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.4.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.4.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.4.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.4.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.4.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.4.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.4.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.4.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.4.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.5.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.5.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.5.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.5.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.5.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.5.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.5.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.5.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.5.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.6.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.6.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.6.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.6.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.6.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.6.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.6.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.6.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.6.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.7.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.7.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.7.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.7.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.7.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.7.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.7.self_attn.k_proj": "scaled_mxfp8",
+    "language_model.model.layers.7.self_attn.o_proj": "scaled_mxfp8",
+    "language_model.model.layers.7.self_attn.q_proj": "scaled_mxfp8",
+    "language_model.model.layers.7.self_attn.v_proj": "scaled_mxfp8",
+    "language_model.model.layers.8.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.8.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.8.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.8.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.8.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.8.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.8.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.8.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.8.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.9.linear_attn.in_proj_qkv": "scaled_mxfp8",
+    "language_model.model.layers.9.linear_attn.in_proj_z": "scaled_mxfp8",
+    "language_model.model.layers.9.linear_attn.out_proj": "scaled_mxfp8",
+    "language_model.model.layers.9.mlp.shared_expert.down_proj": "scaled_nvfp4",
+    "language_model.model.layers.9.mlp.shared_expert.gate_proj": "scaled_nvfp4",
+    "language_model.model.layers.9.mlp.shared_expert.up_proj": "scaled_nvfp4",
+    "language_model.model.layers.9.mlp.switch_mlp.down_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.9.mlp.switch_mlp.gate_proj": "scaled_nvfp4_switch",
+    "language_model.model.layers.9.mlp.switch_mlp.up_proj": "scaled_nvfp4_switch"
+  },
+  "model_file": "modeling_mlx_qwen36_modelopt_hybrid.py",
+  "model_type": "qwen3_5_moe",
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": 248044,
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 32,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "model_type": "qwen3_5_moe_text",
+    "moe_intermediate_size": 512,
+    "mtp_num_hidden_layers": 1,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 16,
+    "num_experts": 256,
+    "num_experts_per_tok": 8,
+    "num_hidden_layers": 40,
+    "num_key_value_heads": 2,
+    "output_router_logits": false,
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "router_aux_loss_coef": 0.001,
+    "shared_expert_intermediate_size": 512,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 248320
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.7.0.dev0",
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 27,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4304,
+    "model_type": "qwen3_5_moe_vision",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 2048,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053,
+  "vlm_model_file": "modeling_mlx_vlm_qwen36_modelopt_hybrid.py"
+}

convert_qwen36_modelopt_hybrid_to_mlx.py

@@ -0,0 +1,568 @@
+#!/usr/bin/env python3
+"""Stream a Qwen3.6 ModelOpt FP8/NVFP4 checkpoint into MLX format.
+
+This is intentionally a format conversion, not a re-quantization:
+
+* FP8 E4M3 weight bytes are packed four-at-a-time into MLX uint32 tensors.
+  Unit E8M0 block scales make MLX's MXFP8 kernel decode the original FP8
+  values exactly; the original ModelOpt tensor scale is retained separately.
+* NVFP4 E2M1 weight nibbles and E4M3 block-scale bytes are repacked without
+  numerical modification.  The original FP32 tensor scales are retained.
+* Expert tensors are stacked into MLX's SwitchGLU layout one layer at a time.
+* The original BF16 vision tower is preserved unchanged in a dedicated shard
+  and loaded by the model-local MLX-VLM runtime.
+* ModelOpt activation scales are recorded as dropped because this runtime uses
+  weight-only quantized kernels and keeps activations in the model dtype.
+
+Each transformer layer is written as its own safetensors shard, bounding peak
+memory to roughly one layer rather than the whole model.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import resource
+import shutil
+import sys
+from collections import Counter, defaultdict
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Dict, Iterable, Mapping
+
+import mlx.core as mx
+
+
+RUNTIME_FILE = "modeling_mlx_qwen36_modelopt_hybrid.py"
+VLM_RUNTIME_FILE = "modeling_mlx_vlm_qwen36_modelopt_hybrid.py"
+EXPERT_RE = re.compile(
+    r"^model\.language_model\.layers\.(\d+)\.mlp\.experts\.(\d+)\."
+    r"(gate_proj|up_proj|down_proj)\."
+    r"(weight|weight_scale|weight_scale_2|input_scale)$"
+)
+LAYER_RE = re.compile(r"^model\.language_model\.layers\.(\d+)\.")
+NORM_SUFFIXES = (
+    ".input_layernorm.weight",
+    ".post_attention_layernorm.weight",
+    "model.norm.weight",
+    ".q_norm.weight",
+    ".k_norm.weight",
+)
+
+
+def log(message: str) -> None:
+    now = datetime.now().strftime("%H:%M:%S")
+    rss = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / (1024**3)
+    print(f"[{now}] [peak RSS {rss:.2f} GiB] {message}", flush=True)
+
+
+def pack_u8x4(x: mx.array) -> mx.array:
+    """Pack four consecutive bytes into one little-endian uint32."""
+    if x.dtype != mx.uint8:
+        raise TypeError(f"Expected uint8 storage, got {x.dtype}")
+    if x.shape[-1] % 4:
+        raise ValueError(f"Last dimension {x.shape[-1]} is not divisible by four")
+    y = x.reshape(*x.shape[:-1], x.shape[-1] // 4, 4).astype(mx.uint32)
+    return y[..., 0] | (y[..., 1] << 8) | (y[..., 2] << 16) | (y[..., 3] << 24)
+
+
+def sanitize_name(key: str) -> str:
+    prefix = "model.language_model."
+    if key.startswith(prefix):
+        return "language_model.model." + key[len(prefix) :]
+    if key.startswith("language_model."):
+        return key
+    return "language_model." + key
+
+
+def should_drop(key: str) -> bool:
+    return (
+        key.startswith("model.visual")
+        or key.startswith("vision_tower")
+        or key.startswith("mtp.")
+        or ".mtp." in key
+    )
+
+
+class SourceWeights:
+    def __init__(self, root: Path, weight_map: Mapping[str, str]):
+        self.root = root
+        self.weight_map = weight_map
+        self._shards: Dict[str, Dict[str, mx.array]] = {}
+
+    def get(self, key: str) -> mx.array:
+        shard_name = self.weight_map[key]
+        if shard_name not in self._shards:
+            log(f"Opening source shard {shard_name} lazily")
+            self._shards[shard_name] = mx.load(str(self.root / shard_name))
+        return self._shards[shard_name][key]
+
+
+def group_id_for_key(key: str) -> int | None:
+    match = LAYER_RE.match(key)
+    return int(match.group(1)) if match else None
+
+
+def group_id_for_prefix(prefix: str) -> int | None:
+    return group_id_for_key(prefix + ".weight")
+
+
+def add_output(
+    output: Dict[str, mx.array],
+    key: str,
+    value: mx.array,
+) -> None:
+    if key in output:
+        raise KeyError(f"Duplicate output tensor {key}")
+    output[key] = value
+
+
+def convert_dense_nvfp4(
+    source: SourceWeights,
+    prefix: str,
+    output: Dict[str, mx.array],
+    quantization: Dict[str, str],
+) -> None:
+    raw = source.get(prefix + ".weight")
+    scales = source.get(prefix + ".weight_scale")
+    tensor_scale = source.get(prefix + ".weight_scale_2")
+    if raw.dtype != mx.uint8 or scales.dtype != mx.uint8:
+        raise TypeError(
+            f"Unexpected NVFP4 storage for {prefix}: {raw.dtype}, {scales.dtype}"
+        )
+    if raw.shape[-1] != scales.shape[-1] * 8:
+        raise ValueError(
+            f"NVFP4 shape mismatch for {prefix}: weight={raw.shape}, scales={scales.shape}"
+        )
+    out_prefix = sanitize_name(prefix)
+    add_output(output, out_prefix + ".weight", pack_u8x4(raw))
+    add_output(output, out_prefix + ".scales", scales)
+    add_output(output, out_prefix + ".global_scale", tensor_scale)
+    quantization[out_prefix] = "scaled_nvfp4"
+
+
+def convert_dense_fp8(
+    source: SourceWeights,
+    prefix: str,
+    output: Dict[str, mx.array],
+    quantization: Dict[str, str],
+) -> None:
+    raw = source.get(prefix + ".weight")
+    tensor_scale = source.get(prefix + ".weight_scale")
+    if raw.dtype != mx.uint8:
+        raise TypeError(f"Unexpected FP8 storage for {prefix}: {raw.dtype}")
+    if raw.shape[-1] % 32:
+        raise ValueError(
+            f"FP8 input dimension for {prefix} is not divisible by 32: {raw.shape}"
+        )
+    out_prefix = sanitize_name(prefix)
+    add_output(output, out_prefix + ".weight", pack_u8x4(raw))
+    # E8M0 byte 0x7f represents exactly 1.0.  With unit scales, MLX's
+    # MXFP8 decoder reproduces ModelOpt's E4M3 weight bytes losslessly.
+    scale_shape = (*raw.shape[:-1], raw.shape[-1] // 32)
+    add_output(
+        output,
+        out_prefix + ".scales",
+        mx.full(scale_shape, 0x7F, dtype=mx.uint8),
+    )
+    add_output(output, out_prefix + ".global_scale", tensor_scale)
+    quantization[out_prefix] = "scaled_mxfp8"
+
+
+def convert_expert_projection(
+    source: SourceWeights,
+    layer: int,
+    projection: str,
+    expert_lookup: Mapping[tuple[int, str, str], Mapping[int, str]],
+    num_experts: int,
+    output: Dict[str, mx.array],
+    quantization: Dict[str, str],
+) -> None:
+    expected = list(range(num_experts))
+    suffix_maps = {
+        suffix: expert_lookup[(layer, projection, suffix)]
+        for suffix in ("weight", "weight_scale", "weight_scale_2")
+    }
+    for suffix, mapping in suffix_maps.items():
+        present = sorted(mapping)
+        if present != expected:
+            missing = sorted(set(expected) - set(present))
+            raise ValueError(
+                f"Layer {layer} {projection} {suffix}: expected {num_experts} "
+                f"experts, missing {missing[:20]}"
+            )
+
+    raw = mx.stack(
+        [source.get(suffix_maps["weight"][expert]) for expert in expected], axis=0
+    )
+    scales = mx.stack(
+        [source.get(suffix_maps["weight_scale"][expert]) for expert in expected],
+        axis=0,
+    )
+    tensor_scales = mx.stack(
+        [source.get(suffix_maps["weight_scale_2"][expert]) for expert in expected],
+        axis=0,
+    )
+    if raw.dtype != mx.uint8 or scales.dtype != mx.uint8:
+        raise TypeError(
+            f"Unexpected expert NVFP4 storage at layer {layer} {projection}: "
+            f"{raw.dtype}, {scales.dtype}"
+        )
+    if raw.shape[-1] != scales.shape[-1] * 8:
+        raise ValueError(
+            f"Expert NVFP4 shape mismatch at layer {layer} {projection}: "
+            f"weight={raw.shape}, scales={scales.shape}"
+        )
+
+    out_prefix = (
+        f"language_model.model.layers.{layer}.mlp.switch_mlp.{projection}"
+    )
+    add_output(output, out_prefix + ".weight", pack_u8x4(raw))
+    add_output(output, out_prefix + ".scales", scales)
+    add_output(output, out_prefix + ".global_scales", tensor_scales)
+    quantization[out_prefix] = "scaled_nvfp4_switch"
+
+
+def transform_standard_weight(
+    key: str,
+    value: mx.array,
+    shift_norm_weights: bool,
+) -> tuple[str, mx.array]:
+    out_key = sanitize_name(key)
+    if "conv1d.weight" in out_key and value.shape[-1] != 1:
+        value = value.moveaxis(2, 1)
+    if (
+        shift_norm_weights
+        and value.ndim == 1
+        and any(out_key.endswith(suffix) for suffix in NORM_SUFFIXES)
+    ):
+        value = value + 1.0
+    return out_key, value
+
+
+def write_shard(
+    partial: Path,
+    filename: str,
+    tensors: Dict[str, mx.array],
+    output_weight_map: Dict[str, str],
+) -> int:
+    tensors = dict(sorted(tensors.items()))
+    total_bytes = sum(array.nbytes for array in tensors.values())
+    log(
+        f"Writing {filename}: {len(tensors)} tensors, "
+        f"{total_bytes / (1024**3):.2f} GiB"
+    )
+    mx.save_safetensors(
+        str(partial / filename),
+        tensors,
+        metadata={"format": "mlx"},
+    )
+    for key in tensors:
+        if key in output_weight_map:
+            raise KeyError(f"Tensor {key} was already assigned to a shard")
+        output_weight_map[key] = filename
+    del tensors
+    try:
+        mx.clear_cache()
+    except AttributeError:
+        # Compatibility with older MLX releases.
+        mx.metal.clear_cache()
+    return total_bytes
+
+
+def copy_metadata(source: Path, partial: Path) -> None:
+    names = [
+        "README.md",
+        "LICENSE",
+        "chat_template.jinja",
+        "generation_config.json",
+        "preprocessor_config.json",
+        "video_preprocessor_config.json",
+        "special_tokens_map.json",
+        "tokenizer.json",
+        "tokenizer.model",
+        "tokenizer_config.json",
+        "vocab.json",
+        "merges.txt",
+    ]
+    for name in names:
+        src = source / name
+        if src.exists():
+            shutil.copy2(src, partial / name, follow_symlinks=True)
+    assets = source / "assets"
+    if assets.exists():
+        shutil.copytree(assets, partial / "assets")
+
+
+def convert(
+    source: Path,
+    output: Path,
+    runtime_source: Path,
+    vlm_runtime_source: Path,
+) -> None:
+    source = source.resolve()
+    output = output.resolve()
+    partial = output.with_name(output.name + ".partial")
+    if output.exists():
+        raise FileExistsError(f"Output already exists: {output}")
+    if partial.exists():
+        raise FileExistsError(
+            f"Partial output already exists: {partial}. Remove it or choose another path."
+        )
+    if not (source / "config.json").exists():
+        raise FileNotFoundError(f"Missing config.json in {source}")
+    if not (source / "model.safetensors.index.json").exists():
+        raise FileNotFoundError(f"Missing model.safetensors.index.json in {source}")
+
+    config = json.loads((source / "config.json").read_text())
+    index = json.loads((source / "model.safetensors.index.json").read_text())
+    weight_map: Dict[str, str] = index["weight_map"]
+    all_keys = sorted(weight_map)
+    vision_keys = [key for key in all_keys if key.startswith("model.visual.")]
+    text_config = config.get("text_config", config)
+    num_layers = int(text_config["num_hidden_layers"])
+    num_experts = int(text_config["num_experts"])
+    log(
+        f"Source {source}: {len(all_keys)} tensors, {num_layers} layers, "
+        f"{num_experts} experts"
+    )
+
+    nvfp4_prefixes = {
+        key[: -len(".weight_scale_2")]
+        for key in all_keys
+        if key.endswith(".weight_scale_2") and not should_drop(key)
+    }
+    all_scale_prefixes = {
+        key[: -len(".weight_scale")]
+        for key in all_keys
+        if key.endswith(".weight_scale") and not should_drop(key)
+    }
+    fp8_prefixes = all_scale_prefixes - nvfp4_prefixes
+
+    expert_lookup: dict[tuple[int, str, str], dict[int, str]] = defaultdict(dict)
+    expert_keys = set()
+    for key in all_keys:
+        match = EXPERT_RE.match(key)
+        if not match:
+            continue
+        layer, expert, projection, suffix = match.groups()
+        expert_lookup[(int(layer), projection, suffix)][int(expert)] = key
+        expert_keys.add(key)
+
+    expert_prefix_marker = ".mlp.experts."
+    dense_nvfp4 = sorted(
+        prefix for prefix in nvfp4_prefixes if expert_prefix_marker not in prefix
+    )
+    dense_fp8 = sorted(
+        prefix for prefix in fp8_prefixes if expert_prefix_marker not in prefix
+    )
+    log(
+        f"Detected {len(dense_fp8)} dense FP8 modules, "
+        f"{len(dense_nvfp4)} dense NVFP4 modules, and "
+        f"{len(expert_keys)} expert component tensors"
+    )
+
+    layer_standard_keys: dict[int, list[str]] = defaultdict(list)
+    global_standard_keys: list[str] = []
+    quantized_prefixes = nvfp4_prefixes | fp8_prefixes
+    quant_metadata_suffixes = (
+        ".input_scale",
+        ".weight_scale",
+        ".weight_scale_2",
+    )
+    for key in all_keys:
+        if should_drop(key) or key in expert_keys:
+            continue
+        if key.endswith(quant_metadata_suffixes):
+            continue
+        if key.endswith(".weight") and key[: -len(".weight")] in quantized_prefixes:
+            continue
+        group = group_id_for_key(key)
+        if group is None:
+            global_standard_keys.append(key)
+        else:
+            layer_standard_keys[group].append(key)
+
+    has_mtp = any(key.startswith("mtp.") or ".mtp." in key for key in all_keys)
+    has_unsanitized_conv = False
+    source_weights = SourceWeights(source, weight_map)
+    for key in all_keys:
+        if "conv1d.weight" in key and not should_drop(key):
+            if source_weights.get(key).shape[-1] != 1:
+                has_unsanitized_conv = True
+                break
+    shift_norm_weights = has_mtp or has_unsanitized_conv
+    log(
+        f"Qwen sanitizer flags: has_mtp={has_mtp}, "
+        f"unsanitized_conv1d={has_unsanitized_conv}, "
+        f"shift_norm_weights={shift_norm_weights}"
+    )
+
+    partial.mkdir(parents=True)
+    copy_metadata(source, partial)
+    shutil.copy2(runtime_source, partial / RUNTIME_FILE)
+    shutil.copy2(vlm_runtime_source, partial / VLM_RUNTIME_FILE)
+
+    quantization: Dict[str, str] = {}
+    output_weight_map: Dict[str, str] = {}
+    total_size = 0
+    shard_count = num_layers + 1 + bool(vision_keys)
+
+    # Global tensors: embeddings, final norm, and LM head.
+    global_output: Dict[str, mx.array] = {}
+    for prefix in dense_fp8:
+        if group_id_for_prefix(prefix) is None:
+            convert_dense_fp8(source_weights, prefix, global_output, quantization)
+    for prefix in dense_nvfp4:
+        if group_id_for_prefix(prefix) is None:
+            convert_dense_nvfp4(source_weights, prefix, global_output, quantization)
+    for key in global_standard_keys:
+        out_key, value = transform_standard_weight(
+            key, source_weights.get(key), shift_norm_weights
+        )
+        add_output(global_output, out_key, value)
+    total_size += write_shard(
+        partial,
+        f"model-{1:05d}-of-{shard_count:05d}.safetensors",
+        global_output,
+        output_weight_map,
+    )
+
+    for layer in range(num_layers):
+        layer_output: Dict[str, mx.array] = {}
+        log(f"Converting transformer layer {layer + 1}/{num_layers}")
+        for prefix in dense_fp8:
+            if group_id_for_prefix(prefix) == layer:
+                convert_dense_fp8(source_weights, prefix, layer_output, quantization)
+        for prefix in dense_nvfp4:
+            if group_id_for_prefix(prefix) == layer:
+                convert_dense_nvfp4(source_weights, prefix, layer_output, quantization)
+        for projection in ("gate_proj", "up_proj", "down_proj"):
+            convert_expert_projection(
+                source_weights,
+                layer,
+                projection,
+                expert_lookup,
+                num_experts,
+                layer_output,
+                quantization,
+            )
+        for key in layer_standard_keys[layer]:
+            out_key, value = transform_standard_weight(
+                key, source_weights.get(key), shift_norm_weights
+            )
+            add_output(layer_output, out_key, value)
+
+        total_size += write_shard(
+            partial,
+            f"model-{layer + 2:05d}-of-{shard_count:05d}.safetensors",
+            layer_output,
+            output_weight_map,
+        )
+        log(f"Finished transformer layer {layer + 1}/{num_layers}")
+
+    # Vision is unchanged by the AntiLoop adapter. Keep the original BF16
+    # tensors under their source names; the MLX-VLM runtime maps and sanitizes
+    # them. Arrays returned by mx.load remain lazy until the safetensors writer
+    # consumes them, so this does not materialize the source shard twice.
+    if vision_keys:
+        vision_output = {key: source_weights.get(key) for key in vision_keys}
+        total_size += write_shard(
+            partial,
+            f"model-{shard_count:05d}-of-{shard_count:05d}.safetensors",
+            vision_output,
+            output_weight_map,
+        )
+        log(f"Preserved {len(vision_keys)} vision tensors without requantization")
+
+    output_index = {
+        "metadata": {"total_size": total_size},
+        "weight_map": dict(sorted(output_weight_map.items())),
+    }
+    (partial / "model.safetensors.index.json").write_text(
+        json.dumps(output_index, indent=2, sort_keys=True) + "\n"
+    )
+
+    original_quantization = config.pop("quantization_config", None)
+    config.pop("quantization", None)
+    if isinstance(config.get("text_config"), dict):
+        config["text_config"].pop("quantization_config", None)
+        config["text_config"].pop("quantization", None)
+    config["model_file"] = RUNTIME_FILE
+    config["vlm_model_file"] = VLM_RUNTIME_FILE
+    config["mlx_modelopt_quantization"] = dict(sorted(quantization.items()))
+    config["mlx_hybrid_format"] = {
+        "format": "modelopt_fp8_nvfp4_v1",
+        "fp8_storage": "mxfp8_carrier_with_unit_e8m0_scales",
+        "nvfp4_storage": "native_e2m1_e4m3_with_output_tensor_scale",
+        "activations": "model_dtype_weight_only_quantized_matmul",
+        "source_activation_scales_retained": False,
+        "source_quantization_config": original_quantization,
+    }
+    (partial / "config.json").write_text(
+        json.dumps(config, indent=2, sort_keys=True) + "\n"
+    )
+
+    input_scales_dropped = sum(key.endswith(".input_scale") for key in all_keys)
+    manifest = {
+        "source": str(source),
+        "output": str(output),
+        "created_at": datetime.now(timezone.utc).isoformat(),
+        "converter": str(Path(__file__).resolve()),
+        "runtime_file": RUNTIME_FILE,
+        "vlm_runtime_file": VLM_RUNTIME_FILE,
+        "num_layers": num_layers,
+        "num_experts": num_experts,
+        "source_tensor_count": len(all_keys),
+        "output_tensor_count": len(output_weight_map),
+        "output_shards": shard_count,
+        "output_total_size_bytes": total_size,
+        "quantized_module_counts": dict(Counter(quantization.values())),
+        "input_scales_dropped": input_scales_dropped,
+        "weight_bytes_requantized": False,
+        "vision_tensor_count": len(vision_keys),
+        "vision_weight_bytes_requantized": False,
+        "norm_weights_shifted": shift_norm_weights,
+    }
+    (partial / "mlx_conversion_manifest.json").write_text(
+        json.dumps(manifest, indent=2, sort_keys=True) + "\n"
+    )
+
+    partial.rename(output)
+    log(
+        f"Conversion complete: {output} ({total_size / (1024**3):.2f} GiB, "
+        f"{len(output_weight_map)} tensors)"
+    )
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--source", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument(
+        "--runtime-source",
+        type=Path,
+        default=Path(__file__).with_name(RUNTIME_FILE),
+    )
+    parser.add_argument(
+        "--vlm-runtime-source",
+        type=Path,
+        default=Path(__file__).with_name(VLM_RUNTIME_FILE),
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    convert(
+        args.source,
+        args.output,
+        args.runtime_source.resolve(),
+        args.vlm_runtime_source.resolve(),
+    )
+
+
+if __name__ == "__main__":
+    main()

generation_config.json

@@ -0,0 +1,12 @@
+{
+    "bos_token_id": 248044,
+    "do_sample": true,
+    "eos_token_id": [
+        248046,
+        248044
+    ],
+    "pad_token_id": 248044,
+    "temperature": 1.0,
+    "top_k": 20,
+    "top_p": 0.95
+}

mlx_conversion_manifest.json

@@ -0,0 +1,27 @@
+{
+  "converter": "convert_qwen36_modelopt_hybrid_to_mlx.py",
+  "created_at": "2026-07-10T03:37:44.758817+00:00",
+  "input_scales_dropped": 30971,
+  "multimodal_upgrade_at": "2026-07-10T04:02:39.296349+00:00",
+  "norm_weights_shifted": true,
+  "num_experts": 256,
+  "num_layers": 40,
+  "output": "mlx-community/Qwen3.6-35B-A3B-AntiLoop-NVFP4",
+  "output_shards": 42,
+  "output_tensor_count": 1808,
+  "output_total_size_bytes": 21758152396,
+  "quantized_module_counts": {
+    "scaled_mxfp8": 130,
+    "scaled_nvfp4": 121,
+    "scaled_nvfp4_switch": 120
+  },
+  "runtime_file": "modeling_mlx_qwen36_modelopt_hybrid.py",
+  "source": "N8Programs/Qwen3.6-35B-A3B-AntiLoop-NVFP4@1fc377564024dce4e8e7f2bdc04d34cd869f928f",
+  "source_tensor_count": 124468,
+  "vision_shard": "model-vision.safetensors",
+  "vision_tensor_count": 333,
+  "vision_tensor_data_bytes": 893142496,
+  "vision_weight_bytes_requantized": false,
+  "vlm_runtime_file": "modeling_mlx_vlm_qwen36_modelopt_hybrid.py",
+  "weight_bytes_requantized": false
+}

mlx_vlm_model_file_loader.py

@@ -0,0 +1,121 @@
+"""Opt-in loader for model-local MLX-VLM architecture files.
+
+MLX-VLM 0.6.4 does not yet consult ``vlm_model_file`` in a model config.
+This module adds that one lookup in-process.  It never edits the installed
+``mlx_vlm`` package, and it only executes model-local code when the caller has
+explicitly passed ``trust_remote_code=True``.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from contextvars import ContextVar
+from pathlib import Path
+from types import ModuleType
+from typing import Any
+
+
+_ACTIVE_MODEL_PATH: ContextVar[Path | None] = ContextVar(
+    "mlx_vlm_model_file_path", default=None
+)
+_ACTIVE_TRUST_REMOTE_CODE: ContextVar[bool] = ContextVar(
+    "mlx_vlm_model_file_trust", default=False
+)
+_MODULE_CACHE: dict[Path, ModuleType] = {}
+
+
+def _load_local_module(model_path: Path, filename: str) -> ModuleType:
+    root = model_path.resolve()
+    module_path = (root / filename).resolve()
+
+    try:
+        module_path.relative_to(root)
+    except ValueError as exc:
+        raise ValueError(
+            f"vlm_model_file must stay inside the model directory: {filename!r}"
+        ) from exc
+
+    if module_path.suffix != ".py" or not module_path.is_file():
+        raise FileNotFoundError(f"Model-local MLX-VLM runtime not found: {module_path}")
+
+    cached = _MODULE_CACHE.get(module_path)
+    if cached is not None:
+        return cached
+
+    module_name = f"mlx_vlm_remote_{abs(hash(str(module_path))):x}"
+    spec = importlib.util.spec_from_file_location(module_name, module_path)
+    if spec is None or spec.loader is None:
+        raise ImportError(f"Could not import model-local runtime: {module_path}")
+
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+    try:
+        spec.loader.exec_module(module)
+    except Exception:
+        sys.modules.pop(module_name, None)
+        raise
+
+    _MODULE_CACHE[module_path] = module
+    return module
+
+
+def install() -> None:
+    """Install the model-file lookup into the current Python process."""
+
+    import mlx_vlm.utils as utils
+
+    if getattr(utils.get_model_and_args, "_model_file_loader_installed", False):
+        return
+
+    original_get_model_and_args = utils.get_model_and_args
+    original_load_model = utils.load_model
+
+    def get_model_and_args(config: dict, *args: Any, **kwargs: Any):
+        filename = config.get("vlm_model_file")
+        model_path = kwargs.get("model_path") or _ACTIVE_MODEL_PATH.get()
+
+        # Calls made later by processor setup do not carry a model path in
+        # MLX-VLM 0.6.4; let its native Qwen implementation handle those.
+        if not filename or model_path is None:
+            return original_get_model_and_args(config, *args, **kwargs)
+
+        trusted = bool(
+            kwargs.get("trust_remote_code", False)
+            or _ACTIVE_TRUST_REMOTE_CODE.get()
+        )
+        if not trusted:
+            raise PermissionError(
+                "This checkpoint includes a model-local MLX-VLM runtime. "
+                "Re-run with --trust-remote-code (or trust_remote_code=True)."
+            )
+
+        module = _load_local_module(Path(model_path), str(filename))
+        return module, f"model-local:{filename}"
+
+    def load_model(model_path: Path, lazy: bool = False, **kwargs: Any):
+        path_token = _ACTIVE_MODEL_PATH.set(Path(model_path))
+        trust_token = _ACTIVE_TRUST_REMOTE_CODE.set(
+            bool(kwargs.get("trust_remote_code", False))
+        )
+        try:
+            return original_load_model(model_path, lazy=lazy, **kwargs)
+        finally:
+            _ACTIVE_TRUST_REMOTE_CODE.reset(trust_token)
+            _ACTIVE_MODEL_PATH.reset(path_token)
+
+    get_model_and_args._model_file_loader_installed = True  # type: ignore[attr-defined]
+    load_model._model_file_loader_installed = True  # type: ignore[attr-defined]
+    utils.get_model_and_args = get_model_and_args
+    utils.load_model = load_model
+
+
+def load(path_or_hf_repo: str, **kwargs: Any):
+    """Programmatic convenience wrapper around :func:`mlx_vlm.load`."""
+
+    install()
+    kwargs.setdefault("trust_remote_code", True)
+    from mlx_vlm import load as mlx_vlm_load
+
+    return mlx_vlm_load(path_or_hf_repo, **kwargs)
+

modeling_mlx_qwen36_modelopt_hybrid.py

@@ -0,0 +1,205 @@
+"""MLX runtime for Qwen3.6 ModelOpt hybrid FP8/NVFP4 checkpoints.
+
+The converter stores ModelOpt FP8 weights losslessly in MLX's packed MXFP8
+carrier with unit E8M0 block scales, then applies the original per-tensor
+ModelOpt scale to the output.  ModelOpt NVFP4 weights and E4M3 block scales
+are likewise retained bit-for-bit; their FP32 tensor scale is applied after
+the matrix multiplication.
+
+Activations remain in the model dtype.  This avoids adding a second lossy
+activation requantization scheme while still using MLX's native quantized
+weight kernels.
+"""
+
+from dataclasses import dataclass, field
+from typing import Dict
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from mlx_lm.models.base import BaseModelArgs
+from mlx_lm.models.qwen3_5_moe import Model as BaseModel
+from mlx_lm.models.switch_layers import SwitchLinear
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: dict
+    mlx_modelopt_quantization: Dict[str, str] = field(default_factory=dict)
+
+
+class ScaledQuantizedLinear(nn.Module):
+    """Weight-quantized dense linear with an additional tensor scale."""
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        *,
+        group_size: int,
+        bits: int,
+        mode: str,
+        bias: bool = False,
+    ):
+        super().__init__()
+        if input_dims % group_size:
+            raise ValueError(
+                f"input_dims={input_dims} is not divisible by group_size={group_size}"
+            )
+        if (input_dims * bits) % 32:
+            raise ValueError(
+                f"input_dims={input_dims}, bits={bits} cannot be packed into uint32"
+            )
+
+        self.group_size = group_size
+        self.bits = bits
+        self.mode = mode
+        self.weight = mx.zeros(
+            (output_dims, input_dims * bits // 32), dtype=mx.uint32
+        )
+        self.scales = mx.zeros(
+            (output_dims, input_dims // group_size), dtype=mx.uint8
+        )
+        self.global_scale = mx.ones((), dtype=mx.float32)
+        if bias:
+            self.bias = mx.zeros((output_dims,))
+        self.freeze()
+
+    @classmethod
+    def from_linear(cls, linear: nn.Module, kind: str):
+        output_dims, input_dims = linear.weight.shape
+        has_bias = linear.get("bias") is not None
+        if kind == "scaled_mxfp8":
+            params = dict(group_size=32, bits=8, mode="mxfp8")
+        elif kind == "scaled_nvfp4":
+            params = dict(group_size=16, bits=4, mode="nvfp4")
+        else:
+            raise ValueError(f"Unsupported dense quantization kind: {kind}")
+        return cls(input_dims, output_dims, bias=has_bias, **params)
+
+    def __call__(self, x):
+        y = mx.quantized_matmul(
+            x,
+            self["weight"],
+            self["scales"],
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+            mode=self.mode,
+        )
+        # Avoid promoting the residual stream to float32.
+        y = y * self["global_scale"].astype(y.dtype)
+        if "bias" in self:
+            y = y + self["bias"]
+        return y
+
+
+class ScaledNVFP4SwitchLinear(nn.Module):
+    """Expert linear using MLX gather_qmm and per-expert tensor scales."""
+
+    group_size = 16
+    bits = 4
+    mode = "nvfp4"
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        num_experts: int,
+        *,
+        bias: bool = False,
+    ):
+        super().__init__()
+        if input_dims % self.group_size:
+            raise ValueError(
+                f"input_dims={input_dims} is not divisible by {self.group_size}"
+            )
+        self.weight = mx.zeros(
+            (num_experts, output_dims, input_dims * self.bits // 32),
+            dtype=mx.uint32,
+        )
+        self.scales = mx.zeros(
+            (num_experts, output_dims, input_dims // self.group_size),
+            dtype=mx.uint8,
+        )
+        self.global_scales = mx.ones((num_experts,), dtype=mx.float32)
+        if bias:
+            self.bias = mx.zeros((num_experts, output_dims))
+        self.freeze()
+
+    @classmethod
+    def from_switch_linear(cls, linear: SwitchLinear):
+        num_experts, output_dims, input_dims = linear.weight.shape
+        has_bias = linear.get("bias") is not None
+        return cls(
+            input_dims,
+            output_dims,
+            num_experts,
+            bias=has_bias,
+        )
+
+    @property
+    def input_dims(self):
+        return self.scales.shape[2] * self.group_size
+
+    @property
+    def output_dims(self):
+        return self.weight.shape[1]
+
+    @property
+    def num_experts(self):
+        return self.weight.shape[0]
+
+    def __call__(self, x, indices, sorted_indices=False):
+        y = mx.gather_qmm(
+            x,
+            self["weight"],
+            self["scales"],
+            rhs_indices=indices,
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+            mode=self.mode,
+            sorted_indices=sorted_indices,
+        )
+        scale = self["global_scales"][indices].astype(y.dtype)[..., None, None]
+        y = y * scale
+        if "bias" in self:
+            y = y + mx.expand_dims(self["bias"][indices], -2)
+        return y
+
+
+def _replace_quantized_modules(model: nn.Module, quantization: Dict[str, str]):
+    leaves = dict(
+        tree_flatten(model.leaf_modules(), is_leaf=lambda m: isinstance(m, nn.Module))
+    )
+    missing = sorted(set(quantization) - set(leaves))
+    if missing:
+        preview = "\n  ".join(missing[:20])
+        raise ValueError(f"Quantized module paths are absent from the model:\n  {preview}")
+
+    for path, kind in quantization.items():
+        module = leaves[path]
+        if kind in ("scaled_mxfp8", "scaled_nvfp4"):
+            if not isinstance(module, nn.Linear):
+                raise TypeError(f"{path} is {type(module).__name__}, expected Linear")
+            leaves[path] = ScaledQuantizedLinear.from_linear(module, kind)
+        elif kind == "scaled_nvfp4_switch":
+            if not isinstance(module, SwitchLinear):
+                raise TypeError(
+                    f"{path} is {type(module).__name__}, expected SwitchLinear"
+                )
+            leaves[path] = ScaledNVFP4SwitchLinear.from_switch_linear(module)
+        else:
+            raise ValueError(f"Unknown quantization kind {kind!r} for {path}")
+
+    model.update_modules(tree_unflatten(list(leaves.items())))
+
+
+class Model(BaseModel):
+    def __init__(self, args: ModelArgs):
+        super().__init__(args)
+        _replace_quantized_modules(self, args.mlx_modelopt_quantization)
+

modeling_mlx_vlm_qwen36_modelopt_hybrid.py

@@ -0,0 +1,217 @@
+"""MLX-VLM runtime for Qwen3.6 ModelOpt hybrid FP8/NVFP4 checkpoints.
+
+The language model keeps the lossless ModelOpt-to-MLX representation used by
+``modeling_mlx_qwen36_modelopt_hybrid.py``.  The vision tower remains in its
+original BF16 representation and is delegated to MLX-VLM's native Qwen3.5 MoE
+vision implementation.
+
+This module intentionally exports the same public symbols as an MLX-VLM model
+package so a model-local loader can select it without changing ``model_type``.
+"""
+
+from dataclasses import dataclass, field
+from typing import Dict
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.utils import tree_flatten, tree_unflatten
+
+from mlx_vlm.models.qwen3_5_moe import LanguageModel, TextConfig, VisionConfig
+from mlx_vlm.models.qwen3_5_moe import Model as BaseModel
+from mlx_vlm.models.qwen3_5_moe import ModelConfig as BaseModelConfig
+from mlx_vlm.models.qwen3_5_moe import VisionModel
+from mlx_vlm.models.switch_layers import SwitchLinear
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    mlx_modelopt_quantization: Dict[str, str] = field(default_factory=dict)
+
+
+class ScaledQuantizedLinear(nn.Module):
+    """Weight-quantized dense linear with a ModelOpt tensor scale."""
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        *,
+        group_size: int,
+        bits: int,
+        mode: str,
+        bias: bool = False,
+    ):
+        super().__init__()
+        if input_dims % group_size:
+            raise ValueError(
+                f"input_dims={input_dims} is not divisible by group_size={group_size}"
+            )
+        if (input_dims * bits) % 32:
+            raise ValueError(
+                f"input_dims={input_dims}, bits={bits} cannot be packed into uint32"
+            )
+
+        self.group_size = group_size
+        self.bits = bits
+        self.mode = mode
+        self.weight = mx.zeros(
+            (output_dims, input_dims * bits // 32), dtype=mx.uint32
+        )
+        self.scales = mx.zeros(
+            (output_dims, input_dims // group_size), dtype=mx.uint8
+        )
+        self.global_scale = mx.ones((), dtype=mx.float32)
+        if bias:
+            self.bias = mx.zeros((output_dims,))
+        self.freeze()
+
+    @classmethod
+    def from_linear(cls, linear: nn.Module, kind: str):
+        output_dims, input_dims = linear.weight.shape
+        has_bias = linear.get("bias") is not None
+        if kind == "scaled_mxfp8":
+            params = dict(group_size=32, bits=8, mode="mxfp8")
+        elif kind == "scaled_nvfp4":
+            params = dict(group_size=16, bits=4, mode="nvfp4")
+        else:
+            raise ValueError(f"Unsupported dense quantization kind: {kind}")
+        return cls(input_dims, output_dims, bias=has_bias, **params)
+
+    def __call__(self, x):
+        y = mx.quantized_matmul(
+            x,
+            self["weight"],
+            self["scales"],
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+            mode=self.mode,
+        )
+        y = y * self["global_scale"].astype(y.dtype)
+        if "bias" in self:
+            y = y + self["bias"]
+        return y
+
+
+class ScaledNVFP4SwitchLinear(nn.Module):
+    """Expert linear using MLX gather_qmm and per-expert tensor scales."""
+
+    group_size = 16
+    bits = 4
+    mode = "nvfp4"
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        num_experts: int,
+        *,
+        bias: bool = False,
+    ):
+        super().__init__()
+        if input_dims % self.group_size:
+            raise ValueError(
+                f"input_dims={input_dims} is not divisible by {self.group_size}"
+            )
+        self.weight = mx.zeros(
+            (num_experts, output_dims, input_dims * self.bits // 32),
+            dtype=mx.uint32,
+        )
+        self.scales = mx.zeros(
+            (num_experts, output_dims, input_dims // self.group_size),
+            dtype=mx.uint8,
+        )
+        self.global_scales = mx.ones((num_experts,), dtype=mx.float32)
+        if bias:
+            self.bias = mx.zeros((num_experts, output_dims))
+        self.freeze()
+
+    @classmethod
+    def from_switch_linear(cls, linear: SwitchLinear):
+        num_experts, output_dims, input_dims = linear.weight.shape
+        has_bias = linear.get("bias") is not None
+        return cls(input_dims, output_dims, num_experts, bias=has_bias)
+
+    @property
+    def input_dims(self):
+        return self.scales.shape[2] * self.group_size
+
+    @property
+    def output_dims(self):
+        return self.weight.shape[1]
+
+    @property
+    def num_experts(self):
+        return self.weight.shape[0]
+
+    def __call__(self, x, indices, sorted_indices=False):
+        y = mx.gather_qmm(
+            x,
+            self["weight"],
+            self["scales"],
+            rhs_indices=indices,
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+            mode=self.mode,
+            sorted_indices=sorted_indices,
+        )
+        scale = self["global_scales"][indices].astype(y.dtype)[..., None, None]
+        y = y * scale
+        if "bias" in self:
+            y = y + mx.expand_dims(self["bias"][indices], -2)
+        return y
+
+
+def _replace_quantized_modules(model: nn.Module, quantization: Dict[str, str]):
+    leaves = dict(
+        tree_flatten(model.leaf_modules(), is_leaf=lambda m: isinstance(m, nn.Module))
+    )
+    missing = sorted(set(quantization) - set(leaves))
+    if missing:
+        preview = "\n  ".join(missing[:20])
+        raise ValueError(f"Quantized module paths are absent from the model:\n  {preview}")
+
+    for path, kind in quantization.items():
+        module = leaves[path]
+        if kind in ("scaled_mxfp8", "scaled_nvfp4"):
+            if not isinstance(module, nn.Linear):
+                raise TypeError(f"{path} is {type(module).__name__}, expected Linear")
+            leaves[path] = ScaledQuantizedLinear.from_linear(module, kind)
+        elif kind == "scaled_nvfp4_switch":
+            if not isinstance(module, SwitchLinear):
+                raise TypeError(
+                    f"{path} is {type(module).__name__}, expected SwitchLinear"
+                )
+            leaves[path] = ScaledNVFP4SwitchLinear.from_switch_linear(module)
+        else:
+            raise ValueError(f"Unknown quantization kind {kind!r} for {path}")
+
+    model.update_modules(tree_unflatten(list(leaves.items())))
+
+
+class Model(BaseModel):
+    def __init__(self, config: ModelConfig):
+        super().__init__(config)
+        _replace_quantized_modules(self, config.mlx_modelopt_quantization)
+
+    def sanitize(self, weights):
+        """Map only raw source keys; converted language keys are already sanitized."""
+        sanitized = {}
+        for key, value in weights.items():
+            if "mtp." in key:
+                continue
+            if key.startswith("model.language_model.visual"):
+                key = key.replace(
+                    "model.language_model.visual", "vision_tower", 1
+                )
+            elif key.startswith("model.language_model"):
+                key = key.replace(
+                    "model.language_model", "language_model.model", 1
+                )
+            elif key.startswith("model.visual"):
+                key = key.replace("model.visual", "vision_tower", 1)
+            elif key.startswith("lm_head"):
+                key = key.replace("lm_head", "language_model.lm_head", 1)
+            sanitized[key] = value
+        return sanitized

preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "size": {
+        "longest_edge": 16777216,
+        "shortest_edge": 65536
+    },
+    "patch_size": 16,
+    "temporal_patch_size": 2,
+    "merge_size": 2,
+    "image_mean": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "image_std": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "processor_class": "Qwen3VLProcessor",
+    "image_processor_type": "Qwen2VLImageProcessorFast"
+}

run_mlx_vlm.py

@@ -0,0 +1,12 @@
+"""Run the MLX-VLM generation CLI with model-local runtime support."""
+
+from mlx_vlm_model_file_loader import install
+
+install()
+
+from mlx_vlm.generate.cli import main  # noqa: E402
+
+
+if __name__ == "__main__":
+    main()
+

run_mlx_vlm_server.py

@@ -0,0 +1,12 @@
+"""Run the MLX-VLM HTTP server with model-local runtime support."""
+
+from mlx_vlm_model_file_loader import install
+
+install()
+
+from mlx_vlm.server.cli import main  # noqa: E402
+
+
+if __name__ == "__main__":
+    main()
+

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42
+size 12807982

tokenizer_config.json

@@ -0,0 +1,305 @@
+{
+    "add_prefix_space": false,
+    "added_tokens_decoder": {
+        "248044": {
+            "content": "<|endoftext|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248045": {
+            "content": "<|im_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248046": {
+            "content": "<|im_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248047": {
+            "content": "<|object_ref_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248048": {
+            "content": "<|object_ref_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248049": {
+            "content": "<|box_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248050": {
+            "content": "<|box_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248051": {
+            "content": "<|quad_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248052": {
+            "content": "<|quad_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248053": {
+            "content": "<|vision_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248054": {
+            "content": "<|vision_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248055": {
+            "content": "<|vision_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248056": {
+            "content": "<|image_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248057": {
+            "content": "<|video_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248058": {
+            "content": "<tool_call>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248059": {
+            "content": "</tool_call>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248060": {
+            "content": "<|fim_prefix|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248061": {
+            "content": "<|fim_middle|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248062": {
+            "content": "<|fim_suffix|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248063": {
+            "content": "<|fim_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248064": {
+            "content": "<|repo_name|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248065": {
+            "content": "<|file_sep|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248066": {
+            "content": "<tool_response>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248067": {
+            "content": "</tool_response>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248068": {
+            "content": "<think>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248069": {
+            "content": "</think>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248070": {
+            "content": "<|audio_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248071": {
+            "content": "<|audio_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248072": {
+            "content": "<tts_pad>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248073": {
+            "content": "<tts_text_bos>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248074": {
+            "content": "<tts_text_eod>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248075": {
+            "content": "<tts_text_bos_single>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248076": {
+            "content": "<|audio_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        }
+    },
+    "additional_special_tokens": [
+        "<|im_start|>",
+        "<|im_end|>",
+        "<|object_ref_start|>",
+        "<|object_ref_end|>",
+        "<|box_start|>",
+        "<|box_end|>",
+        "<|quad_start|>",
+        "<|quad_end|>",
+        "<|vision_start|>",
+        "<|vision_end|>",
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>"
+    ],
+    "bos_token": null,
+    "chat_template": "{%- set image_count = namespace(value=0) %}\n{%- set video_count = namespace(value=0) %}\n{%- macro render_content(content, do_vision_count, is_system_content=false) %}\n    {%- if content is string %}\n        {{- content }}\n    {%- elif content is iterable and content is not mapping %}\n        {%- for item in content %}\n            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}\n                {%- if is_system_content %}\n                    {{- raise_exception('System message cannot contain images.') }}\n                {%- endif %}\n                {%- if do_vision_count %}\n                    {%- set image_count.value = image_count.value + 1 %}\n                {%- endif %}\n                {%- if add_vision_id %}\n                    {{- 'Picture ' ~ image_count.value ~ ': ' }}\n                {%- endif %}\n                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}\n            {%- elif 'video' in item or item.type == 'video' %}\n                {%- if is_system_content %}\n                    {{- raise_exception('System message cannot contain videos.') }}\n                {%- endif %}\n                {%- if do_vision_count %}\n                    {%- set video_count.value = video_count.value + 1 %}\n                {%- endif %}\n                {%- if add_vision_id %}\n                    {{- 'Video ' ~ video_count.value ~ ': ' }}\n                {%- endif %}\n                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}\n            {%- elif 'text' in item %}\n                {{- item.text }}\n            {%- else %}\n                {{- raise_exception('Unexpected item type in content.') }}\n            {%- endif %}\n        {%- endfor %}\n    {%- elif content is none or content is undefined %}\n        {{- '' }}\n    {%- else %}\n        {{- raise_exception('Unexpected content type.') }}\n    {%- endif %}\n{%- endmacro %}\n{%- if not messages %}\n    {{- raise_exception('No messages provided.') }}\n{%- endif %}\n{%- if tools and tools is iterable and tools is not mapping %}\n    {{- '<|im_start|>system\\n' }}\n    {{- \"# Tools\\n\\nYou have access to the following functions:\\n\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\" }}\n    {{- '\\n\\nIf you choose to call a function ONLY reply in the following format with NO suffix:\\n\\n<tool_call>\\n<function=example_function_name>\\n<parameter=example_parameter_1>\\nvalue_1\\n</parameter>\\n<parameter=example_parameter_2>\\nThis is the value for the second parameter\\nthat can span\\nmultiple lines\\n</parameter>\\n</function>\\n</tool_call>\\n\\n<IMPORTANT>\\nReminder:\\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\\n- Required parameters MUST be specified\\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\\n</IMPORTANT>' }}\n    {%- if messages[0].role == 'system' %}\n        {%- set content = render_content(messages[0].content, false, true)|trim %}\n        {%- if content %}\n            {{- '\\n\\n' + content }}\n        {%- endif %}\n    {%- endif %}\n    {{- '<|im_end|>\\n' }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {%- set content = render_content(messages[0].content, false, true)|trim %}\n        {{- '<|im_start|>system\\n' + content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" %}\n        {%- set content = render_content(message.content, false)|trim %}\n        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}\n            {%- set ns.multi_step_tool = false %}\n            {%- set ns.last_query_index = index %}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if ns.multi_step_tool %}\n    {{- raise_exception('No user query found in messages.') }}\n{%- endif %}\n{%- for message in messages %}\n    {%- set content = render_content(message.content, true)|trim %}\n    {%- if message.role == \"system\" %}\n        {%- if not loop.first %}\n            {{- raise_exception('System message must be at the beginning.') }}\n        {%- endif %}\n    {%- elif message.role == \"user\" %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- set reasoning_content = reasoning_content|trim %}\n        {%- if (preserve_thinking is defined and preserve_thinking is true) or (loop.index0 > ns.last_query_index) %}\n            {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content + '\\n</think>\\n\\n' + content }}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if tool_call.function is defined %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {%- if loop.first %}\n                    {%- if content|trim %}\n                        {{- '\\n\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                    {%- else %}\n                        {{- '<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                    {%- endif %}\n                {%- else %}\n                    {{- '\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                {%- endif %}\n                {%- if tool_call.arguments is defined %}\n                    {%- for args_name, args_value in tool_call.arguments|items %}\n                        {{- '<parameter=' + args_name + '>\\n' }}\n                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}\n                        {{- args_value }}\n                        {{- '\\n</parameter>\\n' }}\n                    {%- endfor %}\n                {%- endif %}\n                {{- '</function>\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.previtem and loop.previtem.role != \"tool\" %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if not loop.last and loop.nextitem.role != \"tool\" %}\n            {{- '<|im_end|>\\n' }}\n        {%- elif loop.last %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- else %}\n        {{- raise_exception('Unexpected message role.') }}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- else %}\n        {{- '<think>\\n' }}\n    {%- endif %}\n{%- endif %}",
+    "clean_up_tokenization_spaces": false,
+    "eos_token": "<|im_end|>",
+    "errors": "replace",
+    "model_max_length": 262144,
+    "pad_token": "<|endoftext|>",
+    "split_special_tokens": false,
+    "tokenizer_class": "Qwen2Tokenizer",
+    "unk_token": null,
+    "add_bos_token": false,
+    "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+    "extra_special_tokens": {
+        "audio_bos_token": "<|audio_start|>",
+        "audio_eos_token": "<|audio_end|>",
+        "audio_token": "<|audio_pad|>",
+        "image_token": "<|image_pad|>",
+        "video_token": "<|video_pad|>",
+        "vision_bos_token": "<|vision_start|>",
+        "vision_eos_token": "<|vision_end|>"
+    }
+}

video_preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "size": {
+        "longest_edge": 25165824,
+        "shortest_edge": 4096
+    },
+    "patch_size": 16,
+    "temporal_patch_size": 2,
+    "merge_size": 2,
+    "image_mean": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "image_std": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "processor_class": "Qwen3VLProcessor",
+    "video_processor_type": "Qwen3VLVideoProcessor"
+}