Add Muse Glimmer AutoQuant recipe (#2269)

### What does this PR do?

Type of change: ? <!-- Use one of the following: Bug fix, new feature,
new example, new tests, documentation. -->

<!-- Details about the change. -->

### Usage

```python
# Add a code snippet demonstrating how to use this
```

### Testing
<!-- Mention how have you tested your change if applicable. -->

### Before your PR is "*Ready for review*"

Make sure you read and follow [Contributor
guidelines](https://github.com/NVIDIA/Model-Optimizer/blob/main/CONTRIBUTING.md)
and your commits are signed (`git commit -s -S`).

Make sure you read and follow the [Security Best
Practices](https://github.com/NVIDIA/Model-Optimizer/blob/main/SECURITY.md#security-coding-practices-for-contributors)
(e.g. avoiding hardcoded `trust_remote_code=True`, `torch.load(...,
weights_only=False)`, `pickle`, etc.).

- Is this change backward compatible?: ✅ / ❌ / N/A <!--- If ❌, explain
why. -->
- If you copied code from any other sources or added a new PIP
dependency, did you follow guidance in `CONTRIBUTING.md`: ✅ / ❌ / N/A
<!--- Mandatory -->
- Did you write any new necessary tests?: ✅ / ❌ / N/A <!--- Mandatory
for new features or examples. -->
- Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?:
✅ / ❌ / N/A <!--- Very short summary of changes only for new features,
backward breaking changes, deprecations, or fixes for critical bugs
present in previous releases. -->
- Did you get Claude approval on this PR?: ✅ / ❌ / N/A <!--- Run
`/claude review`. NVIDIA org members can self-trigger for complex
changes; orthogonal to CodeRabbit. -->

### Additional Information
<!-- E.g. related issue. -->


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
  * Added a Muse Glimmer mixed-precision AutoQuantize recipe.
* Supports W4A16 NVFP4 Four-Over-Six and FP8 quantization with BF16
fallback at 5.5 effective bits.
  * Keeps vision modules and the key-value cache unquantized.
* Added an Alpamayo QAD example with FSDP2 support and optional
full-checkpoint export.
* **Documentation**
* Added calibration, export, target-module, and fallback-precision
guidance for the recipe.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Signed-off-by: Jenny Chen <jennifchen@nvidia.com>

Author: jenchen13

CHANGELOG.rst

@@ -8,6 +8,7 @@ Changelog
 
 *Quantization*
 
+- Add a Muse Glimmer AutoQuantize recipe that searches language-model MLP projections, self-attention projections, and ``lm_head`` over W4A16 NVFP4 Four-Over-Six, FP8, and BF16 fallback at 5.5 effective bits while leaving the vision tower unquantized.
 - Add ``examples/alpamayo/qad.py``, which runs quantization-aware distillation on the quantized Alpamayo checkpoint produced by ``examples/alpamayo/quantize.py``. It distills the quantized VLM against the original FP16 VLM with ``QADTrainer``, supports FSDP2 for multi-GPU runs, and ``--export`` reassembles the trained VLM into a full AlpamayoR1 checkpoint that ``AlpamayoR1.from_pretrained`` can reload.
 - Add a calibration-free streaming Kimi-K3 converter and checkpoint-mirror recipe for NVFP4 routed experts with ``input_scale=1.0`` and 128x128 block-FP8 KDA/MLA attention weights. The converter operates shard-by-shard on the source checkpoint's packed MXFP4 experts instead of loading the 2.8T model through the in-memory ``hf_ptq.py`` path.
 - Add ``mtq.temporarily_fold_weights`` for repeated frozen-weight inference and ``mtq.preserve_quantizer_attributes_context`` for restoring temporary quantizer property and type changes. Temporary folding snapshots affected fake-quant weights on a configurable device and restores them with their quantizer state; retained pre-quant scales are inactive, while shared weights, shared quantizers, and ``SequentialQuantizer`` weights are unsupported.

modelopt_recipes/models/meta-models/Muse-Glimmer-30B/auto_quantize/README.md

@@ -0,0 +1,19 @@
+# Muse Glimmer AutoQuantize recipes
+
+`w4a16_nvfp4_4o6_mixed` searches the Muse Glimmer
+language-model MLP projections, self-attention projections, and `lm_head` over
+W4A16 NVFP4 Four-Over-Six, FP8, and BF16 fallback at 5.5 effective bits. The
+vision tower and unmatched modules remain BF16.
+
+Use unquantized KV cache and representative text calibration:
+
+```bash
+python examples/hf_ptq/hf_ptq.py \
+  --pyt_ckpt_path <muse-glimmer-checkpoint> \
+  --recipe models/meta-models/Muse-Glimmer-30B/auto_quantize/w4a16_nvfp4_4o6_mixed \
+  --kv_cache_qformat none \
+  --dataset nemotron-post-training-v3 \
+  --calib_size 512 \
+  --calib_seq 2048 \
+  --export_path <output-path>
+```

modelopt_recipes/models/meta-models/Muse-Glimmer-30B/auto_quantize/w4a16_nvfp4_4o6_mixed.yaml

@@ -0,0 +1,81 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Muse Glimmer language-model search over W4A16 NVFP4 Four-Over-Six, FP8,
+# and BF16 fallback. Vision and unmatched module families remain BF16.
+
+# modelopt-schema: modelopt.recipe.config.ModelOptAutoQuantizeRecipe
+imports:
+  base_disable_all: configs/ptq/units/base_disable_all
+  base_disabled_layers: configs/auto_quantize/units/base_disabled_layers
+  base_cost_excluded_layers: configs/auto_quantize/units/base_cost_excluded_layers
+  default_disabled_quantizers: configs/ptq/units/default_disabled_quantizers
+  fp8_quantizers: configs/ptq/units/w8a8_fp8_fp8
+  nvfp4_four_over_six: configs/numerics/nvfp4_four_over_six
+
+metadata:
+  recipe_type: auto_quantize
+  description: >-
+    Muse Glimmer language-model W4A16 NVFP4 Four-Over-Six, FP8, and BF16
+    fallback search at 5.5 weight-effective bits. Use --kv_cache_qformat none.
+
+# Unmatched modules retain this BF16-by-default baseline. The searched language-model
+# families start from W4A16 NVFP4 and are reassigned by AutoQuantize below.
+quantize:
+  effective_bits: 4.5
+  algorithm:
+    method: mse
+    fp8_scale_sweep: false
+    start_multiplier: 1.0
+    stop_multiplier: 1.5
+    step_size: 0.5
+  quant_cfg:
+    - $import: base_disable_all
+    - quantizer_name: "*language_model.layers.*.mlp*.weight_quantizer"
+      cfg:
+        $import: nvfp4_four_over_six
+    - quantizer_name: "*language_model.layers.*.self_attn.*_proj.weight_quantizer"
+      cfg:
+        $import: nvfp4_four_over_six
+    - quantizer_name: "*lm_head*.weight_quantizer"
+      cfg:
+        $import: nvfp4_four_over_six
+    - $import: default_disabled_quantizers
+
+auto_quantize:
+  constraints:
+    effective_bits: 5.5
+    cost_model: weight
+
+  module_search_spaces:
+    - module_name_patterns:
+        - "*language_model.layers.*.mlp*"
+        - "*language_model.layers.*.self_attn.*_proj"
+        - "*lm_head*"
+      candidate_formats:
+        - effective_bits: 4.5
+          algorithm:
+            method: mse
+            fp8_scale_sweep: false
+            start_multiplier: 1.0
+            stop_multiplier: 1.5
+            step_size: 0.5
+          quant_cfg:
+            - $import: base_disable_all
+            - quantizer_name: "*weight_quantizer"
+              cfg:
+                $import: nvfp4_four_over_six
+        - algorithm: max
+          quant_cfg:
+            - $import: base_disable_all
+            - $import: fp8_quantizers
+      allow_no_quant: true
+
+  auto_quantize_method: gradient
+  score_size: 128
+
+  disabled_layers:
+    - $import: base_disabled_layers
+
+  cost_excluded_layers:
+    - $import: base_cost_excluded_layers