Add LAQ NVFP4 export support

[realAsma]

Summary

Adds LAQ export support for NVFP4 unified HF export.

• Export packs FP4 weights with the LAQ pre scale and serializes the LAQ post scale for dequantization.
• Supports LAQ learnable modes covered by the LAQ config/tests, including tied pre/post and quantize_pre_scale=False.
• Carries the LAQ quantize_pre_scale config plumbing and dtype/FSDP2 test updates from the LAQ branch worktree base.

Testing

• pre-commit run --files modelopt/torch/export/quant_utils.py modelopt/torch/export/unified_export_hf.py modelopt/torch/quantization/config.py modelopt/torch/quantization/model_calib.py modelopt/torch/quantization/nn/modules/tensor_quantizer.py tests/unit/torch/quantization/test_nvfp4_static_export_cpu.py tests/unit/torch/quantization/test_laq.py tests/gpu/torch/quantization/test_laq_cuda.py tests/gpu/torch/quantization/test_nvfp4_static_quantizer_cuda.py tests/gpu/torch/quantization/test_fsdp2.py
• pytest_pwd tests/unit/torch/quantization/test_laq.py tests/unit/torch/quantization/test_nvfp4_static_export_cpu.py -q
• Qwen3-8B LAQ export validation on omniml-a8.nvidia.com, CUDA_VISIBLE_DEVICES=1,2,3,4, recipe nvfp4_laq_post_unquantized_pre_scale-mse_init-fp8_kv.yml with quantize_pre_scale: false: exported config.json, hf_quant_config.json, .quant_summary.txt, tokenizer files, and one 5.2 GB model.safetensors shard under /home/scratch.akuriparambi_coreai/Model-Optimizer-LAQ-export/agent_art/laq_qwen3_8b_export_a8_20260628T1815/qwen3_8b_laq_post_unquantized_pre_scale_export.
• Reran the same Qwen3-8B LAQ export on omniml-a8.nvidia.com, CUDA_VISIBLE_DEVICES=1,2,3,4, after source ~/.bashrc and conda activate black; log confirms python=/home/akuriparambi/anaconda3/envs/black/bin/python, modelopt_file=/home/scratch.akuriparambi_coreai/Model-Optimizer-LAQ-export/modelopt/__init__.py, and pythonpath_head=/home/scratch.akuriparambi_coreai/Model-Optimizer-LAQ-export. Export succeeded with hf_quant_config.json, .quant_summary.txt, tokenizer/config files, and one 5.5 GB model.safetensors under /home/scratch.akuriparambi_coreai/Model-Optimizer-LAQ-export/agent_art/laq_qwen3_8b_export_a8_black_20260629T0005/qwen3_8b_laq_post_unquantized_pre_scale_export.

Notes

The first a8 attempt reused a stale shared modelopt_cuda_ext_fp8 Torch extension cache and failed with CUDA error: no kernel image is available for execution on the device. Passing runs used artifact-local TORCH_EXTENSIONS_DIR values and TORCH_CUDA_ARCH_LIST=8.9 so the FP8 extension rebuilt for Ada.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 7263985c-fe05-4e89-82a2-c9cc999209af

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch asma/laq-export-support

---

_{Comment @coderabbitai help to get the list of available commands.}