[6529944] Fix dynamic NVFP4 TensorRT compatibility

[ajrasane]

What does this PR do?

Type of change: Bug fix

This PR improves TensorRT compatibility for dynamic NVFP4 ONNX exports:

• Preserves the W4A4 ONNX graph and requires TensorRT 11.0 or newer only when compiling dynamic NVFP4 models.
• Adds an early, actionable diagnostic for TensorRT 10 with the explicit --quantize_mode=nvfp4 --recipe=w4a16_nvfp4 fallback.
• Applies the same compatibility check to the ONNX evaluation entry point.
• Fixes weight-only NVFP4 detection, initializer metadata placement, FP16 conversion, nested operator detection, and TensorRT 11 execution-context compatibility.
• Moves the torch_onnx CI lane to TensorRT 11 while keeping diffusers and torch_trt on TensorRT 10.
• Adds CPU unit coverage and Blackwell integration coverage for W4A4 and W4A16 builds.

Usage

# Dynamic W4A4 NVFP4 with TensorRT 11.0+
python examples/torch_onnx/torch_quant_to_onnx.py \
    --timm_model_name=vit_small_patch16_224 \
    --quantize_mode=nvfp4 \
    --onnx_save_path=vit_small_patch16_224.nvfp4.onnx \
    --trt_build

# Weight-only W4A16 fallback for TensorRT 10.16
python examples/torch_onnx/torch_quant_to_onnx.py \
    --timm_model_name=vit_small_patch16_224 \
    --quantize_mode=nvfp4 \
    --recipe=w4a16_nvfp4 \
    --onnx_save_path=vit_small_patch16_224.w4a16_nvfp4.onnx \
    --trt_build

Testing

• Ran the affected CPU unit suite with CUDA_VISIBLE_DEVICES="": 113 passed.
• Ran all applicable pre-commit hooks.
• Ran the complete tests/examples/torch_onnx suite on Blackwell with TensorRT 11.0: 25 passed.
• Verified the full W4A4 and W4A16 ViT TensorRT builds.
• Verified TensorRT 10.16 rejects dynamic W4A4 before model or calibration work and successfully builds and deserializes the documented W4A16 fallback.
• Verified the pretrained NVFP4 export, 512-sample calibration, TensorRT 11 build, 100-image evaluation, and engine deserialization workflow.

Before your PR is "Ready for review"

Make sure you read and follow Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded trust_remote_code=True, torch.load(..., weights_only=False), pickle, etc.).

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: N/A

Additional Information

Tracking: [6529944].

🤖 Generated by Codex (AI agent).

[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

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA.github.io/Model-Optimizer/pr-preview/pr-2333/
Built to branch gh-pages at 2026-09-04 16:00 UTC. Preview will be ready when the GitHub Pages deployment is complete.