[None][test] keep MXFP4 block-scale unswizzle on GPU in create_weights

[xxi-nv]

Summary

• Remove redundant .cpu() before block_scale_interleave_reverse in MXFP4MXFP8QuantizeUtil.create_weights (and inherited MXFP4FP8QuantizeUtil).
• fp4_quantize already produces GPU block scales; block_scale_interleave_reverse supports CUDA tensors, matching W4A8NVFP4FP8QuantizeUtil.

Test plan

• OCI GB200: test_moe_backend filtered to W4A8_MXFP4_* quant algos
• CI

Summary by CodeRabbit

• Tests
  • Improved quantization test coverage for block-scale handling.
  • Updated scale reshaping behavior to better match expected results during weight processing.

[coderabbitai]

📝 Walkthrough

Walkthrough

In the MXFP4 test quantization utility, the weight creation logic for w1/w2/w3 block-scale tensors was changed to reshape directly on the current device, removing intermediate .cpu() calls before view().

Changes

MXFP4 Scale Reshape Update

Layer / File(s)	Summary
Remove CPU move before reshape tests/unittest/_torch/modules/moe/quantize_utils.py	w1/w2/w3_sf_block_unswizzled tensors are now reshaped directly via block_scale_interleave_reverse(...).view(...) without first moving to CPU.

Estimated code review effort: 1 (Trivial) | ~3 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#15714: Also modifies the same MoE MXFP4 quantization test utility to adjust scale tensor handling and reshaping for GPU processing.

Suggested reviewers: leslie-fang25

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly reflects the GPU unswizzle change in create_weights and is concise.
Description check	✅ Passed	The description summarizes the change and test plan, and only omits the optional PR Checklist section.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[xxi-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #57904 [ run ] triggered by Bot. Commit: c517357 Link to invocation

[tensorrt-cicd]

PR_Github #57904 [ run ] completed with state FAILURE. Commit: c517357
/LLM/main/L0_MergeRequest_PR pipeline #46598 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[xxi-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #57939 [ run ] triggered by Bot. Commit: c517357 Link to invocation

[tensorrt-cicd]

PR_Github #57939 [ run ] completed with state SUCCESS. Commit: c517357
/LLM/main/L0_MergeRequest_PR pipeline #46624 completed with status: 'SUCCESS'

CI Report

Link to invocation