[None][perf] Allow CUDA-core NVFP4 GEMM for M up to 16

[mihai-chiorean]

Summary

• Increase CUDA-core NVFP4 GEMM autotuner eligibility from M <= 8 to M <= 16, matching the C++ cudaCoreGemmTemplateMaxM limit.
• Update CUDA-core NVFP4 docs/errors and unit coverage for the M=16 case.
• This does not force CUDA-core selection: CUTLASS/cuBLASLt remain available for larger shapes, and cuda_core-only configurations still fail fast when a shape exceeds CUDA-core support.

Performance

Spark SM121, Qwen3.6 35B A3B NVFP4, no-MTP graph+overlap smoke:

• Baseline: 37.270 ms TPOT / 26.831 tok/s
• M<=16 temp patch: 33.501 ms TPOT / 29.850 tok/s
• Repeat M<=16 temp patch: 33.565 ms TPOT / 29.793 tok/s

Testing

• uvx pre-commit run --files tensorrt_llm/_torch/custom_ops/torch_custom_ops.py tests/unittest/_torch/thop/parallel/test_fp4_linear.py
• Spark SM121: pytest -q tests/unittest/_torch/thop/parallel/test_fp4_linear.py -k test_fp4_linear_cuda_core via TensorRT bindings bootstrap: 4 passed
• Spark SM121: pytest -q tests/unittest/_torch/thop/parallel/test_fp4_linear.py -k "test_fp4_gemm_bias_per_backend and cuda_core" via TensorRT bindings bootstrap: 1 passed, 3 skipped

Summary by CodeRabbit

• New Features

  • Expanded CUDA Core NVFP4 GEMM support for small-batch workloads with matrix dimension M up to 16.

• Bug Fixes

  • Updated FP4 GEMM validation to correctly allow supported M=16 workloads and bias parity checks.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 7d6dff96-9a11-42a9-857d-32eb73dfb0fa

📥 Commits

Reviewing files that changed from the base of the PR and between 54d484f and 8e5cce1.

📒 Files selected for processing (2)

• tensorrt_llm/_torch/custom_ops/torch_custom_ops.py
• tests/unittest/_torch/thop/parallel/test_fp4_linear.py

---

📝 Walkthrough

Walkthrough

CUDA Core NVFP4 GEMM eligibility is expanded from M ≤ 8 to M ≤ 16. Runtime documentation, the shape limit, and FP4 linear tests are updated to cover the larger dimension.

Changes

NVFP4 CUDA Core eligibility

Layer / File(s)	Summary
Expand CUDA Core limit and validation tensorrt_llm/_torch/custom_ops/torch_custom_ops.py, tests/unittest/_torch/thop/parallel/test_fp4_linear.py	The CUDA Core limit and GEMM documentation now specify M <= 16; tests add M=16 coverage and relax the bias parity skip threshold.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: hyukn, bo-nv, xuantengh, pengbowang-nv, nv-guomingz

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 75.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the repository template and clearly summarizes the main change to CUDA-core NVFP4 GEMM eligibility.
Description check	✅ Passed	The description follows the template well with summary, performance, and testing sections that explain what changed and how it was verified.
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.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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