[TRTLLM-12720][feat] Support nvfp4 w4a16 on sm120

[pamelap-nvidia]

Dev Engineer Review

The changes add SM120/121 PyTorch support for W4A16 NVFP4 inference across dense linear, LM-head, and MoE paths.

• Adds CUDA-core, Triton, Marlin, CUTLASS, and CuteDSL backend support.
• Updates Marlin build and runtime guards for SM90 and SM120.
• Adds compressed-tensors and ModelOpt configuration handling.
• Adds checkpoint mapping for packed weights and global scales.
• Adds backend-specific derived-state caching and Marlin weight transformation.
• Updates MoE backend selection and W4A16 NVFP4 scale handling.
• Restricts ReLU2 fusion to eligible static NVFP4 activation inputs.
• Clarifies NVFP4 activation-quantization checks in DeepSeek models.
• Adds SM12x support to Marlin GEMM test gating.

The reported local validation passes 74 merge-focused tests and 69 W4A16/model regression tests. Six CI runs failed or had L0 pipeline failures. Failure analysis and a new NVIDIA-triggered run are required.

QA Engineer Review

The PR adds test functions and coverage for:

• Nemotron H MoE, MTP, and checkpoint mapping.
• CuteDSL B12x W4A16 NVFP4 backend selection and scale transformation.
• W4A16 NVFP4 linear, Marlin, Triton, LM-head, fusion, shape, and quantization behavior.
• Model configuration and compressed-tensors parsing.
• Marlin SM90/SM12x GEMM gating.
• MoE activation-quantization behavior.

No files under tests/integration/test_lists/, test-db/, qa/, or waives.txt were modified. Coverage registration for the added test functions is not shown.

Verdict: needs follow-up.

Summary

Adds end-to-end support for W4A16 NVFP4 checkpoints (packed FP4 weights + per-group FP8
scales, high-precision activations) in the PyTorch backend, targeting Qwen3.5/3.6 and
Nemotron-H NVFP4 exports on SM120/121 (RTX 5090/6000, GB10) and Hopper.

The core problem this solves: TRT-LLM previously conflated "has NVFP4 weights" with
"quantizes activations to NVFP4". W4A16 breaks that assumption — it has FP4 weights but
consumes BF16/FP16 activations and has no calibrated input_scale. Every FP4 fusion site
keyed on has_nvfp4 would therefore feed an Fp4QuantizedTensor into a layer that can't
consume one, or read a None activation scale.

Changes

Activation-quantization capability, made explicit

• New LinearMethodBase.quantizes_nvfp4_activations ClassVar, surfaced as
  Linear.has_nvfp4_activation_quantization and MoE.has_nvfp4_activation_quantization.

Linear

• W4A16NVFP4LinearMethod: on-the-fly Triton weight dequant. Leaves input_scale /
  inv_input_scale / alpha as None from create_weights onward; the parent
  process_weights_after_loading_* now skip scales whose destination was never allocated.
• MarlinNVFP4LinearMethod: Marlin-backed W4A16, selected by default for BF16 on
  SM90/120/121. Handles its own K->64 / N->128 padding.
• Per-block scales stay in swizzled layout; the unswizzled view is a derived non-persistent
  buffer, which also makes the GMS cache_derived_state refresh idempotent (the previous
  in-place unswizzle would double-apply on re-materialization).

MoE

• CuteDslB12xFusedMoE accepts W4A16_NVFP4 and stays on the b12x path for both prefill
  and decode (quant_mode="w4a16"); falls back to CUTLASS for EP / attention-DP or
  non-SM120/121.
• W4A16NVFP4CutlassFusedMoEMethod on the CUTLASS path; shared per-layer quant-config
  lookup so mixed-precision checkpoints resolve expert configs correctly.
• Nemotron-H: promotes experts to W4A4 NVFP4 on non-CuteDSL backends, and routes BF16 MTP
  body sublayers to CUTLASS.

C++

• Build the dense Marlin NVFP4 GEMM/repack kernels for 120f in addition to 90; relax the
  runtime SM asserts to 9.x || 12.x. No kernel logic changes.

No LLM API surface change — llm_args.py is untouched, so no golden-manifest
regeneration.

Behavior changes for existing users

• Hopper NVFP4 BF16 linears now use the Marlin GEMM backend by default (previously opt-in
  via nvfp4_gemm_config.allowed_backends: [..., marlin]). Attention FP4 output and the
  torch.compile inplace op are disabled on those layers as a consequence.
• MoE backend AUTO resolves to CUTEDSL for W4A16 NVFP4 on SM120/121.

Test coverage

test_w4a16_nvfp4_linear.py (new) covers method selection across SM90/100/120/121 and dtype,
the Marlin-vs-Triton numerical match on GPU (including K/N padding shapes), scale-cache buffer
semantics, the mixed-precision use_fused_gemm_allreduce rebind, and every fusion gate that
changed. Plus new checkpoint-parsing tests (test_model_config.py,
test_quant_config_utils.py, test_kv_cache_dtype_override.py), Nemotron-H MoE/MTP config
tests, the weight-mapper tests, and W4A16 cases added to the b12x MoE backend suite.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[pamelap-nvidia]

/bot run

[pamelap-nvidia]

/bot run

[pamelap-nvidia]

/bot run

[tensorrt-cicd]

PR_Github #64397 [ run ] triggered by Bot. Commit: 7882e58 Link to invocation

[tensorrt-cicd]

PR_Github #64397 [ run ] completed with state FAILURE. Commit: 7882e58
/LLM/main/L0_MergeRequest_PR pipeline #52286 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

[pamelap-nvidia]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #64434 [ run ] triggered by Bot. Commit: 7882e58 Link to invocation

[pamelap-nvidia]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #64502 [ run ] triggered by Bot. Commit: 8316a98 Link to invocation

[tensorrt-cicd]

PR_Github #64434 [ run ] completed with state ABORTED. Commit: 7882e58

Link to invocation

[tensorrt-cicd]

PR_Github #64502 [ run ] completed with state FAILURE. Commit: 8316a98
/LLM/main/L0_MergeRequest_PR pipeline #52382 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

[Wanli-Jiang]

/bot run --disable-fail-fast

[pamelap-nvidia]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #64623 [ run ] triggered by Bot. Commit: 8316a98 Link to invocation

[tensorrt-cicd]

PR_Github #64623 [ run ] completed with state SUCCESS. Commit: 8316a98
/LLM/main/L0_MergeRequest_PR pipeline #52483 completed with status: 'UNSTABLE'

CI Report

⚠️ Multi-GPU Label Required:
Multi-GPU tests require the ci: full pre-merge approved label on this PR. Ask a member of NVIDIA/trt-llm-ci-approvers to add the label, then re-trigger CI with the same bot command (no rebase needed).

⚠️ 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

Link to invocation

[pamelap-nvidia]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #64669 [ run ] triggered by Bot. Commit: 8316a98 Link to invocation

[tensorrt-cicd]

PR_Github #64669 [ run ] completed with state FAILURE. Commit: 8316a98
/LLM/main/L0_MergeRequest_PR pipeline #52527 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

[pamelap-nvidia]

/bot run

[tensorrt-cicd]

PR_Github #64707 [ run ] triggered by Bot. Commit: 8316a98 Link to invocation

[tensorrt-cicd]

PR_Github #64707 [ run ] completed with state SUCCESS. Commit: 8316a98
/LLM/main/L0_MergeRequest_PR pipeline #52563 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

[schetlur-nv]

/bot skip --comment "Failure analysis: the failing SBSA multi-GPU Slurm job completed successfully with exit code 0:0; the pipeline failed in CI Slurm log-tracking/teardown after tail -f lost its transport endpoint and kill $tailPid errored. https://pbss.s8k.io/v1/AUTH_svc_tensorrt/sw-tensorrt-ci-analysis/LLM/main/L0_MergeRequest_PR/52563/failure_analysis.html "

[tensorrt-cicd]

PR_Github #64716 [ skip ] triggered by Bot. Commit: 8316a98 Link to invocation

[tensorrt-cicd]

PR_Github #64716 [ skip ] completed with state SUCCESS. Commit: 8316a98
Skipping testing for commit 8316a98

Link to invocation