[None][fix] Handle DeepSeek-V4 fused A scale shape

[lfr-0531]

@coderabbitai summary

Description

Fix DeepSeek-V4 NVFP4 expert checkpoint loading when the checkpoint fused A projection scale has the FP8 block-scale shape (16, 56), while the module can be initialized with a stale/mismatched (16, 16) weight_scale.

The loader now preserves the existing oversized slice-copy behavior for padded kv_a_proj_with_mqa modules, and rebuilds the weight_scale parameter plus tensor metadata only when the checkpoint scale matches the expected FP8 block-scale shape for the fused A projection.

Test Coverage

• Built TensorRT-LLM wheel: tensorrt_llm-1.3.0rc15-cp312-cp312-linux_x86_64.whl.
• Installed the built wheel, then installed the current source in editable mode for Python-layer validation.
• pre-commit run --files tensorrt_llm/_torch/models/modeling_deepseekv4.py tests/unittest/_torch/modeling/test_modeling_deepseekv4.py
• PYTHONNOUSERSITE=1 pytest tests/unittest/_torch/modeling/test_modeling_deepseekv4.py -k "fused_a_weight_scale or routed_moe_quant_config or weight_remap" -q (8 passed)
• Manually validated the NVFP4 checkpoint at /home/scratch.fanrongl_coreai/models/deepseek_v4/pro-nvfp4-experts-v3.5: checkpoint fused scale is (16, 56) and the loader helper rebuilds module weight_scale from (16, 16) to (16, 56).
• Full GPU model run was not started because all B300 GPUs on the shared machine were already occupied with ~144-146 GiB used per GPU.

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)

• 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.

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48405 [ run ] triggered by Bot. Commit: 42ad633 Link to invocation

[tensorrt-cicd]

PR_Github #48405 [ run ] completed with state SUCCESS. Commit: 42ad633
/LLM/main/L0_MergeRequest_PR pipeline #38208 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

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48478 [ run ] triggered by Bot. Commit: 42ad633 Link to invocation

[tensorrt-cicd]

PR_Github #48478 [ run ] completed with state SUCCESS. Commit: 42ad633
/LLM/main/L0_MergeRequest_PR pipeline #38274 completed with status: 'SUCCESS'

CI Report

Link to invocation

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48506 [ run ] triggered by Bot. Commit: 13bf121 Link to invocation

[tensorrt-cicd]

PR_Github #48506 [ run ] completed with state SUCCESS. Commit: 13bf121
/LLM/main/L0_MergeRequest_PR pipeline #38304 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

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48575 [ run ] triggered by Bot. Commit: 13bf121 Link to invocation

[tensorrt-cicd]

PR_Github #48575 [ run ] completed with state SUCCESS. Commit: 13bf121
/LLM/main/L0_MergeRequest_PR pipeline #38362 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

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48733 [ run ] triggered by Bot. Commit: 13bf121 Link to invocation

[tensorrt-cicd]

PR_Github #48733 [ run ] completed with state ABORTED. Commit: 13bf121
/LLM/main/L0_MergeRequest_PR pipeline #38502 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

Link to invocation

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48771 [ run ] triggered by Bot. Commit: 13bf121 Link to invocation

[tensorrt-cicd]

PR_Github #48771 [ run ] completed with state FAILURE. Commit: 13bf121
/LLM/main/L0_MergeRequest_PR pipeline #38538 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

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48831 [ run ] triggered by Bot. Commit: 13bf121 Link to invocation

[tensorrt-cicd]

PR_Github #48831 [ run ] completed with state SUCCESS. Commit: 13bf121
/LLM/main/L0_MergeRequest_PR pipeline #38591 completed with status: 'SUCCESS'

CI Report

Link to invocation