[peihu-nv]

Description

Adds GB300 (sm_103) MiniMax-M3.1 low-latency decode targets and shared native kernels, including guarded DSpark integration. Base: feat/minimax-m3.1-flash at 98660c55. This draft contains one signed-off commit, 40ca75b4b2.

TP8 fused decode covers K7, concurrency 1, up to 8 flattened rows; TP4 covers up to 32 rows. Supported paths default on and expose per-item TRTLLM_M31_DISABLE_* controls, including TOPP_DET, SOFTMAX_MC, REJ_MC, and MCAG. Mixed/context and larger-row iterations use the target's supported fallback paths; piecewise execution is not supported. Scope is TP8 and TP4 decode only.

Historical K7 measurements were 1.8344× the M3-formula SOL for TP8 (3,809.74 µs) and 1.6271× for TP4 (4,722.92 µs), where lower is better. These are historical measurements, not results from this commit. New-base build, full-model parity, per-rank engagement, determinism, synth and E2E validation remain outstanding. In particular, GB multicast-gather determinism remains unverified; component gates must not be read as a blanket bitwise-parity claim.

Test Coverage

• Passed 14 local source-routing/scope tests: test_m31_decode_scope_source.py, test_m31_tp4_routing_source.py, and test_m31_tp8_routing_source.py.
• Python AST checks and applicable changed-file pre-commit checks passed, including Ruff, formatting and source test-list checks. The CI-equivalent lightweight mypy check passes all 17 configured source files; full type checking with compiled bindings remains pending.
• Internal campaign/job references were removed. Restored the TP4 MoE kernel required by its existing loader from its original source; the loader's executable AST matches that source.
• The source build and GPU validation are in progress. Review is pending; this draft is not merge-ready.

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.

[peihu-nv]

/bot run

[peihu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80276 [ run ] triggered by Bot. Commit: b969ed5 Link to invocation

[peihu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80279 [ run ] triggered by Bot. Commit: 40ca75b Link to invocation

[tensorrt-cicd]

PR_Github #80280 [ run ] triggered by Bot. Commit: 40ca75b Link to invocation

[tensorrt-cicd]

PR_Github #80276 [ run ] completed with state ABORTED. Commit: b969ed5

Link to invocation

[tensorrt-cicd]

PR_Github #80279 [ run ] completed with state ABORTED. Commit: 40ca75b

Link to invocation

[tensorrt-cicd]

PR_Github #80280 [ run ] completed with state FAILURE. Commit: 40ca75b
/LLM/main/L0_MergeRequest_PR pipeline #66073 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

[peihu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80296 [ run ] triggered by Bot. Commit: 647ee5c Link to invocation

[tensorrt-cicd]

PR_Github #80296 [ run ] completed with state SUCCESS. Commit: 647ee5c
/LLM/main/L0_MergeRequest_PR pipeline #66086 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

[peihu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80336 [ run ] triggered by Bot. Commit: 1113060 Link to invocation

[tensorrt-cicd]

PR_Github #80336 [ run ] completed with state FAILURE. Commit: 1113060
/LLM/main/L0_MergeRequest_PR pipeline #66121 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

[peihu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80405 [ run ] triggered by Bot. Commit: 9d97565 Link to invocation

[tensorrt-cicd]

PR_Github #80405 [ run ] completed with state FAILURE. Commit: 9d97565
/LLM/main/L0_MergeRequest_PR pipeline #66181 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