[https://nvbugs/6432948][fix] Exclude TRTLLM-Gen small tileN (8/16) for all FP8 block-scale MoE tactic selection

[trtllm-agent]

Description

The small-tile (tileN 8/16) dynB TRTLLM-Gen batched-GEMM cubins flakily hit an illegal memory access
(garbage TMA-descriptor pointer, MMU fault in the gemm2 K-loop). tileN >= 32 is unaffected
(10/10 clean vs. minutes-to-crash baseline on B300 TP=4).

The existing workaround excluded those tiles only when shared experts were fused into the grouped GEMM
(num_fused_shared_experts > 0). That scoping was wrong: the defect is in the shared small-tile cubins
and is not caused by expert fusion. DeepSeek-R1 FP8 TP=8 (unfused) faults identically during warmup,
where the 1/2/8-token shapes are the only ones that can select tileN 8/16 (12288 tokens gets tileN
64/128 and always passes). The same was already observed in #15297, where the IMA reproduced with
num_fused_shared_experts=0 and cuda-gdb pointed at a tileN-8 dynB_sm100f cubin faulting on
UTMALDG.4D.

Changes:

• Hoist the threshold into a single moeMinTileN() accessor built on common::getIntEnv, replacing two
  independently parsed function-local statics that both used unchecked std::atoi (an unparseable value
  silently disabled the WAR). The knob is renamed to TLLM_MOE_MIN_TILEN, with
  TLLM_MOE_FUSED_MIN_TILEN kept as a deprecated alias since the exclusion is no longer scoped to the
  fused path.
• Precompute mEligibleTileN in the ctor and drive both tactic selection and the tileN heuristic from
  it. Running computeSelectedTileN on the eligible list rather than on mSupportedTileN keeps the
  excluded tiles from consuming the returned neighbourhood: a shape whose heuristic tile is 8 now gets
  {32, 64, 128} instead of being left with 32 as its only candidate.
• Reject a threshold that excludes every supported tile at construction time instead of silently
  returning an empty tactic list.
• Include the problem dimensions in the fused fallback's no-valid-config error so a report from an
  unchecked model is actionable without a repro.

The unfused fallback keeps getDefaultValidConfigIndex. An earlier revision of this PR routed it
through getValidConfigIndices(...).front() as well; that was reverted because the two differ in
ordering, not in validation -- getDefaultValidConfigIndex returns the first entry of the list sorted
by the perf heuristic in KernelRunner.cpp:540-598, whereas getValidConfigIndices(...).front() takes
the first valid pair in raw cartesian order. Switching it would have changed the selected kernel config
for all default-path traffic, which is unrelated to this fix. This PR is now scoped to widening the
tileN exclusion only.

This is a runtime-side workaround; it can be reverted once the kernel-side fix lands.

Test Coverage

No new tests -- this changes tactic selection only and is covered by the existing FP8 block-scale MoE
tests (test_trtllm_fp8_block_scales* in tests/unittest/_torch/modules/moe/test_moe_backend.py, on
l0_b200 / l0_b300).

Crash-freedom:

• DeepSeek-R1 FP8 TP=8 (unfused): previously faulted during warmup, now passes.
• B300 TP=4 with fused shared experts: 10/10 clean runs (unchanged from the previous WAR).

Cost of the exclusion, measured op-level on B200 with DeepSeek-R1 FP8 EP=8 shapes
(hidden=7168, intermediate=2048, num_experts=256, local_experts=32, top_k=8). Each (num_tokens, tileN)
combination runs in its own process; the number reported is the best config for that tile, i.e. what the
autotuner would converge to.

num_tokens	best {8,16}	best {32,64,128}	delta
1	34.8 us	34.7 us	-0.18%
2	34.3 us	34.8 us	+1.56%
4	35.4 us	36.4 us	+2.93%
8	38.1 us	39.1 us	+2.62%
16	47.7 us	50.2 us	+5.29%
32	50.5 us	50.7 us	+0.34%
64	54.9 us	54.5 us	-0.77%

Worst case is +5.3% at 16 tokens, with every other point within +/-3% and two points slightly faster.
Note that the small tiles did not fault during this sweep -- the IMA is flaky and reproduces with
cascaded runs over real captured token_ids -- so these numbers bound the cost of the WAR but say
nothing about its necessity.

TLLM_MOE_MIN_TILEN=0 restores the previous behaviour for A/B experiments.

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.

Links

• Bug: https://nvbugs/6432948

GitHub Bot Help

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

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

Walkthrough

FP8 block-scale MoE tactic selection now excludes undersized tiles for fused and unfused workloads. Fallback selection validates configurations, skips undersized tiles, prioritizes the heuristic tile, and errors when no valid configuration exists.

Changes

FP8 MoE tactic selection

Layer / File(s)	Summary
Universal tile filtering and fallback validation cpp/tensorrt_llm/thop/fp8BlockScaleMoe.cpp	The minimum tile-size filter now applies to fused and unfused workloads. Fallback selection uses validated configurations, skips undersized tiles, prioritizes the heuristic tile, and reports an error when no valid configuration exists.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Suggested reviewers: barry-delaney

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Out of Scope Changes check	⚠️ Warning	The stated objectives describe an unrelated FMHA JIT warmup flag, while the changes implement FP8 MoE tile exclusion.	Align the PR objectives with the implemented FP8 block-scale MoE workaround, or change the code and description to implement the stated warmup fix.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Title check	✅ Passed	The title clearly describes the FP8 block-scale MoE tileN exclusion and includes the valid bug reference and fix type.
Description check	✅ Passed	The description explains the defect, scope, implementation, testing, performance impact, fallback behavior, and checklist status.
Linked Issues check	✅ Passed	The PR references NVBugs 6432948 in both the title and the Links section.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[zhaoyangwang-nvidia]

/bot run

[tensorrt-cicd]

PR_Github #68644 [ run ] triggered by Bot. Commit: 47e9b5e Link to invocation

[tensorrt-cicd]

PR_Github #68644 [ run ] completed with state FAILURE. Commit: 47e9b5e
/LLM/main/L0_MergeRequest_PR pipeline #56048 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

[zhaoyangwang-nvidia]

The PR fix was not the root case, #18133 fix it, closed this PR.