[None][test] Handle MiniMax-M3 FP8 rounding variance

[peihu-nv]

@coderabbitai summary

Description

The MiniMax-M3 FP8 indexer and direct KV-insertion tests compared Q/K outputs byte-for-byte with the general fused QK-norm/RoPE reference. The specialized kernels use powf for RoPE frequencies, while the optimized reference uses the mathematically equivalent exp2f/log2f formulation. A small number of values can consequently round to adjacent E4M3 values and fail byte equality despite satisfying the intended FP8 numerical contract.

Compare only normalized/RoPE Q and K numerically using the tolerance already established by the horizontal-producer test. V conversion, cache destinations, guard pages, and CUDA Graph replay updates remain exact. Add the standalone indexer, direct KV-insertion, and active horizontal-producer tests to the B200 MiniMax-M3 pre-merge test list.

Test Coverage

• Focused B300 validation reproduced every reported shape and confirmed that Q/K differences remain within the established FP8 tolerance; V outputs and cache guard pages remain byte-exact, with no NaNs.
• CUDA Graph replay validation confirms outputs update from replay inputs and cache writes remain correct.
• Changed-files pre-commit checks pass, including Python formatting/linting, YAML validation, and test-list entry validation.

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

[tensorrt-cicd]

PR_Github #63569 [ run ] triggered by Bot. Commit: a427a7e Link to invocation

[tensorrt-cicd]

PR_Github #63569 [ run ] completed with state SUCCESS. Commit: a427a7e
/LLM/main/L0_MergeRequest_PR pipeline #51534 completed with status: 'SUCCESS'

CI Report

Link to invocation

[brb-nv]

LGTM.