[None][fix] Honor MSA block-index strides for q_len <= 32

[peihu-nv]

@coderabbitai summary

Description

MiniMax-M3's legacy MSA sparse-attention path for q_len <= 32 can receive
kv_block_indexes as a non-contiguous, head-major tensor view. The SM100
kernel treated the underlying pointer as a contiguous
[query, KV head, top-k] allocation, which could select incorrect KV pages
and produce incorrect or non-finite attention output.

This change passes the tensor's query, KV-head, and top-k strides through the
existing MSA native interface and uses them for every sparse block-index load.
It also updates the bounds-check extent for the strided view. The attention
algorithm is unchanged, and no materialization, new kernel, or additional
buffer is introduced. The feature branch's existing downstream MSA patch and
TensorRT-LLM-collected test carry the change until it lands in MSA directly.

Test Coverage

• Focused SM100 regression on GB300: contiguous and non-contiguous head-major
  selectors are finite and bit-identical for q_len 1 and 5 with one and two
  KV splits (4/4 combinations passed).
• Five matched full AgentX pairs on GB300 using the MiniMax-M3 feature stack:
  all ten 1,800-second, default-sampling profiles completed validly with zero
  serving errors and no non-finite marker.
• Focused CUDA timing measured 23.284 us for stride-aware attention and
  23.361 us for the compact reference; avoiding materialization saved
  11.161 us in the selector-plus-consumer path.
• The combined downstream patch applies cleanly to the pinned MSA commit, and
  the six patched native-source hashes match the E2E-tested implementation.

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 #63953 [ run ] triggered by Bot. Commit: 02db4d9 Link to invocation

[tensorrt-cicd]

PR_Github #63953 [ run ] completed with state SUCCESS. Commit: 02db4d9
/LLM/main/L0_MergeRequest_PR pipeline #51887 completed with status: 'SUCCESS'

CI Report

Link to invocation