cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_64_32_S_q_paged_kv_96_sm89.cubin.cpp (2)

1-3: Silence Clang/compile for LFS .cpp pointers to avoid false errors.

Static analysis errors are expected here since these files aren’t C++ sources. Exclude them from compilation/tidy.

Examples:

• CMake

# Mark LFS pointer stubs as non-compilable sources
file(GLOB CUBIN_POINTERS
  "${CMAKE_SOURCE_DIR}/cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/*.cubin.cpp")
set_source_files_properties(${CUBIN_POINTERS} PROPERTIES HEADER_FILE_ONLY ON)

• Or exclude the cubin directory from your clang-tidy/run-clang-tidy target (path filter/globs).

---

1-1: License header on .cpp (optional).

These .cpp pointer stubs lack the NVIDIA Apache-2.0 header. Either add a brief header comment or formally exempt this path from the header check to satisfy policy. As per coding guidelines.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_64_32_S_q_kv_128_sm89.cubin.cpp (1)

1-3: Optional: add a CMake guard to fail fast when LFS content isn’t present.

To avoid hard-to-read compiler errors, add a small configure-time check that errors if a .cubin.cpp still contains the LFS header.

Example snippet you can adapt:

# In the CMake for this target:
function(check_lfs_resolved src)
  file(READ "${src}" _contents LIMIT 64)
  string(FIND "${_contents}" "git-lfs.github.com/spec/v1" _pos)
  if(NOT _pos EQUAL -1)
    message(FATAL_ERROR "Unresolved Git LFS pointer detected in ${src}. Run: git lfs fetch --all && git lfs checkout")
  endif()
endfunction()

# Call for each *.cubin.cpp you add
check_lfs_resolved(${CMAKE_CURRENT_SOURCE_DIR}/cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_64_32_S_q_kv_128_sm89.cubin.cpp)

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_bf16_64_128_S_qkv_128_tma_ws_sm90.cubin.cpp (1)

1-3: Optional: document style exception for generated cubin sources and exclude from linters.

File naming deviates from lowerCamelCase and content is machine‑generated; consider:

• Documenting an exception for cubin-generated .cpp under cubin/ in coding guidelines.
• Excluding these from clang‑tidy/format checks and static analysis to reduce noise and build time.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_fp16_64_128_S_q_kv_128_softmax_tma_ws_sm90.cubin.cpp (1)

2-3: Binary blob pointer bumped — verify Git LFS smudge in CI.

No functional/code changes. Please ensure CI smudges LFS before compilation and fails if pointer text is present; clang errors shown by static analysis are false positives when analyzing the pointer file.

Use the validation script provided in a sibling comment to gate builds on missing LFS smudge.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_fp16_fp32_64_128_S_qkv_128_softcapping_sm90.cubin.cpp (1)

2-3: LFS-only update confirmed.

Good to go. Please keep the CI guard to detect any lingering LFS pointers pre-build and confirm .gitattributes tracks these files under LFS.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_64_32_S_qkv_128_sage_64_32_32_output_fp16_sm89.cubin.cpp (1)

2-3: Pointer refresh only; ensure packaging picks up new blobs.

Approve. In addition to the LFS CI guard, verify wheel/package/install rules still include this path so the updated cubin ships with release/1.1.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_bf16_64_128_S_q_paged_kv_128_softcapping_tma_ws_sm90.cubin.cpp (1)

2-3: Paged‑KV cubin pointer updated — run a quick smoke test.

Change is LFS-only. Please run a minimal paged‑KV FMHA smoke test (sm90) to confirm loadability and no version skew with the runtime loader, in addition to the LFS CI guard.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_64_32_S_qkv_80_sage_64_32_32_output_bf16_sm89.cubin.cpp (1)

2-3: Ada (sm89) variant pointer bump — LGTM.

Approve the LFS pointer bump. Keep the LFS placeholder check in CI to prevent accidental compilation of pointer text.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_64_256_S_qkv_128_alibi_tma_ws_sm90.cubin.cpp (1)

2-3: Add a pre-build guard to catch LFS pointer text early.

Consider a CI step that fails if any .cubin.cpp begins with the LFS header to avoid false clang errors.

#!/bin/bash
set -e
files=$(rg -l '^\s*version https://git-lfs.github.com/spec/v1' --glob '**/*.cubin.cpp' || true)
test -z "$files" || { echo "LFS pointers detected in:"; echo "$files"; exit 1; }

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_fp16_64_128_S_qkv_128_sm90.cubin.cpp (1)

2-3: Approved — LFS metadata only.

No code/API change. Ensure CI pulls LFS to avoid analyzing stubs.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_bf16_64_128_S_q_paged_kv_128_tma_ws_sm90.cubin.cpp (1)

2-3: LGTM: metadata-only change.

No runtime/control-flow changes. Ensure CI ignores .cubin.cpp files for C++ compilation and license headers to avoid false positives.

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/fmha_v2_flash_attention_e4m3_fp32_128_128_S_qkv_32_sm89.cubin.cpp (1)

2-2: Silence static-analysis false positives for LFS stubs.

The clang errors are expected for LFS pointer text. Exclude path cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/cubin/*.cubin.cpp from compilation/analysis or mark as HEADER_FILE_ONLY in CMake to keep CI green.