Skip to content

[Perf] Ship the qualified SM70 long-context layout and drop its context fences #609

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yangzhuxinyzx merged 2 commits into from
Sep 12, 2026
Merged

[Perf] Ship the qualified SM70 long-context layout and drop its context fences #609

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c38d180
[Perf] Ship the qualified SM70 long-context layout and drop its conte…
yangzhuxinyzx Sep 12, 2026
9461060
[Perf] Route the SM70 long-context attention on the served window, no…
yangzhuxinyzx Sep 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion .pre-commit-config.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,10 @@ repos:
rev: v21.1.2
hooks:
- id: clang-format
exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
# The sm70_grouped_long kernels are ported verbatim from qualified native
# candidates and are checked against the source digests recorded in their
# manifests, so they stay byte-comparable instead of being reformatted.
exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh)|attention/sm70_grouped_long/kernel/(grouped|scalar)-attention.cu)|vllm/third_party/.*'
types_or: [c++, cuda]
args: [--style=file, --verbose]
- repo: https://github.com/DavidAnson/markdownlint-cli2
Expand Down
3 changes: 2 additions & 1 deletion cmake/external_projects/vllm_flash_attn.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,8 @@ if(VLLM_FLASH_ATTN_SM70 AND TARGET _vllm_fa2_C)
set(SM70_GROUPED_LONG_DIR
"${CMAKE_CURRENT_LIST_DIR}/../../csrc/attention/sm70_grouped_long")
set(SM70_GROUPED_LONG_SRC
"${SM70_GROUPED_LONG_DIR}/kernel/grouped-attention.cu")
"${SM70_GROUPED_LONG_DIR}/kernel/grouped-attention.cu"
"${SM70_GROUPED_LONG_DIR}/kernel/scalar-attention.cu")
# Flags mirror the manifest the operator was qualified with. As with v37, the
# properties must be set in the target scope or SM70 silently loses them.
set_source_files_properties(${SM70_GROUPED_LONG_SRC}
Expand Down
1,001 changes: 835 additions & 166 deletions csrc/attention/sm70_grouped_long/kernel/grouped-attention.cu
View file
Open in desktop

Large diffs are not rendered by default.

Loading
Loading