Skip to content

update vllm kernel benchmark scripts - #562

Open
1pikachu wants to merge 2 commits into
vllm-project:mainfrom
1pikachu:main
Open

update vllm kernel benchmark scripts#562
1pikachu wants to merge 2 commits into
vllm-project:mainfrom
1pikachu:main

Conversation

@1pikachu

@1pikachu 1pikachu commented Sep 1, 2026

Copy link
Copy Markdown
Collaborator

Update the FA decode test scope for trend analysis

Signed-off-by: jundu <jun.du@intel.com>
Copilot AI lite review requested due to automatic review settings September 1, 2026 09:27
@1pikachu

1pikachu commented Sep 1, 2026

Copy link
Copy Markdown
Collaborator Author

trigger benchmark

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Updates the benchmark configuration generator for CUTLASS FlashAttention decode performance so trend analysis can be run over a more structured, uniform KV-length sweep across representative decode shapes.

Changes:

  • Replaces the previous ad-hoc seq_lens list with a uniform KV-length sweep (kv_len_sweep) across multiple batch sizes.
  • Introduces a curated set of decode shapes (hardcoded + derived from model_lists) and filters configs to fit a KV-cache memory budget.
  • Makes config generation deterministic by de-duplicating via sets and sorting by (num_heads, head_size, block_size, batch, kv_len).

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment on lines +311 to +313
# Size the pool so no block is aliased within the batch,
# otherwise KV reads hit cache and inflate bandwidth.
blocks_per_seq = (kv_len + block_size - 1) // block_size
@1pikachu

1pikachu commented Sep 2, 2026

Copy link
Copy Markdown
Collaborator Author

trigger benchmark

@1pikachu 1pikachu added the need_benchmark will trigger a benchmark run. label Sep 2, 2026
@1pikachu 1pikachu removed the need_benchmark will trigger a benchmark run. label Sep 4, 2026
@1pikachu

1pikachu commented Sep 4, 2026

Copy link
Copy Markdown
Collaborator Author

trigger benchmark

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants