[Perf] Integrate flash-maxsim Triton kernels for late-interaction scoring#40337

Open

roipony wants to merge 10 commits intovllm-project:mainfrom

roipony:flash-maxsim-integration

Commits on Apr 20, 2026

[Perf] Integrate flash-maxsim Triton kernels for late-interaction scoring
roi.pony
committed
[mypy] Fix type errors flagged by pre-commit
roi.pony
committed
[Lint] Comply with vllm import/API policies
roi.pony
committed
[Lint] Finish migrating side-effect torch.cuda APIs to torch.accelerator
roi.pony
committed
[Lint] ruff format + SPDX headers
roi.pony
committed
[Docs] Add return-type annotation on pack_pairs
roi.pony
committed
[Slim] Trim PR to the production path only (<1K LOC)
roi.pony
committed

Commits on Apr 21, 2026

[Docs] Add return-type annotation on pack_docs
roi.pony
committed

Commits on Apr 28, 2026

[Perf] Widen flash-maxsim warmup to cover Lq bucket 512/1024

roi.pony
and
claude
committed
[Bugfix] project_batch dtype-cast for non-Linear (Sequential) projectors

roi.pony
and
claude
committed