vLLM issue #59362: FA4 ignores num_splits on SM90 in v0.30.0, up to 48% slower decode
forum
Fields
id
sources/gotcha-vllm-issue-59362-fa4-num-splits-sm90Publisher
vLLM GitHub issues
Kind
forum
Published
2026-09-30
Accessed
2026-10-03
notes
Includes a kernel-only num_splits sweep (v0.26.0 385->20 us vs v0.30.0 flat ~360 us) and the full root-cause isolation table.
Notes
Includes a kernel-only num_splits sweep (v0.26.0 385->20 us vs v0.30.0 flat ~360 us) and the full root-cause isolation table.
Cross-references
This record references no other records.