biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[FlashInfer] Revert block_size 16 + head_size 256 workaround on Blackwell (#36987)

Browse Source 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
This commit is contained in:
Vadim Gimpelson
2026-03-16 13:04:29 +04:00
committed by GitHub
parent 912fbe9555
commit 8374387bd8
2 changed files with 0 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/attention/backends/flashinfer.py
View File

@@ -630,15 +630,6 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
self.paged_kv_indices = self._make_buffer(max_num_pages)
self.paged_kv_last_page_len = self._make_buffer(max_num_reqs)
if self.head_dim == 256 and current_platform.is_device_capability_family(100):
# https://github.com/flashinfer-ai/flashinfer/issues/1993 reports that
# head size 256 and block size 16 is not supported on blackwell.
assert kv_cache_spec.block_size != 16, (
"There is a bug in FlashInfer "
"block_size 16 head size 256 support. Please avoid this combination by "
"passing --block-size 32 or --block-size 64."
)
def _make_buffer(
self, *size: int | torch.SymInt, dtype: torch.dtype = torch.int32
) -> CpuGpuBuffer: