Skip to content

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

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
hmellor merged 1 commit into from
Mar 16, 2026
Merged

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

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
12 changes: 0 additions & 12 deletions vllm/model_executor/models/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,6 @@

from vllm.logger import init_logger
from vllm.model_executor.models import ModelRegistry
from vllm.platforms import current_platform
from vllm.utils.math_utils import cdiv, round_up
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.attention.backends.registry import AttentionBackendEnum
Expand Down Expand Up @@ -148,17 +147,6 @@ def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
).page_size_bytes
else:
kernel_block_alignment_size = 16
if (
current_platform.is_device_capability_family(100)
and model_config.get_head_size() == 256
and (
attention_config.backend is None
or attention_config.backend == AttentionBackendEnum.FLASHINFER
)
):
# https://github.com/flashinfer-ai/flashinfer/issues/1993 reports that`
# head size 256 and block size 16 is not supported on blackwell.
kernel_block_alignment_size = 32
attn_page_size_1_token = FullAttentionSpec(
block_size=1,
num_kv_heads=model_config.get_num_kv_heads(parallel_config),
Expand Down
9 changes: 0 additions & 9 deletions vllm/v1/attention/backends/flashinfer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -630,15 +630,6 @@ def __init__(
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:
Expand Down
Loading