Skip to content

[fix] [AutoDeploy] flashinfer usage on H100 #4162

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
lucaslie merged 1 commit into from
May 8, 2025
Merged

[fix] [AutoDeploy] flashinfer usage on H100 #4162

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
11 changes: 9 additions & 2 deletions tensorrt_llm/_torch/auto_deploy/custom_ops/flashinfer_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,8 +70,13 @@ def init_workspace(self, workspace_buffer: torch.Tensor):
self.__init__() # reset all state

self.workspace_buffer = workspace_buffer
# NOTE (lucaslie): flashinfer fa3 backend has accuracy issue + illegal memory access issues
# on H100 PCIe, see https://github.com/NVIDIA/TensorRT-LLM/pull/3686 and
# https://github.com/flashinfer-ai/flashinfer/issues/924
self.prefill_wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
self.workspace_buffer, "NHD"
self.workspace_buffer,
"NHD",
backend="fa2",
)
self.decode_wrapper = self._init_decode_wrapper()

Expand Down Expand Up @@ -383,7 +388,9 @@ def _get_cache(si: SequenceInfo):
@classmethod
def get_global_buffer_initializers(cls, source_attn_node: Node) -> BufferInitializerDict:
def _init_workspace(si: SequenceInfo) -> torch.Tensor:
buffer = torch.empty(128 * 1024 * 1024, dtype=torch.uint8, device=si.device)
# NOTE (lucaslie): avoid OOM for many cudagraphs,
# see https://github.com/NVIDIA/TensorRT-LLM/pull/3686
buffer = torch.empty(320 * 1024 * 1024, dtype=torch.uint8, device=si.device)
cls._get_planner().init_workspace(buffer)
return buffer

Expand Down