EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix fa3 cuda graph page_size > 1 precision and page_size=1 speed (#4855)

Browse Source
This commit is contained in:
Qingquan Song
2025-03-28 01:35:59 -07:00
committed by GitHub
parent 47e6628aae
commit 6ffb6bd47a
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
python/sglang/srt/layers/attention/flashattention_backend.py
View File

@@ -322,10 +322,13 @@ class FlashAttentionBackend(AttentionBackend):
torch.cumsum(seq_lens_in_batch, dim=0, dtype=torch.int32), (1, 0)
)
metadata.page_table = self.req_to_token[
:, self.decode_cuda_graph_metadata["strided_indices"]
max_seq_pages = (metadata.max_seq_len_k + self.page_size - 1) // self.page_size
page_indices = self.req_to_token[
:, self.decode_cuda_graph_metadata["strided_indices"][:max_seq_pages]
]
metadata.page_table = metadata.page_table[req_pool_indices[:bs]]
page_indices = page_indices[req_pool_indices[:bs]] // self.page_size
metadata.page_table[:, :max_seq_pages].copy_(page_indices)
metadata.page_table[:, max_seq_pages:].fill_(0)
self.forward_metadata = metadata
def get_cuda_graph_seq_len_fill_value(self):