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

Super tiny fix typo (#8046)

Browse Source
This commit is contained in:
fzyzcjy
2025-07-17 21:15:51 +08:00
committed by GitHub
parent af1cc8fe2d
commit 519ff5c8e6
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/attention/flashattention_backend.py
View File

@@ -1617,7 +1617,7 @@ class FlashAttentionBackend(AttentionBackend):
metadata.max_seq_len_k + self.page_size - 1 metadata.max_seq_len_k + self.page_size - 1
) // self.page_size ) // self.page_size
normal_decode_set_medadata( normal_decode_set_metadata(
metadata.cache_seqlens_int32, metadata.cache_seqlens_int32,
metadata.cu_seqlens_k, metadata.cu_seqlens_k,
metadata.page_table, metadata.page_table,
@@ -1666,7 +1666,7 @@ class FlashAttentionBackend(AttentionBackend):
max_seq_pages = (max_len + self.page_size - 1) // self.page_size max_seq_pages = (max_len + self.page_size - 1) // self.page_size
metadata.max_seq_len_k = max_len metadata.max_seq_len_k = max_len
normal_decode_set_medadata( normal_decode_set_metadata(
metadata.cache_seqlens_int32, metadata.cache_seqlens_int32,
metadata.cu_seqlens_k, metadata.cu_seqlens_k,
metadata.page_table, metadata.page_table,
@@ -2089,7 +2089,7 @@ class FlashAttentionMultiStepBackend:
# @torch.compile(dynamic=True, backend=get_compiler_backend()) # @torch.compile(dynamic=True, backend=get_compiler_backend())
# TODO: fuse these kernels # TODO: fuse these kernels
# NOTE: torch.compile makes it slower in speculative decoding # NOTE: torch.compile makes it slower in speculative decoding
def normal_decode_set_medadata( def normal_decode_set_metadata(
cache_seqlens_int32: torch.Tensor, cache_seqlens_int32: torch.Tensor,
cu_seqlens_k: torch.Tensor, cu_seqlens_k: torch.Tensor,
page_table: torch.Tensor, page_table: torch.Tensor,