xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 17:55:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[V1] Support sliding window attention (#9679)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2024-10-24 22:20:37 -07:00 committed by GitHub
parent a6f3721861
commit 9645b9f646
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/v1/attention/backends/flash_attn.py
View File

@ -82,8 +82,10 @@ class FlashAttentionImpl(AttentionImpl):
if alibi_slopes is not None:
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.alibi_slopes = alibi_slopes
self.sliding_window = ((sliding_window, sliding_window)
if sliding_window is not None else (-1, -1))
if sliding_window is None:
self.sliding_window = (-1, -1)
else:
self.sliding_window = (sliding_window - 1, 0)
self.kv_cache_dtype = kv_cache_dtype
if logits_soft_cap is None:
# In flash-attn, setting logits_soft_cap as 0 means no soft cap.
@ -93,12 +95,6 @@ class FlashAttentionImpl(AttentionImpl):
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
if sliding_window is not None:
# NOTE(woosuk): flash-attn's sliding window does not work with
# paged KV cache.
raise ValueError(
"Sliding window is not supported in FlashAttention.")
support_head_sizes = FlashAttentionBackend.get_supported_head_sizes()
if head_size not in support_head_sizes:
raise ValueError(