xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 18:25:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix][Qwen][DCA] fixes bug in dual-chunk-flash-attn backend for qwen 1m models. (#21364)

Browse Source 
Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
This commit is contained in:
Tao He 2025-07-23 21:34:37 +08:00 committed by GitHub
parent f59ec35b7f
commit 7c734ee09b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/attention/backends/dual_chunk_flash_attn.py
View File

@ -1055,7 +1055,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_intra, v_states_intra,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=True, causal=True,
block_table=block_table,
stage="intra", stage="intra",
vertical_indices=vertical_buffer, vertical_indices=vertical_buffer,
slash_indices=slash_buffer, slash_indices=slash_buffer,
@ -1070,7 +1069,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_intra, v_states_intra,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=True, causal=True,
block_table=block_table,
stage="intra", stage="intra",
vertical_indices=intra_vertical_indices, vertical_indices=intra_vertical_indices,
slash_indices=intra_slash_indices, slash_indices=intra_slash_indices,
@ -1085,7 +1083,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_succ, v_states_succ,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=False, causal=False,
block_table=block_table,
stage="succ", stage="succ",
vertical_indices=succ_vertical_buffer, vertical_indices=succ_vertical_buffer,
slash_indices=succ_slash_buffer, slash_indices=succ_slash_buffer,
@ -1100,7 +1097,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_succ, v_states_succ,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=False, causal=False,
block_table=block_table,
stage="succ", stage="succ",
vertical_indices=succ_vertical_indices, vertical_indices=succ_vertical_indices,
slash_indices=succ_slash_indices, slash_indices=succ_slash_indices,
@ -1115,7 +1111,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_inter, v_states_inter,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=False, causal=False,
block_table=block_table,
stage="inter", stage="inter",
vertical_indices=inter_vertical_buffer, vertical_indices=inter_vertical_buffer,
slash_indices=inter_slash_buffer, slash_indices=inter_slash_buffer,
@ -1130,7 +1125,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
v_states_inter, v_states_inter,
softmax_scale=softmax_scale, softmax_scale=softmax_scale,
causal=False, causal=False,
block_table=block_table,
stage="inter", stage="inter",
vertical_indices=inter_vertical_indices, vertical_indices=inter_vertical_indices,
slash_indices=inter_slash_indices, slash_indices=inter_slash_indices,
@ -1151,7 +1145,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
value_states: torch.Tensor, value_states: torch.Tensor,
softmax_scale: float, softmax_scale: float,
causal: bool = True, causal: bool = True,
block_table: torch.Tensor = None,
max_seqlen_k: Optional[int] = None, max_seqlen_k: Optional[int] = None,
stage: str = "intra", stage: str = "intra",
vertical_indices: Optional[torch.Tensor] = None, vertical_indices: Optional[torch.Tensor] = None,
@ -1230,7 +1223,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
device=query_states.device), device=query_states.device),
max_seqlen_k=max_seqlen_k, max_seqlen_k=max_seqlen_k,
causal=causal, causal=causal,
block_table=block_table.unsqueeze(0),
return_softmax_lse=True, return_softmax_lse=True,
) )
softmax_lse = softmax_lse.view(q_len, q_heads, 1).transpose(0, softmax_lse = softmax_lse.view(q_len, q_heads, 1).transpose(0,