A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[Bugfix][Qwen][DCA] fixes bug in dual-chunk-flash-attn backend for qwen 1m models. (#21364)
Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
Tao He committed
Jul 23, 2025 at 21:34 UTC
7c734ee09b0a40681ba49d3d7ef5517ddb106074
1 file changed
-8
vllm/attention/backends/dual_chunk_flash_attn.py
-8
@@ -1055,7 +1055,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1055
v_states_intra,
1056
softmax_scale=softmax_scale,
1057
causal=True,
1058
- block_table=block_table,
1058
stage="intra",
1059
vertical_indices=vertical_buffer,
1060
slash_indices=slash_buffer,
@@ -1070,7 +1069,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1069
v_states_intra,
1070
softmax_scale=softmax_scale,
1071
causal=True,
1073
- block_table=block_table,
1072
stage="intra",
1073
vertical_indices=intra_vertical_indices,
1074
slash_indices=intra_slash_indices,
@@ -1085,7 +1083,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1083
v_states_succ,
1084
softmax_scale=softmax_scale,
1085
causal=False,
1088
- block_table=block_table,
1086
stage="succ",
1087
vertical_indices=succ_vertical_buffer,
1088
slash_indices=succ_slash_buffer,
@@ -1100,7 +1097,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1097
v_states_succ,
1098
softmax_scale=softmax_scale,
1099
causal=False,
1103
- block_table=block_table,
1100
stage="succ",
1101
vertical_indices=succ_vertical_indices,
1102
slash_indices=succ_slash_indices,
@@ -1115,7 +1111,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1111
v_states_inter,
1112
softmax_scale=softmax_scale,
1113
causal=False,
1118
- block_table=block_table,
1114
stage="inter",
1115
vertical_indices=inter_vertical_buffer,
1116
slash_indices=inter_slash_buffer,
@@ -1130,7 +1125,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1125
v_states_inter,
1126
softmax_scale=softmax_scale,
1127
causal=False,
1133
- block_table=block_table,
1128
stage="inter",
1129
vertical_indices=inter_vertical_indices,
1130
slash_indices=inter_slash_indices,
@@ -1151,7 +1145,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1145
value_states: torch.Tensor,
1146
softmax_scale: float,
1147
causal: bool = True,
1154
- block_table: torch.Tensor = None,
1148
max_seqlen_k: Optional[int] = None,
1149
stage: str = "intra",
1150
vertical_indices: Optional[torch.Tensor] = None,
@@ -1230,7 +1223,6 @@ class DualChunkFlashAttentionImpl(FlashAttentionImpl):
1223
device=query_states.device),
1224
max_seqlen_k=max_seqlen_k,
1225
causal=causal,
1233
- block_table=block_table.unsqueeze(0),
1226
return_softmax_lse=True,
1227
)
1228
softmax_lse = softmax_lse.view(q_len, q_heads, 1).transpose(0,