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 7, 14:00 UTC
[Bugfix] Fix DCP + FA3 crash due to missing num_splits in _forward_with_dcp (#35082)
Signed-off-by: haosdent <haosdent@gmail.com>
haosdent committed
Feb 27, 2026 at 22:27 UTC
6d4f9d3ad5aa3750697edcf013ad080619ae25e9
1 file changed
+2
vllm/v1/attention/backends/flash_attn.py
+2
@@ -847,6 +847,7 @@ class FlashAttentionImpl(AttentionImpl):
847
q_descale=q_descale,
848
k_descale=k_descale,
849
v_descale=v_descale,
850
+ num_splits=attn_metadata.max_num_splits,
851
)
852
# FA returns LSE in shape [ H, B ] but cp_lse_ag_out_rs wants [ B, H ]
853
context_attn_out_cor, context_lse_cor = cp_lse_ag_out_rs(
@@ -876,6 +877,7 @@ class FlashAttentionImpl(AttentionImpl):
877
q_descale=q_descale,
878
k_descale=k_descale,
879
v_descale=v_descale,
880
+ num_splits=attn_metadata.max_num_splits,
881
)
882
assert context_attn_out_cor.shape == query_attn_out.shape
883
assert context_lse_cor.shape == query_lse.shape