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 15, 13:30 UTC
[Bugfix] Triton FA function takes no keyword arguments (#16902)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Apr 23, 2025 at 12:35 UTC
047797ef904fd77b13af3a972e92ef6de37a36db
1 file changed
+8
-1
vllm/attention/backends/mla/common.py
+8
-1
@@ -1091,7 +1091,14 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
1091
q,
1092
k,
1093
maybe_padded_v,
1094
- **kwargs,
1094
+ None, # output
1095
+ kwargs["cu_seqlens_q"],
1096
+ kwargs["cu_seqlens_k"],
1097
+ kwargs["max_seqlen_q"],
1098
+ kwargs["max_seqlen_k"],
1099
+ kwargs["causal"],
1100
+ softmax_scale,
1101
+ None, # bias
1102
)
1103
if is_vllm_fa:
1104
attn_out = self.flash_attn_varlen_func(