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 13, 07:00 UTC
[Attention][Bugfix] Fix FA sink support (#28660)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Nov 13, 2025 at 12:20 UTC
f9f3b596f374c4a01acef275ee1f35398bb05164
1 file changed
+6
vllm/v1/attention/backends/flash_attn.py
+6
@@ -130,6 +130,12 @@ class FlashAttentionBackend(AttentionBackend):
130
return flash_attn_supports_fp8()
131
return kv_cache_dtype in ["auto"]
132
133
+ @classmethod
134
+ def supports_sink(cls) -> bool:
135
+ if not is_flash_attn_varlen_func_available():
136
+ return False
137
+ return flash_attn_supports_sinks()
138
+
139
@classmethod
140
def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
141
return capability >= DeviceCapability(8, 0)