support sinks

965934b8 · zhuwenwen · 8b1077ba · 965934b8
Commit 965934b8 authored Jan 05, 2026 by zhuwenwen
Show whitespace changes
Inline Side-by-side

Showing with 3 additions and 2 deletions

vllm/v1/attention/backends/flash_attn.py vllm/v1/attention/backends/flash_attn.py +3 -2

No files found.
--- a/vllm/v1/attention/backends/flash_attn.py
+++ b/vllm/v1/attention/backends/flash_attn.py
@@ -453,6 +453,7 @@ class FlashAttentionImpl(AttentionImpl):

        self.sinks = sinks
        if self.sinks is not None:
+            if not current_platform.is_rocm():
                assert self.vllm_flash_attn_version == 3, (
                    "Sinks are only supported in FlashAttention 3")
            assert self.sinks.shape[0] == num_heads, (