[v1][attention] Support Hybrid Allocator + FlashInfer (#21412)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>
2025-07-29 18:45:29 -07:00
parent 0e36abf993
commit 555e7225bc
16 changed files with 85 additions and 57 deletions
--- a/vllm/v1/attention/backends/triton_attn.py
+++ b/vllm/v1/attention/backends/triton_attn.py
@@ -59,8 +59,8 @@ class TritonAttentionMetadataBuilder(
        AttentionMetadataBuilder[TritonAttentionMetadata]):
    full_cudagraph_supported: ClassVar[bool] = True

-    def __init__(self, kv_cache_spec: AttentionSpec, vllm_config: VllmConfig,
-                 device: torch.device):
+    def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
+                 vllm_config: VllmConfig, device: torch.device):
        self.device = device
        self.block_size = kv_cache_spec.block_size
        self.kv_cache_spec = kv_cache_spec