comments

Signed-off-by: Huamin Li <[email protected]>

Author: hl475

vllm/attention/backends/abstract.py

@@ -142,6 +142,17 @@ def supports_sink(cls) -> bool:
     def is_sparse(cls) -> bool:
         return False
 
+    @classmethod
+    def supports_attn_type(cls, attn_type: str) -> bool:
+        """Check if backend supports a given attention type.
+
+        By default, only supports decoder attention.
+        Backends should override this to support other attention types.
+        """
+        from vllm.attention import AttentionType
+
+        return attn_type == AttentionType.DECODER
+
     @classmethod
     def supports_compute_capability(cls, capability: "DeviceCapability") -> bool:
         return True
@@ -171,6 +182,7 @@ def validate_configuration(
         has_sink: bool,
         use_sparse: bool,
         device_capability: "DeviceCapability",
+        attn_type: str,
     ) -> list[str]:
         invalid_reasons = []
         if not cls.supports_head_size(head_size):
@@ -195,6 +207,8 @@ def validate_configuration(
                 invalid_reasons.append("non-sparse not supported")
         if not cls.supports_compute_capability(device_capability):
             invalid_reasons.append("compute capability not supported")
+        if not cls.supports_attn_type(attn_type):
+            invalid_reasons.append(f"attention type {attn_type} not supported")
         combination_reason = cls.supports_combination(
             head_size,
             dtype,

vllm/attention/layer.py

@@ -291,6 +291,7 @@ def __init__(
                 block_size,
                 use_mla=False,
                 has_sink=self.has_sink,
+                attn_type=attn_type,
             )
         else:
             self.attn_backend = attn_backend

vllm/attention/layers/encoder_only_attention.py

@@ -74,7 +74,11 @@ def __init__(
             block_size = 16
 
         underlying_attn_backend = get_attn_backend(
-            head_size, dtype, kv_cache_dtype, block_size
+            head_size,
+            dtype,
+            kv_cache_dtype,
+            block_size,
+            attn_type=AttentionType.ENCODER_ONLY,
         )
 
         attn_backend = create_encoder_only_attention_backend(underlying_attn_backend)

vllm/attention/selector.py

@@ -76,6 +76,7 @@ def get_attn_backend(
     use_mla: bool = False,
     has_sink: bool = False,
     use_sparse: bool = False,
+    attn_type: str | None = None,
 ) -> type[AttentionBackend]:
     """Selects which attention backend to use and lazily imports it."""
 
@@ -94,6 +95,7 @@ def get_attn_backend(
         use_mla=use_mla,
         has_sink=has_sink,
         use_sparse=use_sparse,
+        attn_type=attn_type,
     )
 
 
@@ -106,6 +108,7 @@ def _cached_get_attn_backend(
     use_mla: bool = False,
     has_sink: bool = False,
     use_sparse: bool = False,
+    attn_type: str | None = None,
 ) -> type[AttentionBackend]:
     # Check whether a particular choice of backend was
     # previously forced.

vllm/platforms/cpu.py

@@ -134,6 +134,7 @@ def get_attn_backend_cls(
         use_mla: bool,
         has_sink: bool,
         use_sparse: bool,
+        attn_type: str | None = None,
     ) -> str:
         from vllm.attention.backends.registry import AttentionBackendEnum
 

vllm/platforms/cuda.py

@@ -298,6 +298,7 @@ def get_valid_backends(
         has_sink,
         use_sparse,
         device_capability,
+        attn_type,
     ) -> tuple[
         list[tuple["AttentionBackendEnum", int]],
         dict["AttentionBackendEnum", list[str]],
@@ -318,6 +319,7 @@ def get_valid_backends(
                     has_sink,
                     use_sparse,
                     device_capability,
+                    attn_type,
                 )
             except ImportError:
                 invalid_reasons_i = ["ImportError"]
@@ -339,7 +341,13 @@ def get_attn_backend_cls(
         use_mla: bool,
         has_sink: bool,
         use_sparse: bool,
+        attn_type: str | None = None,
     ) -> str:
+        from vllm.attention import AttentionType
+
+        if attn_type is None:
+            attn_type = AttentionType.DECODER
+
         device_capability = cls.get_device_capability()
         assert device_capability is not None
 
@@ -356,6 +364,7 @@ def get_attn_backend_cls(
                     has_sink,
                     use_sparse,
                     device_capability,
+                    attn_type,
                 )
             except ImportError:
                 invalid_reasons = ["ImportError"]
@@ -379,6 +388,7 @@ def get_attn_backend_cls(
             has_sink,
             use_sparse,
             device_capability,
+            attn_type,
         )
         reasons_str = (
             "{"

vllm/platforms/interface.py

@@ -218,6 +218,7 @@ def get_attn_backend_cls(
         use_mla: bool,
         has_sink: bool,
         use_sparse: bool,
+        attn_type: str | None = None,
     ) -> str:
         """Get the attention backend class of a device."""
         return ""

vllm/platforms/rocm.py

@@ -216,6 +216,7 @@ def get_attn_backend_cls(
         use_mla,
         has_sink,
         use_sparse,
+        attn_type: str | None = None,
     ) -> str:
         from vllm._aiter_ops import rocm_aiter_ops
         from vllm.attention.backends.registry import AttentionBackendEnum

vllm/platforms/tpu.py

@@ -61,6 +61,7 @@ def get_attn_backend_cls(
         use_mla: bool,
         has_sink,
         use_sparse,
+        attn_type: str | None = None,
     ) -> str:
         from vllm.attention.backends.registry import AttentionBackendEnum
 

vllm/platforms/xpu.py

@@ -51,6 +51,7 @@ def get_attn_backend_cls(
         use_mla: bool,
         has_sink: bool,
         use_sparse,
+        attn_type: str | None = None,
     ) -> str:
         from vllm.v1.attention.backends.utils import set_kv_cache_layout