[Attention] Refactor AttentionMetadata Preparation for Encoder-only Models

tests/v1/worker/test_gpu_model_runner.py

@@ -680,17 +680,20 @@ def test_init_kv_cache_with_kv_sharing_valid():
         kv_cache_spec[layer_0].page_size_bytes
 
     runner.initialize_kv_cache(kv_cache_config)
+    kv_cache_config_after_init = runner.kv_cache_config
 
     layer_0_kv = vllm_ctx[layer_0].kv_cache[0]
     layer_1_kv = vllm_ctx[layer_1].kv_cache[0]
     # check layer 1 kv cache shares memory with layer 0
     assert id(layer_1_kv) == id(layer_0_kv)
 
     # check layer 1 added to kv cache group's layer names
-    assert len(kv_cache_config.kv_cache_groups) == 1
-    assert len(kv_cache_config.kv_cache_groups[0].layer_names) == 2
-    assert kv_cache_config.kv_cache_groups[0].layer_names[0] == layer_0
-    assert kv_cache_config.kv_cache_groups[0].layer_names[1] == layer_1
+    assert len(kv_cache_config_after_init.kv_cache_groups) == 1
+    assert len(kv_cache_config_after_init.kv_cache_groups[0].layer_names) == 2
+    assert kv_cache_config_after_init.kv_cache_groups[0].layer_names[
+        0] == layer_0
+    assert kv_cache_config_after_init.kv_cache_groups[0].layer_names[
+        1] == layer_1
 
 
 def test_hybrid_attention_mamba_tensor_shapes(monkeypatch):

vllm/attention/layers/chunked_local_attention.py

@@ -6,12 +6,13 @@
 import torch
 
 from vllm import envs
-from vllm.attention.backends.abstract import AttentionBackend
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata)
 from vllm.attention.selector import get_attn_backend
 from vllm.config import CacheConfig, QuantizationConfig
 from vllm.v1.attention.backends.utils import (
     CommonAttentionMetadata, make_local_attention_virtual_batches,
-    subclass_attention_backend, subclass_attention_metadata_builder)
+    subclass_attention_backend)
 
 from ..layer import Attention
 
@@ -24,21 +25,23 @@ def create_chunked_local_attention_backend(
 ) -> type[AttentionBackend]:
     prefix = f"ChunkedLocalAttention_{attention_chunk_size}_{block_size}_"
 
-    def build_preprocess_fn(cm: CommonAttentionMetadata):
-        return make_local_attention_virtual_batches(attention_chunk_size, cm,
-                                                    block_size)
+    underlying_builder = underlying_attn_backend.get_builder_cls()
+
+    class ChunkedLocalAttentionBuilder(underlying_builder):  # type: ignore
+
+        def build(self,
+                  common_prefix_len: int,
+                  common_attn_metadata: CommonAttentionMetadata,
+                  fast_build: bool = False) -> AttentionMetadata:
+            common_attn_metadata = make_local_attention_virtual_batches(
+                attention_chunk_size, common_attn_metadata, block_size)
+            return super().build(common_prefix_len, common_attn_metadata,
+                                 fast_build)
 
-    # Dynamically create a new attention backend that wraps the
-    # underlying attention backend but applies
-    # `make_local_attention_virtual_batches` before calling `build(...)`
-    builder_cls = subclass_attention_metadata_builder(
-        name_prefix=prefix,
-        builder_cls=underlying_attn_backend.get_builder_cls(),
-        build_preprocess_fn=build_preprocess_fn)
     attn_backend = subclass_attention_backend(
         name_prefix=prefix,
         attention_backend_cls=underlying_attn_backend,
-        builder_cls=builder_cls)
+        builder_cls=ChunkedLocalAttentionBuilder)
 
     return attn_backend
 

vllm/attention/layers/encoder_only_attention.py

@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import functools
+from copy import copy
+from typing import Optional
+
+import torch
+from transformers import CacheConfig
+
+from vllm import envs
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.layer import Attention
+from vllm.attention.selector import get_attn_backend
+from vllm.v1.attention.backends.utils import (CommonAttentionMetadata,
+                                              subclass_attention_backend)
+
+
+@functools.lru_cache
+def create_encoder_only_attention_backend(
+    underlying_attn_backend: AttentionBackend, ) -> type[AttentionBackend]:
+    prefix = "EncoderOnlyAttention_"
+    underlying_builder = underlying_attn_backend.get_builder_cls()
+
+    class EncoderOnlyAttentionBuilder(underlying_builder):  # type: ignore
+
+        def build(self,
+                  common_prefix_len: int,
+                  common_attn_metadata: CommonAttentionMetadata,
+                  fast_build: bool = False) -> AttentionMetadata:
+            new_common_attn_metadata = copy(common_attn_metadata)
+            new_common_attn_metadata.causal = False
+            return super().build(common_prefix_len, new_common_attn_metadata,
+                                 fast_build)
+
+    attn_backend = subclass_attention_backend(
+        name_prefix=prefix,
+        attention_backend_cls=underlying_attn_backend,
+        builder_cls=EncoderOnlyAttentionBuilder)
+
+    return attn_backend
+
+
+class EncoderOnlyAttention(Attention):
+    """
+    Encoder attention is a special case that doesn't need a KV Cache.
+    """
+
+    def __init__(self,
+                 num_heads: int,
+                 head_size: int,
+                 scale: float,
+                 cache_config: Optional[CacheConfig] = None,
+                 attn_type: Optional[str] = None,
+                 **kwargs):
+        dtype = torch.get_default_dtype()
+
+        if cache_config is not None:
+            kv_cache_dtype = cache_config.cache_dtype
+            block_size = cache_config.block_size
+        else:
+            kv_cache_dtype = "auto"
+            block_size = 16
+
+        if envs.VLLM_USE_V1:
+            underlying_attn_backend = get_attn_backend(head_size, dtype,
+                                                       kv_cache_dtype,
+                                                       block_size)
+
+            attn_backend = create_encoder_only_attention_backend(
+                underlying_attn_backend)
+        else:
+            # in v0 encoder only attention is handled inside the backends
+            attn_backend = None
+
+        if attn_type is not None:
+            assert attn_type == AttentionType.ENCODER_ONLY, \
+                "EncoderOnlyAttention only supports AttentionType.ENCODER_ONLY"
+
+        super().__init__(num_heads=num_heads,
+                         head_size=head_size,
+                         scale=scale,
+                         cache_config=cache_config,
+                         attn_backend=attn_backend,
+                         attn_type=AttentionType.ENCODER_ONLY,
+                         **kwargs)

vllm/model_executor/models/bert.py

@@ -8,7 +8,7 @@
 from torch import nn
 from transformers import BertConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, PoolerConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
@@ -239,14 +239,13 @@ def __init__(
             quant_config=quant_config,
             prefix=f"{prefix}.qkv_proj")
 
-        self.attn = Attention(num_heads=self.num_heads,
-                              head_size=self.head_dim,
-                              scale=self.scaling,
-                              num_kv_heads=self.num_kv_heads,
-                              cache_config=cache_config,
-                              quant_config=quant_config,
-                              prefix=f"{prefix}.attn",
-                              attn_type=AttentionType.ENCODER_ONLY)
+        self.attn = EncoderOnlyAttention(num_heads=self.num_heads,
+                                         head_size=self.head_dim,
+                                         scale=self.scaling,
+                                         num_kv_heads=self.num_kv_heads,
+                                         cache_config=cache_config,
+                                         quant_config=quant_config,
+                                         prefix=f"{prefix}.attn")
 
     def forward(
         self,

vllm/model_executor/models/bert_with_rope.py

@@ -7,7 +7,7 @@
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (divide, get_tensor_model_parallel_rank,
@@ -119,14 +119,13 @@ def __init__(
 
         self.rotary_emb = get_rope(**rotary_kwargs)
 
-        self.attn = Attention(num_heads=self.num_heads,
-                              head_size=self.head_dim,
-                              scale=self.scaling,
-                              num_kv_heads=self.num_kv_heads,
-                              cache_config=cache_config,
-                              quant_config=quant_config,
-                              prefix=f"{prefix}.attn",
-                              attn_type=AttentionType.ENCODER_ONLY)
+        self.attn = EncoderOnlyAttention(num_heads=self.num_heads,
+                                         head_size=self.head_dim,
+                                         scale=self.scaling,
+                                         num_kv_heads=self.num_kv_heads,
+                                         cache_config=cache_config,
+                                         quant_config=quant_config,
+                                         prefix=f"{prefix}.attn")
 
         self.out_proj = RowParallelLinear(input_size=hidden_size,
                                           output_size=hidden_size,

vllm/model_executor/models/llama.py

@@ -31,6 +31,7 @@
 from transformers import LlamaConfig
 
 from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
@@ -173,7 +174,10 @@ def __init__(
             if is_sliding:
                 sliding_window = config.sliding_window
 
-        self.attn = Attention(
+        attn_cls = (EncoderOnlyAttention
+                    if attn_type == AttentionType.ENCODER_ONLY else Attention)
+
+        self.attn = attn_cls(
             self.num_heads,
             self.head_dim,
             self.scaling,

vllm/model_executor/models/modernbert.py

@@ -7,7 +7,7 @@
 from torch import nn
 from transformers import ModernBertConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
@@ -104,12 +104,12 @@ def __init__(self,
                                                     head_size=self.head_dim,
                                                     dim=self.head_dim,
                                                     base=rope_theta)
-        self.attn = Attention(self.num_heads,
-                              self.head_dim,
-                              self.scaling,
-                              prefix=f"{layer_id}.attn",
-                              attn_type=AttentionType.ENCODER_ONLY,
-                              per_layer_sliding_window=sliding_window)
+        self.attn = EncoderOnlyAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            prefix=f"{layer_id}.attn",
+            per_layer_sliding_window=sliding_window)
         self.Wo = RowParallelLinear(config.hidden_size,
                                     config.hidden_size,
                                     bias=config.attention_bias)

vllm/model_executor/models/qwen2.py

@@ -32,6 +32,7 @@
 from transformers import Qwen2Config
 
 from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
@@ -159,7 +160,9 @@ def __init__(
             rope_scaling=rope_scaling,
             dual_chunk_attention_config=dual_chunk_attention_config,
         )
-        self.attn = Attention(
+        attn_cls = (EncoderOnlyAttention
+                    if attn_type == AttentionType.ENCODER_ONLY else Attention)
+        self.attn = attn_cls(
             self.num_heads,
             self.head_dim,
             self.scaling,

vllm/v1/attention/backends/utils.py

@@ -5,8 +5,7 @@
 import functools
 from abc import abstractmethod
 from dataclasses import dataclass, make_dataclass
-from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Generic, Optional,
-                    TypeVar)
+from typing import TYPE_CHECKING, Any, ClassVar, Generic, Optional, TypeVar
 
 import numpy as np
 import torch
@@ -543,35 +542,6 @@ def make_local_attention_virtual_batches(
     )
 
 
-def subclass_attention_metadata_builder(
-    name_prefix: str,
-    builder_cls: type[AttentionMetadataBuilder[M]],
-    build_preprocess_fn: Callable[[CommonAttentionMetadata],
-                                  CommonAttentionMetadata],
-) -> type[AttentionMetadataBuilder[M]]:
-    """
-    Return a new subclass of `builder_cls` whose .build(...) method
-    first calls build_preprocess_fn(common_attn_metadata) on the metadata.
-    """
-    name: str = name_prefix + builder_cls.__name__  # type: ignore
-
-    def build(self,
-              common_prefix_len: int,
-              common_attn_metadata: CommonAttentionMetadata,
-              fast_build: bool = False):
-        return builder_cls.build(self, common_prefix_len,
-                                 build_preprocess_fn(common_attn_metadata),
-                                 fast_build)
-
-    Wrapped = type(
-        name,
-        (builder_cls, ),  # inherit from the original
-        {
-            "build": build,
-        })
-    return Wrapped  # type: ignore
-
-
 def subclass_attention_backend(
         name_prefix: str, attention_backend_cls: type[AttentionBackend],
         builder_cls: type[AttentionMetadataBuilder[M]]

vllm/v1/kv_cache_interface.py

@@ -203,6 +203,14 @@ def max_memory_usage_bytes(self, vllm_config: VllmConfig) -> int:
         return self.page_size_bytes
 
 
+@dataclass(frozen=True)
+class EncoderOnlyAttentionSpec(AttentionSpec):
+
+    def max_memory_usage_bytes(self, vllm_config: VllmConfig) -> int:
+        # Encoder-only layers do not need KV cache
+        return 0
+
+
 @dataclass
 class KVCacheTensor:
     """