| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import functools |
| 4 | |
| 5 | import torch |
| 6 | |
| 7 | from vllm.config import CacheConfig, VllmConfig |
| 8 | from vllm.forward_context import ForwardContext, get_forward_context |
| 9 | from vllm.logger import init_logger |
| 10 | from vllm.model_executor.custom_op import CustomOp |
| 11 | from vllm.model_executor.layers.attention import Attention |
| 12 | from vllm.utils.math_utils import cdiv |
| 13 | from vllm.utils.torch_utils import ( |
| 14 | LayerNameType, |
| 15 | _encode_layer_name, |
| 16 | _resolve_layer_name, |
| 17 | direct_register_custom_op, |
| 18 | ) |
| 19 | from vllm.v1.attention.backend import ( |
| 20 | AttentionBackend, |
| 21 | AttentionMetadata, |
| 22 | AttentionType, |
| 23 | CommonAttentionMetadata, |
| 24 | subclass_attention_backend, |
| 25 | ) |
| 26 | from vllm.v1.attention.ops.triton_reshape_and_cache_flash import ( |
| 27 | triton_reshape_and_cache_flash_diffkv, |
| 28 | ) |
| 29 | from vllm.v1.attention.selector import get_attn_backend |
| 30 | from vllm.v1.kv_cache_interface import ( |
| 31 | AttentionSpec, |
| 32 | KVCacheSpec, |
| 33 | SinkFullAttentionSpec, |
| 34 | get_kv_quant_mode, |
| 35 | ) |
| 36 | |
| 37 | logger = init_logger(__name__) |
| 38 | |
| 39 | |
| 40 | @functools.lru_cache |
| 41 | def create_static_sink_attention_backend( |
| 42 | underlying_attn_backend: type[AttentionBackend], |
| 43 | sink_len: int = 0, |
| 44 | ) -> type[AttentionBackend]: |
| 45 | prefix = "StaticSink_" |
| 46 | underlying_builder = underlying_attn_backend.get_builder_cls() |
| 47 | |
| 48 | class StaticSinkAttentionBuilder(underlying_builder): # type: ignore |
| 49 | def __init__( |
| 50 | self, |
| 51 | kv_cache_spec: AttentionSpec, |
| 52 | layer_names: list[str], |
| 53 | vllm_config: VllmConfig, |
| 54 | device: torch.device, |
| 55 | ): |
| 56 | super().__init__(kv_cache_spec, layer_names, vllm_config, device) |
| 57 | model_config = vllm_config.model_config |
| 58 | scheduler_config = vllm_config.scheduler_config |
| 59 | self.sink_len = sink_len |
| 60 | self.block_size = vllm_config.cache_config.block_size |
| 61 | self.num_sink_blocks = self.sink_len // vllm_config.cache_config.block_size |
| 62 | self.max_num_blocks = cdiv( |
| 63 | model_config.max_model_len, vllm_config.cache_config.block_size |
| 64 | ) |
| 65 | self.block_table_with_sink = torch.zeros( |
| 66 | ( |
| 67 | scheduler_config.max_num_seqs, |
| 68 | self.max_num_blocks + self.num_sink_blocks, |
| 69 | ), |
| 70 | device=device, |
| 71 | dtype=torch.int32, |
| 72 | ) |
| 73 | self.block_table_with_sink[:, : self.num_sink_blocks] = torch.arange( |
| 74 | 1, |
| 75 | self.num_sink_blocks + 1, |
| 76 | device=device, |
| 77 | dtype=torch.int32, |
| 78 | ) |
| 79 | |
| 80 | def build( |
| 81 | self, |
| 82 | common_prefix_len: int, |
| 83 | common_attn_metadata: CommonAttentionMetadata, |
| 84 | fast_build: bool = False, |
| 85 | ) -> AttentionMetadata: |
| 86 | common_attn_metadata.seq_lens[:] = ( |
| 87 | common_attn_metadata.seq_lens + self.sink_len |
| 88 | ) |
| 89 | common_attn_metadata.seq_lens[ |
| 90 | common_attn_metadata.seq_lens == self.sink_len |
| 91 | ] = 0 |
| 92 | common_attn_metadata.max_seq_len = ( |
| 93 | common_attn_metadata.max_seq_len + self.sink_len |
| 94 | ) |
| 95 | max_num_blocks = cdiv(common_attn_metadata.max_seq_len, self.block_size) |
| 96 | num_reqs = common_attn_metadata.num_reqs |
| 97 | self.block_table_with_sink[ |
| 98 | :num_reqs, self.num_sink_blocks : self.num_sink_blocks + max_num_blocks |
| 99 | ] = common_attn_metadata.block_table_tensor[:, :max_num_blocks] |
| 100 | common_attn_metadata.block_table_tensor = self.block_table_with_sink[ |
| 101 | :num_reqs |
| 102 | ] |
| 103 | |
| 104 | return super().build(common_prefix_len, common_attn_metadata, fast_build) |
| 105 | |
| 106 | attn_backend = subclass_attention_backend( |
| 107 | name_prefix=prefix, |
| 108 | attention_backend_cls=underlying_attn_backend, |
| 109 | builder_cls=StaticSinkAttentionBuilder, |
| 110 | ) |
| 111 | |
| 112 | return attn_backend |
| 113 | |
| 114 | |
| 115 | @CustomOp.register("static_sink_attention") |
| 116 | class StaticSinkAttention(Attention, CustomOp): |
| 117 | """Attention with static sink tokens.""" |
| 118 | |
| 119 | def __init__( |
| 120 | self, |
| 121 | num_heads: int, |
| 122 | head_size: int, |
| 123 | scale: float, |
| 124 | sink_len: int, |
| 125 | attn_backend: type[AttentionBackend] | None = None, |
| 126 | cache_config: CacheConfig | None = None, |
| 127 | **kwargs, |
| 128 | ): |
| 129 | dtype = torch.get_default_dtype() |
| 130 | |
| 131 | if cache_config is not None: |
| 132 | kv_cache_dtype = cache_config.cache_dtype |
| 133 | else: |
| 134 | kv_cache_dtype = "auto" |
| 135 | |
| 136 | if attn_backend is not None: |
| 137 | underlying_attn_backend = attn_backend |
| 138 | else: |
| 139 | underlying_attn_backend = get_attn_backend(head_size, dtype, kv_cache_dtype) |
| 140 | attn_backend = create_static_sink_attention_backend( |
| 141 | underlying_attn_backend, # type: ignore[arg-type] |
| 142 | sink_len=sink_len, |
| 143 | ) |
| 144 | Attention.__init__( |
| 145 | self=self, |
| 146 | num_heads=num_heads, |
| 147 | head_size=head_size, |
| 148 | scale=scale, |
| 149 | cache_config=cache_config, |
| 150 | attn_backend=attn_backend, |
| 151 | **kwargs, |
| 152 | ) |
| 153 | self.sink_len = sink_len |
| 154 | self.sink_populated = False |
| 155 | self.sink_key = None |
| 156 | self.sink_value = None |
| 157 | |
| 158 | def update_sink_kv(self, sink_key, sink_value) -> None: |
| 159 | self.sink_key = sink_key |
| 160 | self.sink_value = sink_value |
| 161 | |
| 162 | def forward_native( |
| 163 | self, |
| 164 | query: torch.Tensor, |
| 165 | key: torch.Tensor, |
| 166 | value: torch.Tensor, |
| 167 | output_shape: torch.Size | None = None, |
| 168 | ) -> torch.Tensor: |
| 169 | assert self.sink_key is not None and self.sink_value is not None, ( |
| 170 | "sink_key and sink_value have not been prepared" |
| 171 | ) |
| 172 | if not self.sink_populated: |
| 173 | self_kv_cache = self.kv_cache |
| 174 | torch.ops.vllm.maybe_populate_sink( |
| 175 | self_kv_cache, _encode_layer_name(self.layer_name) |
| 176 | ) |
| 177 | |
| 178 | return super().forward(query, key, value, output_shape) |
| 179 | |
| 180 | def forward_cuda( |
| 181 | self, |
| 182 | query: torch.Tensor, |
| 183 | key: torch.Tensor, |
| 184 | value: torch.Tensor, |
| 185 | output_shape: torch.Size | None = None, |
| 186 | ) -> torch.Tensor: |
| 187 | return self.forward_native(query, key, value, output_shape) |
| 188 | |
| 189 | def forward(self, *args, **kwargs): |
| 190 | return self._forward_method(*args, **kwargs) |
| 191 | |
| 192 | def populate_sink_kv(self, self_kv_cache): |
| 193 | sink_kv_slot_mapping = torch.arange( |
| 194 | self.block_size, |
| 195 | self.sink_len + self.block_size, |
| 196 | device=torch.accelerator.current_device_index(), |
| 197 | dtype=torch.long, |
| 198 | ) |
| 199 | triton_reshape_and_cache_flash_diffkv( |
| 200 | self.sink_key, |
| 201 | self.sink_value, |
| 202 | self_kv_cache, |
| 203 | sink_kv_slot_mapping, |
| 204 | self.kv_cache_dtype, |
| 205 | self._k_scale, |
| 206 | self._v_scale, |
| 207 | ) |
| 208 | # We only populate the sink_key and sink_value once |
| 209 | self.sink_populated = True |
| 210 | |
| 211 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 212 | # Block size may get updated after model loading, refresh it |
| 213 | self.block_size = vllm_config.cache_config.block_size |
| 214 | # Should not be called for enc-dec or encoder-only attention. |
| 215 | assert self.attn_type == AttentionType.DECODER |
| 216 | |
| 217 | return SinkFullAttentionSpec( |
| 218 | block_size=self.block_size, |
| 219 | num_kv_heads=self.num_kv_heads, |
| 220 | head_size=self.head_size, |
| 221 | head_size_v=self.head_size_v, |
| 222 | sink_len=self.sink_len, |
| 223 | dtype=self.kv_cache_torch_dtype, |
| 224 | kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype), |
| 225 | ) |
| 226 | |
| 227 | |
| 228 | def maybe_populate_sink( |
| 229 | self_kv_cache: torch.Tensor, |
| 230 | layer_name: LayerNameType, |
| 231 | ) -> None: |
| 232 | layer_name = _resolve_layer_name(layer_name) |
| 233 | forward_context: ForwardContext = get_forward_context() |
| 234 | self = forward_context.no_compile_layers[layer_name] |
| 235 | if self.sink_populated or self_kv_cache.numel() == 0: |
| 236 | return |
| 237 | self.populate_sink_kv(self_kv_cache) |
| 238 | |
| 239 | |
| 240 | direct_register_custom_op( |
| 241 | op_name="maybe_populate_sink", |
| 242 | op_func=maybe_populate_sink, |
| 243 | mutates_args=["self_kv_cache"], |
| 244 | ) |