| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from typing import TYPE_CHECKING, Any, cast |
| 5 | |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | |
| 9 | import vllm.envs as envs |
| 10 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 11 | from vllm.config import ( |
| 12 | CacheConfig, |
| 13 | get_current_vllm_config, |
| 14 | ) |
| 15 | from vllm.config.vllm import VllmConfig |
| 16 | from vllm.forward_context import ForwardContext, get_forward_context |
| 17 | from vllm.logger import init_logger |
| 18 | from vllm.model_executor.layers.attention.kv_transfer_utils import ( |
| 19 | maybe_transfer_kv_layer, |
| 20 | ) |
| 21 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 22 | from vllm.model_executor.layers.linear import ( |
| 23 | UnquantizedLinearMethod, |
| 24 | ) |
| 25 | from vllm.model_executor.layers.quantization import ( |
| 26 | QuantizationConfig, |
| 27 | resolve_quant_method, |
| 28 | ) |
| 29 | from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase |
| 30 | from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8 |
| 31 | from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod |
| 32 | from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape |
| 33 | from vllm.platforms import current_platform |
| 34 | from vllm.utils.torch_utils import ( |
| 35 | LayerNameType, |
| 36 | _encode_layer_name, |
| 37 | _resolve_layer_name, |
| 38 | direct_register_custom_op, |
| 39 | kv_cache_dtype_str_to_dtype, |
| 40 | ) |
| 41 | from vllm.v1.attention.backend import ( |
| 42 | AttentionBackend, |
| 43 | AttentionMetadata, |
| 44 | AttentionType, |
| 45 | ) |
| 46 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 47 | from vllm.v1.attention.selector import get_attn_backend |
| 48 | from vllm.v1.kv_cache_interface import ( |
| 49 | FullAttentionSpec, |
| 50 | KVCacheSpec, |
| 51 | SlidingWindowSpec, |
| 52 | get_kv_quant_mode, |
| 53 | ) |
| 54 | |
| 55 | if TYPE_CHECKING: |
| 56 | from vllm.model_executor.layers.attention import MLAAttention |
| 57 | |
| 58 | logger = init_logger(__name__) |
| 59 | |
| 60 | |
| 61 | def validate_kv_sharing_target( |
| 62 | current_layer_name, target_layer_name, static_forward_context |
| 63 | ): |
| 64 | error_msg = ( |
| 65 | f"Specified KV sharing target layer for {current_layer_name} " |
| 66 | f"is not valid: target layer {target_layer_name} " |
| 67 | ) |
| 68 | |
| 69 | if current_layer_name == target_layer_name: |
| 70 | raise ValueError(error_msg + "cannot be the same as the current layer.") |
| 71 | |
| 72 | if target_layer_name not in static_forward_context: |
| 73 | from vllm.model_executor.models.utils import extract_layer_index |
| 74 | |
| 75 | # If target layer name is not in the static fwd context, it means either |
| 76 | # a) the target layer does not come BEFORE the current layer, or |
| 77 | # b) the target layer is not an Attention layer that exists in the model |
| 78 | current_layer_idx = extract_layer_index(current_layer_name) |
| 79 | target_layer_idx = extract_layer_index(target_layer_name) |
| 80 | if current_layer_idx <= target_layer_idx: |
| 81 | raise ValueError(error_msg + "must come before the current layer.") |
| 82 | else: |
| 83 | raise ValueError(error_msg + "is not a valid Attention layer in the model.") |
| 84 | |
| 85 | # Currently KV sharing is only supported between layers of the same type |
| 86 | target_layer_attn_type = static_forward_context[target_layer_name].attn_type |
| 87 | expected = static_forward_context[current_layer_name].attn_type |
| 88 | if target_layer_attn_type != expected: |
| 89 | raise ValueError( |
| 90 | error_msg + f"must be the same type as the current layer ({expected})." |
| 91 | ) |
| 92 | |
| 93 | |
| 94 | def should_load_quant_weights(quant_method: QuantizeMethodBase | None) -> bool: |
| 95 | """Returns whether the quantization method should load quantized weights.""" |
| 96 | return quant_method is not None and not isinstance( |
| 97 | quant_method, UnquantizedLinearMethod |
| 98 | ) |
| 99 | |
| 100 | |
| 101 | def _largest_kernel_block_within( |
| 102 | attn_backend: "type[AttentionBackend]", |
| 103 | per_token_bytes: int, |
| 104 | page_budget: int, |
| 105 | fallback: int, |
| 106 | kv_cache_spec: KVCacheSpec | None = None, |
| 107 | ) -> int: |
| 108 | """Largest supported kernel block size whose page fits in ``page_budget``. |
| 109 | |
| 110 | A padded spec (e.g. skip-quant layer) that pads its page up to a large shared page |
| 111 | wastes ``page_budget - block*per_token`` bytes per block. Picking the largest kernel |
| 112 | block whose natural page still fits under ``page_budget`` minimizes that waste. |
| 113 | ``MultipleOf`` declarations are expanded to the largest aligned block that fits. |
| 114 | Falls back to the smallest supported block when nothing fits. |
| 115 | """ |
| 116 | from vllm.v1.attention.backend import MultipleOf |
| 117 | |
| 118 | sizes = attn_backend.get_supported_kernel_block_sizes(kv_cache_spec) |
| 119 | max_block_size = page_budget // per_token_bytes |
| 120 | candidates = [s for s in sizes if isinstance(s, int)] |
| 121 | candidates.extend( |
| 122 | max(s.base, max_block_size // s.base * s.base) |
| 123 | for s in sizes |
| 124 | if isinstance(s, MultipleOf) |
| 125 | ) |
| 126 | if not candidates: |
| 127 | return fallback |
| 128 | smallest = min(candidates) |
| 129 | fitting = [b for b in candidates if b * per_token_bytes <= page_budget] |
| 130 | return max(fitting) if fitting else smallest |
| 131 | |
| 132 | |
| 133 | def set_default_quant_scales(layer: nn.Module, register_buffer: bool = False) -> None: |
| 134 | """Sets default quantization scales for the layer.""" |
| 135 | if register_buffer: |
| 136 | layer.register_buffer("_k_scale", torch.tensor(1.0, dtype=torch.float32)) |
| 137 | layer.register_buffer("_v_scale", torch.tensor(1.0, dtype=torch.float32)) |
| 138 | layer.register_buffer("_q_scale", torch.tensor(1.0, dtype=torch.float32)) |
| 139 | layer.register_buffer("_prob_scale", torch.tensor(1.0, dtype=torch.float32)) |
| 140 | else: |
| 141 | layer._k_scale.fill_(1.0) |
| 142 | layer._v_scale.fill_(1.0) |
| 143 | layer._q_scale.fill_(1.0) |
| 144 | layer._prob_scale.fill_(1.0) |
| 145 | |
| 146 | # We also keep q/k/v_scale on host (cpu) memory for attention |
| 147 | # backends that require the scales to be on host instead of on device. |
| 148 | # e.g. Flashinfer & AITER |
| 149 | layer._q_scale_float = 1.0 |
| 150 | layer._k_scale_float = 1.0 |
| 151 | layer._v_scale_float = 1.0 |
| 152 | layer._k_scale_cpu = torch.tensor(1.0, dtype=torch.float32) |
| 153 | layer._v_scale_cpu = torch.tensor(1.0, dtype=torch.float32) |
| 154 | layer._prob_scale_float = 1.0 |
| 155 | |
| 156 | |
| 157 | def _init_kv_cache_quant( |
| 158 | layer: nn.Module, |
| 159 | quant_config: QuantizationConfig | None, |
| 160 | prefix: str, |
| 161 | ) -> None: |
| 162 | """Initializes KV cache scaling factors and quantization method. |
| 163 | |
| 164 | This helper function sets up the KV cache quantization attributes that are |
| 165 | shared between Attention and MLAAttention layers. It initializes scale |
| 166 | tensors for query, key, value, and probability, and configures the |
| 167 | quantization method if applicable. |
| 168 | |
| 169 | Args: |
| 170 | layer: The attention layer instance to initialize. |
| 171 | quant_config: Optional quantization configuration. |
| 172 | prefix: Layer name prefix for quantization method lookup. |
| 173 | |
| 174 | """ |
| 175 | # Note [Register q/k/v/prob scales in state dict] |
| 176 | # When calling model.to(device), only parameters/buffers in state dict are |
| 177 | # moved. If not registering q/k/v/prob scales in state dict, there would |
| 178 | # be an IMA error when a cuda kernel (e.g., quant_fp8) accesses the tensor |
| 179 | # on cpu. |
| 180 | # Registering in state dict means it interacts with weight loading. One edge |
| 181 | # case is when quant_method is None, or quant_method is UnquantizedLinearMethod |
| 182 | # (i.e., should_load_quant_weights(quant_method) == False). |
| 183 | # In this case, the checkpoint does not have the scales. We need to |
| 184 | # initialize the scales to 1.0 and update the scales after weight loading. |
| 185 | # This is espectially important when we load dummy weights first (providing |
| 186 | # wrong scales) and then load real weights (which misses scales and keeps the |
| 187 | # wrong scales from dummy load). |
| 188 | set_default_quant_scales(layer, register_buffer=True) |
| 189 | |
| 190 | # The output scale on host memory. This should be the input scale of |
| 191 | # the quant op after this attention layer. |
| 192 | layer._o_scale_float = None |
| 193 | |
| 194 | quant_method = ( |
| 195 | resolve_quant_method(quant_config, layer, prefix=prefix) |
| 196 | if quant_config |
| 197 | else None |
| 198 | ) |
| 199 | |
| 200 | # See [Note: Register q/k/v/prob scales in state dict] |
| 201 | if should_load_quant_weights(quant_method): |
| 202 | assert isinstance(quant_method, BaseKVCacheMethod) |
| 203 | # TODO (mgoin): kv cache dtype should be specified in the FP8 |
| 204 | # checkpoint config and become the "auto" behavior |
| 205 | if layer.kv_cache_dtype == "fp8_e5m2": |
| 206 | # A compressed-tensors checkpoint stores fp8 KV scales only when it |
| 207 | # declares a kv_cache_scheme; weight-only ones declare none and must |
| 208 | # keep fp8_e5m2, the only fp8 KV dtype usable on Ampere. |
| 209 | from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501 |
| 210 | CompressedTensorsConfig, |
| 211 | CompressedTensorsKVCacheMethod, |
| 212 | ) |
| 213 | |
| 214 | if not isinstance(quant_method, CompressedTensorsKVCacheMethod) or ( |
| 215 | cast(CompressedTensorsConfig, quant_method.quant_config).kv_cache_scheme |
| 216 | is not None |
| 217 | ): |
| 218 | raise ValueError( |
| 219 | "fp8_e5m2 kv-cache is not supported with fp8 checkpoints." |
| 220 | ) |
| 221 | # If quantization is enabled, we make "k_scale" and "v_scale" |
| 222 | # parameters so that it can be loaded from the model checkpoint. |
| 223 | # The k/v_scale will then be converted back to native float32 |
| 224 | # values after weight loading. |
| 225 | layer.quant_method = quant_method |
| 226 | layer.quant_method.create_weights(layer) |
| 227 | |
| 228 | |
| 229 | class Attention(nn.Module, AttentionLayerBase): |
| 230 | """Attention layer. |
| 231 | |
| 232 | This class takes query, key, and value tensors as input. The input tensors |
| 233 | can either contain prompt tokens or generation tokens. |
| 234 | The class does the following: |
| 235 | |
| 236 | 1. Store the input key and value tensors in the KV cache. |
| 237 | 2. Perform (multi-head/multi-query/grouped-query) attention. |
| 238 | 3. Return the output tensor. |
| 239 | """ |
| 240 | |
| 241 | def __init__( |
| 242 | self, |
| 243 | num_heads: int, |
| 244 | head_size: int, |
| 245 | scale: float, |
| 246 | num_kv_heads: int | None = None, |
| 247 | alibi_slopes: list[float] | None = None, |
| 248 | use_alibi_sqrt: bool | None = None, |
| 249 | cache_config: CacheConfig | None = None, |
| 250 | quant_config: QuantizationConfig | None = None, |
| 251 | logits_soft_cap: float | None = None, |
| 252 | per_layer_sliding_window: int | None = None, |
| 253 | prefix: str = "", |
| 254 | attn_type: str = AttentionType.DECODER, |
| 255 | kv_sharing_target_layer_name: str | None = None, |
| 256 | mm_prefix_clamp_sliding_window: bool = False, |
| 257 | attn_backend: type[AttentionBackend] | None = None, |
| 258 | head_size_v: int | None = None, |
| 259 | **extra_impl_args, |
| 260 | ) -> None: |
| 261 | """The KV cache is stored inside this class and is accessed via |
| 262 | `self.kv_cache`. |
| 263 | """ |
| 264 | super().__init__() |
| 265 | sliding_window: int | None |
| 266 | if per_layer_sliding_window is not None: |
| 267 | # per-layer sliding window |
| 268 | sliding_window = per_layer_sliding_window |
| 269 | elif cache_config is not None: |
| 270 | # model-level sliding window |
| 271 | sliding_window = cache_config.sliding_window |
| 272 | else: |
| 273 | sliding_window = None |
| 274 | |
| 275 | vllm_config = get_current_vllm_config() |
| 276 | if cache_config is not None: |
| 277 | kv_cache_dtype = cache_config.cache_dtype |
| 278 | else: |
| 279 | kv_cache_dtype = "auto" |
| 280 | |
| 281 | # llm-compressor models declare an FP8 KV-cache scheme in their |
| 282 | # checkpoint config. Honor it only when the user did not explicitly |
| 283 | # pick a kv_cache_dtype; an explicit choice (e.g. bfloat16) must win. |
| 284 | # The "auto" case is normally resolved upstream in |
| 285 | # resolve_kv_cache_dtype_string, but we re-apply here defensively in |
| 286 | # case anything bypassed that path. |
| 287 | kv_cache_scheme = getattr(quant_config, "kv_cache_scheme", None) |
| 288 | if kv_cache_scheme is not None and kv_cache_dtype == "auto": |
| 289 | kv_cache_dtype = "fp8" |
| 290 | if cache_config is not None: |
| 291 | cache_config.cache_dtype = "fp8" |
| 292 | |
| 293 | # Check if per-head quant scales are required based on kv_cache_scheme |
| 294 | use_per_head_quant_scales = ( |
| 295 | kv_cache_scheme is not None |
| 296 | and kv_cache_scheme.get("strategy") == "attn_head" |
| 297 | ) |
| 298 | |
| 299 | # Skip quantization for specified layers |
| 300 | if cache_config is not None and cache_config.kv_cache_dtype_skip_layers: |
| 301 | from vllm.model_executor.models.utils import extract_layer_index |
| 302 | |
| 303 | skip = False |
| 304 | # Check attention type |
| 305 | if ( |
| 306 | sliding_window is not None |
| 307 | and "sliding_window" in cache_config.kv_cache_dtype_skip_layers |
| 308 | ): |
| 309 | skip = True |
| 310 | # Check layer index |
| 311 | layer_idx = extract_layer_index(prefix) |
| 312 | if str(layer_idx) in cache_config.kv_cache_dtype_skip_layers: |
| 313 | skip = True |
| 314 | if skip: |
| 315 | kv_cache_dtype = "auto" |
| 316 | logger.debug( |
| 317 | "Layer %s: kv_cache_dtype=%s, sliding_window=%s", |
| 318 | prefix, |
| 319 | kv_cache_dtype, |
| 320 | sliding_window, |
| 321 | ) |
| 322 | |
| 323 | self.kv_cache_torch_dtype = kv_cache_dtype_str_to_dtype( |
| 324 | kv_cache_dtype, vllm_config.model_config |
| 325 | ) |
| 326 | self.kv_cache_dtype = kv_cache_dtype |
| 327 | if num_kv_heads is None: |
| 328 | num_kv_heads = num_heads |
| 329 | assert num_heads % num_kv_heads == 0, ( |
| 330 | f"num_heads ({num_heads}) is not divisible by num_kv_heads ({num_kv_heads})" |
| 331 | ) |
| 332 | self.quant_config = quant_config |
| 333 | self.layer_name = prefix |
| 334 | |
| 335 | self.num_heads = num_heads |
| 336 | self.head_size = head_size |
| 337 | self.head_size_v = self.head_size if head_size_v is None else head_size_v |
| 338 | self.num_kv_heads = num_kv_heads |
| 339 | self.sliding_window = sliding_window |
| 340 | self.has_sink = extra_impl_args.get("sinks") is not None |
| 341 | |
| 342 | # NOTE: model_config may be None during certain tests |
| 343 | model_config = vllm_config.model_config |
| 344 | self.use_mm_prefix = model_config is not None and model_config.is_mm_prefix_lm |
| 345 | |
| 346 | # During model initialization, the default dtype is set as the model |
| 347 | # weight and activation dtype. |
| 348 | dtype = torch.get_default_dtype() |
| 349 | if attn_backend is None: |
| 350 | self.attn_backend = get_attn_backend( |
| 351 | head_size, |
| 352 | dtype, |
| 353 | kv_cache_dtype, |
| 354 | use_mla=False, |
| 355 | has_sink=self.has_sink, |
| 356 | use_mm_prefix=self.use_mm_prefix, |
| 357 | use_per_head_quant_scales=use_per_head_quant_scales, |
| 358 | attn_type=attn_type, |
| 359 | has_sliding_window=sliding_window is not None, |
| 360 | ) |
| 361 | else: |
| 362 | self.attn_backend = attn_backend |
| 363 | backend_supports_alibi_sqrt = self.attn_backend.supports_alibi_sqrt() |
| 364 | use_alibi_sqrt = use_alibi_sqrt if use_alibi_sqrt else False |
| 365 | if use_alibi_sqrt and not backend_supports_alibi_sqrt: |
| 366 | raise ValueError( |
| 367 | f"use_alibi_sqrt is not supported by backend " |
| 368 | f"{self.attn_backend.get_name()}." |
| 369 | ) |
| 370 | self.use_alibi_sqrt = bool(use_alibi_sqrt) |
| 371 | if backend_supports_alibi_sqrt: |
| 372 | extra_impl_args["use_alibi_sqrt"] = self.use_alibi_sqrt |
| 373 | # prefix caching + batch invariance is currently not supported for |
| 374 | # FLASHINFER and TRITON_MLA. |
| 375 | if ( |
| 376 | cache_config is not None |
| 377 | and cache_config.enable_prefix_caching |
| 378 | and envs.VLLM_BATCH_INVARIANT |
| 379 | and ( |
| 380 | self.attn_backend.get_name() == "FLASHINFER" |
| 381 | or self.attn_backend.get_name() == "TRITON_MLA" |
| 382 | ) |
| 383 | ): |
| 384 | logger.warning_once( |
| 385 | "Disabling prefix caching for FLASHINFER/TRITON_MLA " |
| 386 | "with batch invariance, as it is not yet supported.", |
| 387 | ) |
| 388 | cache_config.enable_prefix_caching = False |
| 389 | |
| 390 | if extra_impl_args.get("chunk_lookback", -1) > -1: |
| 391 | assert self.attn_backend.get_name() == "TRITON_ATTN", ( |
| 392 | f"Chunked attention with lookback requires the Triton backend, " |
| 393 | f"but got {self.attn_backend.get_name()}." |
| 394 | ) |
| 395 | |
| 396 | if self.attn_backend.get_name() == "FLEX_ATTENTION": |
| 397 | block_m = vllm_config.attention_config.flex_attn_block_m |
| 398 | block_n = vllm_config.attention_config.flex_attn_block_n |
| 399 | |
| 400 | if envs.VLLM_BATCH_INVARIANT and cache_config is not None: |
| 401 | if block_m is not None and block_m > cache_config.block_size: |
| 402 | raise ValueError( |
| 403 | f"flex_attn_block_m ({block_m}) must be " |
| 404 | f"<= cache block size ({cache_config.block_size}) for " |
| 405 | f"batch invariance" |
| 406 | ) |
| 407 | if block_n is not None and block_n > cache_config.block_size: |
| 408 | raise ValueError( |
| 409 | f"flex_attn_block_n ({block_n}) must be " |
| 410 | f"<= cache block size ({cache_config.block_size}) for " |
| 411 | f"batch invariance" |
| 412 | ) |
| 413 | |
| 414 | if block_m is not None: |
| 415 | extra_impl_args.setdefault("block_m", block_m) |
| 416 | if block_n is not None: |
| 417 | extra_impl_args.setdefault("block_n", block_n) |
| 418 | |
| 419 | impl_cls = self.attn_backend.get_impl_cls() |
| 420 | self.impl = impl_cls( # type: ignore[assignment] # impl_cls always returns an AttentionImpl subclass |
| 421 | num_heads, |
| 422 | head_size, |
| 423 | scale, |
| 424 | num_kv_heads, |
| 425 | alibi_slopes, |
| 426 | sliding_window, |
| 427 | kv_cache_dtype, |
| 428 | logits_soft_cap, |
| 429 | attn_type, |
| 430 | kv_sharing_target_layer_name, |
| 431 | **extra_impl_args, |
| 432 | ) |
| 433 | self.backend = AttentionBackendEnum[self.attn_backend.get_name()] |
| 434 | self.dtype = dtype |
| 435 | |
| 436 | # For cuda-alike (CUDA and ROCM) and cpu platforms, we control how |
| 437 | # torch.compile works by registering the attention as one giant |
| 438 | # opaque custom op. For other platforms, we directly call them |
| 439 | # and let torch.compile handle them. |
| 440 | self.use_direct_call = not current_platform.opaque_attention_op() |
| 441 | |
| 442 | compilation_config = vllm_config.compilation_config |
| 443 | if prefix in compilation_config.static_forward_context: |
| 444 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 445 | compilation_config.static_forward_context[prefix] = self |
| 446 | self.attn_type = attn_type |
| 447 | |
| 448 | if kv_sharing_target_layer_name is not None: |
| 449 | validate_kv_sharing_target( |
| 450 | prefix, |
| 451 | kv_sharing_target_layer_name, |
| 452 | compilation_config.static_forward_context, |
| 453 | ) |
| 454 | self.kv_sharing_target_layer_name = kv_sharing_target_layer_name |
| 455 | # Gemma4: clamp mm_prefix bidirectional ranges by the sliding window |
| 456 | # (read by the Triton backend impl). Default False for all other models. |
| 457 | self.mm_prefix_clamp_sliding_window = mm_prefix_clamp_sliding_window |
| 458 | |
| 459 | # use a placeholder kv cache tensor during init, which will be replaced |
| 460 | # by bind_kv_cache |
| 461 | # this variable will not be accessed if use_direct_call is True |
| 462 | self.kv_cache = torch.tensor([]) |
| 463 | |
| 464 | # Initialize KV cache quantization attributes |
| 465 | _init_kv_cache_quant(self, quant_config, prefix) |
| 466 | |
| 467 | # for attn backends supporting query quantization |
| 468 | self.query_quant = None |
| 469 | if ( |
| 470 | self.impl.supports_quant_query_input |
| 471 | and ( |
| 472 | self.kv_cache_dtype.startswith("fp8") |
| 473 | or self.kv_cache_dtype.startswith("nvfp4") |
| 474 | ) |
| 475 | and not self.kv_cache_dtype.endswith("per_token_head") |
| 476 | ): |
| 477 | is_per_head = ( |
| 478 | hasattr(self, "q_scale") and self.q_scale.numel() == self.num_kv_heads |
| 479 | ) |
| 480 | block_size = self.head_size * self.num_heads // self.num_kv_heads |
| 481 | self.query_quant = QuantFP8( |
| 482 | static=True, |
| 483 | group_shape=GroupShape(-1, block_size) |
| 484 | if is_per_head |
| 485 | else GroupShape.PER_TENSOR, |
| 486 | ) |
| 487 | |
| 488 | def forward( |
| 489 | self, |
| 490 | query: torch.Tensor, |
| 491 | key: torch.Tensor, |
| 492 | value: torch.Tensor, |
| 493 | # For some alternate attention backends like MLA the attention output |
| 494 | # shape does not match the query shape, so we optionally let the model |
| 495 | # definition specify the output tensor shape. |
| 496 | output_shape: torch.Size | None = None, |
| 497 | output_dtype: torch.dtype | None = None, |
| 498 | ) -> torch.Tensor: |
| 499 | """The KV cache is stored inside this class and is accessed via |
| 500 | `self.kv_cache`. |
| 501 | |
| 502 | Attention metadata (`attn_metadata`) is set using a context manager in |
| 503 | the model runner's `execute_model` method. It is accessed via forward |
| 504 | context using |
| 505 | `vllm.forward_context.get_forward_context().attn_metadata`. |
| 506 | """ |
| 507 | if output_dtype is None: |
| 508 | output_dtype = query.dtype |
| 509 | if self.query_quant is not None: |
| 510 | # quantizing with a simple torch operation enables |
| 511 | # torch.compile to fuse this into previous ops |
| 512 | # which reduces overheads during decoding. |
| 513 | # Otherwise queries are quantized using custom ops |
| 514 | # which causes decoding overheads |
| 515 | assert self.kv_cache_dtype in {"fp8", "fp8_e4m3"} or ( |
| 516 | self.kv_cache_dtype.startswith("nvfp4") |
| 517 | ) |
| 518 | |
| 519 | # check if query quantization is supported |
| 520 | if self.impl.supports_quant_query_input: |
| 521 | query, _ = self.query_quant(query, self._q_scale) |
| 522 | |
| 523 | if output_shape is None: |
| 524 | # Handle both 2D [num_tokens, hidden] and |
| 525 | # 3D [num_tokens, heads, head_dim] query |
| 526 | num_tokens = query.shape[0] |
| 527 | output_shape = torch.Size((num_tokens, self.num_heads * self.head_size_v)) |
| 528 | output = torch.empty(output_shape, dtype=output_dtype, device=query.device) |
| 529 | hidden_size = output_shape[-1] |
| 530 | # Reshape the query, key, and value tensors. |
| 531 | # NOTE(woosuk): We do this outside the custom op to minimize the |
| 532 | # CPU overheads from the non-CUDA-graph regions. |
| 533 | query = query.view(-1, self.num_heads, self.head_size) |
| 534 | output = output.view(-1, self.num_heads, self.head_size_v) |
| 535 | if key is not None: |
| 536 | key = key.view(-1, self.num_kv_heads, self.head_size) |
| 537 | if value is not None: |
| 538 | value = value.view(-1, self.num_kv_heads, self.head_size_v) |
| 539 | kv_cache_dummy_dep = None |
| 540 | if self.use_direct_call: |
| 541 | # Skip this if sharing KV cache with an earlier attention layer. |
| 542 | if ( |
| 543 | not self.attn_backend.forward_includes_kv_cache_update |
| 544 | and self.kv_sharing_target_layer_name is None |
| 545 | and key is not None |
| 546 | and value is not None |
| 547 | ): |
| 548 | kv_cache_dummy_dep = unified_kv_cache_update( |
| 549 | key, value, self.layer_name |
| 550 | ) |
| 551 | unified_attention_with_output( |
| 552 | query, |
| 553 | key, |
| 554 | value, |
| 555 | output, |
| 556 | self.layer_name, |
| 557 | kv_cache_dummy_dep=kv_cache_dummy_dep, |
| 558 | ) |
| 559 | else: |
| 560 | # Skip this if sharing KV cache with an earlier attention layer. |
| 561 | encoded = _encode_layer_name(self.layer_name) |
| 562 | if ( |
| 563 | not self.attn_backend.forward_includes_kv_cache_update |
| 564 | and self.kv_sharing_target_layer_name is None |
| 565 | and key is not None |
| 566 | and value is not None |
| 567 | ): |
| 568 | kv_cache_dummy_dep = torch.ops.vllm.unified_kv_cache_update( |
| 569 | key, value, encoded |
| 570 | ) |
| 571 | torch.ops.vllm.unified_attention_with_output( |
| 572 | query, |
| 573 | key, |
| 574 | value, |
| 575 | output, |
| 576 | encoded, |
| 577 | kv_cache_dummy_dep=kv_cache_dummy_dep, |
| 578 | ) |
| 579 | return output.view(-1, hidden_size) |
| 580 | |
| 581 | def extra_repr(self) -> str: |
| 582 | s = f"head_size={self.impl.head_size}" # type: ignore |
| 583 | s += f", num_heads={self.impl.num_heads}" # type: ignore |
| 584 | s += f", num_kv_heads={self.impl.num_kv_heads}" # type: ignore |
| 585 | s += f", scale={self.impl.scale}" # type: ignore |
| 586 | s += f", backend={self.impl.__class__.__name__}" |
| 587 | return s |
| 588 | |
| 589 | def process_weights_after_loading(self, act_dtype: torch.dtype): |
| 590 | self.impl.process_weights_after_loading(act_dtype) |
| 591 | |
| 592 | # If we should not load quant weights, we initialize the scales to 1.0 |
| 593 | # as the default value. See [Note: Register q/k/v/prob scales in state dict] |
| 594 | # for more details. |
| 595 | quant_method = ( |
| 596 | resolve_quant_method(self.quant_config, self, prefix=self.layer_name) |
| 597 | if self.quant_config |
| 598 | else None |
| 599 | ) |
| 600 | if not should_load_quant_weights(quant_method): |
| 601 | set_default_quant_scales(self, register_buffer=False) |
| 602 | |
| 603 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 604 | return self.attn_backend |
| 605 | |
| 606 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None: |
| 607 | # Block size may get updated after model loading, refresh it |
| 608 | block_size = vllm_config.cache_config.block_size |
| 609 | # Encoder-only attention is prefill-only and keeps no autoregressive KV |
| 610 | # cache. In hybrid models (e.g. Qwen3.5 / ColQwen3.5: GatedDeltaNet |
| 611 | # linear_attention interleaved with full_attention) the runner iterates |
| 612 | # every attention module to build the KV-cache spec, so an ENCODER_ONLY |
| 613 | # full_attention layer reaches here; it contributes no KV cache group. |
| 614 | if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER): |
| 615 | return None |
| 616 | # Should not be called for enc-dec attention. |
| 617 | assert self.attn_type == AttentionType.DECODER |
| 618 | quant_mode = get_kv_quant_mode(self.kv_cache_dtype) |
| 619 | if self.sliding_window is not None: |
| 620 | assert not self.attn_backend.is_mla(), ( |
| 621 | "MLA is not supported for sliding window" |
| 622 | ) |
| 623 | # SW chooses its own block_size, decoupled from the user's |
| 624 | # ``--block-size`` (which only constrains primary attention). |
| 625 | # When this SW layer is a padded spec (skip-quant: its page is |
| 626 | # padded up to ``skip_page_size_padded``), pick the largest kernel |
| 627 | # block that still fits the shared page so we waste fewer padding |
| 628 | # bytes per block. Otherwise (page_size_padded is None) take the |
| 629 | # primary block size when the backend can run it unsplit: if this |
| 630 | # page does not divide the primary page, ``unify`` then pads it |
| 631 | # (padded pages cannot be split) instead of scaling a small block |
| 632 | # to a size coprime with the primary one, which inflates the |
| 633 | # scheduler LCM (e.g. a 1024 B/token SWA draft next to a 1152 |
| 634 | # B/token MLA target: 1728 vs 1536 gives LCM 13824). Backends that |
| 635 | # cannot run the primary block start from their smallest block and |
| 636 | # ``unify`` scales it up by an integer ratio. |
| 637 | shared_page = vllm_config.cache_config.skip_page_size_padded |
| 638 | # The backend owns its packing |
| 639 | kv_cache_spec = self.attn_backend.customize_spec( |
| 640 | SlidingWindowSpec( |
| 641 | block_size=1, |
| 642 | num_kv_heads=self.num_kv_heads, |
| 643 | head_size=self.head_size, |
| 644 | head_size_v=self.head_size_v, |
| 645 | dtype=self.kv_cache_torch_dtype, |
| 646 | kv_quant_mode=quant_mode, |
| 647 | sliding_window=self.sliding_window, |
| 648 | ) |
| 649 | ) |
| 650 | sw_per_token = kv_cache_spec.real_page_size_bytes |
| 651 | page_budget = shared_page or sw_per_token * block_size |
| 652 | sw_block_size = _largest_kernel_block_within( |
| 653 | self.attn_backend, |
| 654 | sw_per_token, |
| 655 | page_budget, |
| 656 | block_size, |
| 657 | kv_cache_spec, |
| 658 | ) |
| 659 | return SlidingWindowSpec( |
| 660 | block_size=sw_block_size, |
| 661 | num_kv_heads=self.num_kv_heads, |
| 662 | head_size=self.head_size, |
| 663 | head_size_v=self.head_size_v, |
| 664 | dtype=self.kv_cache_torch_dtype, |
| 665 | kv_quant_mode=quant_mode, |
| 666 | sliding_window=self.sliding_window, |
| 667 | page_size_padded=shared_page, |
| 668 | ) |
| 669 | else: |
| 670 | return FullAttentionSpec( |
| 671 | block_size=block_size, |
| 672 | num_kv_heads=self.num_kv_heads, |
| 673 | head_size=self.head_size, |
| 674 | head_size_v=self.head_size_v, |
| 675 | dtype=self.kv_cache_torch_dtype, |
| 676 | kv_quant_mode=quant_mode, |
| 677 | ) |
| 678 | |
| 679 | |
| 680 | def get_attention_context( |
| 681 | layer_name: str, |
| 682 | ) -> tuple[Any, "Attention | MLAAttention", torch.Tensor, torch.Tensor]: |
| 683 | """Extract attention context for a given layer. |
| 684 | |
| 685 | This helper function extracts the attention metadata, attention layer |
| 686 | instance, KV cache tensor, and slot mapping for a specific layer. |
| 687 | |
| 688 | Args: |
| 689 | layer_name: The name/identifier of the attention layer. |
| 690 | |
| 691 | Returns: |
| 692 | A tuple containing: |
| 693 | - attn_metadata: Attention metadata for this specific layer, or None if |
| 694 | no metadata available |
| 695 | - attn_layer: The attention layer instance (Attention or MLAAttention) |
| 696 | - kv_cache: The KV cache tensor for current forward pass |
| 697 | - slot_mapping: The slot mapping for this specific layer |
| 698 | |
| 699 | Note: attn_metadata may be None, but attn_layer and kv_cache are always |
| 700 | extracted from the forward context. |
| 701 | |
| 702 | """ |
| 703 | forward_context: ForwardContext = get_forward_context() |
| 704 | attn_metadata_raw = forward_context.attn_metadata |
| 705 | attn_metadata: AttentionMetadata |
| 706 | if isinstance(attn_metadata_raw, dict): |
| 707 | attn_metadata = attn_metadata_raw[layer_name] |
| 708 | elif isinstance(attn_metadata_raw, list): |
| 709 | # list[dict[str, AttentionMetadata]]: used in speculative decoding |
| 710 | # where [0] is the base-model (non-speculative) metadata dict. |
| 711 | attn_metadata = attn_metadata_raw[0][layer_name] |
| 712 | else: |
| 713 | attn_metadata = attn_metadata_raw |
| 714 | attn_layer: Attention | MLAAttention = forward_context.no_compile_layers[layer_name] |
| 715 | kv_cache = attn_layer.kv_cache |
| 716 | slot_mapping = forward_context.slot_mapping |
| 717 | assert isinstance(slot_mapping, dict), ( |
| 718 | f"Expected slot_mapping to be a dict, got {type(slot_mapping)}. " |
| 719 | ) |
| 720 | layer_slot_mapping = slot_mapping.get(layer_name) |
| 721 | return attn_metadata, attn_layer, kv_cache, layer_slot_mapping |
| 722 | |
| 723 | |
| 724 | def unified_kv_cache_update( |
| 725 | key: torch.Tensor, |
| 726 | value: torch.Tensor, |
| 727 | layer_name: LayerNameType, |
| 728 | ) -> torch.Tensor: |
| 729 | """Returns a dummy that is passed to unified_attention to signal a side effect and |
| 730 | the data dependency between them to ensure torch.compile preserves ordering. |
| 731 | """ |
| 732 | layer_name = _resolve_layer_name(layer_name) |
| 733 | _, attn_layer, kv_cache, layer_slot_mapping = get_attention_context(layer_name) |
| 734 | if layer_slot_mapping is not None: |
| 735 | assert hasattr(attn_layer.impl, "do_kv_cache_update"), ( |
| 736 | f"{attn_layer.impl.__class__.__name__} does not support kv cache update" |
| 737 | ) |
| 738 | attn_layer.impl.do_kv_cache_update( # type: ignore[attr-defined] |
| 739 | attn_layer, |
| 740 | key, |
| 741 | value, |
| 742 | kv_cache, |
| 743 | layer_slot_mapping, |
| 744 | ) |
| 745 | |
| 746 | return key.new_empty(0) |
| 747 | |
| 748 | |
| 749 | def unified_kv_cache_update_fake( |
| 750 | key: torch.Tensor, |
| 751 | value: torch.Tensor, |
| 752 | layer_name: LayerNameType, |
| 753 | ) -> torch.Tensor: |
| 754 | return torch.empty(0, device=key.device, dtype=key.dtype) |
| 755 | |
| 756 | |
| 757 | direct_register_custom_op( |
| 758 | op_name="unified_kv_cache_update", |
| 759 | op_func=unified_kv_cache_update, |
| 760 | fake_impl=unified_kv_cache_update_fake, |
| 761 | mutates_args=[], |
| 762 | ) |
| 763 | |
| 764 | |
| 765 | @eager_break_during_capture |
| 766 | @maybe_transfer_kv_layer |
| 767 | def unified_attention_with_output( |
| 768 | query: torch.Tensor, |
| 769 | key: torch.Tensor, |
| 770 | value: torch.Tensor, |
| 771 | output: torch.Tensor, |
| 772 | layer_name: LayerNameType, |
| 773 | output_scale: torch.Tensor | None = None, |
| 774 | output_block_scale: torch.Tensor | None = None, |
| 775 | kv_cache_dummy_dep: torch.Tensor | None = None, |
| 776 | ) -> None: |
| 777 | # kv_cache_dummy_dep is not used but accepting it creates a data dependency |
| 778 | # that ensures torch.compile preserves ordering between KV cache update and |
| 779 | # attention forward. |
| 780 | del kv_cache_dummy_dep |
| 781 | layer_name = _resolve_layer_name(layer_name) |
| 782 | attn_metadata, self, kv_cache, _ = get_attention_context(layer_name) |
| 783 | |
| 784 | self.impl.forward( |
| 785 | self, |
| 786 | query, |
| 787 | key, |
| 788 | value, |
| 789 | kv_cache, |
| 790 | attn_metadata, |
| 791 | output=output, |
| 792 | output_scale=output_scale, |
| 793 | output_block_scale=output_block_scale, |
| 794 | ) |
| 795 | |
| 796 | |
| 797 | direct_register_custom_op( |
| 798 | op_name="unified_attention_with_output", |
| 799 | op_func=unified_attention_with_output, |
| 800 | mutates_args=["output", "output_block_scale"], |
| 801 | ) |