| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import functools |
| 4 | from dataclasses import dataclass |
| 5 | from typing import TYPE_CHECKING, ClassVar |
| 6 | |
| 7 | if TYPE_CHECKING: |
| 8 | from vllm.config.cache import CacheDType |
| 9 | |
| 10 | import torch |
| 11 | |
| 12 | from vllm import _custom_ops as ops |
| 13 | from vllm import envs |
| 14 | from vllm.config import ( |
| 15 | VllmConfig, |
| 16 | get_current_vllm_config, |
| 17 | get_layers_from_vllm_config, |
| 18 | ) |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.model_executor.layers.attention import Attention |
| 21 | from vllm.platforms import CpuArchEnum, current_platform |
| 22 | from vllm.utils.torch_utils import is_quantized_kv_cache |
| 23 | from vllm.v1.attention.backend import ( |
| 24 | AttentionBackend, |
| 25 | AttentionImpl, |
| 26 | AttentionLayer, |
| 27 | AttentionMetadataBuilder, |
| 28 | AttentionType, |
| 29 | CommonAttentionMetadata, |
| 30 | MultipleOf, |
| 31 | ) |
| 32 | from vllm.v1.attention.backends.utils import ( |
| 33 | get_num_attention_heads_from_layers, |
| 34 | ) |
| 35 | from vllm.v1.attention.backends.zentorch_sdpa import ( |
| 36 | should_use_zentorch_sdpa, |
| 37 | zentorch_sdpa_attn, |
| 38 | ) |
| 39 | from vllm.v1.kv_cache_interface import ( |
| 40 | AttentionSpec, |
| 41 | CrossAttentionSpec, |
| 42 | EncoderOnlyAttentionSpec, |
| 43 | KVCacheLayout, |
| 44 | ) |
| 45 | |
| 46 | logger = init_logger(__name__) |
| 47 | |
| 48 | |
| 49 | class CPUAttentionBackend(AttentionBackend): |
| 50 | forward_includes_kv_cache_update: bool = False |
| 51 | |
| 52 | supported_dtypes: ClassVar[list[torch.dtype]] = [ |
| 53 | torch.float16, |
| 54 | torch.bfloat16, |
| 55 | torch.float32, |
| 56 | ] |
| 57 | supported_kv_cache_dtypes: ClassVar[list["CacheDType"]] = [ |
| 58 | "auto", |
| 59 | "fp8", |
| 60 | "fp8_e4m3", |
| 61 | "fp8_e5m2", |
| 62 | ] |
| 63 | |
| 64 | @staticmethod |
| 65 | def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: |
| 66 | return [MultipleOf(32)] |
| 67 | |
| 68 | @classmethod |
| 69 | def get_supported_head_sizes(cls) -> list[int]: |
| 70 | return [32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512] |
| 71 | |
| 72 | @classmethod |
| 73 | def supported_kv_cache_layouts(cls) -> tuple[KVCacheLayout, ...]: |
| 74 | # The CPU backend only reads head-major block interiors. |
| 75 | return (KVCacheLayout.LBHNC,) |
| 76 | |
| 77 | @staticmethod |
| 78 | def get_name() -> str: |
| 79 | return "CPU_ATTN" |
| 80 | |
| 81 | @classmethod |
| 82 | def supports_non_causal(cls) -> bool: |
| 83 | return True |
| 84 | |
| 85 | @classmethod |
| 86 | def supports_sliding_window(cls) -> bool: |
| 87 | return True |
| 88 | |
| 89 | @classmethod |
| 90 | def supports_attn_type(cls, attn_type: str) -> bool: |
| 91 | """CPU attention supports decoder, |
| 92 | encoder-only and encoder-decoder attention.""" |
| 93 | return attn_type in ( |
| 94 | AttentionType.DECODER, |
| 95 | AttentionType.ENCODER, |
| 96 | AttentionType.ENCODER_ONLY, |
| 97 | AttentionType.ENCODER_DECODER, |
| 98 | ) |
| 99 | |
| 100 | @staticmethod |
| 101 | def get_impl_cls() -> type["CPUAttentionBackendImpl"]: |
| 102 | return CPUAttentionBackendImpl |
| 103 | |
| 104 | @staticmethod |
| 105 | def get_builder_cls() -> type["CPUAttentionMetadataBuilder"]: |
| 106 | return CPUAttentionMetadataBuilder |
| 107 | |
| 108 | @staticmethod |
| 109 | def use_cascade_attention(*args, **kwargs) -> bool: |
| 110 | return False |
| 111 | |
| 112 | |
| 113 | @dataclass |
| 114 | class CPUAttentionMetadata: |
| 115 | num_actual_tokens: int # Number of tokens excluding padding. |
| 116 | max_query_len: int |
| 117 | query_start_loc: torch.Tensor |
| 118 | max_seq_len: int |
| 119 | seq_lens: torch.Tensor |
| 120 | block_table: torch.Tensor |
| 121 | slot_mapping: torch.Tensor |
| 122 | scheduler_metadata: torch.Tensor | None |
| 123 | causal: bool = True |
| 124 | dynamic_causal: torch.Tensor | None = None |
| 125 | |
| 126 | # can be removed after deprecate sdpa |
| 127 | use_sdpa_prefill: bool = False |
| 128 | num_decode_tokens: int = 0 |
| 129 | sdpa_attn_masks: list[torch.Tensor | None] | None = None |
| 130 | sdpa_start_loc: torch.Tensor | None = None |
| 131 | |
| 132 | encoder_cache: torch.Tensor | None = None |
| 133 | |
| 134 | |
| 135 | class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]): |
| 136 | def __init__( |
| 137 | self, |
| 138 | kv_cache_spec: AttentionSpec, |
| 139 | layer_names: list[str], |
| 140 | vllm_config: VllmConfig, |
| 141 | device: torch.device, |
| 142 | ) -> None: |
| 143 | super().__init__(kv_cache_spec, layer_names, vllm_config, device) |
| 144 | |
| 145 | self.kv_cache_spec = kv_cache_spec |
| 146 | self.vllm_config = vllm_config |
| 147 | |
| 148 | parallel_config = vllm_config.parallel_config |
| 149 | self.num_kv_heads = kv_cache_spec.num_kv_heads |
| 150 | # The scheduler metadata built here sizes a scratchpad from the query |
| 151 | # head count, so it must come from this group's layers: the model-wide |
| 152 | # count is wrong for models that vary it per layer (e.g. Laguna). |
| 153 | self.num_heads = get_num_attention_heads_from_layers( |
| 154 | vllm_config, layer_names |
| 155 | ) or vllm_config.model_config.get_num_attention_heads(parallel_config) |
| 156 | self.head_dim = kv_cache_spec.head_size |
| 157 | self.dtype = vllm_config.model_config.dtype |
| 158 | self.window_size = self._group_sliding_window() |
| 159 | self.block_size = kv_cache_spec.block_size |
| 160 | self.kv_cache_dtype = vllm_config.cache_config.cache_dtype |
| 161 | self.isa = _get_attn_isa( |
| 162 | self.dtype, |
| 163 | self.block_size, |
| 164 | self.head_dim, |
| 165 | self.kv_cache_dtype, |
| 166 | ) |
| 167 | self._set_isa_to_layers(layer_names) |
| 168 | self.is_cross_attention = isinstance(kv_cache_spec, CrossAttentionSpec) |
| 169 | self.is_encoder_only_attention = isinstance( |
| 170 | kv_cache_spec, EncoderOnlyAttentionSpec |
| 171 | ) |
| 172 | |
| 173 | def _set_isa_to_layers(self, layer_names: list[str]) -> None: |
| 174 | attn_layers = get_layers_from_vllm_config( |
| 175 | self.vllm_config, |
| 176 | Attention, |
| 177 | layer_names, |
| 178 | ) |
| 179 | for layer in attn_layers.values(): |
| 180 | layer.isa = self.isa # type: ignore |
| 181 | |
| 182 | def _group_sliding_window(self) -> int: |
| 183 | """The window shared by every layer in this group, else -1 (no window). |
| 184 | |
| 185 | Taken from the layers rather than the group spec: one KV cache group can |
| 186 | hold both windowed and global layers (e.g. Gemma-3 with the hybrid KV |
| 187 | cache manager disabled), and the scheduler metadata built here is shared |
| 188 | by the whole group, so it may only assume a window all of them agree on. |
| 189 | """ |
| 190 | layers = get_layers_from_vllm_config( |
| 191 | self.vllm_config, Attention, self.layer_names |
| 192 | ) |
| 193 | windows = { |
| 194 | layer.impl.sliding_window |
| 195 | for layer in layers.values() |
| 196 | if isinstance(layer.impl, CPUAttentionBackendImpl) |
| 197 | } |
| 198 | if len(windows) != 1: |
| 199 | return -1 |
| 200 | window = windows.pop() |
| 201 | return -1 if window is None else window |
| 202 | |
| 203 | def build( |
| 204 | self, |
| 205 | common_prefix_len: int, |
| 206 | common_attn_metadata: CommonAttentionMetadata, |
| 207 | fast_build: bool = False, |
| 208 | ) -> CPUAttentionMetadata: |
| 209 | num_reqs = common_attn_metadata.num_reqs |
| 210 | num_actual_tokens = common_attn_metadata.num_actual_tokens |
| 211 | max_query_len = common_attn_metadata.max_query_len |
| 212 | max_seq_len = common_attn_metadata.max_seq_len |
| 213 | query_start_loc = common_attn_metadata.query_start_loc |
| 214 | seq_lens = common_attn_metadata.seq_lens |
| 215 | block_table_tensor = common_attn_metadata.block_table_tensor |
| 216 | slot_mapping = common_attn_metadata.slot_mapping |
| 217 | is_dynamic_casual = isinstance(common_attn_metadata.causal, torch.Tensor) |
| 218 | dynamic_casual = None |
| 219 | if is_dynamic_casual: |
| 220 | dynamic_casual = common_attn_metadata.causal |
| 221 | |
| 222 | causal = ( |
| 223 | False |
| 224 | if self.is_cross_attention or is_dynamic_casual |
| 225 | else common_attn_metadata.causal |
| 226 | ) |
| 227 | |
| 228 | encoder_cache_tensor = None |
| 229 | if self.is_encoder_only_attention: |
| 230 | block_nums = (seq_lens + self.block_size - 1) // self.block_size |
| 231 | start_block_ids = torch.zeros_like(seq_lens) |
| 232 | torch.cumsum(block_nums[:-1], 0, out=start_block_ids[1:]) |
| 233 | total_block_num: int = block_nums.sum().item() |
| 234 | max_block_num = block_nums.max().item() |
| 235 | block_offsets = torch.arange( |
| 236 | 0, max_block_num, dtype=block_table_tensor.dtype |
| 237 | ) |
| 238 | encoder_block_table = start_block_ids[:, None] + block_offsets[None, :] |
| 239 | torch.ops._C.compute_slot_mapping_kernel_impl( |
| 240 | query_start_loc, |
| 241 | common_attn_metadata.positions, |
| 242 | encoder_block_table, |
| 243 | slot_mapping, |
| 244 | self.block_size, |
| 245 | ) |
| 246 | encoder_cache_tensor = torch.zeros( |
| 247 | ( |
| 248 | total_block_num, |
| 249 | self.num_kv_heads, |
| 250 | self.block_size, |
| 251 | 2 * self.head_dim, |
| 252 | ), |
| 253 | dtype=self.dtype, |
| 254 | ) |
| 255 | block_table_tensor = encoder_block_table |
| 256 | |
| 257 | scheduler_metadata = ops.cpu_attn_get_scheduler_metadata( |
| 258 | num_reqs=num_reqs, |
| 259 | num_heads=self.num_heads, |
| 260 | num_kv_heads=self.num_kv_heads, |
| 261 | head_dim=self.head_dim, |
| 262 | seq_lens=seq_lens, |
| 263 | dtype=self.dtype, |
| 264 | query_start_loc=query_start_loc, |
| 265 | causal=causal, |
| 266 | sliding_window_size=self.window_size, |
| 267 | isa=self.isa, |
| 268 | enable_kv_split=envs.VLLM_CPU_ATTN_SPLIT_KV, |
| 269 | dynamic_causal=dynamic_casual, |
| 270 | kv_cache_dtype=self.kv_cache_dtype, |
| 271 | ) |
| 272 | |
| 273 | attn_metadata = CPUAttentionMetadata( |
| 274 | num_actual_tokens=num_actual_tokens, |
| 275 | max_query_len=max_query_len, |
| 276 | query_start_loc=query_start_loc, |
| 277 | max_seq_len=max_seq_len, |
| 278 | seq_lens=seq_lens, |
| 279 | block_table=block_table_tensor, |
| 280 | slot_mapping=slot_mapping, |
| 281 | scheduler_metadata=scheduler_metadata, |
| 282 | causal=causal, |
| 283 | encoder_cache=encoder_cache_tensor, |
| 284 | dynamic_causal=dynamic_casual, |
| 285 | ) |
| 286 | |
| 287 | return attn_metadata |
| 288 | |
| 289 | |
| 290 | class CPUAttentionBackendImpl(AttentionImpl): |
| 291 | def __init__( |
| 292 | self, |
| 293 | num_heads: int, |
| 294 | head_size: int, |
| 295 | scale: float, |
| 296 | num_kv_heads: int, |
| 297 | alibi_slopes: list[float] | None, |
| 298 | sliding_window: int | None, |
| 299 | kv_cache_dtype: str, |
| 300 | logits_soft_cap: float | None = None, |
| 301 | attn_type: str = AttentionType.DECODER, |
| 302 | kv_sharing_target_layer_name: str | None = None, |
| 303 | sinks: torch.Tensor | None = None, |
| 304 | ) -> None: |
| 305 | self.kv_sharing_target_layer_name = kv_sharing_target_layer_name |
| 306 | self.num_heads = num_heads |
| 307 | self.head_size = head_size |
| 308 | self.scale = float(scale) |
| 309 | if logits_soft_cap is not None and attn_type in ( |
| 310 | AttentionType.ENCODER, |
| 311 | AttentionType.ENCODER_ONLY, |
| 312 | ): |
| 313 | logger.warning_once( |
| 314 | "CPU_ATTN does not support logits softcap for" |
| 315 | " ENCODER and ENCODER_ONLY, outputs may be slightly off" |
| 316 | ) |
| 317 | if logits_soft_cap is None: |
| 318 | logits_soft_cap = 0 |
| 319 | self.logits_soft_cap = logits_soft_cap |
| 320 | |
| 321 | self.num_kv_heads = num_kv_heads |
| 322 | if alibi_slopes is not None: |
| 323 | alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32) |
| 324 | self.alibi_slopes = alibi_slopes |
| 325 | if sliding_window is None: |
| 326 | self.sliding_window = -1 |
| 327 | else: |
| 328 | self.sliding_window = sliding_window |
| 329 | self.kv_cache_dtype = kv_cache_dtype |
| 330 | self.num_queries_per_kv = self.num_heads // self.num_kv_heads |
| 331 | |
| 332 | self.is_fp8_kv_cache = is_quantized_kv_cache(kv_cache_dtype) |
| 333 | self.attn_type = attn_type |
| 334 | |
| 335 | self.sinks = sinks |
| 336 | if self.sinks is not None: |
| 337 | assert self.sinks.shape[0] == num_heads, ( |
| 338 | "Sinks must have the same number of heads as the number of " |
| 339 | "heads in the layer" |
| 340 | ) |
| 341 | |
| 342 | vllm_config = get_current_vllm_config() |
| 343 | self.use_zentorch_sdpa = should_use_zentorch_sdpa( |
| 344 | attn_type, |
| 345 | vllm_config.model_config.dtype, |
| 346 | ) |
| 347 | |
| 348 | def forward( |
| 349 | self, |
| 350 | layer: AttentionLayer, |
| 351 | query: torch.Tensor, |
| 352 | key: torch.Tensor, |
| 353 | value: torch.Tensor, |
| 354 | kv_cache: torch.Tensor, |
| 355 | attn_metadata: CPUAttentionMetadata | None, |
| 356 | output: torch.Tensor, |
| 357 | output_scale: torch.Tensor | None = None, |
| 358 | output_block_scale: torch.Tensor | None = None, |
| 359 | ) -> torch.Tensor: |
| 360 | """Forward pass for CPU attention backend. |
| 361 | |
| 362 | Args: |
| 363 | layer: The attention layer, providing the q/k/v quantization scales. |
| 364 | query: shape = [num_tokens, num_heads, head_size] |
| 365 | key: shape = [num_tokens, num_kv_heads, head_size] |
| 366 | value: shape = [num_tokens, num_kv_heads, head_size] |
| 367 | kv_cache: shape = |
| 368 | [num_blocks, num_kv_heads, block_size, 2 * head_size] |
| 369 | attn_metadata: Metadata for attention. |
| 370 | output: Tensor that the attention result is written into. |
| 371 | output_scale: Scale for fused output quantization; not supported |
| 372 | by this backend. |
| 373 | output_block_scale: Block scale for fused output quantization; |
| 374 | not supported by this backend. |
| 375 | |
| 376 | Returns: |
| 377 | shape = [num_tokens, num_heads * head_size] |
| 378 | |
| 379 | """ |
| 380 | if output_scale is not None or output_block_scale is not None: |
| 381 | raise NotImplementedError( |
| 382 | "fused output quantization is not yet supported" |
| 383 | " for CPUAttentionBackendImpl" |
| 384 | ) |
| 385 | |
| 386 | # For warming-up |
| 387 | if attn_metadata is None: |
| 388 | return output |
| 389 | |
| 390 | num_actual_tokens = attn_metadata.num_actual_tokens |
| 391 | |
| 392 | is_encoder_attention = self.attn_type in ( |
| 393 | AttentionType.ENCODER_ONLY, |
| 394 | AttentionType.ENCODER, |
| 395 | ) |
| 396 | if is_encoder_attention: |
| 397 | if self.use_zentorch_sdpa: |
| 398 | # Encoder attention never reads the KV cache back, so the |
| 399 | # zentorch path attends the packed QKV directly instead of |
| 400 | # staging it through the scratch encoder cache. |
| 401 | zentorch_sdpa_attn( |
| 402 | query[:num_actual_tokens], |
| 403 | key[:num_actual_tokens], |
| 404 | value[:num_actual_tokens], |
| 405 | output[:num_actual_tokens], |
| 406 | attn_metadata, |
| 407 | self.scale, |
| 408 | self.sliding_window, |
| 409 | self.alibi_slopes, |
| 410 | ) |
| 411 | return output |
| 412 | # For encoder attention, |
| 413 | kv_cache = attn_metadata.encoder_cache |
| 414 | |
| 415 | # KV cache size are [num_blocks, num_kv_heads, block_size, |
| 416 | # 2 * head_size]. Make a view [num_blocks, num_kv_heads, |
| 417 | # block_size * 2, head_size]. Then slice KV at dim 2 |
| 418 | num_blocks, num_kv_heads, block_size, _ = kv_cache.size() |
| 419 | kv_cache = kv_cache.view((num_blocks, num_kv_heads, block_size * 2, -1)) |
| 420 | key_cache, value_cache = kv_cache.chunk(2, dim=2) |
| 421 | |
| 422 | if is_encoder_attention: |
| 423 | ops.cpu_attn_reshape_and_cache( |
| 424 | key, |
| 425 | value, |
| 426 | key_cache, |
| 427 | value_cache, |
| 428 | attn_metadata.slot_mapping, |
| 429 | layer.isa, # type: ignore |
| 430 | k_scale=layer._k_scale_float, |
| 431 | v_scale=layer._v_scale_float, |
| 432 | kv_cache_dtype=self.kv_cache_dtype, |
| 433 | ) |
| 434 | |
| 435 | # The CPU kernel executes attention sinks natively in bf16. If the |
| 436 | # sinks tensor is anything other than bf16, cast it to fp32 so it is |
| 437 | # executed in full float precision (done lazily here, after weights |
| 438 | # are loaded, rather than at __init__ time). |
| 439 | if self.sinks is not None and self.sinks.dtype not in [ |
| 440 | torch.bfloat16, |
| 441 | torch.float32, |
| 442 | ]: |
| 443 | self.sinks = self.sinks.to(torch.float32) |
| 444 | ops.cpu_attention_with_kv_cache( |
| 445 | query=query[:num_actual_tokens], |
| 446 | key_cache=key_cache, |
| 447 | value_cache=value_cache, |
| 448 | output=output[:num_actual_tokens], # type: ignore |
| 449 | query_start_loc=attn_metadata.query_start_loc, |
| 450 | seq_lens=attn_metadata.seq_lens, |
| 451 | scale=self.scale, |
| 452 | causal=attn_metadata.causal, |
| 453 | alibi_slopes=self.alibi_slopes, # type: ignore |
| 454 | sliding_window=self.sliding_window, |
| 455 | block_table=attn_metadata.block_table, |
| 456 | softcap=self.logits_soft_cap, |
| 457 | scheduler_metadata=attn_metadata.scheduler_metadata, |
| 458 | s_aux=self.sinks, |
| 459 | dynamic_causal=attn_metadata.dynamic_causal, |
| 460 | k_scale=layer._k_scale_float, |
| 461 | v_scale=layer._v_scale_float, |
| 462 | kv_cache_dtype=self.kv_cache_dtype, |
| 463 | ) |
| 464 | |
| 465 | return output |
| 466 | |
| 467 | def do_kv_cache_update( |
| 468 | self, |
| 469 | layer: torch.nn.Module, |
| 470 | key: torch.Tensor, |
| 471 | value: torch.Tensor, |
| 472 | kv_cache: torch.Tensor, |
| 473 | slot_mapping: torch.Tensor, |
| 474 | ) -> None: |
| 475 | if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER): |
| 476 | return |
| 477 | |
| 478 | num_blocks, num_kv_heads, block_size, _ = kv_cache.size() |
| 479 | kv_cache = kv_cache.view((num_blocks, num_kv_heads, block_size * 2, -1)) |
| 480 | key_cache, value_cache = kv_cache.chunk(2, dim=2) |
| 481 | ops.cpu_attn_reshape_and_cache( |
| 482 | key, |
| 483 | value, |
| 484 | key_cache, |
| 485 | value_cache, |
| 486 | slot_mapping, |
| 487 | layer.isa, # type: ignore |
| 488 | k_scale=layer._k_scale_float, |
| 489 | v_scale=layer._v_scale_float, |
| 490 | kv_cache_dtype=self.kv_cache_dtype, |
| 491 | ) |
| 492 | |
| 493 | |
| 494 | @functools.lru_cache(maxsize=1) |
| 495 | def _riscv_supports_rvv() -> bool: |
| 496 | """Whether the C++ RVV attention path is usable. |
| 497 | |
| 498 | The kernel in csrc/cpu/cpu_attn_rvv.hpp uses VLEN-agnostic RVVI() |
| 499 | macros and supports VLEN=128 and VLEN=256. CMake auto-detects the |
| 500 | largest zvl<N>b from /proc/cpuinfo and passes it via -mrvv-vector-bits. |
| 501 | The RVV path is compiled whenever __riscv_v_min_vlen is defined, so |
| 502 | we check that at least one supported zvl<N>b is advertised. |
| 503 | """ |
| 504 | # The C++ compile-time check is the ground truth: it knows which |
| 505 | # VLEN the binary was actually compiled for. The cpuinfo check |
| 506 | # below is only a fast-path shortcut. |
| 507 | try: |
| 508 | import torch |
| 509 | |
| 510 | if torch.ops._C.cpu_attn_has_isa("rvv"): |
| 511 | return True |
| 512 | except Exception: |
| 513 | pass |
| 514 | |
| 515 | # Fallback: check /proc/cpuinfo for zvl128b/zvl256b. |
| 516 | try: |
| 517 | with open("/proc/cpuinfo") as f: |
| 518 | cpuinfo = f.read() |
| 519 | except OSError: |
| 520 | return False |
| 521 | return any(f"zvl{n}b" in cpuinfo for n in (128, 256)) |
| 522 | |
| 523 | |
| 524 | def _get_attn_isa( |
| 525 | dtype: torch.dtype, |
| 526 | block_size: int, |
| 527 | head_size: int | None = None, |
| 528 | kv_cache_dtype: str | None = None, |
| 529 | ) -> str: |
| 530 | fp8_kv = is_quantized_kv_cache(kv_cache_dtype) if kv_cache_dtype else False |
| 531 | if head_size is not None and head_size % 32 != 0 and head_size % 16 == 0: |
| 532 | if fp8_kv: |
| 533 | raise NotImplementedError( |
| 534 | "FP8 KV cache requires head_size divisible by 32 on CPU." |
| 535 | ) |
| 536 | return "vec16" |
| 537 | supports_amx = torch.cpu._is_amx_tile_supported() |
| 538 | arch = current_platform.get_cpu_architecture() |
| 539 | supports_arm = arch == CpuArchEnum.ARM |
| 540 | supports_vxe = arch == CpuArchEnum.S390X |
| 541 | supports_riscv = arch == CpuArchEnum.RISCV |
| 542 | supports_vsx = arch == CpuArchEnum.POWERPC |
| 543 | supports_avx512 = torch.cpu._is_avx512_supported() |
| 544 | if fp8_kv and not supports_amx and not supports_avx512: |
| 545 | raise NotImplementedError( |
| 546 | "FP8 KV cache on CPU requires x86 with AVX-512 or AMX." |
| 547 | ) |
| 548 | if supports_amx and dtype in (torch.bfloat16,) and block_size % 32 == 0: |
| 549 | return "amx" |
| 550 | elif supports_arm: |
| 551 | return "neon" |
| 552 | elif block_size % 32 == 0: |
| 553 | if supports_riscv and _riscv_supports_rvv(): |
| 554 | return "rvv" |
| 555 | elif supports_vxe: |
| 556 | return "vxe" |
| 557 | elif supports_vsx: |
| 558 | return "vsx" |
| 559 | else: |
| 560 | return "vec" |
| 561 | else: |
| 562 | return "vec16" |