| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import functools |
| 4 | import logging |
| 5 | import math |
| 6 | from dataclasses import replace |
| 7 | from fractions import Fraction |
| 8 | from functools import partial |
| 9 | from typing import ClassVar, Protocol |
| 10 | |
| 11 | import torch |
| 12 | import torch.nn.functional as F |
| 13 | from torch import nn |
| 14 | |
| 15 | from vllm.config import CacheConfig, VllmConfig |
| 16 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 17 | from vllm.model_executor.layers.attention import Attention |
| 18 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 19 | from vllm.model_executor.layers.linear import ( |
| 20 | QKVParallelLinear, |
| 21 | RowParallelLinear, |
| 22 | ) |
| 23 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 24 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 25 | from vllm.model_executor.models.mistral import MistralMLP |
| 26 | from vllm.model_executor.models.whisper import WhisperPosEmbedType |
| 27 | from vllm.utils.math_utils import cdiv |
| 28 | from vllm.v1.attention.backend import ( |
| 29 | AttentionBackend, |
| 30 | AttentionCGSupport, |
| 31 | AttentionMetadata, |
| 32 | AttentionType, |
| 33 | CommonAttentionMetadata, |
| 34 | subclass_attention_backend_with_overrides, |
| 35 | ) |
| 36 | from vllm.v1.attention.backends.cpu_attn import CPUAttentionBackend |
| 37 | from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend |
| 38 | from vllm.v1.attention.backends.rocm_attn import RocmAttentionBackend |
| 39 | from vllm.v1.attention.backends.triton_attn import TritonAttentionBackend |
| 40 | from vllm.v1.attention.selector import get_attn_backend |
| 41 | from vllm.v1.kv_cache_interface import AttentionSpec, SlidingWindowSpec |
| 42 | |
| 43 | from .utils import make_layers |
| 44 | |
| 45 | AiterFlashAttentionBackend: type[AttentionBackend] | None = None |
| 46 | try: |
| 47 | from vllm.v1.attention.backends import rocm_aiter_fa |
| 48 | |
| 49 | AiterFlashAttentionBackend = rocm_aiter_fa.AiterFlashAttentionBackend |
| 50 | except ImportError: |
| 51 | pass |
| 52 | |
| 53 | logger = logging.getLogger(__name__) |
| 54 | |
| 55 | CausalRMSNorm = partial(RMSNorm, eps=1e-5) |
| 56 | |
| 57 | |
| 58 | class _BlockPoolingAttentionMetadata(Protocol): |
| 59 | @property |
| 60 | def slot_mapping(self) -> torch.Tensor: ... |
| 61 | |
| 62 | |
| 63 | def _pad1d( |
| 64 | x: torch.Tensor, |
| 65 | paddings: tuple[int, int], |
| 66 | mode: str = "constant", |
| 67 | value: float = 0.0, |
| 68 | ) -> torch.Tensor: |
| 69 | """Tiny wrapper around F.pad, just to allow for |
| 70 | reflect padding on small input. |
| 71 | If this is the case, we insert extra 0 padding |
| 72 | to the right before the reflection happen. |
| 73 | """ |
| 74 | length = x.shape[-1] |
| 75 | padding_left, padding_right = paddings |
| 76 | assert padding_left >= 0 and padding_right >= 0, (padding_left, padding_right) |
| 77 | if mode == "reflect": |
| 78 | max_pad = max(padding_left, padding_right) |
| 79 | extra_pad = 0 |
| 80 | if length <= max_pad: |
| 81 | extra_pad = max_pad - length + 1 |
| 82 | x = F.pad(x, (0, extra_pad)) |
| 83 | padded = F.pad(x, paddings, mode, value) |
| 84 | end = padded.shape[-1] - extra_pad |
| 85 | return padded[..., :end] |
| 86 | else: |
| 87 | return F.pad(x, paddings, mode, value) |
| 88 | |
| 89 | |
| 90 | class WhisperCausalConv1d(nn.Conv1d): |
| 91 | def __init__( |
| 92 | self, |
| 93 | in_channels: int, |
| 94 | out_channels: int, |
| 95 | kernel_size: int, |
| 96 | stride: int = 1, |
| 97 | padding: int = 0, |
| 98 | bias: bool = True, |
| 99 | ) -> None: |
| 100 | super().__init__( |
| 101 | in_channels, |
| 102 | out_channels, |
| 103 | kernel_size, |
| 104 | stride=stride, |
| 105 | padding=padding, |
| 106 | bias=bias, |
| 107 | ) |
| 108 | self._stride = self.stride[0] |
| 109 | self._effective_kernel_size = (kernel_size - 1) * self.dilation[0] + 1 |
| 110 | self._padding_total = self._effective_kernel_size - self._stride |
| 111 | |
| 112 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 113 | n_frames = ( |
| 114 | x.shape[-1] - self._effective_kernel_size + self._padding_total |
| 115 | ) / self._stride + 1 |
| 116 | target_length = (math.ceil(n_frames) - 1) * self._stride + ( |
| 117 | self._effective_kernel_size - self._padding_total |
| 118 | ) |
| 119 | extra_padding = target_length - x.shape[-1] |
| 120 | x = _pad1d(x, (self._padding_total, extra_padding), mode="constant") |
| 121 | return super().forward(x) |
| 122 | |
| 123 | |
| 124 | @functools.lru_cache |
| 125 | def create_whisper_attention_backend_with_block_pooling( |
| 126 | underlying_attn_backend: type[AttentionBackend], |
| 127 | block_pool_size: int, |
| 128 | sliding_window: int | None = None, |
| 129 | ) -> type[AttentionBackend]: |
| 130 | prefix = "WhisperCausalAttentionWithBlockPooling_" |
| 131 | underlying_builder = underlying_attn_backend.get_builder_cls() |
| 132 | underlying_impl = underlying_attn_backend.get_impl_cls() |
| 133 | |
| 134 | class WhisperCausalAttentionWithBlockPoolingBuilder(underlying_builder): # type: ignore |
| 135 | # Full cudagraphs only for uniform single-token decode: capture bakes in |
| 136 | # tensor addresses, so `build` writes metadata into persistent buffers |
| 137 | # (below). Prefill/mixed batches fall back to piecewise. |
| 138 | _cudagraph_support: ClassVar[AttentionCGSupport] = ( |
| 139 | AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE |
| 140 | ) |
| 141 | # Disable the update_block_table fast path: it would bypass block pooling |
| 142 | # by splicing the raw slot_mapping/block_table into cached metadata. |
| 143 | supports_update_block_table: bool = False |
| 144 | |
| 145 | def __init__( |
| 146 | self, |
| 147 | kv_cache_spec: AttentionSpec, |
| 148 | layer_names: list[str], |
| 149 | vllm_config: VllmConfig, |
| 150 | device: torch.device, |
| 151 | ): |
| 152 | kv_cache_spec = replace( |
| 153 | kv_cache_spec, |
| 154 | block_size=kv_cache_spec.block_size * block_pool_size, |
| 155 | tokens_per_state=1, |
| 156 | ) |
| 157 | if isinstance(kv_cache_spec, SlidingWindowSpec): |
| 158 | # The manager keeps `sliding_window` in pooled units; the kernel |
| 159 | # runs on the expanded sequence, so give it the model's window in |
| 160 | # unpooled units (`get_kv_cache_spec` sizes the pooled window so |
| 161 | # this window always stays resident). |
| 162 | assert sliding_window is not None |
| 163 | kv_cache_spec = replace(kv_cache_spec, sliding_window=sliding_window) |
| 164 | super().__init__(kv_cache_spec, layer_names, vllm_config, device) |
| 165 | # Override model_config-derived values with the actual |
| 166 | # encoder values from kv_cache_spec |
| 167 | self.num_heads_kv = kv_cache_spec.num_kv_heads |
| 168 | self.headdim = kv_cache_spec.head_size |
| 169 | # num_heads_q for the encoder is the same as num_kv_heads |
| 170 | # (no GQA in whisper encoder) |
| 171 | self.num_heads_q = kv_cache_spec.num_kv_heads |
| 172 | # Persistent capture-safe buffers for pooled metadata; lazily sized |
| 173 | # on first `build` to match the incoming tensors' dtype/device. |
| 174 | self._pool_buffers_ready = False |
| 175 | |
| 176 | def _maybe_init_pool_buffers( |
| 177 | self, common_attn_metadata: CommonAttentionMetadata |
| 178 | ) -> None: |
| 179 | if self._pool_buffers_ready: |
| 180 | return |
| 181 | device = common_attn_metadata.slot_mapping.device |
| 182 | sched = self.vllm_config.scheduler_config |
| 183 | self._qsl_buf = torch.empty( |
| 184 | sched.max_num_seqs + 1, |
| 185 | dtype=common_attn_metadata.query_start_loc.dtype, |
| 186 | device=device, |
| 187 | ) |
| 188 | self._seq_buf = torch.empty( |
| 189 | sched.max_num_seqs, |
| 190 | dtype=common_attn_metadata.seq_lens.dtype, |
| 191 | device=device, |
| 192 | ) |
| 193 | self._slot_buf = torch.empty( |
| 194 | sched.max_num_batched_tokens * block_pool_size, |
| 195 | dtype=common_attn_metadata.slot_mapping.dtype, |
| 196 | device=device, |
| 197 | ) |
| 198 | self._pool_arange = torch.arange( |
| 199 | block_pool_size, |
| 200 | dtype=common_attn_metadata.slot_mapping.dtype, |
| 201 | device=device, |
| 202 | ) |
| 203 | self._pool_buffers_ready = True |
| 204 | |
| 205 | def build( |
| 206 | self, |
| 207 | common_prefix_len: int, |
| 208 | common_attn_metadata: CommonAttentionMetadata, |
| 209 | fast_build: bool = False, |
| 210 | ) -> AttentionMetadata: |
| 211 | cm = common_attn_metadata |
| 212 | self._maybe_init_pool_buffers(cm) |
| 213 | |
| 214 | # Scale query_start_loc / seq_lens into the persistent buffers. |
| 215 | n = cm.query_start_loc.numel() |
| 216 | assert n <= self._qsl_buf.numel(), ( |
| 217 | f"{n=} exceeds pooled query_start_loc buffer {self._qsl_buf.numel()}" |
| 218 | ) |
| 219 | qsl_buf = self._qsl_buf[:n] |
| 220 | qsl_buf.copy_(cm.query_start_loc) |
| 221 | qsl_buf.mul_(block_pool_size) |
| 222 | |
| 223 | r = cm.seq_lens.numel() |
| 224 | assert r <= self._seq_buf.numel(), ( |
| 225 | f"{r=} exceeds pooled seq_lens buffer {self._seq_buf.numel()}" |
| 226 | ) |
| 227 | seq_buf = self._seq_buf[:r] |
| 228 | seq_buf.copy_(cm.seq_lens) |
| 229 | seq_buf.mul_(block_pool_size) |
| 230 | |
| 231 | # Expand each slot into block_pool_size contiguous slots. Padding |
| 232 | # slots (-1) stay -1 after clamp so the KV write is skipped. |
| 233 | m = cm.slot_mapping.numel() |
| 234 | assert m * block_pool_size <= self._slot_buf.numel(), ( |
| 235 | f"pooled slot_mapping {m * block_pool_size} exceeds buffer " |
| 236 | f"{self._slot_buf.numel()}" |
| 237 | ) |
| 238 | slot_buf = self._slot_buf[: m * block_pool_size] |
| 239 | slot_buf.view(m, block_pool_size).copy_( |
| 240 | cm.slot_mapping.unsqueeze(1) * block_pool_size + self._pool_arange |
| 241 | ) |
| 242 | slot_buf.clamp_(min=-1) |
| 243 | |
| 244 | # CPU tensors aren't captured; fresh scaled copies avoid mutating |
| 245 | # the runner's shared buffers. |
| 246 | qsl_cpu = cm.query_start_loc_cpu * block_pool_size |
| 247 | seq_lens_cpu_upper_bound = ( |
| 248 | cm.seq_lens_cpu_upper_bound * block_pool_size |
| 249 | if cm.seq_lens_cpu_upper_bound is not None |
| 250 | else None |
| 251 | ) |
| 252 | |
| 253 | new_common_attn_metadata = cm.replace( |
| 254 | query_start_loc=qsl_buf, |
| 255 | query_start_loc_cpu=qsl_cpu, |
| 256 | seq_lens=seq_buf, |
| 257 | seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound, |
| 258 | num_actual_tokens=cm.num_actual_tokens * block_pool_size, |
| 259 | max_query_len=cm.max_query_len * block_pool_size, |
| 260 | max_seq_len=cm.max_seq_len * block_pool_size, |
| 261 | slot_mapping=slot_buf, |
| 262 | ) |
| 263 | return super().build( |
| 264 | common_prefix_len * block_pool_size, |
| 265 | new_common_attn_metadata, |
| 266 | fast_build, |
| 267 | ) |
| 268 | |
| 269 | # NOTE: We need a custom impl so we can use the transformed slot_mapping |
| 270 | # computed by `WhisperCausalAttentionWithBlockPoolingBuilder` instead of |
| 271 | # the one from `forward_context.slot_mapping` (gpu_model_runner). |
| 272 | # This follows the same pattern as CrossAttentionImpl. |
| 273 | class WhisperCausalAttentionWithBlockPoolingImpl(underlying_impl): # type: ignore[valid-type,misc] |
| 274 | def forward( |
| 275 | self, |
| 276 | layer: torch.nn.Module, |
| 277 | query: torch.Tensor, |
| 278 | key: torch.Tensor, |
| 279 | value: torch.Tensor, |
| 280 | kv_cache: torch.Tensor, |
| 281 | attn_metadata: _BlockPoolingAttentionMetadata | None, |
| 282 | output: torch.Tensor, |
| 283 | output_scale: torch.Tensor | None = None, |
| 284 | output_block_scale: torch.Tensor | None = None, |
| 285 | ) -> torch.Tensor: |
| 286 | if ( |
| 287 | not underlying_attn_backend.forward_includes_kv_cache_update |
| 288 | and attn_metadata is not None |
| 289 | and layer.kv_sharing_target_layer_name is None |
| 290 | and key is not None |
| 291 | and value is not None |
| 292 | ): |
| 293 | self.do_kv_cache_update( |
| 294 | layer, key, value, kv_cache, attn_metadata.slot_mapping |
| 295 | ) |
| 296 | |
| 297 | return super().forward( |
| 298 | layer, |
| 299 | query, |
| 300 | key, |
| 301 | value, |
| 302 | kv_cache, |
| 303 | attn_metadata, |
| 304 | output, |
| 305 | output_scale, |
| 306 | output_block_scale, |
| 307 | ) |
| 308 | |
| 309 | _SUPPORTED_BACKENDS = tuple( |
| 310 | b |
| 311 | for b in ( |
| 312 | AiterFlashAttentionBackend, |
| 313 | CPUAttentionBackend, |
| 314 | FlashAttentionBackend, |
| 315 | RocmAttentionBackend, |
| 316 | TritonAttentionBackend, |
| 317 | ) |
| 318 | if b is not None |
| 319 | ) |
| 320 | |
| 321 | if not issubclass(underlying_attn_backend, _SUPPORTED_BACKENDS): |
| 322 | raise NotImplementedError( |
| 323 | f"{underlying_attn_backend} is not yet supported." |
| 324 | "Contributions to support more backends are much " |
| 325 | "appreciated." |
| 326 | ) |
| 327 | |
| 328 | if not issubclass( |
| 329 | underlying_attn_backend, (CPUAttentionBackend, FlashAttentionBackend) |
| 330 | ): |
| 331 | logger.info( |
| 332 | "Using %s for Whisper causal attention with block pooling. " |
| 333 | "This backend was recently enabled for this model. " |
| 334 | "If you encounter any accuracy or performance issues, " |
| 335 | "please open an issue at " |
| 336 | "https://github.com/vllm-project/vllm/issues " |
| 337 | "with the [ROCm] tag so it can be triaged by the " |
| 338 | "appropriate team.", |
| 339 | underlying_attn_backend.get_name(), |
| 340 | ) |
| 341 | |
| 342 | attn_backend = subclass_attention_backend_with_overrides( |
| 343 | name_prefix=prefix, |
| 344 | attention_backend_cls=underlying_attn_backend, |
| 345 | overrides={ |
| 346 | "get_builder_cls": lambda: WhisperCausalAttentionWithBlockPoolingBuilder, |
| 347 | "get_impl_cls": lambda: WhisperCausalAttentionWithBlockPoolingImpl, |
| 348 | "forward_includes_kv_cache_update": True, |
| 349 | }, |
| 350 | ) |
| 351 | |
| 352 | return attn_backend |
| 353 | |
| 354 | |
| 355 | class WhisperCausalAttentionWithBlockPooling(Attention): |
| 356 | """Attention layer with block pooling.""" |
| 357 | |
| 358 | def __init__( |
| 359 | self, |
| 360 | num_heads: int, |
| 361 | head_size: int, |
| 362 | scale: float, |
| 363 | num_kv_heads: int | None = None, |
| 364 | alibi_slopes: list[float] | None = None, |
| 365 | cache_config: CacheConfig | None = None, |
| 366 | quant_config: QuantizationConfig | None = None, |
| 367 | logits_soft_cap: float | None = None, |
| 368 | per_layer_sliding_window: int | None = None, |
| 369 | prefix: str = "", |
| 370 | attn_type: str = AttentionType.DECODER, |
| 371 | kv_sharing_target_layer_name: str | None = None, |
| 372 | block_pool_size: int = 1, |
| 373 | attn_backend: type[AttentionBackend] | None = None, |
| 374 | **extra_impl_args, |
| 375 | ) -> None: |
| 376 | self.block_pool_size = block_pool_size |
| 377 | dtype = torch.get_default_dtype() |
| 378 | |
| 379 | if cache_config is not None: |
| 380 | kv_cache_dtype = cache_config.cache_dtype |
| 381 | else: |
| 382 | kv_cache_dtype = "auto" |
| 383 | |
| 384 | underlying_attn_backend = get_attn_backend( |
| 385 | head_size, |
| 386 | dtype, |
| 387 | kv_cache_dtype, |
| 388 | attn_type=attn_type, |
| 389 | ) |
| 390 | attn_backend = create_whisper_attention_backend_with_block_pooling( |
| 391 | underlying_attn_backend, block_pool_size, per_layer_sliding_window |
| 392 | ) |
| 393 | |
| 394 | super().__init__( |
| 395 | num_heads=num_heads, |
| 396 | head_size=head_size, |
| 397 | scale=scale, |
| 398 | num_kv_heads=num_kv_heads, |
| 399 | alibi_slopes=alibi_slopes, |
| 400 | cache_config=cache_config, |
| 401 | quant_config=quant_config, |
| 402 | logits_soft_cap=logits_soft_cap, |
| 403 | per_layer_sliding_window=per_layer_sliding_window, |
| 404 | prefix=prefix, |
| 405 | attn_type=attn_type, |
| 406 | kv_sharing_target_layer_name=kv_sharing_target_layer_name, |
| 407 | attn_backend=attn_backend, |
| 408 | **extra_impl_args, |
| 409 | ) |
| 410 | |
| 411 | def get_kv_cache_spec(self, vllm_config: VllmConfig): |
| 412 | kv_cache_spec = super().get_kv_cache_spec(vllm_config) |
| 413 | assert isinstance(kv_cache_spec, AttentionSpec) |
| 414 | kv_cache_spec = replace( |
| 415 | kv_cache_spec, |
| 416 | tokens_per_state=Fraction(1, self.block_pool_size), |
| 417 | ) |
| 418 | if isinstance(kv_cache_spec, SlidingWindowSpec): |
| 419 | # The manager counts blocks in pooled units, so express the window in |
| 420 | # pooled units to avoid reserving `block_pool_size`x too many blocks. |
| 421 | # The `+1` is an eviction margin: block-aligned eviction only keeps |
| 422 | # `pooled - 1` pooled tokens, so this guarantees the kernel's full |
| 423 | # (unpooled) window still stays resident. |
| 424 | pooled = cdiv(kv_cache_spec.sliding_window, self.block_pool_size) + 1 |
| 425 | kv_cache_spec = replace(kv_cache_spec, sliding_window=pooled) |
| 426 | return kv_cache_spec |
| 427 | |
| 428 | |
| 429 | class WhisperCausalAttention(nn.Module): |
| 430 | def __init__( |
| 431 | self, |
| 432 | embed_dim: int, |
| 433 | num_heads: int, |
| 434 | head_dim: int, |
| 435 | max_position_embeddings: int, |
| 436 | bias: bool = True, |
| 437 | attn_type: AttentionType = AttentionType.DECODER, |
| 438 | per_layer_sliding_window: int | None = None, |
| 439 | block_pool_size: int = 1, |
| 440 | cache_config: CacheConfig | None = None, |
| 441 | quant_config: QuantizationConfig | None = None, |
| 442 | prefix: str = "", |
| 443 | ): |
| 444 | super().__init__() |
| 445 | self.embed_dim = embed_dim |
| 446 | tp_size = get_tensor_model_parallel_world_size() |
| 447 | self.total_num_heads = num_heads |
| 448 | assert self.total_num_heads % tp_size == 0 |
| 449 | self.num_heads = self.total_num_heads // tp_size |
| 450 | if self.total_num_heads >= tp_size: |
| 451 | # Number of heads is greater than TP size, so we partition |
| 452 | # the KV heads across multiple tensor parallel GPUs. |
| 453 | assert self.total_num_heads % tp_size == 0 |
| 454 | else: |
| 455 | # Number of heads is less than TP size, so we replicate |
| 456 | # the KV heads across multiple tensor parallel GPUs. |
| 457 | assert tp_size % self.total_num_heads == 0 |
| 458 | self.num_kv_heads = max(1, self.total_num_heads // tp_size) |
| 459 | self.head_dim = head_dim |
| 460 | self.q_size = self.num_heads * self.head_dim |
| 461 | self.kv_size = self.num_kv_heads * self.head_dim |
| 462 | self.attn_type = attn_type |
| 463 | |
| 464 | self.scaling = self.head_dim**-0.5 |
| 465 | |
| 466 | self._init_qkv(embed_dim, bias, quant_config, prefix=prefix) |
| 467 | self.out_proj = RowParallelLinear( |
| 468 | input_size=self.total_num_heads * self.head_dim, |
| 469 | output_size=embed_dim, |
| 470 | bias=bias, |
| 471 | quant_config=quant_config, |
| 472 | prefix=f"{prefix}.out_proj", |
| 473 | ) |
| 474 | assert block_pool_size > 1, ( |
| 475 | f"Causal attention only supports block_pool_size>1, not {block_pool_size}." |
| 476 | ) |
| 477 | self.attn = WhisperCausalAttentionWithBlockPooling( |
| 478 | self.num_heads, |
| 479 | self.head_dim, |
| 480 | self.scaling, |
| 481 | num_kv_heads=self.num_kv_heads, |
| 482 | cache_config=cache_config, |
| 483 | quant_config=quant_config, |
| 484 | prefix=f"{prefix}.attn", |
| 485 | attn_type=AttentionType.DECODER, |
| 486 | per_layer_sliding_window=per_layer_sliding_window, |
| 487 | block_pool_size=block_pool_size, |
| 488 | ) |
| 489 | |
| 490 | assert per_layer_sliding_window is not None, ( |
| 491 | "rope can only used in combination with a sliding window" |
| 492 | ) |
| 493 | self._init_rotary_emb(max_position_embeddings) |
| 494 | |
| 495 | def _init_rotary_emb(self, max_position_embeddings: int) -> None: |
| 496 | self.rotary_emb = get_rope( |
| 497 | self.head_dim, |
| 498 | max_position=max_position_embeddings, |
| 499 | is_neox_style=False, |
| 500 | rope_parameters={"rope_theta": 1e6}, |
| 501 | ) |
| 502 | |
| 503 | def _init_qkv( |
| 504 | self, |
| 505 | embed_dim: int, |
| 506 | bias: bool = True, |
| 507 | quant_config: QuantizationConfig | None = None, |
| 508 | prefix: str = "", |
| 509 | ) -> None: |
| 510 | self.qkv_proj = QKVParallelLinear( |
| 511 | hidden_size=embed_dim, |
| 512 | head_size=self.head_dim, |
| 513 | total_num_heads=self.total_num_heads, |
| 514 | total_num_kv_heads=self.total_num_heads, |
| 515 | bias=bias, |
| 516 | quant_config=quant_config, |
| 517 | prefix=f"{prefix}.qkv_proj", |
| 518 | ) |
| 519 | |
| 520 | def forward( |
| 521 | self, |
| 522 | hidden_states: torch.Tensor, |
| 523 | positions: torch.Tensor | None = None, |
| 524 | ): |
| 525 | qkv, _ = self.qkv_proj(hidden_states) |
| 526 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 527 | |
| 528 | assert positions is not None |
| 529 | q, k = self.rotary_emb(positions, q, k) |
| 530 | |
| 531 | attn_output = self.attn(q, k, v) |
| 532 | |
| 533 | output, _ = self.out_proj(attn_output) |
| 534 | |
| 535 | return output |
| 536 | |
| 537 | |
| 538 | class WhisperCausalEncoderLayer(nn.Module): |
| 539 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 540 | super().__init__() |
| 541 | config = vllm_config.model_config.hf_config |
| 542 | sliding_window = getattr(config, "sliding_window", None) |
| 543 | block_pool_size = config.block_pool_size |
| 544 | assert block_pool_size > 1 |
| 545 | |
| 546 | cache_config = vllm_config.cache_config |
| 547 | quant_config = vllm_config.quant_config |
| 548 | |
| 549 | self.embed_dim = config.d_model |
| 550 | self.head_dim = self.embed_dim // config.encoder_attention_heads |
| 551 | self.self_attn = WhisperCausalAttention( |
| 552 | embed_dim=self.embed_dim, |
| 553 | num_heads=config.encoder_attention_heads, |
| 554 | head_dim=config.encoder_head_dim, |
| 555 | max_position_embeddings=config.max_position_embeddings, |
| 556 | block_pool_size=block_pool_size, |
| 557 | per_layer_sliding_window=sliding_window, |
| 558 | cache_config=cache_config, |
| 559 | quant_config=quant_config, |
| 560 | prefix=f"{prefix}.self_attn", |
| 561 | ) |
| 562 | self.self_attn_layer_norm = CausalRMSNorm(self.embed_dim) |
| 563 | |
| 564 | self.mlp = MistralMLP( |
| 565 | hidden_size=config.d_model, |
| 566 | intermediate_size=config.encoder_ffn_dim, |
| 567 | hidden_act="silu", |
| 568 | quant_config=quant_config, |
| 569 | bias=True, |
| 570 | gate_up_proj_bias=False, |
| 571 | prefix=f"{prefix}.mlp", |
| 572 | ) |
| 573 | self.final_layer_norm = CausalRMSNorm(self.embed_dim) |
| 574 | |
| 575 | def forward( |
| 576 | self, |
| 577 | hidden_states: torch.Tensor, |
| 578 | positions: torch.Tensor | None = None, |
| 579 | ): |
| 580 | residual = hidden_states |
| 581 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 582 | hidden_states = self.self_attn(hidden_states=hidden_states, positions=positions) |
| 583 | hidden_states = residual + hidden_states |
| 584 | residual = hidden_states |
| 585 | hidden_states = self.final_layer_norm(hidden_states) |
| 586 | hidden_states = self.mlp(hidden_states) |
| 587 | hidden_states = residual + hidden_states |
| 588 | |
| 589 | return hidden_states |
| 590 | |
| 591 | |
| 592 | class WhisperCausalEncoder(nn.Module): |
| 593 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 594 | super().__init__() |
| 595 | config = vllm_config.model_config.hf_config |
| 596 | embed_dim = config.d_model |
| 597 | |
| 598 | assert WhisperPosEmbedType(config.pos_embed) == WhisperPosEmbedType.ROPE |
| 599 | assert config.is_causal |
| 600 | |
| 601 | self.num_mel_bins = config.num_mel_bins |
| 602 | self.max_source_positions = config.max_source_positions |
| 603 | self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0 |
| 604 | |
| 605 | self.conv1 = WhisperCausalConv1d(self.num_mel_bins, embed_dim, kernel_size=3) |
| 606 | self.conv2 = WhisperCausalConv1d(embed_dim, embed_dim, stride=2, kernel_size=3) |
| 607 | |
| 608 | self.total_stride = self.conv1.stride[0] * self.conv2.stride[0] |
| 609 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 610 | config.encoder_layers, |
| 611 | lambda prefix: WhisperCausalEncoderLayer( |
| 612 | vllm_config=vllm_config, prefix=f"{prefix}.layers" |
| 613 | ), |
| 614 | prefix=f"{prefix}.layers", |
| 615 | ) |
| 616 | self.layer_norm = CausalRMSNorm(config.d_model) |
| 617 | |
| 618 | def forward_conv( |
| 619 | self, input_features: torch.Tensor | list[torch.Tensor] |
| 620 | ) -> torch.Tensor: |
| 621 | hidden_states = [] |
| 622 | for features in input_features: |
| 623 | embeds = nn.functional.gelu(self.conv1(features)) |
| 624 | embeds = nn.functional.gelu(self.conv2(embeds)) |
| 625 | |
| 626 | embeds = embeds.transpose(-1, -2).to(embeds.dtype) |
| 627 | |
| 628 | hidden_states.append(embeds) |
| 629 | |
| 630 | hidden_states = torch.cat(hidden_states) |
| 631 | |
| 632 | return hidden_states |
| 633 | |
| 634 | def forward( |
| 635 | self, hidden_states: torch.Tensor, positions: torch.Tensor |
| 636 | ) -> torch.Tensor: |
| 637 | for encoder_layer in self.layers: |
| 638 | hidden_states = encoder_layer(hidden_states, positions) |
| 639 | |
| 640 | hidden_states = self.layer_norm(hidden_states) |
| 641 | return hidden_states |