| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Sequence |
| 5 | from dataclasses import dataclass |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn.functional as F |
| 9 | from torch.nn.parameter import Parameter |
| 10 | |
| 11 | import vllm.envs as envs |
| 12 | from vllm import _custom_ops as ops |
| 13 | from vllm.distributed import ( |
| 14 | GroupCoordinator, |
| 15 | divide, |
| 16 | get_tensor_model_parallel_rank, |
| 17 | get_tensor_model_parallel_world_size, |
| 18 | tensor_model_parallel_all_reduce, |
| 19 | ) |
| 20 | from vllm.model_executor.custom_op import PluggableLayer |
| 21 | from vllm.model_executor.determinism.batch_invariant import ( |
| 22 | linear_batch_invariant, |
| 23 | ) |
| 24 | from vllm.model_executor.layers.quantization.base_config import ( |
| 25 | QuantizationConfig, |
| 26 | QuantizeMethodBase, |
| 27 | method_has_implemented_embedding, |
| 28 | resolve_quant_method, |
| 29 | ) |
| 30 | from vllm.model_executor.layers.utils import dispatch_unquantized_gemm |
| 31 | from vllm.model_executor.parameter import BasevLLMParameter |
| 32 | from vllm.model_executor.utils import set_weight_attrs |
| 33 | from vllm.platforms import current_platform |
| 34 | |
| 35 | DEFAULT_VOCAB_PADDING_SIZE = 64 |
| 36 | |
| 37 | |
| 38 | class UnquantizedEmbeddingMethod(QuantizeMethodBase): |
| 39 | """Unquantized method for embeddings.""" |
| 40 | |
| 41 | supports_pre_processed_weights = True |
| 42 | |
| 43 | def create_weights( |
| 44 | self, |
| 45 | layer: torch.nn.Module, |
| 46 | input_size_per_partition: int, |
| 47 | output_partition_sizes: list[int], |
| 48 | input_size: int, |
| 49 | output_size: int, |
| 50 | params_dtype: torch.dtype, |
| 51 | **extra_weight_attrs, |
| 52 | ): |
| 53 | """Create weights for embedding layer.""" |
| 54 | weight = Parameter( |
| 55 | torch.empty( |
| 56 | sum(output_partition_sizes), |
| 57 | input_size_per_partition, |
| 58 | dtype=params_dtype, |
| 59 | ), |
| 60 | requires_grad=False, |
| 61 | ) |
| 62 | set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0}) |
| 63 | layer.register_parameter("weight", weight) |
| 64 | set_weight_attrs(weight, extra_weight_attrs) |
| 65 | |
| 66 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 67 | if current_platform.is_cpu(): |
| 68 | from vllm.model_executor.layers.utils import dispatch_cpu_unquantized_gemm |
| 69 | |
| 70 | dispatch_cpu_unquantized_gemm(layer, remove_weight=False) |
| 71 | |
| 72 | def apply( |
| 73 | self, |
| 74 | layer: torch.nn.Module, |
| 75 | x: torch.Tensor, |
| 76 | bias: torch.Tensor | None = None, |
| 77 | ) -> torch.Tensor: |
| 78 | if envs.VLLM_BATCH_INVARIANT and ( |
| 79 | current_platform.is_cuda_alike() or current_platform.is_xpu() |
| 80 | ): |
| 81 | return linear_batch_invariant(x, layer.weight, bias) |
| 82 | return dispatch_unquantized_gemm()(layer, x, layer.weight, bias) |
| 83 | |
| 84 | def embedding(self, layer: torch.nn.Module, input_: torch.Tensor) -> torch.Tensor: |
| 85 | return F.embedding(input_, layer.weight) |
| 86 | |
| 87 | def tie_weights( |
| 88 | self, layer: torch.nn.Module, embed_tokens: "VocabParallelEmbedding" |
| 89 | ): |
| 90 | layer.weight = embed_tokens.weight |
| 91 | return layer |
| 92 | |
| 93 | |
| 94 | def pad_vocab_size(vocab_size: int, pad_to: int = DEFAULT_VOCAB_PADDING_SIZE) -> int: |
| 95 | """Pad the vocab size to the given value.""" |
| 96 | return ((vocab_size + pad_to - 1) // pad_to) * pad_to |
| 97 | |
| 98 | |
| 99 | def vocab_range_from_per_partition_vocab_size( |
| 100 | per_partition_vocab_size: int, rank: int, offset: int = 0 |
| 101 | ) -> Sequence[int]: |
| 102 | index_f = rank * per_partition_vocab_size |
| 103 | index_l = index_f + per_partition_vocab_size |
| 104 | return index_f + offset, index_l + offset |
| 105 | |
| 106 | |
| 107 | def vocab_range_from_global_vocab_size( |
| 108 | global_vocab_size: int, rank: int, world_size: int, offset: int = 0 |
| 109 | ) -> Sequence[int]: |
| 110 | per_partition_vocab_size = divide(global_vocab_size, world_size) |
| 111 | return vocab_range_from_per_partition_vocab_size( |
| 112 | per_partition_vocab_size, rank, offset=offset |
| 113 | ) |
| 114 | |
| 115 | |
| 116 | @dataclass |
| 117 | class VocabParallelEmbeddingShardIndices: |
| 118 | """Indices for a shard of a vocab parallel embedding.""" |
| 119 | |
| 120 | padded_org_vocab_start_index: int |
| 121 | padded_org_vocab_end_index: int |
| 122 | padded_added_vocab_start_index: int |
| 123 | padded_added_vocab_end_index: int |
| 124 | |
| 125 | org_vocab_start_index: int |
| 126 | org_vocab_end_index: int |
| 127 | added_vocab_start_index: int |
| 128 | added_vocab_end_index: int |
| 129 | |
| 130 | @property |
| 131 | def num_org_elements(self) -> int: |
| 132 | return self.org_vocab_end_index - self.org_vocab_start_index |
| 133 | |
| 134 | @property |
| 135 | def num_added_elements(self) -> int: |
| 136 | return self.added_vocab_end_index - self.added_vocab_start_index |
| 137 | |
| 138 | @property |
| 139 | def num_org_elements_padded(self) -> int: |
| 140 | return self.padded_org_vocab_end_index - self.padded_org_vocab_start_index |
| 141 | |
| 142 | @property |
| 143 | def num_added_elements_padded(self) -> int: |
| 144 | return self.padded_added_vocab_end_index - self.padded_added_vocab_start_index |
| 145 | |
| 146 | @property |
| 147 | def num_org_vocab_padding(self) -> int: |
| 148 | return self.num_org_elements_padded - self.num_org_elements |
| 149 | |
| 150 | @property |
| 151 | def num_added_vocab_padding(self) -> int: |
| 152 | return self.num_added_elements_padded - self.num_added_elements |
| 153 | |
| 154 | @property |
| 155 | def num_elements_padded(self) -> int: |
| 156 | return self.num_org_elements_padded + self.num_added_elements_padded |
| 157 | |
| 158 | def __post_init__(self): |
| 159 | # sanity checks |
| 160 | assert self.padded_org_vocab_start_index <= self.padded_org_vocab_end_index |
| 161 | assert self.padded_added_vocab_start_index <= self.padded_added_vocab_end_index |
| 162 | |
| 163 | assert self.org_vocab_start_index <= self.org_vocab_end_index |
| 164 | assert self.added_vocab_start_index <= self.added_vocab_end_index |
| 165 | |
| 166 | assert self.org_vocab_start_index <= self.padded_org_vocab_start_index |
| 167 | assert self.added_vocab_start_index <= self.padded_added_vocab_start_index |
| 168 | assert self.org_vocab_end_index <= self.padded_org_vocab_end_index |
| 169 | assert self.added_vocab_end_index <= self.padded_added_vocab_end_index |
| 170 | |
| 171 | assert self.num_org_elements <= self.num_org_elements_padded |
| 172 | assert self.num_added_elements <= self.num_added_elements_padded |
| 173 | |
| 174 | |
| 175 | @torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) |
| 176 | def get_masked_input_and_mask( |
| 177 | input_: torch.Tensor, |
| 178 | org_vocab_start_index: int, |
| 179 | org_vocab_end_index: int, |
| 180 | num_org_vocab_padding: int, |
| 181 | added_vocab_start_index: int, |
| 182 | added_vocab_end_index: int, |
| 183 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 184 | # torch.compile will fuse all of the pointwise ops below |
| 185 | # into a single kernel, making it very fast |
| 186 | org_vocab_mask = (input_ >= org_vocab_start_index) & (input_ < org_vocab_end_index) |
| 187 | added_vocab_mask = (input_ >= added_vocab_start_index) & ( |
| 188 | input_ < added_vocab_end_index |
| 189 | ) |
| 190 | added_offset = ( |
| 191 | added_vocab_start_index |
| 192 | - (org_vocab_end_index - org_vocab_start_index) |
| 193 | - num_org_vocab_padding |
| 194 | ) |
| 195 | valid_offset = (org_vocab_start_index * org_vocab_mask) + ( |
| 196 | added_offset * added_vocab_mask |
| 197 | ) |
| 198 | vocab_mask = org_vocab_mask | added_vocab_mask |
| 199 | input_ = vocab_mask * (input_ - valid_offset) |
| 200 | return input_, ~vocab_mask |
| 201 | |
| 202 | |
| 203 | # --8<-- [start:vocab_parallel_embedding] |
| 204 | @PluggableLayer.register("vocab_parallel_embedding") |
| 205 | class VocabParallelEmbedding(PluggableLayer): |
| 206 | """Embedding parallelized in the vocabulary dimension. |
| 207 | |
| 208 | Adapted from torch.nn.Embedding, note that we pad the vocabulary size to |
| 209 | make sure it is divisible by the number of model parallel GPUs. |
| 210 | |
| 211 | In order to support various loading methods, we ensure that LoRA-added |
| 212 | embeddings are always at the end of TP-sharded tensors. In other words, |
| 213 | we shard base embeddings and LoRA embeddings separately (both padded), |
| 214 | and place them in the same tensor. |
| 215 | In this example, we will have the original vocab size = 1010, |
| 216 | added vocab size = 16 and padding to 64. Therefore, the total |
| 217 | vocab size with padding will be 1088 (because we first pad 1010 to |
| 218 | 1024, add 16, and then pad to 1088). |
| 219 | Therefore, the tensor format looks like the following: |
| 220 | TP1, rank 0 (no sharding): |
| 221 | |< --------BASE-------- >|< -BASE PADDING-- >|< -----LORA------ >|< -LORA PADDING-- >| |
| 222 | corresponding token_id: | 0 | 1 | ... | 1009 | -1 | ... | -1 | 1010 | ... | 1025 | -1 | ... | -1 | |
| 223 | index: | 0 | 1 | ... | 1009 | 1010 | ... | 1023 | 1024 | ... | 1039 | 1040 | ... | 1087 | |
| 224 | |
| 225 | TP2, rank 0: |
| 226 | |< --------------------BASE--------------------- >|< -----LORA------ >|< -LORA PADDING- >| |
| 227 | corresponding token_id: | 0 | 1 | 2 | ... | 497 | 498 | ... | 511 | 1010 | ... | 1025 | -1 | ... | -1 | |
| 228 | index: | 0 | 1 | 2 | ... | 497 | 498 | ... | 511 | 512 | ... | 527 | 528 | ... | 543 | |
| 229 | TP2, rank 1: |
| 230 | |< -----------BASE----------- >|< -BASE PADDING- >|< -----------LORA PADDING----------- >| |
| 231 | corresponding token_id: | 512 | 513 | 514 | ... | 1009 | -1 | ... | -1 | -1 | ... | -1 | -1 | ... | -1 | |
| 232 | index: | 0 | 1 | 2 | ... | 497 | 498 | ... | 511 | 512 | ... | 527 | 528 | ... | 543 | |
| 233 | |
| 234 | Args: |
| 235 | num_embeddings: vocabulary size. |
| 236 | embedding_dim: size of hidden state. |
| 237 | params_dtype: type of the parameters. |
| 238 | org_num_embeddings: original vocabulary size (without LoRA). |
| 239 | padding_size: padding size for the vocabulary. |
| 240 | quant_config: quant config for the layer |
| 241 | prefix: full name of the layer in the state dict |
| 242 | disable_tp: If true, tensor parallelism will be disabled for this layer. |
| 243 | quant_method: Preselected quantization method for model-specific layers. |
| 244 | parallel_group: Process group used to shard and reduce the embedding. |
| 245 | |
| 246 | """ # noqa: E501 |
| 247 | |
| 248 | # --8<-- [end:vocab_parallel_embedding] |
| 249 | |
| 250 | def __init__( |
| 251 | self, |
| 252 | num_embeddings: int, |
| 253 | embedding_dim: int, |
| 254 | params_dtype: torch.dtype | None = None, |
| 255 | org_num_embeddings: int | None = None, |
| 256 | padding_size: int = DEFAULT_VOCAB_PADDING_SIZE, |
| 257 | quant_config: QuantizationConfig | None = None, |
| 258 | prefix: str = "", |
| 259 | *, |
| 260 | disable_tp: bool = False, |
| 261 | quant_method: QuantizeMethodBase | None = None, |
| 262 | parallel_group: GroupCoordinator | None = None, |
| 263 | ): |
| 264 | super().__init__() |
| 265 | |
| 266 | # Keep the input dimensions. |
| 267 | self.disable_tp = disable_tp |
| 268 | self.parallel_group = parallel_group |
| 269 | if disable_tp: |
| 270 | tp_rank, self.tp_size = 0, 1 |
| 271 | elif parallel_group is not None: |
| 272 | tp_rank = parallel_group.rank_in_group |
| 273 | self.tp_size = parallel_group.world_size |
| 274 | else: |
| 275 | tp_rank = get_tensor_model_parallel_rank() |
| 276 | self.tp_size = get_tensor_model_parallel_world_size() |
| 277 | self.tp_rank = tp_rank |
| 278 | self.num_embeddings = num_embeddings |
| 279 | self.padding_size = padding_size |
| 280 | self.org_vocab_size = org_num_embeddings or num_embeddings |
| 281 | self.num_added_embeddings = num_embeddings - self.org_vocab_size |
| 282 | self.org_vocab_size_padded = pad_vocab_size( |
| 283 | self.org_vocab_size, self.padding_size |
| 284 | ) |
| 285 | self.num_embeddings_padded = pad_vocab_size( |
| 286 | self.org_vocab_size_padded + self.num_added_embeddings, self.padding_size |
| 287 | ) |
| 288 | assert self.org_vocab_size_padded <= self.num_embeddings_padded |
| 289 | |
| 290 | self.shard_indices = self._get_indices( |
| 291 | self.num_embeddings_padded, |
| 292 | self.org_vocab_size_padded, |
| 293 | self.num_embeddings, |
| 294 | self.org_vocab_size, |
| 295 | tp_rank, |
| 296 | self.tp_size, |
| 297 | ) |
| 298 | self.embedding_dim = embedding_dim |
| 299 | # Divide the weight matrix along the vocabulary dimension. |
| 300 | self.num_embeddings_per_partition = divide( |
| 301 | self.num_embeddings_padded, self.tp_size |
| 302 | ) |
| 303 | |
| 304 | # Quantization methods share the same weight factory as linear layers, |
| 305 | # so setup standard linear metadata. |
| 306 | self.input_size = self.input_size_per_partition = embedding_dim |
| 307 | self.output_size = self.num_embeddings_padded |
| 308 | self.output_size_per_partition = self.num_embeddings_per_partition |
| 309 | self.output_partition_sizes = [self.output_size_per_partition] |
| 310 | self.prefix = prefix |
| 311 | |
| 312 | # Avoid overriding a preselected model-specific method with generic |
| 313 | # config-based dispatch. |
| 314 | if quant_method is None and quant_config is not None: |
| 315 | quant_method = resolve_quant_method(quant_config, self, prefix=prefix) |
| 316 | if quant_method is None: |
| 317 | quant_method = UnquantizedEmbeddingMethod() |
| 318 | |
| 319 | # If we are making an embedding layer, then our quantization linear |
| 320 | # method must implement the embedding operation. If we are another |
| 321 | # layer type like ParallelLMHead, this is not important. |
| 322 | is_embedding_layer = not isinstance(self, ParallelLMHead) |
| 323 | quant_method_implements_embedding = method_has_implemented_embedding( |
| 324 | type(quant_method) |
| 325 | ) |
| 326 | if is_embedding_layer and not quant_method_implements_embedding: |
| 327 | raise NotImplementedError( |
| 328 | f"The class {type(quant_method).__name__} must implement " |
| 329 | "the 'embedding' method, see UnquantizedEmbeddingMethod." |
| 330 | ) |
| 331 | |
| 332 | self.quant_method: QuantizeMethodBase = quant_method |
| 333 | |
| 334 | if params_dtype is None: |
| 335 | params_dtype = torch.get_default_dtype() |
| 336 | self.params_dtype = params_dtype |
| 337 | assert ( |
| 338 | self.shard_indices.num_elements_padded == self.num_embeddings_per_partition |
| 339 | ) |
| 340 | self.num_org_embeddings_per_partition = ( |
| 341 | self.shard_indices.org_vocab_end_index |
| 342 | - self.shard_indices.org_vocab_start_index |
| 343 | ) |
| 344 | self.num_added_embeddings_per_partition = ( |
| 345 | self.shard_indices.added_vocab_end_index |
| 346 | - self.shard_indices.added_vocab_start_index |
| 347 | ) |
| 348 | |
| 349 | # The fused kernel masks, shifts and gathers in one pass; it assumes a |
| 350 | # plain row-major shard, so it only applies to unquantized weights. |
| 351 | self.use_fused_embedding = ( |
| 352 | self.tp_size > 1 |
| 353 | and current_platform.is_cuda() |
| 354 | and isinstance(quant_method, UnquantizedEmbeddingMethod) |
| 355 | ) |
| 356 | |
| 357 | self.quant_method.create_weights( |
| 358 | self, |
| 359 | self.input_size_per_partition, |
| 360 | self.output_partition_sizes, |
| 361 | self.input_size, |
| 362 | self.output_size, |
| 363 | params_dtype=params_dtype, |
| 364 | weight_loader=self.weight_loader, |
| 365 | ) |
| 366 | self.update_param_tp_status() |
| 367 | |
| 368 | def update_param_tp_status(self): |
| 369 | for param in self.parameters(): |
| 370 | if isinstance(param, BasevLLMParameter): |
| 371 | param.tp_rank = self.tp_rank |
| 372 | param.tp_size = self.tp_size |
| 373 | |
| 374 | @classmethod |
| 375 | def _get_indices( |
| 376 | cls, |
| 377 | vocab_size_padded: int, |
| 378 | org_vocab_size_padded: int, |
| 379 | vocab_size: int, |
| 380 | org_vocab_size: int, |
| 381 | tp_rank: int, |
| 382 | tp_size: int, |
| 383 | ) -> VocabParallelEmbeddingShardIndices: |
| 384 | """Get start and end indices for vocab parallel embedding, following the |
| 385 | layout outlined in the class docstring, based on the given tp_rank and |
| 386 | tp_size.""" |
| 387 | num_added_embeddings_padded = vocab_size_padded - org_vocab_size_padded |
| 388 | padded_org_vocab_start_index, padded_org_vocab_end_index = ( |
| 389 | vocab_range_from_global_vocab_size(org_vocab_size_padded, tp_rank, tp_size) |
| 390 | ) |
| 391 | padded_added_vocab_start_index, padded_added_vocab_end_index = ( |
| 392 | vocab_range_from_global_vocab_size( |
| 393 | num_added_embeddings_padded, tp_rank, tp_size, offset=org_vocab_size |
| 394 | ) |
| 395 | ) |
| 396 | # remove padding |
| 397 | org_vocab_start_index = min(padded_org_vocab_start_index, org_vocab_size) |
| 398 | org_vocab_end_index = min(padded_org_vocab_end_index, org_vocab_size) |
| 399 | added_vocab_start_index = min(padded_added_vocab_start_index, vocab_size) |
| 400 | added_vocab_end_index = min(padded_added_vocab_end_index, vocab_size) |
| 401 | return VocabParallelEmbeddingShardIndices( |
| 402 | padded_org_vocab_start_index, |
| 403 | padded_org_vocab_end_index, |
| 404 | padded_added_vocab_start_index, |
| 405 | padded_added_vocab_end_index, |
| 406 | org_vocab_start_index, |
| 407 | org_vocab_end_index, |
| 408 | added_vocab_start_index, |
| 409 | added_vocab_end_index, |
| 410 | ) |
| 411 | |
| 412 | def get_sharded_to_full_mapping(self) -> list[int] | None: |
| 413 | """Get a mapping that can be used to reindex the gathered |
| 414 | logits for sampling. |
| 415 | |
| 416 | During sampling, we gather logits from all ranks. The relationship |
| 417 | of index->token_id will follow the same format as outlined in the class |
| 418 | docstring. However, after the gather, we want to reindex the final |
| 419 | logits tensor to map index->token_id one-to-one (the index is always |
| 420 | equal the token_id it corresponds to). The indices returned by this |
| 421 | method allow us to do that. |
| 422 | """ |
| 423 | if self.tp_size < 2: |
| 424 | return None |
| 425 | |
| 426 | base_embeddings: list[int] = [] |
| 427 | added_embeddings: list[int] = [] |
| 428 | padding: list[int] = [] |
| 429 | for tp_rank in range(self.tp_size): |
| 430 | shard_indices = self._get_indices( |
| 431 | self.num_embeddings_padded, |
| 432 | self.org_vocab_size_padded, |
| 433 | self.num_embeddings, |
| 434 | self.org_vocab_size, |
| 435 | tp_rank, |
| 436 | self.tp_size, |
| 437 | ) |
| 438 | range_start = self.num_embeddings_per_partition * tp_rank |
| 439 | range_end = self.num_embeddings_per_partition * (tp_rank + 1) |
| 440 | base_embeddings.extend( |
| 441 | range(range_start, range_start + shard_indices.num_org_elements) |
| 442 | ) |
| 443 | padding.extend( |
| 444 | range( |
| 445 | range_start + shard_indices.num_org_elements, |
| 446 | range_start + shard_indices.num_org_elements_padded, |
| 447 | ) |
| 448 | ) |
| 449 | added_embeddings.extend( |
| 450 | range( |
| 451 | range_start + shard_indices.num_org_elements_padded, |
| 452 | range_start |
| 453 | + shard_indices.num_org_elements_padded |
| 454 | + shard_indices.num_added_elements, |
| 455 | ) |
| 456 | ) |
| 457 | padding.extend( |
| 458 | range( |
| 459 | range_start |
| 460 | + shard_indices.num_org_elements_padded |
| 461 | + shard_indices.num_added_elements, |
| 462 | range_start |
| 463 | + shard_indices.num_org_elements_padded |
| 464 | + shard_indices.num_added_elements_padded, |
| 465 | ) |
| 466 | ) |
| 467 | assert ( |
| 468 | range_start |
| 469 | + shard_indices.num_org_elements_padded |
| 470 | + shard_indices.num_added_elements_padded |
| 471 | == range_end |
| 472 | ) |
| 473 | ret = base_embeddings + added_embeddings + padding |
| 474 | assert len(ret) == self.num_embeddings_padded |
| 475 | return ret |
| 476 | |
| 477 | def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor): |
| 478 | output_dim = getattr(param, "output_dim", None) |
| 479 | packed_dim = getattr(param, "packed_dim", None) |
| 480 | |
| 481 | # Parameters without an output dimension are copied onto all GPUs. |
| 482 | if output_dim is None: |
| 483 | if ( |
| 484 | loaded_weight.ndim == 0 |
| 485 | and param.data.ndim == 1 |
| 486 | and param.data.numel() == 1 |
| 487 | ): |
| 488 | loaded_weight = loaded_weight.reshape(1) |
| 489 | assert param.data.shape == loaded_weight.shape |
| 490 | param.data.copy_(loaded_weight) |
| 491 | return |
| 492 | |
| 493 | # Shard indexes for loading the weight |
| 494 | start_idx = self.shard_indices.org_vocab_start_index |
| 495 | shard_size = self.shard_indices.org_vocab_end_index - start_idx |
| 496 | |
| 497 | # If param packed on the same dim we are sharding on, then |
| 498 | # need to adjust offsets of loaded weight by pack_factor. |
| 499 | if packed_dim is not None and packed_dim == output_dim: |
| 500 | packed_factor = ( |
| 501 | param.packed_factor |
| 502 | if isinstance(param, BasevLLMParameter) |
| 503 | else param.pack_factor |
| 504 | ) |
| 505 | assert loaded_weight.shape[output_dim] == ( |
| 506 | self.org_vocab_size // param.packed_factor |
| 507 | ) |
| 508 | start_idx = start_idx // packed_factor |
| 509 | shard_size = shard_size // packed_factor |
| 510 | else: |
| 511 | assert loaded_weight.shape[output_dim] == self.org_vocab_size |
| 512 | |
| 513 | # Copy the data. Select chunk corresponding to current shard. |
| 514 | loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size) |
| 515 | param[: loaded_weight.shape[0]].data.copy_(loaded_weight) |
| 516 | param[loaded_weight.shape[0] :].data.fill_(0) |
| 517 | |
| 518 | def forward(self, input_): |
| 519 | if self.tp_size == 1: |
| 520 | return self.quant_method.embedding(self, input_.long()) |
| 521 | |
| 522 | if self.use_fused_embedding: |
| 523 | output_parallel = ops.vocab_parallel_embedding( |
| 524 | input_ if input_.ndim == 1 else input_.reshape(-1), |
| 525 | self.weight, |
| 526 | self.shard_indices.org_vocab_start_index, |
| 527 | self.shard_indices.org_vocab_end_index, |
| 528 | self.shard_indices.num_org_vocab_padding, |
| 529 | self.shard_indices.added_vocab_start_index, |
| 530 | self.shard_indices.added_vocab_end_index, |
| 531 | ) |
| 532 | if input_.ndim != 1: |
| 533 | output_parallel = output_parallel.view( |
| 534 | *input_.shape, self.embedding_dim |
| 535 | ) |
| 536 | else: |
| 537 | masked_input, input_mask = get_masked_input_and_mask( |
| 538 | input_, |
| 539 | self.shard_indices.org_vocab_start_index, |
| 540 | self.shard_indices.org_vocab_end_index, |
| 541 | self.shard_indices.num_org_vocab_padding, |
| 542 | self.shard_indices.added_vocab_start_index, |
| 543 | self.shard_indices.added_vocab_end_index, |
| 544 | ) |
| 545 | output_parallel = self.quant_method.embedding(self, masked_input.long()) |
| 546 | if output_parallel.dtype in ( |
| 547 | torch.float8_e4m3fn, |
| 548 | torch.float8_e5m2, |
| 549 | ): |
| 550 | # Each vocab token has one owner, so FP8 bytes can use int8 SUM. |
| 551 | comm_output = output_parallel.view(torch.int8) |
| 552 | comm_output.masked_fill_(input_mask.unsqueeze(-1), 0) |
| 553 | output = ( |
| 554 | self.parallel_group.all_reduce(comm_output) |
| 555 | if self.parallel_group is not None |
| 556 | else tensor_model_parallel_all_reduce(comm_output) |
| 557 | ) |
| 558 | return output.view(output_parallel.dtype) |
| 559 | output_parallel.masked_fill_(input_mask.unsqueeze(-1), 0) |
| 560 | # Reduce across all the model parallel GPUs. |
| 561 | if self.parallel_group is not None: |
| 562 | return self.parallel_group.all_reduce(output_parallel) |
| 563 | return tensor_model_parallel_all_reduce(output_parallel) |
| 564 | |
| 565 | def extra_repr(self) -> str: |
| 566 | s = f"num_embeddings={self.num_embeddings}" |
| 567 | s += f", num_embeddings_per_partition={self.num_embeddings_per_partition}" |
| 568 | s += f", embedding_dim={self.embedding_dim}" |
| 569 | s += f", org_vocab_size={self.org_vocab_size}" |
| 570 | s += f", num_embeddings_padded={self.num_embeddings_padded}" |
| 571 | s += f", tp_size={self.tp_size}" |
| 572 | return s |
| 573 | |
| 574 | |
| 575 | # --8<-- [start:parallel_lm_head] |
| 576 | @PluggableLayer.register("parallel_lm_head") |
| 577 | class ParallelLMHead(VocabParallelEmbedding): |
| 578 | """Parallelized LM head. |
| 579 | |
| 580 | Output logits weight matrices used in the Sampler. The weight and bias |
| 581 | tensors are padded to make sure they are divisible by the number of |
| 582 | model parallel GPUs. |
| 583 | |
| 584 | Args: |
| 585 | num_embeddings: vocabulary size. |
| 586 | embedding_dim: size of hidden state. |
| 587 | bias: whether to use bias. |
| 588 | params_dtype: type of the parameters. |
| 589 | org_num_embeddings: original vocabulary size (without LoRA). |
| 590 | padding_size: padding size for the vocabulary. |
| 591 | disable_tp: If true, tensor parallelism will be disabled for this layer. |
| 592 | |
| 593 | """ |
| 594 | |
| 595 | # --8<-- [end:parallel_lm_head] |
| 596 | |
| 597 | def __init__( |
| 598 | self, |
| 599 | num_embeddings: int, |
| 600 | embedding_dim: int, |
| 601 | bias: bool = False, |
| 602 | params_dtype: torch.dtype | None = None, |
| 603 | org_num_embeddings: int | None = None, |
| 604 | padding_size: int = DEFAULT_VOCAB_PADDING_SIZE, |
| 605 | quant_config: QuantizationConfig | None = None, |
| 606 | prefix: str = "", |
| 607 | *, |
| 608 | disable_tp: bool = False, |
| 609 | ): |
| 610 | self.has_bias = bias |
| 611 | super().__init__( |
| 612 | num_embeddings, |
| 613 | embedding_dim, |
| 614 | params_dtype, |
| 615 | org_num_embeddings, |
| 616 | padding_size, |
| 617 | quant_config, |
| 618 | prefix, |
| 619 | disable_tp=disable_tp, |
| 620 | ) |
| 621 | self.quant_config = quant_config |
| 622 | if bias: |
| 623 | self._register_bias() |
| 624 | else: |
| 625 | self.register_parameter("bias", None) |
| 626 | |
| 627 | def _register_bias(self): |
| 628 | data = torch.empty(self.num_embeddings_per_partition, dtype=self.params_dtype) |
| 629 | self.bias = Parameter(data, requires_grad=False) |
| 630 | weight_attrs = dict(output_dim=0, weight_loader=self.weight_loader) |
| 631 | set_weight_attrs(weight=self.bias, weight_attrs=weight_attrs) |
| 632 | |
| 633 | def tie_weights(self, embed_tokens: VocabParallelEmbedding): |
| 634 | """Tie the weights with word embeddings.""" |
| 635 | return self.quant_method.tie_weights(self, embed_tokens) |
| 636 | |
| 637 | def forward(self, input_): |
| 638 | del input_ |
| 639 | raise RuntimeError("LMHead's weights should be used in the sampler.") |