| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/huggingface/transformers/blob/a5cc30d72ae2dc19af534e4b35c986cc28db1275/src/transformers/models/falcon/modeling_falcon.py |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # Copyright 2023 the Falcon authors and HuggingFace Inc. team. All rights |
| 8 | # reserved. |
| 9 | # |
| 10 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 11 | # you may not use this file except in compliance with the License. |
| 12 | # You may obtain a copy of the License at |
| 13 | # |
| 14 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 15 | # |
| 16 | # Unless required by applicable law or agreed to in writing, software |
| 17 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 18 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 19 | # See the License for the specific language governing permissions and |
| 20 | # limitations under the License. |
| 21 | """PyTorch Falcon model.""" |
| 22 | |
| 23 | import math |
| 24 | from collections.abc import Iterable |
| 25 | from itertools import islice |
| 26 | from typing import TypeAlias |
| 27 | |
| 28 | import torch |
| 29 | from torch import nn |
| 30 | from torch.nn import LayerNorm |
| 31 | from transformers import FalconConfig as HF_FalconConfig |
| 32 | |
| 33 | from vllm.compilation.decorators import support_torch_compile |
| 34 | from vllm.config import CacheConfig, VllmConfig |
| 35 | from vllm.distributed import ( |
| 36 | get_pp_group, |
| 37 | get_tensor_model_parallel_rank, |
| 38 | get_tensor_model_parallel_world_size, |
| 39 | tensor_model_parallel_all_reduce, |
| 40 | ) |
| 41 | from vllm.model_executor.layers.activation import get_act_fn |
| 42 | from vllm.model_executor.layers.attention import Attention |
| 43 | from vllm.model_executor.layers.linear import ( |
| 44 | ColumnParallelLinear, |
| 45 | QKVParallelLinear, |
| 46 | RowParallelLinear, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 49 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 50 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 51 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 52 | ParallelLMHead, |
| 53 | VocabParallelEmbedding, |
| 54 | ) |
| 55 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 56 | from vllm.sequence import IntermediateTensors |
| 57 | from vllm.transformers_utils.configs.falcon import RWConfig |
| 58 | |
| 59 | from .interfaces import SupportsPP |
| 60 | from .utils import ( |
| 61 | AutoWeightsLoader, |
| 62 | is_pp_missing_parameter, |
| 63 | make_empty_intermediate_tensors_factory, |
| 64 | make_layers, |
| 65 | maybe_prefix, |
| 66 | ) |
| 67 | |
| 68 | FalconConfig: TypeAlias = HF_FalconConfig | RWConfig |
| 69 | |
| 70 | |
| 71 | def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor: |
| 72 | closest_power_of_2 = 2 ** math.floor(math.log2(total_num_heads)) |
| 73 | base = torch.tensor( |
| 74 | 2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), dtype=torch.float32 |
| 75 | ) |
| 76 | powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32) |
| 77 | slopes = torch.pow(base, powers) |
| 78 | |
| 79 | if closest_power_of_2 != total_num_heads: |
| 80 | extra_base = torch.tensor( |
| 81 | 2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), dtype=torch.float32 |
| 82 | ) |
| 83 | num_remaining_heads = min( |
| 84 | closest_power_of_2, total_num_heads - closest_power_of_2 |
| 85 | ) |
| 86 | extra_powers = torch.arange( |
| 87 | 1, 1 + 2 * num_remaining_heads, 2, dtype=torch.int32 |
| 88 | ) |
| 89 | slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0) |
| 90 | |
| 91 | return slopes |
| 92 | |
| 93 | |
| 94 | class FalconAttention(nn.Module): |
| 95 | def __init__( |
| 96 | self, |
| 97 | config: FalconConfig, |
| 98 | cache_config: CacheConfig | None = None, |
| 99 | quant_config: QuantizationConfig | None = None, |
| 100 | prefix: str = "", |
| 101 | ): |
| 102 | super().__init__() |
| 103 | |
| 104 | self.hidden_size = config.hidden_size |
| 105 | tp_size = get_tensor_model_parallel_world_size() |
| 106 | |
| 107 | self.total_num_heads = config.num_attention_heads |
| 108 | assert self.total_num_heads % tp_size == 0 |
| 109 | self.num_heads = self.total_num_heads // tp_size |
| 110 | self.head_dim = self.hidden_size // self.total_num_heads |
| 111 | assert self.head_dim * self.total_num_heads == self.hidden_size |
| 112 | |
| 113 | self.new_decoder_architecture = config.new_decoder_architecture |
| 114 | self.multi_query = config.multi_query |
| 115 | |
| 116 | if self.new_decoder_architecture: |
| 117 | self.total_num_kv_heads = config.num_kv_heads |
| 118 | elif self.multi_query: |
| 119 | self.total_num_kv_heads = 1 |
| 120 | else: |
| 121 | self.total_num_kv_heads = self.total_num_heads |
| 122 | if self.total_num_kv_heads >= tp_size: |
| 123 | # Number of KV heads is greater than TP size, so we partition |
| 124 | # the KV heads across multiple tensor parallel GPUs. |
| 125 | assert self.total_num_kv_heads % tp_size == 0 |
| 126 | else: |
| 127 | # Number of KV heads is less than TP size, so we replicate |
| 128 | # the KV heads across multiple tensor parallel GPUs. |
| 129 | assert tp_size % self.total_num_kv_heads == 0 |
| 130 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 131 | |
| 132 | self.query_key_value = QKVParallelLinear( |
| 133 | self.hidden_size, |
| 134 | self.head_dim, |
| 135 | self.total_num_heads, |
| 136 | self.total_num_kv_heads, |
| 137 | bias=config.bias, |
| 138 | skip_bias_add=True, |
| 139 | quant_config=quant_config, |
| 140 | prefix=f"{prefix}.query_key_value", |
| 141 | ) |
| 142 | self.q_size = self.num_heads * self.head_dim |
| 143 | self.kv_size = self.num_kv_heads * self.head_dim |
| 144 | |
| 145 | # Layer-wise attention scaling |
| 146 | self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim) |
| 147 | self.reduce_row_parallel_results = not ( |
| 148 | config.new_decoder_architecture or config.parallel_attn |
| 149 | ) |
| 150 | self.dense = RowParallelLinear( |
| 151 | self.hidden_size, |
| 152 | self.hidden_size, |
| 153 | bias=config.bias, |
| 154 | skip_bias_add=True, |
| 155 | quant_config=quant_config, |
| 156 | reduce_results=self.reduce_row_parallel_results, |
| 157 | prefix=f"{prefix}.dense", |
| 158 | ) |
| 159 | |
| 160 | self.use_rotary = config.rotary |
| 161 | self.use_alibi = config.alibi |
| 162 | assert not (self.use_rotary and self.use_alibi), ( |
| 163 | "Rotary and alibi are mutually exclusive." |
| 164 | ) |
| 165 | |
| 166 | if self.use_rotary: |
| 167 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 168 | self.rotary_emb = get_rope( |
| 169 | self.head_dim, |
| 170 | max_position=max_position_embeddings, |
| 171 | rope_parameters=config.rope_parameters, |
| 172 | ) |
| 173 | self.attn = Attention( |
| 174 | self.num_heads, |
| 175 | self.head_dim, |
| 176 | self.inv_norm_factor, |
| 177 | num_kv_heads=self.num_kv_heads, |
| 178 | quant_config=quant_config, |
| 179 | prefix=f"{prefix}.attn", |
| 180 | ) |
| 181 | elif self.use_alibi: |
| 182 | tp_rank = get_tensor_model_parallel_rank() |
| 183 | head_start = tp_rank * self.num_heads |
| 184 | head_end = (tp_rank + 1) * self.num_heads |
| 185 | alibi_slopes = ( |
| 186 | _get_alibi_slopes(self.total_num_heads) * self.inv_norm_factor |
| 187 | ) |
| 188 | alibi_slopes = alibi_slopes[head_start:head_end].tolist() |
| 189 | self.attn = Attention( |
| 190 | self.num_heads, |
| 191 | self.head_dim, |
| 192 | self.inv_norm_factor, |
| 193 | num_kv_heads=self.num_kv_heads, |
| 194 | alibi_slopes=alibi_slopes, |
| 195 | quant_config=quant_config, |
| 196 | prefix=f"{prefix}.attn", |
| 197 | ) |
| 198 | else: |
| 199 | self.attn = Attention( |
| 200 | self.num_heads, |
| 201 | self.head_dim, |
| 202 | scale=self.inv_norm_factor, |
| 203 | num_kv_heads=self.num_kv_heads, |
| 204 | cache_config=cache_config, |
| 205 | quant_config=quant_config, |
| 206 | prefix=f"{prefix}.attn", |
| 207 | ) |
| 208 | |
| 209 | def forward( |
| 210 | self, |
| 211 | positions: torch.Tensor, |
| 212 | hidden_states: torch.Tensor, |
| 213 | ) -> torch.Tensor: |
| 214 | qkv, bias = self.query_key_value(hidden_states) |
| 215 | if bias is not None: |
| 216 | qkv += bias |
| 217 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 218 | if self.use_rotary: |
| 219 | q, k = self.rotary_emb(positions, q, k) |
| 220 | attn_output = self.attn(q, k, v) |
| 221 | attn_output, bias = self.dense(attn_output) |
| 222 | return attn_output, bias |
| 223 | |
| 224 | |
| 225 | class FalconMLP(nn.Module): |
| 226 | def __init__( |
| 227 | self, |
| 228 | config: FalconConfig, |
| 229 | quant_config: QuantizationConfig | None = None, |
| 230 | prefix: str = "", |
| 231 | ): |
| 232 | super().__init__() |
| 233 | hidden_size = config.hidden_size |
| 234 | |
| 235 | self.dense_h_to_4h = ColumnParallelLinear( |
| 236 | hidden_size, |
| 237 | 4 * hidden_size, |
| 238 | bias=config.bias, |
| 239 | skip_bias_add=True, |
| 240 | quant_config=quant_config, |
| 241 | prefix=f"{prefix}.dense_h_to_4h", |
| 242 | ) |
| 243 | self.act = get_act_fn("gelu") |
| 244 | self.reduce_row_parallel_results = not ( |
| 245 | config.new_decoder_architecture or config.parallel_attn |
| 246 | ) |
| 247 | self.dense_4h_to_h = RowParallelLinear( |
| 248 | 4 * hidden_size, |
| 249 | hidden_size, |
| 250 | bias=config.bias, |
| 251 | skip_bias_add=True, |
| 252 | reduce_results=self.reduce_row_parallel_results, |
| 253 | quant_config=quant_config, |
| 254 | prefix=f"{prefix}.dense_4h_to_h", |
| 255 | ) |
| 256 | |
| 257 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 258 | # NOTE(zhuohan): Following huggingface, we do not fuse bias add here. |
| 259 | x, bias = self.dense_h_to_4h(x) |
| 260 | if bias is not None: |
| 261 | x += bias |
| 262 | x = self.act(x) |
| 263 | x, bias = self.dense_4h_to_h(x) |
| 264 | return x, bias |
| 265 | |
| 266 | |
| 267 | class FalconDecoderLayer(nn.Module): |
| 268 | def __init__( |
| 269 | self, |
| 270 | config: FalconConfig, |
| 271 | cache_config: CacheConfig | None = None, |
| 272 | quant_config: QuantizationConfig | None = None, |
| 273 | prefix: str = "", |
| 274 | ): |
| 275 | super().__init__() |
| 276 | hidden_size = config.hidden_size |
| 277 | self.num_heads = config.num_attention_heads |
| 278 | self.self_attention = FalconAttention( |
| 279 | config, cache_config, quant_config, prefix=f"{prefix}.self_attention" |
| 280 | ) |
| 281 | self.mlp = FalconMLP(config, quant_config, prefix=f"{prefix}.mlp") |
| 282 | self.config = config |
| 283 | |
| 284 | if not hasattr(config, "num_ln_in_parallel_attn"): |
| 285 | config.num_ln_in_parallel_attn = None |
| 286 | |
| 287 | if config.num_ln_in_parallel_attn is None and config.new_decoder_architecture: |
| 288 | config.num_ln_in_parallel_attn = 2 |
| 289 | |
| 290 | if not config.parallel_attn: |
| 291 | self.post_attention_layernorm = LayerNorm( |
| 292 | hidden_size, eps=config.layer_norm_epsilon |
| 293 | ) |
| 294 | self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) |
| 295 | else: |
| 296 | if config.num_ln_in_parallel_attn == 2: |
| 297 | # The layer norm before self-attention |
| 298 | self.ln_attn = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) |
| 299 | # The layer norm before the MLP |
| 300 | self.ln_mlp = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) |
| 301 | else: |
| 302 | self.input_layernorm = LayerNorm( |
| 303 | hidden_size, eps=config.layer_norm_epsilon |
| 304 | ) |
| 305 | |
| 306 | self.reduce_row_parallel_results = not ( |
| 307 | config.new_decoder_architecture or config.parallel_attn |
| 308 | ) |
| 309 | |
| 310 | def forward( |
| 311 | self, |
| 312 | positions: torch.Tensor, |
| 313 | hidden_states: torch.Tensor, |
| 314 | ) -> torch.Tensor: |
| 315 | residual = hidden_states |
| 316 | |
| 317 | if self.config.num_ln_in_parallel_attn == 2: |
| 318 | attention_layernorm_out = self.ln_attn(hidden_states) |
| 319 | mlp_layernorm_out = self.ln_mlp(hidden_states) |
| 320 | else: |
| 321 | attention_layernorm_out = self.input_layernorm(hidden_states) |
| 322 | |
| 323 | # Self attention. |
| 324 | attention_output, attention_bias = self.self_attention( |
| 325 | positions=positions, |
| 326 | hidden_states=attention_layernorm_out, |
| 327 | ) |
| 328 | if self.reduce_row_parallel_results and attention_bias is not None: |
| 329 | attention_output += attention_bias |
| 330 | |
| 331 | if not self.config.new_decoder_architecture: |
| 332 | if self.config.parallel_attn: |
| 333 | mlp_layernorm_out = attention_layernorm_out |
| 334 | else: |
| 335 | residual += attention_output |
| 336 | mlp_layernorm_out = self.post_attention_layernorm(residual) |
| 337 | |
| 338 | if ( |
| 339 | self.config.new_decoder_architecture |
| 340 | and self.config.parallel_attn |
| 341 | and self.config.num_ln_in_parallel_attn == 1 |
| 342 | ): |
| 343 | mlp_layernorm_out = attention_layernorm_out |
| 344 | |
| 345 | # MLP. |
| 346 | mlp_output, mlp_bias = self.mlp(mlp_layernorm_out) |
| 347 | if self.reduce_row_parallel_results and mlp_bias is not None: |
| 348 | mlp_output += mlp_bias |
| 349 | |
| 350 | if not self.reduce_row_parallel_results: |
| 351 | # When MLP and Attention layers are parallel, we can use |
| 352 | # only one all-reduce operator to reduce the results from |
| 353 | # both MLP and Attention layers. |
| 354 | mlp_output += attention_output |
| 355 | mlp_output = tensor_model_parallel_all_reduce(mlp_output) |
| 356 | if attention_bias is not None: |
| 357 | mlp_output += attention_bias |
| 358 | if mlp_bias is not None: |
| 359 | mlp_output += mlp_bias |
| 360 | |
| 361 | output = mlp_output + residual |
| 362 | return output |
| 363 | |
| 364 | |
| 365 | @support_torch_compile |
| 366 | class FalconModel(nn.Module): |
| 367 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 368 | super().__init__() |
| 369 | |
| 370 | config = vllm_config.model_config.hf_config |
| 371 | cache_config = vllm_config.cache_config |
| 372 | quant_config = vllm_config.quant_config |
| 373 | |
| 374 | self.config = config |
| 375 | self.embed_dim = config.hidden_size |
| 376 | self.num_heads = config.num_attention_heads |
| 377 | self.use_alibi = config.alibi |
| 378 | |
| 379 | # Embedding + LN Embedding |
| 380 | self.word_embeddings = VocabParallelEmbedding( |
| 381 | config.vocab_size, |
| 382 | self.embed_dim, |
| 383 | ) |
| 384 | |
| 385 | # Transformer blocks |
| 386 | self.start_layer, self.end_layer, self.h = make_layers( |
| 387 | config.num_hidden_layers, |
| 388 | lambda prefix: FalconDecoderLayer( |
| 389 | config, cache_config, quant_config, prefix=prefix |
| 390 | ), |
| 391 | prefix=f"{prefix}.h", |
| 392 | ) |
| 393 | |
| 394 | # Final Layer Norm |
| 395 | self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon) |
| 396 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 397 | ["hidden_states"], config.hidden_size |
| 398 | ) |
| 399 | |
| 400 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 401 | return self.word_embeddings(input_ids) |
| 402 | |
| 403 | def forward( |
| 404 | self, |
| 405 | input_ids: torch.Tensor | None, |
| 406 | positions: torch.Tensor, |
| 407 | intermediate_tensors: IntermediateTensors | None, |
| 408 | inputs_embeds: torch.Tensor | None = None, |
| 409 | ) -> torch.Tensor | IntermediateTensors: |
| 410 | if get_pp_group().is_first_rank: |
| 411 | if inputs_embeds is not None: |
| 412 | hidden_states = inputs_embeds |
| 413 | else: |
| 414 | hidden_states = self.embed_input_ids(input_ids) |
| 415 | else: |
| 416 | assert intermediate_tensors is not None |
| 417 | hidden_states = intermediate_tensors["hidden_states"] |
| 418 | for layer in islice(self.h, self.start_layer, self.end_layer): |
| 419 | hidden_states = layer(positions, hidden_states) |
| 420 | if not get_pp_group().is_last_rank: |
| 421 | return IntermediateTensors({"hidden_states": hidden_states}) |
| 422 | hidden_states = self.ln_f(hidden_states) |
| 423 | return hidden_states |
| 424 | |
| 425 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 426 | total_num_heads = self.config.num_attention_heads |
| 427 | if self.config.new_decoder_architecture: |
| 428 | total_num_kv_heads = self.config.num_kv_heads |
| 429 | elif self.config.multi_query: |
| 430 | total_num_kv_heads = 1 |
| 431 | else: |
| 432 | total_num_kv_heads = total_num_heads |
| 433 | num_query_heads_per_kv_head = total_num_heads // total_num_kv_heads |
| 434 | params_dict = dict(self.named_parameters(remove_duplicate=False)) |
| 435 | loaded_params: set[str] = set() |
| 436 | for name, loaded_weight in weights: |
| 437 | # Skip loading extra bias for GPTQ models. |
| 438 | if name.endswith(".bias") and name not in params_dict: |
| 439 | continue |
| 440 | if is_pp_missing_parameter(name, self): |
| 441 | continue |
| 442 | param = params_dict[name] |
| 443 | if "query_key_value" in name: |
| 444 | output_dim = getattr(param, "output_dim", None) |
| 445 | loaded_weight_shape = loaded_weight.shape |
| 446 | if output_dim is not None: |
| 447 | loaded_weight = loaded_weight.view( |
| 448 | loaded_weight_shape[:output_dim] |
| 449 | + (total_num_kv_heads, num_query_heads_per_kv_head + 2, -1) |
| 450 | + loaded_weight_shape[output_dim + 1 :] |
| 451 | ) |
| 452 | wq = loaded_weight.narrow( |
| 453 | output_dim + 1, 0, num_query_heads_per_kv_head |
| 454 | ).reshape( |
| 455 | *loaded_weight_shape[:output_dim], |
| 456 | -1, |
| 457 | *loaded_weight_shape[output_dim + 1 :], |
| 458 | ) |
| 459 | wk = loaded_weight.narrow( |
| 460 | output_dim + 1, num_query_heads_per_kv_head, 1 |
| 461 | ).reshape( |
| 462 | *loaded_weight_shape[:output_dim], |
| 463 | -1, |
| 464 | *loaded_weight_shape[output_dim + 1 :], |
| 465 | ) |
| 466 | wv = loaded_weight.narrow( |
| 467 | output_dim + 1, num_query_heads_per_kv_head + 1, 1 |
| 468 | ).reshape( |
| 469 | *loaded_weight_shape[:output_dim], |
| 470 | -1, |
| 471 | *loaded_weight_shape[output_dim + 1 :], |
| 472 | ) |
| 473 | loaded_weight = torch.cat([wq, wk, wv], dim=output_dim) |
| 474 | |
| 475 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 476 | weight_loader(param, loaded_weight) |
| 477 | loaded_params.add(name) |
| 478 | return loaded_params |
| 479 | |
| 480 | |
| 481 | class FalconForCausalLM(nn.Module, SupportsPP): |
| 482 | packed_modules_mapping = { |
| 483 | "query_key_value": ["query_key_value"], |
| 484 | } |
| 485 | |
| 486 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 487 | super().__init__() |
| 488 | config = vllm_config.model_config.hf_config |
| 489 | quant_config = vllm_config.quant_config |
| 490 | self.config = config |
| 491 | self.quant_config = quant_config |
| 492 | self.transformer = FalconModel( |
| 493 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer") |
| 494 | ) |
| 495 | # only Falcon-11B doesn't share lm_head weight with word embeddings |
| 496 | # and previous Falcon model doesn't have tie_word_embeddings config |
| 497 | # so we set tie_word_embeddings to True by default |
| 498 | self.tie_word_embeddings = ( |
| 499 | config.tie_word_embeddings |
| 500 | if config.tie_word_embeddings is not None |
| 501 | else True |
| 502 | ) |
| 503 | self.lm_head = ParallelLMHead( |
| 504 | config.vocab_size, |
| 505 | config.hidden_size, |
| 506 | quant_config=quant_config, |
| 507 | prefix=maybe_prefix(prefix, "lm_head"), |
| 508 | ) |
| 509 | if self.tie_word_embeddings: |
| 510 | self.lm_head = self.lm_head.tie_weights(self.transformer.word_embeddings) |
| 511 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 512 | self.make_empty_intermediate_tensors = ( |
| 513 | self.transformer.make_empty_intermediate_tensors |
| 514 | ) |
| 515 | |
| 516 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 517 | return self.transformer.embed_input_ids(input_ids) |
| 518 | |
| 519 | def forward( |
| 520 | self, |
| 521 | input_ids: torch.LongTensor, |
| 522 | positions: torch.Tensor, |
| 523 | intermediate_tensors: IntermediateTensors | None = None, |
| 524 | inputs_embeds: torch.Tensor | None = None, |
| 525 | ) -> torch.Tensor: |
| 526 | hidden_states = self.transformer( |
| 527 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 528 | ) |
| 529 | return hidden_states |
| 530 | |
| 531 | def compute_logits( |
| 532 | self, |
| 533 | hidden_states: torch.Tensor, |
| 534 | ) -> torch.Tensor | None: |
| 535 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 536 | return logits |
| 537 | |
| 538 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 539 | loader = AutoWeightsLoader(self) |
| 540 | return loader.load_weights(weights) |