| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # |
| 4 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | # you may not use this file except in compliance with the License. |
| 6 | # You may obtain a copy of the License at |
| 7 | # |
| 8 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | # |
| 10 | # Unless required by applicable law or agreed to in writing, software |
| 11 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | # See the License for the specific language governing permissions and |
| 14 | # limitations under the License. |
| 15 | """Inference-only LoopCoder model compatible with HuggingFace weights.""" |
| 16 | |
| 17 | from __future__ import annotations |
| 18 | |
| 19 | from collections.abc import Callable, Iterable |
| 20 | from dataclasses import replace |
| 21 | from typing import Any |
| 22 | |
| 23 | import torch |
| 24 | from torch import nn |
| 25 | from transformers import PreTrainedConfig |
| 26 | |
| 27 | from vllm.compilation.decorators import support_torch_compile |
| 28 | from vllm.config import CacheConfig, VllmConfig |
| 29 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 30 | from vllm.model_executor.layers.attention import Attention |
| 31 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 32 | from vllm.model_executor.layers.linear import ( |
| 33 | ColumnParallelLinear, |
| 34 | QKVParallelLinear, |
| 35 | RowParallelLinear, |
| 36 | ) |
| 37 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 38 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 39 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 40 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 41 | ParallelLMHead, |
| 42 | VocabParallelEmbedding, |
| 43 | ) |
| 44 | from vllm.model_executor.model_loader.weight_utils import ( |
| 45 | default_weight_loader, |
| 46 | maybe_remap_kv_scale_name, |
| 47 | ) |
| 48 | from vllm.model_executor.models.llama import LlamaMLP |
| 49 | from vllm.sequence import IntermediateTensors |
| 50 | from vllm.v1.attention.backend import AttentionType |
| 51 | |
| 52 | from .utils import ( |
| 53 | AutoWeightsLoader, |
| 54 | extract_layer_index, |
| 55 | make_layers, |
| 56 | maybe_prefix, |
| 57 | ) |
| 58 | |
| 59 | |
| 60 | class LoopCoderAttention(nn.Module): |
| 61 | def __init__( |
| 62 | self, |
| 63 | config: PreTrainedConfig, |
| 64 | hidden_size: int, |
| 65 | num_heads: int, |
| 66 | num_kv_heads: int, |
| 67 | max_position: int = 4096 * 32, |
| 68 | cache_config: CacheConfig | None = None, |
| 69 | quant_config: QuantizationConfig | None = None, |
| 70 | prefix: str = "", |
| 71 | attn_type: str = AttentionType.DECODER, |
| 72 | dual_chunk_attention_config: dict[str, Any] | None = None, |
| 73 | layer_idx: int = 0, |
| 74 | ) -> None: |
| 75 | super().__init__() |
| 76 | self.layer_idx = layer_idx |
| 77 | self.hidden_size = hidden_size |
| 78 | tp_size = get_tensor_model_parallel_world_size() |
| 79 | self.total_num_heads = num_heads |
| 80 | assert self.total_num_heads % tp_size == 0 |
| 81 | self.num_heads = self.total_num_heads // tp_size |
| 82 | self.total_num_kv_heads = num_kv_heads |
| 83 | if self.total_num_kv_heads >= tp_size: |
| 84 | # Number of KV heads is greater than TP size, so we partition |
| 85 | # the KV heads across multiple tensor parallel GPUs. |
| 86 | assert self.total_num_kv_heads % tp_size == 0 |
| 87 | else: |
| 88 | # Number of KV heads is less than TP size, so we replicate |
| 89 | # the KV heads across multiple tensor parallel GPUs. |
| 90 | assert tp_size % self.total_num_kv_heads == 0 |
| 91 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 92 | self.head_dim = hidden_size // self.total_num_heads |
| 93 | self.q_size = self.num_heads * self.head_dim |
| 94 | self.kv_size = self.num_kv_heads * self.head_dim |
| 95 | self.scaling = self.head_dim**-0.5 |
| 96 | self.dual_chunk_attention_config = dual_chunk_attention_config |
| 97 | |
| 98 | # Get loop_num from config, default to 2 if not specified |
| 99 | self.loop_num = getattr(config, "loop_num", 2) |
| 100 | |
| 101 | self.loop_window_size = getattr(config, "loop_window_size", 64) |
| 102 | |
| 103 | # Use total number of hidden layers instead of hardcoded 24 |
| 104 | total_layers = config.num_hidden_layers |
| 105 | |
| 106 | self.qkv_proj = QKVParallelLinear( |
| 107 | hidden_size, |
| 108 | self.head_dim, |
| 109 | self.total_num_heads, |
| 110 | self.total_num_kv_heads, |
| 111 | bias=False, |
| 112 | quant_config=quant_config, |
| 113 | prefix=f"{prefix}.qkv_proj", |
| 114 | ) |
| 115 | self.o_proj = RowParallelLinear( |
| 116 | self.total_num_heads * self.head_dim, |
| 117 | hidden_size, |
| 118 | bias=False, |
| 119 | quant_config=quant_config, |
| 120 | prefix=f"{prefix}.o_proj", |
| 121 | ) |
| 122 | |
| 123 | self.rotary_emb = get_rope( |
| 124 | self.head_dim, |
| 125 | max_position=max_position, |
| 126 | rope_parameters=config.rope_parameters, |
| 127 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 128 | ) |
| 129 | self.attn = nn.ModuleList() |
| 130 | |
| 131 | base_cache_config = cache_config |
| 132 | |
| 133 | for loop_idx in range(self.loop_num): |
| 134 | base_layer_idx = extract_layer_index(prefix) |
| 135 | unique_layer_idx = loop_idx * total_layers + base_layer_idx |
| 136 | |
| 137 | unique_prefix = prefix.replace( |
| 138 | f"layers.{base_layer_idx}", f"layers.{unique_layer_idx}" |
| 139 | ) |
| 140 | |
| 141 | if loop_idx == 0: |
| 142 | loop_cache_config = cache_config |
| 143 | else: |
| 144 | if base_cache_config is not None: |
| 145 | loop_cache_config = replace( |
| 146 | base_cache_config, |
| 147 | sliding_window=self.loop_window_size, |
| 148 | ) |
| 149 | else: |
| 150 | loop_cache_config = CacheConfig( |
| 151 | sliding_window=self.loop_window_size, |
| 152 | cache_dtype="auto", |
| 153 | ) |
| 154 | |
| 155 | self.attn.append( |
| 156 | Attention( |
| 157 | self.num_heads, |
| 158 | self.head_dim, |
| 159 | self.scaling, |
| 160 | num_kv_heads=self.num_kv_heads, |
| 161 | cache_config=loop_cache_config, |
| 162 | quant_config=quant_config, |
| 163 | attn_type=attn_type, |
| 164 | prefix=f"{unique_prefix}.attn", |
| 165 | **{ |
| 166 | "layer_idx": unique_layer_idx, |
| 167 | "dual_chunk_attention_config": dual_chunk_attention_config, |
| 168 | } |
| 169 | if dual_chunk_attention_config and loop_idx == 0 |
| 170 | else {}, |
| 171 | ) |
| 172 | ) |
| 173 | |
| 174 | def forward( |
| 175 | self, |
| 176 | positions: torch.Tensor, |
| 177 | hidden_states: torch.Tensor, |
| 178 | loop_idx: int, |
| 179 | gate_proj: LoopGateProjection | None = None, |
| 180 | ) -> torch.Tensor: |
| 181 | if loop_idx == 0: |
| 182 | attn = self.attn[0] |
| 183 | qkv, _ = self.qkv_proj(hidden_states) |
| 184 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 185 | q, k = self.rotary_emb(positions, q, k) |
| 186 | attn_output = attn(q, k, v) |
| 187 | output, _ = self.o_proj(attn_output) |
| 188 | return output |
| 189 | else: |
| 190 | global_attn = self.attn[0] |
| 191 | local_attn = self.attn[loop_idx] |
| 192 | qkv, _ = self.qkv_proj(hidden_states) |
| 193 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 194 | q, k = self.rotary_emb(positions, q, k) |
| 195 | num_tokens, _ = q.shape |
| 196 | num_heads = self.num_heads |
| 197 | head_dim = self.head_dim |
| 198 | |
| 199 | q_reshaped = q.view(num_tokens, num_heads, head_dim).transpose(0, 1) |
| 200 | |
| 201 | global_attn_output = global_attn(q, None, None) |
| 202 | local_attn_output = local_attn(q, k, v) |
| 203 | assert gate_proj is not None, "gate_proj must be provided for loop_idx > 0" |
| 204 | gate = gate_proj(q_reshaped) |
| 205 | output = global_attn_output * gate + local_attn_output * (1 - gate) |
| 206 | output, _ = self.o_proj(output) |
| 207 | return output |
| 208 | |
| 209 | |
| 210 | class LoopCoderDecoderLayer(nn.Module): |
| 211 | def __init__( |
| 212 | self, |
| 213 | config: PreTrainedConfig, |
| 214 | cache_config: CacheConfig | None = None, |
| 215 | quant_config: QuantizationConfig | None = None, |
| 216 | prefix: str = "", |
| 217 | layer_idx: int = 0, |
| 218 | ) -> None: |
| 219 | super().__init__() |
| 220 | self.hidden_size = config.hidden_size |
| 221 | dual_chunk_attention_config = getattr( |
| 222 | config, "dual_chunk_attention_config", None |
| 223 | ) |
| 224 | self.layer_idx = layer_idx |
| 225 | if getattr(config, "is_causal", True): |
| 226 | attn_type = AttentionType.DECODER |
| 227 | else: |
| 228 | attn_type = AttentionType.ENCODER_ONLY |
| 229 | |
| 230 | self.self_attn = LoopCoderAttention( |
| 231 | config=config, |
| 232 | hidden_size=self.hidden_size, |
| 233 | num_heads=config.num_attention_heads, |
| 234 | max_position=config.max_position_embeddings, |
| 235 | num_kv_heads=config.num_key_value_heads, |
| 236 | cache_config=cache_config, |
| 237 | quant_config=quant_config, |
| 238 | prefix=f"{prefix}.self_attn", |
| 239 | attn_type=attn_type, |
| 240 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 241 | layer_idx=self.layer_idx, |
| 242 | ) |
| 243 | self.mlp = LlamaMLP( |
| 244 | hidden_size=self.hidden_size, |
| 245 | intermediate_size=config.intermediate_size, |
| 246 | hidden_act=config.hidden_act, |
| 247 | quant_config=quant_config, |
| 248 | prefix=f"{prefix}.mlp", |
| 249 | ) |
| 250 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 251 | self.post_attention_layernorm = RMSNorm( |
| 252 | config.hidden_size, eps=config.rms_norm_eps |
| 253 | ) |
| 254 | |
| 255 | def forward( |
| 256 | self, |
| 257 | positions: torch.Tensor, |
| 258 | hidden_states: torch.Tensor, |
| 259 | loop_idx: int, |
| 260 | gate_proj: LoopGateProjection | None = None, |
| 261 | ) -> torch.Tensor: |
| 262 | residual = hidden_states |
| 263 | hidden_states = self.input_layernorm(hidden_states) |
| 264 | hidden_states = self.self_attn( |
| 265 | positions=positions, |
| 266 | hidden_states=hidden_states, |
| 267 | loop_idx=loop_idx, |
| 268 | gate_proj=gate_proj, |
| 269 | ) |
| 270 | hidden_states = hidden_states + residual |
| 271 | residual = hidden_states |
| 272 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 273 | hidden_states = self.mlp(hidden_states) |
| 274 | hidden_states = hidden_states + residual |
| 275 | |
| 276 | return hidden_states |
| 277 | |
| 278 | |
| 279 | class LoopGateProjection(nn.Module): |
| 280 | """Gate projection for mixed attention in Loop 2+. |
| 281 | |
| 282 | Computes: g = sigmoid(linear(Q)) for each head independently. |
| 283 | This gate determines how much to use Loop1's KV (global) vs current |
| 284 | loop's KV (local). |
| 285 | |
| 286 | Supports tensor parallelism: each GPU handles a subset of heads. |
| 287 | The weight matrix has shape [num_heads, head_dim] and is split along |
| 288 | the head dimension. |
| 289 | """ |
| 290 | |
| 291 | def __init__( |
| 292 | self, |
| 293 | total_num_heads: int, |
| 294 | head_dim: int, |
| 295 | quant_config: QuantizationConfig | None = None, |
| 296 | prefix: str = "", |
| 297 | ): |
| 298 | super().__init__() |
| 299 | self.total_num_heads = total_num_heads |
| 300 | self.head_dim = head_dim |
| 301 | tp_size = get_tensor_model_parallel_world_size() |
| 302 | assert self.total_num_heads % tp_size == 0 |
| 303 | self.num_heads = self.total_num_heads // tp_size |
| 304 | |
| 305 | self.gate_proj = ColumnParallelLinear( |
| 306 | head_dim, |
| 307 | self.total_num_heads, |
| 308 | bias=True, |
| 309 | gather_output=False, |
| 310 | quant_config=quant_config, |
| 311 | prefix=f"{prefix}.gate_proj", |
| 312 | ) |
| 313 | |
| 314 | def forward(self, query: torch.Tensor) -> torch.Tensor: |
| 315 | """Compute gate values from query tensor. |
| 316 | |
| 317 | Args: |
| 318 | query: [num_heads, num_tokens, head_dim] (vLLM flattened format) |
| 319 | where num_heads is the number of heads on this TP rank |
| 320 | and num_tokens = batch * seq_len |
| 321 | |
| 322 | Returns: |
| 323 | gate: [num_tokens, num_heads * head_dim] (flattened format matching q shape) |
| 324 | |
| 325 | """ |
| 326 | num_heads, num_tokens, head_dim = query.shape |
| 327 | |
| 328 | assert num_heads == self.num_heads, ( |
| 329 | f"Expected {self.num_heads} heads, got {num_heads}" |
| 330 | ) |
| 331 | |
| 332 | query_flat = query.reshape(-1, head_dim) |
| 333 | |
| 334 | gate_logits_flat, _ = self.gate_proj(query_flat) |
| 335 | |
| 336 | gate_logits = gate_logits_flat.reshape( |
| 337 | num_heads, num_tokens, self.num_heads |
| 338 | ) # [num_heads, num_tokens, num_heads] |
| 339 | |
| 340 | # Extract diagonal: each head h's query should use output column h |
| 341 | # gate_logits[h, :, h] gives the output for head h at each token |
| 342 | gate_logits = torch.diagonal( |
| 343 | gate_logits, dim1=0, dim2=2 |
| 344 | ) # [num_tokens, num_heads] |
| 345 | gate_logits = gate_logits.transpose(0, 1) # [num_heads, num_tokens] |
| 346 | gate_logits = gate_logits.unsqueeze(-1) # [num_heads, num_tokens, 1] |
| 347 | |
| 348 | # Apply sigmoid |
| 349 | gate = torch.sigmoid(gate_logits) # [num_heads, num_tokens, 1] |
| 350 | |
| 351 | # Expand and reshape to match q shape: [num_tokens, num_heads * head_dim] |
| 352 | gate = gate.transpose(0, 1) # [num_tokens, num_heads, 1] |
| 353 | gate = gate.expand(-1, -1, head_dim) # [num_tokens, num_heads, head_dim] |
| 354 | gate = gate.reshape( |
| 355 | num_tokens, num_heads * head_dim |
| 356 | ) # [num_tokens, num_heads * head_dim] |
| 357 | |
| 358 | return gate |
| 359 | |
| 360 | |
| 361 | @support_torch_compile( |
| 362 | dynamic_arg_dims={ |
| 363 | "input_ids": 0, |
| 364 | "positions": -1, |
| 365 | "intermediate_tensors": 0, |
| 366 | "inputs_embeds": 0, |
| 367 | } |
| 368 | ) |
| 369 | class IQuestLoopCoderModel(nn.Module): |
| 370 | def __init__( |
| 371 | self, |
| 372 | *, |
| 373 | vllm_config: VllmConfig, |
| 374 | prefix: str = "", |
| 375 | decoder_layer_type: type[nn.Module] = LoopCoderDecoderLayer, |
| 376 | ): |
| 377 | super().__init__() |
| 378 | |
| 379 | config = vllm_config.model_config.hf_config |
| 380 | cache_config = vllm_config.cache_config |
| 381 | quant_config = vllm_config.quant_config |
| 382 | |
| 383 | # TODO (@robertgshaw2): see if this can be moved out |
| 384 | if cache_config.sliding_window is not None and hasattr( |
| 385 | config, "max_window_layers" |
| 386 | ): |
| 387 | assert config.max_window_layers == config.num_hidden_layers, ( |
| 388 | "Sliding window for some but all layers is not supported. " |
| 389 | "This model uses sliding window but `max_window_layers` = {} " |
| 390 | "is less than `num_hidden_layers` = {}. Please open an issue " |
| 391 | "to discuss this feature.".format( |
| 392 | config.max_window_layers, |
| 393 | config.num_hidden_layers, |
| 394 | ) |
| 395 | ) |
| 396 | |
| 397 | self.config = config |
| 398 | self.quant_config = quant_config |
| 399 | self.vocab_size = config.vocab_size |
| 400 | |
| 401 | self.embed_tokens = VocabParallelEmbedding( |
| 402 | config.vocab_size, |
| 403 | config.hidden_size, |
| 404 | quant_config=quant_config, |
| 405 | prefix=f"{prefix}.embed_tokens", |
| 406 | ) |
| 407 | |
| 408 | self.loop_num = getattr(self.config, "loop_num", 2) |
| 409 | self.window_size = getattr(self.config, "loop_window_size", 64) |
| 410 | |
| 411 | # Gate projections for Loop 2+ (one per layer) |
| 412 | head_dim = config.hidden_size // config.num_attention_heads |
| 413 | _, _, self.gate_projections = make_layers( |
| 414 | config.num_hidden_layers, |
| 415 | lambda prefix: LoopGateProjection( |
| 416 | total_num_heads=config.num_attention_heads, |
| 417 | head_dim=head_dim, |
| 418 | quant_config=quant_config, |
| 419 | prefix=prefix, |
| 420 | ), |
| 421 | prefix=f"{prefix}.gate_projections", |
| 422 | ) |
| 423 | |
| 424 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 425 | config.num_hidden_layers, |
| 426 | lambda prefix: LoopCoderDecoderLayer( |
| 427 | config=config, |
| 428 | cache_config=cache_config, |
| 429 | quant_config=quant_config, |
| 430 | prefix=prefix, |
| 431 | layer_idx=extract_layer_index(prefix), |
| 432 | ), |
| 433 | prefix=f"{prefix}.layers", |
| 434 | ) |
| 435 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 436 | |
| 437 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 438 | return self.embed_tokens(input_ids) |
| 439 | |
| 440 | def forward( |
| 441 | self, |
| 442 | input_ids: torch.Tensor | None, |
| 443 | positions: torch.Tensor, |
| 444 | intermediate_tensors: IntermediateTensors | None = None, |
| 445 | inputs_embeds: torch.Tensor | None = None, |
| 446 | ) -> torch.Tensor | IntermediateTensors: |
| 447 | if inputs_embeds is not None: |
| 448 | hidden_states = inputs_embeds |
| 449 | else: |
| 450 | hidden_states = self.embed_input_ids(input_ids) |
| 451 | |
| 452 | for loop_idx in range(self.loop_num): |
| 453 | for layer_idx, layer in enumerate( |
| 454 | self.layers[self.start_layer : self.end_layer] |
| 455 | ): |
| 456 | # Get the actual layer index (accounting for pipeline parallelism) |
| 457 | actual_layer_idx = self.start_layer + layer_idx |
| 458 | # Get gate_proj for this layer (only for loop_idx > 0) |
| 459 | gate_proj = ( |
| 460 | self.gate_projections[actual_layer_idx] if loop_idx > 0 else None |
| 461 | ) |
| 462 | hidden_states = layer(positions, hidden_states, loop_idx, gate_proj) |
| 463 | hidden_states = self.norm(hidden_states) |
| 464 | return hidden_states |
| 465 | |
| 466 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 467 | stacked_params_mapping = [ |
| 468 | # (param_name, shard_name, shard_id) |
| 469 | ("qkv_proj", "q_proj", "q"), |
| 470 | ("qkv_proj", "k_proj", "k"), |
| 471 | ("qkv_proj", "v_proj", "v"), |
| 472 | ("gate_up_proj", "gate_proj", 0), |
| 473 | ("gate_up_proj", "up_proj", 1), |
| 474 | ] |
| 475 | params_dict = dict(self.named_parameters(remove_duplicate=False)) |
| 476 | loaded_params: set[str] = set() |
| 477 | for name, loaded_weight in weights: |
| 478 | if "rotary_emb.inv_freq" in name: |
| 479 | continue |
| 480 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 481 | if "gate_projections" in name: |
| 482 | continue |
| 483 | if weight_name not in name: |
| 484 | continue |
| 485 | name = name.replace(weight_name, param_name) |
| 486 | # Skip loading extra bias for GPTQ models. |
| 487 | if name.endswith(".bias") and name not in params_dict: |
| 488 | continue |
| 489 | if name.endswith("scale"): |
| 490 | # Remapping the name of FP8 kv-scale. |
| 491 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 492 | if remapped_name is None: |
| 493 | continue |
| 494 | name = remapped_name |
| 495 | param = params_dict[name] |
| 496 | weight_loader: Callable[..., None] = getattr( |
| 497 | param, "weight_loader", default_weight_loader |
| 498 | ) |
| 499 | if weight_loader == default_weight_loader: |
| 500 | weight_loader(param, loaded_weight) |
| 501 | else: |
| 502 | weight_loader(param, loaded_weight, shard_id) |
| 503 | break |
| 504 | else: |
| 505 | if name.startswith("gate_projections."): |
| 506 | if name.endswith(".weight"): |
| 507 | vllm_name = name.replace(".weight", ".gate_proj.weight") |
| 508 | elif name.endswith(".bias"): |
| 509 | vllm_name = name.replace(".bias", ".gate_proj.bias") |
| 510 | else: |
| 511 | continue |
| 512 | |
| 513 | if vllm_name in params_dict: |
| 514 | param = params_dict[vllm_name] |
| 515 | weight_loader = getattr( |
| 516 | param, "weight_loader", default_weight_loader |
| 517 | ) |
| 518 | weight_loader(param, loaded_weight) |
| 519 | loaded_params.add(vllm_name) |
| 520 | continue |
| 521 | continue |
| 522 | if name.endswith(".bias") and name not in params_dict: |
| 523 | continue |
| 524 | # Remapping the name of FP8 kv-scale. |
| 525 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 526 | if remapped_name is None: |
| 527 | continue |
| 528 | name = remapped_name |
| 529 | param = params_dict[name] |
| 530 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 531 | weight_loader(param, loaded_weight) |
| 532 | loaded_params.add(name) |
| 533 | return loaded_params |
| 534 | |
| 535 | |
| 536 | class IQuestLoopCoderForCausalLM(nn.Module): |
| 537 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 538 | super().__init__() |
| 539 | config = vllm_config.model_config.hf_config |
| 540 | quant_config = vllm_config.quant_config |
| 541 | |
| 542 | self.config = config |
| 543 | |
| 544 | self.quant_config = quant_config |
| 545 | self.model = IQuestLoopCoderModel( |
| 546 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 547 | ) |
| 548 | |
| 549 | self.lm_head = ParallelLMHead( |
| 550 | config.vocab_size, |
| 551 | config.hidden_size, |
| 552 | quant_config=quant_config, |
| 553 | prefix=maybe_prefix(prefix, "lm_head"), |
| 554 | ) |
| 555 | if config.tie_word_embeddings: |
| 556 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 557 | |
| 558 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 559 | |
| 560 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 561 | return self.model.embed_input_ids(input_ids) |
| 562 | |
| 563 | def forward( |
| 564 | self, |
| 565 | input_ids: torch.Tensor | None, |
| 566 | positions: torch.Tensor, |
| 567 | intermediate_tensors: IntermediateTensors | None = None, |
| 568 | inputs_embeds: torch.Tensor | None = None, |
| 569 | ) -> torch.Tensor | IntermediateTensors: |
| 570 | hidden_states = self.model( |
| 571 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 572 | ) |
| 573 | return hidden_states |
| 574 | |
| 575 | def compute_logits( |
| 576 | self, |
| 577 | hidden_states: torch.Tensor, |
| 578 | ) -> torch.Tensor | None: |
| 579 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 580 | return logits |
| 581 | |
| 582 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 583 | loader = AutoWeightsLoader(self) |
| 584 | return loader.load_weights(weights) |