| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """HRM-Text: Hierarchical Reasoning Model — Text variant. |
| 4 | |
| 5 | Reference Hugging Face implementation: |
| 6 | src/transformers/models/hrm_text/modeling_hrm_text.py |
| 7 | |
| 8 | The model performs a hierarchical recurrent forward over two transformer |
| 9 | stacks (``H`` slow, ``L`` fast) inside nested loops. Each recurrence step |
| 10 | gets its own KV cache slot via a unique vLLM-visible layer index. The |
| 11 | PrefixLM attention pattern (prompt bidirectional, response causal) is |
| 12 | realized by reusing ``EncoderOnlyAttention`` (which sets ``causal=False`` |
| 13 | unconditionally on every metadata build) but with ``attn_type=DECODER`` |
| 14 | so the KV cache is allocated; see ``HrmTextAttention`` for usage. |
| 15 | |
| 16 | The on-disk ``attn.gqkv_proj.weight`` (rows concatenated as |
| 17 | ``[gate | q | k | v]``) is loaded by a single |
| 18 | ``MergedColumnParallelLinear`` with four equal-sized output partitions; |
| 19 | its weight loader auto-splits the fused tensor along the output dim by |
| 20 | ``output_sizes`` (the same path used by Phi-3's fused gate_up_proj). |
| 21 | """ |
| 22 | |
| 23 | from collections.abc import Iterable |
| 24 | from typing import Literal |
| 25 | |
| 26 | import torch |
| 27 | from torch import nn |
| 28 | from transformers import PreTrainedConfig |
| 29 | |
| 30 | from vllm.compilation.decorators import support_torch_compile |
| 31 | from vllm.config import CacheConfig, VllmConfig |
| 32 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 33 | from vllm.model_executor.layers.activation import SiluAndMul |
| 34 | from vllm.model_executor.layers.attention import PrefillPrefixLMAttention |
| 35 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 36 | from vllm.model_executor.layers.linear import ( |
| 37 | MergedColumnParallelLinear, |
| 38 | RowParallelLinear, |
| 39 | ) |
| 40 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 41 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 42 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 43 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 44 | ParallelLMHead, |
| 45 | VocabParallelEmbedding, |
| 46 | ) |
| 47 | from vllm.sequence import IntermediateTensors |
| 48 | |
| 49 | from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix |
| 50 | |
| 51 | |
| 52 | class HrmTextMLP(nn.Module): |
| 53 | def __init__( |
| 54 | self, |
| 55 | hidden_size: int, |
| 56 | intermediate_size: int, |
| 57 | hidden_act: str, |
| 58 | bias: bool = False, |
| 59 | quant_config: QuantizationConfig | None = None, |
| 60 | prefix: str = "", |
| 61 | ) -> None: |
| 62 | super().__init__() |
| 63 | if hidden_act != "silu": |
| 64 | raise ValueError( |
| 65 | f"HrmTextMLP only supports hidden_act='silu', got {hidden_act!r}" |
| 66 | ) |
| 67 | self.gate_up_proj = MergedColumnParallelLinear( |
| 68 | hidden_size, |
| 69 | [intermediate_size] * 2, |
| 70 | bias=bias, |
| 71 | quant_config=quant_config, |
| 72 | prefix=f"{prefix}.gate_up_proj", |
| 73 | ) |
| 74 | self.down_proj = RowParallelLinear( |
| 75 | intermediate_size, |
| 76 | hidden_size, |
| 77 | bias=bias, |
| 78 | quant_config=quant_config, |
| 79 | prefix=f"{prefix}.down_proj", |
| 80 | ) |
| 81 | self.act_fn = SiluAndMul() |
| 82 | |
| 83 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 84 | gate_up, _ = self.gate_up_proj(x) |
| 85 | x = self.act_fn(gate_up) |
| 86 | x, _ = self.down_proj(x) |
| 87 | return x |
| 88 | |
| 89 | |
| 90 | class HrmTextAttention(nn.Module): |
| 91 | """One self-attention block; weights shared across recurrence steps. |
| 92 | |
| 93 | HF transformers writes a single fused ``attn.gqkv_proj.weight`` on |
| 94 | disk (per ``transformers/conversion_mapping.py`` ``"hrm_text"`` |
| 95 | mapping; rows are concatenated as ``[gate | q | k | v]`` along |
| 96 | ``dim=0``). We mirror that on the model side with a single |
| 97 | ``MergedColumnParallelLinear`` whose four equal output partitions |
| 98 | are sharded along the head axis under TP; its weight loader |
| 99 | auto-splits the fused tensor (same path used by Phi-3's fused |
| 100 | gate_up_proj). HF's runtime config currently hardcodes MHA |
| 101 | (``num_key_value_groups=1``); GQA would require ``QKVParallelLinear`` |
| 102 | semantics for q/k/v shard replication and is left for a follow-up |
| 103 | if/when HF adds it. |
| 104 | |
| 105 | Holds: |
| 106 | - parameters: gqkv_proj, o_proj, rotary_emb (shared across cycles). |
| 107 | - ``attn_per_step``: a ``nn.ModuleDict`` keyed by recurrence step |
| 108 | (as a string), each value an ``EncoderOnlyAttention`` (with |
| 109 | ``attn_type=DECODER`` so the KV cache is allocated; the |
| 110 | ``EncoderOnlyAttention`` wrapper sets ``causal=False`` on every |
| 111 | metadata build). The L stack steps are |
| 112 | ``[high_cycle_idx*(L_cycles+1)+low_cycle_idx]`` and the H stack |
| 113 | steps are ``[high_cycle_idx*(L_cycles+1)+L_cycles]``; the two |
| 114 | ranges are disjoint so each instance registers a unique vLLM |
| 115 | ``layer_name`` |
| 116 | (``model.{H,L}_module.layers.{global_idx}.self_attn``) and gets |
| 117 | its own KV cache slot. The global layer index per recurrence step |
| 118 | is ``step * num_layers_per_stack + layer_idx_in_stack``, matching |
| 119 | the HF transformers ``cycle_offset`` formula in |
| 120 | ``modeling_hrm_text.py``. |
| 121 | """ |
| 122 | |
| 123 | def __init__( |
| 124 | self, |
| 125 | config: PreTrainedConfig, |
| 126 | layer_idx_in_stack: int, |
| 127 | stack_kind: Literal["L", "H"], |
| 128 | cache_config: CacheConfig | None = None, |
| 129 | quant_config: QuantizationConfig | None = None, |
| 130 | prefix: str = "", |
| 131 | ) -> None: |
| 132 | super().__init__() |
| 133 | self.config = config |
| 134 | self.hidden_size = config.hidden_size |
| 135 | tp_size = get_tensor_model_parallel_world_size() |
| 136 | |
| 137 | self.total_num_heads = config.num_attention_heads |
| 138 | assert self.total_num_heads % tp_size == 0, ( |
| 139 | f"num_attention_heads={self.total_num_heads} must be divisible " |
| 140 | f"by tp_size={tp_size}" |
| 141 | ) |
| 142 | # HF main hardcodes MHA (num_key_value_groups=1). We follow. |
| 143 | self.total_num_kv_heads = config.num_attention_heads |
| 144 | self.num_heads = self.total_num_heads // tp_size |
| 145 | self.num_kv_heads = self.total_num_kv_heads // tp_size |
| 146 | self.head_dim = getattr( |
| 147 | config, "head_dim", self.hidden_size // self.total_num_heads |
| 148 | ) |
| 149 | self.q_size = self.num_heads * self.head_dim |
| 150 | self.kv_size = self.num_kv_heads * self.head_dim |
| 151 | self.scaling = self.head_dim**-0.5 |
| 152 | |
| 153 | bias = getattr(config, "attention_bias", False) |
| 154 | |
| 155 | # gqkv_proj: 4-way fused [gate | q | k | v] matching the on-disk |
| 156 | # `attn.gqkv_proj.weight` row layout. MergedColumnParallelLinear's |
| 157 | # weight_loader auto-splits the fused disk tensor along the output |
| 158 | # dim by `output_sizes` (Phi-3's fused gate_up_proj path). MHA |
| 159 | # only: GQA (num_kv_heads != num_heads) would need |
| 160 | # QKVParallelLinear semantics for q/k/v shard replication. |
| 161 | per_head_size = self.total_num_heads * self.head_dim |
| 162 | self.gqkv_proj = MergedColumnParallelLinear( |
| 163 | input_size=self.hidden_size, |
| 164 | output_sizes=[per_head_size] * 4, |
| 165 | bias=bias, |
| 166 | quant_config=quant_config, |
| 167 | prefix=f"{prefix}.gqkv_proj", |
| 168 | ) |
| 169 | self.o_proj = RowParallelLinear( |
| 170 | input_size=self.total_num_heads * self.head_dim, |
| 171 | output_size=self.hidden_size, |
| 172 | bias=bias, |
| 173 | quant_config=quant_config, |
| 174 | prefix=f"{prefix}.o_proj", |
| 175 | ) |
| 176 | |
| 177 | # vllm get_rope accepts ``rope_parameters`` directly, matching |
| 178 | # the dict-shaped HF config field. |
| 179 | self.rotary_emb = get_rope( |
| 180 | head_size=self.head_dim, |
| 181 | max_position=config.max_position_embeddings, |
| 182 | rope_parameters=config.rope_parameters, |
| 183 | ) |
| 184 | |
| 185 | # Create one Attention instance per recurrence step actually used |
| 186 | # by this stack. L runs at steps {h*(L+1)+l : 0 <= l < L_cycles}, |
| 187 | # H at steps {h*(L+1)+L : 0 <= h < H_cycles}; the sets are |
| 188 | # disjoint, so one global index per (step, layer_in_stack) gives |
| 189 | # each Attention its own ``layer_name`` and KV cache slot. |
| 190 | H_cycles = config.H_cycles |
| 191 | L_cycles = config.L_cycles |
| 192 | num_layers_per_stack = config.num_layers_per_stack |
| 193 | if stack_kind == "L": |
| 194 | steps_used = [ |
| 195 | high_cycle_idx * (L_cycles + 1) + low_cycle_idx |
| 196 | for high_cycle_idx in range(H_cycles) |
| 197 | for low_cycle_idx in range(L_cycles) |
| 198 | ] |
| 199 | else: # "H" |
| 200 | steps_used = [ |
| 201 | high_cycle_idx * (L_cycles + 1) + L_cycles |
| 202 | for high_cycle_idx in range(H_cycles) |
| 203 | ] |
| 204 | |
| 205 | # `PrefillPrefixLMAttention` forces `causal=False` on every metadata |
| 206 | # build, so the prompt attends bidirectionally during prefill (matching |
| 207 | # the HRM-Text training distribution), while `attn_type=DECODER` keeps |
| 208 | # the KV cache allocation needed by the recurrent forward. At |
| 209 | # single-token decode `causal=False` is a no-op. See |
| 210 | # `PrefillPrefixLMAttention`. |
| 211 | self.attn_per_step = nn.ModuleDict() |
| 212 | for step in steps_used: |
| 213 | global_idx = step * num_layers_per_stack + layer_idx_in_stack |
| 214 | unique_prefix = prefix.replace( |
| 215 | f"layers.{layer_idx_in_stack}", f"layers.{global_idx}" |
| 216 | ) |
| 217 | self.attn_per_step[str(step)] = PrefillPrefixLMAttention( |
| 218 | num_heads=self.num_heads, |
| 219 | head_size=self.head_dim, |
| 220 | scale=self.scaling, |
| 221 | num_kv_heads=self.num_kv_heads, |
| 222 | cache_config=cache_config, |
| 223 | quant_config=quant_config, |
| 224 | prefix=f"{unique_prefix}.attn", |
| 225 | ) |
| 226 | |
| 227 | def forward( |
| 228 | self, |
| 229 | positions: torch.Tensor, |
| 230 | hidden_states: torch.Tensor, |
| 231 | current_step: int, |
| 232 | ) -> torch.Tensor: |
| 233 | gqkv, _ = self.gqkv_proj(hidden_states) |
| 234 | g, q, k, v = gqkv.split( |
| 235 | [self.q_size, self.q_size, self.kv_size, self.kv_size], dim=-1 |
| 236 | ) |
| 237 | q, k = self.rotary_emb(positions, q, k) |
| 238 | attn_out = self.attn_per_step[str(current_step)](q, k, v) |
| 239 | # Sigmoid gate. Shapes: attn_out is (..., q_size); g is (..., q_size). |
| 240 | attn_out = torch.sigmoid(g) * attn_out |
| 241 | out, _ = self.o_proj(attn_out) |
| 242 | return out |
| 243 | |
| 244 | |
| 245 | class HrmTextDecoderLayer(nn.Module): |
| 246 | def __init__( |
| 247 | self, |
| 248 | config: PreTrainedConfig, |
| 249 | layer_idx_in_stack: int, |
| 250 | stack_kind: Literal["L", "H"], |
| 251 | cache_config: CacheConfig | None = None, |
| 252 | quant_config: QuantizationConfig | None = None, |
| 253 | prefix: str = "", |
| 254 | ) -> None: |
| 255 | super().__init__() |
| 256 | self.hidden_size = config.hidden_size |
| 257 | # Attribute name `self_attn` matches HF's model class. The on-disk |
| 258 | # `attn.{gqkv_proj,o_proj}.weight` keys are renamed to |
| 259 | # `self_attn.{gqkv_proj,o_proj}.weight` by the `WeightsMapper` in |
| 260 | # `HrmTextForCausalLM` so vLLM's standard `AutoWeightsLoader` |
| 261 | # handles the rest. |
| 262 | self.self_attn = HrmTextAttention( |
| 263 | config=config, |
| 264 | layer_idx_in_stack=layer_idx_in_stack, |
| 265 | stack_kind=stack_kind, |
| 266 | cache_config=cache_config, |
| 267 | quant_config=quant_config, |
| 268 | prefix=f"{prefix}.self_attn", |
| 269 | ) |
| 270 | self.mlp = HrmTextMLP( |
| 271 | hidden_size=config.hidden_size, |
| 272 | intermediate_size=config.intermediate_size, |
| 273 | hidden_act=config.hidden_act, |
| 274 | bias=getattr(config, "mlp_bias", False), |
| 275 | quant_config=quant_config, |
| 276 | prefix=f"{prefix}.mlp", |
| 277 | ) |
| 278 | # Parameterless RMSNorm (HF main: HrmTextRMSNorm has no weight). |
| 279 | self.input_layernorm = RMSNorm( |
| 280 | config.hidden_size, eps=config.rms_norm_eps, has_weight=False |
| 281 | ) |
| 282 | self.post_attention_layernorm = RMSNorm( |
| 283 | config.hidden_size, eps=config.rms_norm_eps, has_weight=False |
| 284 | ) |
| 285 | |
| 286 | def forward( |
| 287 | self, |
| 288 | positions: torch.Tensor, |
| 289 | hidden_states: torch.Tensor, |
| 290 | current_step: int, |
| 291 | ) -> torch.Tensor: |
| 292 | residual = hidden_states |
| 293 | hidden_states = self.input_layernorm(hidden_states) |
| 294 | hidden_states = self.self_attn( |
| 295 | positions=positions, |
| 296 | hidden_states=hidden_states, |
| 297 | current_step=current_step, |
| 298 | ) |
| 299 | hidden_states = residual + hidden_states |
| 300 | |
| 301 | residual = hidden_states |
| 302 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 303 | hidden_states = self.mlp(hidden_states) |
| 304 | hidden_states = residual + hidden_states |
| 305 | return hidden_states |
| 306 | |
| 307 | |
| 308 | class HrmTextStack(nn.Module): |
| 309 | """A single transformer stack — used twice (H and L).""" |
| 310 | |
| 311 | def __init__( |
| 312 | self, |
| 313 | config: PreTrainedConfig, |
| 314 | stack_kind: Literal["L", "H"], |
| 315 | cache_config: CacheConfig | None = None, |
| 316 | quant_config: QuantizationConfig | None = None, |
| 317 | prefix: str = "", |
| 318 | ) -> None: |
| 319 | super().__init__() |
| 320 | self.layers = nn.ModuleList( |
| 321 | [ |
| 322 | HrmTextDecoderLayer( |
| 323 | config=config, |
| 324 | layer_idx_in_stack=i, |
| 325 | stack_kind=stack_kind, |
| 326 | cache_config=cache_config, |
| 327 | quant_config=quant_config, |
| 328 | prefix=f"{prefix}.layers.{i}", |
| 329 | ) |
| 330 | for i in range(config.num_layers_per_stack) |
| 331 | ] |
| 332 | ) |
| 333 | self.final_norm = RMSNorm( |
| 334 | config.hidden_size, eps=config.rms_norm_eps, has_weight=False |
| 335 | ) |
| 336 | |
| 337 | def forward( |
| 338 | self, |
| 339 | positions: torch.Tensor, |
| 340 | hidden_states: torch.Tensor, |
| 341 | current_step_base: int, |
| 342 | ) -> torch.Tensor: |
| 343 | for layer in self.layers: |
| 344 | hidden_states = layer( |
| 345 | positions=positions, |
| 346 | hidden_states=hidden_states, |
| 347 | current_step=current_step_base, |
| 348 | ) |
| 349 | return self.final_norm(hidden_states) |
| 350 | |
| 351 | |
| 352 | @support_torch_compile |
| 353 | class HrmTextModel(nn.Module): |
| 354 | """Hierarchical recurrent transformer body. |
| 355 | |
| 356 | Forward (matches HF main exactly, |
| 357 | src/transformers/models/hrm_text/modeling_hrm_text.py:495-547): |
| 358 | |
| 359 | hidden_states_high_cycle = embed(input_ids) * embedding_scale |
| 360 | hidden_states_low_cycle = z_L_init.expand_as(hidden_states_high_cycle) |
| 361 | for high_cycle_idx in range(H_cycles): |
| 362 | for low_cycle_idx in range(L_cycles): |
| 363 | step = high_cycle_idx * (L_cycles + 1) + low_cycle_idx |
| 364 | hidden_states_low_cycle = L_module( |
| 365 | hidden_states_low_cycle + hidden_states_high_cycle, |
| 366 | current_step=step, |
| 367 | ) |
| 368 | step = high_cycle_idx * (L_cycles + 1) + L_cycles |
| 369 | hidden_states_high_cycle = H_module( |
| 370 | hidden_states_high_cycle + hidden_states_low_cycle, |
| 371 | current_step=step, |
| 372 | ) |
| 373 | return hidden_states_high_cycle |
| 374 | """ |
| 375 | |
| 376 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 377 | super().__init__() |
| 378 | config = vllm_config.model_config.hf_config |
| 379 | cache_config = vllm_config.cache_config |
| 380 | quant_config = vllm_config.quant_config |
| 381 | self.config = config |
| 382 | |
| 383 | self.embed_tokens = VocabParallelEmbedding( |
| 384 | config.vocab_size, |
| 385 | config.hidden_size, |
| 386 | quant_config=quant_config, |
| 387 | prefix=f"{prefix}.embed_tokens", |
| 388 | ) |
| 389 | self.L_module = HrmTextStack( |
| 390 | config=config, |
| 391 | stack_kind="L", |
| 392 | cache_config=cache_config, |
| 393 | quant_config=quant_config, |
| 394 | prefix=f"{prefix}.L_module", |
| 395 | ) |
| 396 | self.H_module = HrmTextStack( |
| 397 | config=config, |
| 398 | stack_kind="H", |
| 399 | cache_config=cache_config, |
| 400 | quant_config=quant_config, |
| 401 | prefix=f"{prefix}.H_module", |
| 402 | ) |
| 403 | # Frozen learned initial L state. HF inits to zeros and sets |
| 404 | # requires_grad_(False); for inference we just load the tensor. |
| 405 | self.z_L_init = nn.Parameter( |
| 406 | torch.zeros(config.hidden_size), requires_grad=False |
| 407 | ) |
| 408 | |
| 409 | # Embedding scale: HF uses config.embedding_scale (default |
| 410 | # 1 / initializer_range = 50.0 when initializer_range=0.02). NOT |
| 411 | # sqrt(hidden_size) like Gemma. |
| 412 | self.embedding_scale = getattr(config, "embedding_scale", None) |
| 413 | if self.embedding_scale is None: |
| 414 | init_range = getattr(config, "initializer_range", 0.02) |
| 415 | self.embedding_scale = 1.0 / init_range |
| 416 | |
| 417 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 418 | return self.embed_tokens(input_ids) * self.embedding_scale |
| 419 | |
| 420 | def forward( |
| 421 | self, |
| 422 | input_ids: torch.Tensor | None, |
| 423 | positions: torch.Tensor, |
| 424 | intermediate_tensors: IntermediateTensors | None = None, |
| 425 | inputs_embeds: torch.Tensor | None = None, |
| 426 | ) -> torch.Tensor: |
| 427 | if inputs_embeds is None: |
| 428 | assert input_ids is not None |
| 429 | inputs_embeds = self.embed_input_ids(input_ids) |
| 430 | |
| 431 | hidden_states_high_cycle = inputs_embeds |
| 432 | hidden_states_low_cycle = self.z_L_init.to( |
| 433 | dtype=hidden_states_high_cycle.dtype, |
| 434 | device=hidden_states_high_cycle.device, |
| 435 | ).expand_as(hidden_states_high_cycle) |
| 436 | |
| 437 | H_cycles = self.config.H_cycles |
| 438 | L_cycles = self.config.L_cycles |
| 439 | for high_cycle_idx in range(H_cycles): |
| 440 | for low_cycle_idx in range(L_cycles): |
| 441 | step = high_cycle_idx * (L_cycles + 1) + low_cycle_idx |
| 442 | hidden_states_low_cycle = self.L_module( |
| 443 | positions=positions, |
| 444 | hidden_states=hidden_states_low_cycle + hidden_states_high_cycle, |
| 445 | current_step_base=step, |
| 446 | ) |
| 447 | step = high_cycle_idx * (L_cycles + 1) + L_cycles |
| 448 | hidden_states_high_cycle = self.H_module( |
| 449 | positions=positions, |
| 450 | hidden_states=hidden_states_high_cycle + hidden_states_low_cycle, |
| 451 | current_step_base=step, |
| 452 | ) |
| 453 | |
| 454 | return hidden_states_high_cycle |
| 455 | |
| 456 | |
| 457 | class HrmTextForCausalLM(nn.Module): |
| 458 | """Hierarchical Reasoning Model — Text variant, causal LM. |
| 459 | |
| 460 | Reference: src/transformers/models/hrm_text/modeling_hrm_text.py |
| 461 | """ |
| 462 | |
| 463 | # On-disk weight key remap: HF stores attention weights as |
| 464 | # `attn.{gqkv_proj,o_proj}.weight`; our model uses `self_attn.*` |
| 465 | # (matching HF's runtime model class). Both `gqkv_proj` (4-way fused |
| 466 | # gate/q/k/v) and `mlp.gate_up_proj` (2-way fused gate/up) are loaded |
| 467 | # directly via MergedColumnParallelLinear's fused-on-disk path; no |
| 468 | # packed_modules_mapping entries are needed. |
| 469 | hf_to_vllm_mapper = WeightsMapper( |
| 470 | orig_to_new_substr={".attn.": ".self_attn."}, |
| 471 | ) |
| 472 | |
| 473 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 474 | super().__init__() |
| 475 | config = vllm_config.model_config.hf_config |
| 476 | quant_config = vllm_config.quant_config |
| 477 | self.config = config |
| 478 | self.quant_config = quant_config |
| 479 | |
| 480 | if vllm_config.parallel_config.pipeline_parallel_size > 1: |
| 481 | raise ValueError( |
| 482 | "HrmTextForCausalLM does not support pipeline parallelism." |
| 483 | ) |
| 484 | |
| 485 | self.model = HrmTextModel( |
| 486 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 487 | ) |
| 488 | |
| 489 | self.lm_head = ParallelLMHead( |
| 490 | config.vocab_size, |
| 491 | config.hidden_size, |
| 492 | quant_config=quant_config, |
| 493 | prefix=maybe_prefix(prefix, "lm_head"), |
| 494 | ) |
| 495 | if config.tie_word_embeddings: |
| 496 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 497 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 498 | |
| 499 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 500 | return self.model.embed_input_ids(input_ids) |
| 501 | |
| 502 | def forward( |
| 503 | self, |
| 504 | input_ids: torch.Tensor | None, |
| 505 | positions: torch.Tensor, |
| 506 | intermediate_tensors: IntermediateTensors | None = None, |
| 507 | inputs_embeds: torch.Tensor | None = None, |
| 508 | ) -> torch.Tensor: |
| 509 | return self.model( |
| 510 | input_ids=input_ids, |
| 511 | positions=positions, |
| 512 | intermediate_tensors=intermediate_tensors, |
| 513 | inputs_embeds=inputs_embeds, |
| 514 | ) |
| 515 | |
| 516 | def compute_logits( |
| 517 | self, |
| 518 | hidden_states: torch.Tensor, |
| 519 | ) -> torch.Tensor | None: |
| 520 | return self.logits_processor(self.lm_head, hidden_states) |
| 521 | |
| 522 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 523 | loader = AutoWeightsLoader(self) |
| 524 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |