| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2024 The vLLM team. |
| 5 | # |
| 6 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 7 | # you may not use this file except in compliance with the License. |
| 8 | # You may obtain a copy of the License at |
| 9 | # |
| 10 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 11 | # |
| 12 | # Unless required by applicable law or agreed to in writing, software |
| 13 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 14 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 15 | # See the License for the specific language governing permissions and |
| 16 | # limitations under the License. |
| 17 | """Transformers modeling backend mixin for causal language models.""" |
| 18 | |
| 19 | from collections.abc import Iterable |
| 20 | from typing import TYPE_CHECKING |
| 21 | |
| 22 | import torch |
| 23 | |
| 24 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 25 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 26 | ParallelLMHead, |
| 27 | VocabParallelEmbedding, |
| 28 | ) |
| 29 | from vllm.model_executor.models.interfaces_base import VllmModelForTextGeneration |
| 30 | from vllm.model_executor.models.utils import PPMissingLayer, maybe_prefix |
| 31 | |
| 32 | from .base import Base |
| 33 | |
| 34 | if TYPE_CHECKING: |
| 35 | from vllm.config import VllmConfig |
| 36 | |
| 37 | |
| 38 | class CausalMixin(VllmModelForTextGeneration, Base): |
| 39 | def __init__(self, *, vllm_config: "VllmConfig", prefix: str = ""): |
| 40 | # Skip VllmModelForTextGeneration.__init__ and call the next class in MRO |
| 41 | super(VllmModelForTextGeneration, self).__init__( |
| 42 | vllm_config=vllm_config, prefix=prefix |
| 43 | ) |
| 44 | |
| 45 | tie_word_embeddings = self._get_tie_word_embeddings() |
| 46 | |
| 47 | if self.pp_group.is_last_rank: |
| 48 | self.lm_head = ParallelLMHead( |
| 49 | self.text_config.vocab_size, |
| 50 | self.text_config.hidden_size, |
| 51 | quant_config=self.quant_config, |
| 52 | prefix=maybe_prefix(prefix, "lm_head"), |
| 53 | ) |
| 54 | if tie_word_embeddings: |
| 55 | for module in self.model.get_input_embeddings().modules(): |
| 56 | if isinstance(module, VocabParallelEmbedding): |
| 57 | self.lm_head = self.lm_head.tie_weights(module) |
| 58 | break |
| 59 | |
| 60 | self.logits_processor = LogitsProcessor( |
| 61 | self.text_config.vocab_size, |
| 62 | scale=getattr(self.text_config, "logit_scale", 1.0), |
| 63 | soft_cap=getattr(self.text_config, "final_logit_softcapping", None), |
| 64 | ) |
| 65 | else: |
| 66 | self.lm_head = PPMissingLayer() |
| 67 | |
| 68 | def load_weights(self, weights: Iterable[tuple[str, "torch.Tensor"]]) -> set[str]: |
| 69 | """A thin wrapper around `Base.load_weights` to handle the lm_head bias.""" |
| 70 | lm_head_bias = set() |
| 71 | load_device = self.vllm_config.load_config.device or self.device_config.device |
| 72 | |
| 73 | def auto_load_lm_head_bias(weights): |
| 74 | for name, weight in weights: |
| 75 | if name.endswith("lm_head.bias") and self.pp_group.is_last_rank: |
| 76 | # load_weights is called outside vLLM's device context, |
| 77 | # so we must enter it again when registering the bias. |
| 78 | with torch.device(load_device): |
| 79 | self.lm_head._register_bias() |
| 80 | self.lm_head.bias.weight_loader(self.lm_head.bias, weight) |
| 81 | lm_head_bias.add(name) |
| 82 | else: |
| 83 | yield name, weight |
| 84 | |
| 85 | return super().load_weights(auto_load_lm_head_bias(weights)) | lm_head_bias |
| 86 | |
| 87 | def compute_logits(self, hidden_states: "torch.Tensor") -> "torch.Tensor | None": |
| 88 | logits = self.logits_processor(self.lm_head, hidden_states, self.lm_head.bias) |
| 89 | return logits |