| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2025 The Baidu team. |
| 5 | # Copyright 2023 The vLLM team. |
| 6 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 7 | # |
| 8 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 9 | # and OPT implementations in this library. It has been modified from its |
| 10 | # original forms to accommodate minor architectural differences compared |
| 11 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 12 | # |
| 13 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 14 | # you may not use this file except in compliance with the License. |
| 15 | # You may obtain a copy of the License at |
| 16 | # |
| 17 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 18 | # |
| 19 | # Unless required by applicable law or agreed to in writing, software |
| 20 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 21 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 22 | # See the License for the specific language governing permissions and |
| 23 | # limitations under the License. |
| 24 | """Inference-only Ernie-MTP model.""" |
| 25 | |
| 26 | from collections.abc import Iterable |
| 27 | |
| 28 | import torch |
| 29 | import torch.nn as nn |
| 30 | |
| 31 | from vllm.config import VllmConfig |
| 32 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 33 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 34 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 35 | ParallelLMHead, |
| 36 | VocabParallelEmbedding, |
| 37 | ) |
| 38 | from vllm.sequence import IntermediateTensors |
| 39 | |
| 40 | from .llama import LlamaDecoderLayer |
| 41 | from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix |
| 42 | |
| 43 | |
| 44 | class ErnieMultiTokenPredictorLayer(nn.Module): |
| 45 | def __init__( |
| 46 | self, |
| 47 | vllm_config: VllmConfig, |
| 48 | prefix: str, |
| 49 | ) -> None: |
| 50 | super().__init__() |
| 51 | config = vllm_config.model_config.hf_config |
| 52 | |
| 53 | self.mtp_emb_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 54 | self.mtp_hidden_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 55 | self.mtp_linear_proj = nn.Linear( |
| 56 | config.hidden_size * 2, config.hidden_size, bias=False |
| 57 | ) |
| 58 | self.mtp_block = LlamaDecoderLayer(vllm_config, prefix) |
| 59 | |
| 60 | def forward( |
| 61 | self, |
| 62 | inputs_embeds: torch.Tensor, |
| 63 | positions: torch.Tensor, |
| 64 | previous_hidden_states: torch.Tensor, |
| 65 | spec_step_index: int = 0, |
| 66 | ) -> torch.Tensor: |
| 67 | assert inputs_embeds is not None |
| 68 | # masking inputs at position 0, as not needed by MTP |
| 69 | inputs_embeds[positions == 0] = 0 |
| 70 | |
| 71 | inputs_embeds = self.mtp_emb_norm(inputs_embeds) |
| 72 | previous_hidden_states = self.mtp_hidden_norm(previous_hidden_states) |
| 73 | |
| 74 | hidden_states = self.mtp_linear_proj( |
| 75 | torch.cat([inputs_embeds, previous_hidden_states], dim=-1) |
| 76 | ) |
| 77 | |
| 78 | hidden_states, residual = self.mtp_block( |
| 79 | positions=positions, hidden_states=hidden_states, residual=None |
| 80 | ) |
| 81 | hidden_states = residual + hidden_states |
| 82 | |
| 83 | return hidden_states |
| 84 | |
| 85 | |
| 86 | class ErnieMultiTokenPredictor(nn.Module): |
| 87 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 88 | super().__init__() |
| 89 | |
| 90 | config = vllm_config.model_config.hf_config |
| 91 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 92 | self.num_mtp_layers = config.num_nextn_predict_layers |
| 93 | # to map the exact layer index from weights |
| 94 | self.layers = torch.nn.ModuleDict( |
| 95 | { |
| 96 | str(idx): ErnieMultiTokenPredictorLayer( |
| 97 | vllm_config, |
| 98 | f"{prefix}.layers.{idx}", |
| 99 | ) |
| 100 | for idx in range( |
| 101 | self.mtp_start_layer_idx, |
| 102 | self.mtp_start_layer_idx + self.num_mtp_layers, |
| 103 | ) |
| 104 | } |
| 105 | ) |
| 106 | self.embed_tokens = VocabParallelEmbedding( |
| 107 | config.vocab_size, |
| 108 | config.hidden_size, |
| 109 | ) |
| 110 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 111 | |
| 112 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 113 | return self.embed_tokens(input_ids) |
| 114 | |
| 115 | def forward( |
| 116 | self, |
| 117 | input_ids: torch.Tensor, |
| 118 | positions: torch.Tensor, |
| 119 | previous_hidden_states: torch.Tensor, |
| 120 | inputs_embeds: torch.Tensor | None = None, |
| 121 | spec_step_idx: int = 0, |
| 122 | ) -> torch.Tensor: |
| 123 | if inputs_embeds is None: |
| 124 | inputs_embeds = self.embed_tokens(input_ids) |
| 125 | return self.layers[str(self.mtp_start_layer_idx + spec_step_idx)]( |
| 126 | inputs_embeds, |
| 127 | positions, |
| 128 | previous_hidden_states, |
| 129 | spec_step_idx, |
| 130 | ) |
| 131 | |
| 132 | def compute_logits( |
| 133 | self, |
| 134 | hidden_states: torch.Tensor, |
| 135 | lm_head: ParallelLMHead, |
| 136 | spec_step_idx: int = 0, |
| 137 | ) -> torch.Tensor: |
| 138 | self.layers[str(self.mtp_start_layer_idx + spec_step_idx)] |
| 139 | logits = self.logits_processor(lm_head, hidden_states) |
| 140 | return logits |
| 141 | |
| 142 | |
| 143 | class ErnieMTP(nn.Module): |
| 144 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 145 | super().__init__() |
| 146 | |
| 147 | self.config = vllm_config.model_config.hf_config |
| 148 | # MTP weights are stored under a flat `mtp_*.0.` block in the |
| 149 | # checkpoint; rewrite them into `model.layers.{spec_layer}.*`. |
| 150 | spec_layer = self.config.num_hidden_layers |
| 151 | self.hf_to_vllm_mapper = WeightsMapper( |
| 152 | orig_to_new_substr={ |
| 153 | "model.mtp_emb_norm.0.": f"model.layers.{spec_layer}.mtp_emb_norm.", |
| 154 | "model.mtp_hidden_norm.0.": ( |
| 155 | f"model.layers.{spec_layer}.mtp_hidden_norm." |
| 156 | ), |
| 157 | "model.mtp_linear_proj.0.": ( |
| 158 | f"model.layers.{spec_layer}.mtp_linear_proj." |
| 159 | ), |
| 160 | "model.mtp_block.0.": f"model.layers.{spec_layer}.mtp_block.", |
| 161 | }, |
| 162 | orig_to_new_stacked={ |
| 163 | ".q_proj": (".qkv_proj", "q"), |
| 164 | ".k_proj": (".qkv_proj", "k"), |
| 165 | ".v_proj": (".qkv_proj", "v"), |
| 166 | ".gate_proj": (".gate_up_proj", 0), |
| 167 | ".up_proj": (".gate_up_proj", 1), |
| 168 | }, |
| 169 | ) |
| 170 | self.model = ErnieMultiTokenPredictor( |
| 171 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 172 | ) |
| 173 | self.lm_head = ParallelLMHead( |
| 174 | self.config.vocab_size, |
| 175 | self.config.hidden_size, |
| 176 | prefix=maybe_prefix(prefix, "lm_head"), |
| 177 | ) |
| 178 | |
| 179 | if self.config.tie_word_embeddings: |
| 180 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 181 | |
| 182 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 183 | return self.model.embed_input_ids(input_ids) |
| 184 | |
| 185 | def forward( |
| 186 | self, |
| 187 | input_ids: torch.Tensor | None, |
| 188 | positions: torch.Tensor, |
| 189 | hidden_states: torch.Tensor, |
| 190 | intermediate_tensors: IntermediateTensors | None = None, |
| 191 | inputs_embeds: torch.Tensor | None = None, |
| 192 | spec_step_idx: int = 0, |
| 193 | ) -> torch.Tensor: |
| 194 | assert spec_step_idx == 0, "ernie_mtp only support predict one token" |
| 195 | hidden_states = self.model( |
| 196 | input_ids, positions, hidden_states, inputs_embeds, spec_step_idx |
| 197 | ) |
| 198 | return hidden_states |
| 199 | |
| 200 | def compute_logits( |
| 201 | self, |
| 202 | hidden_states: torch.Tensor, |
| 203 | spec_step_idx: int = 0, |
| 204 | ) -> torch.Tensor | None: |
| 205 | return self.model.compute_logits(hidden_states, self.lm_head, spec_step_idx) |
| 206 | |
| 207 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 208 | # Checkpoint bundles the full base model; only MTP, embed_tokens and |
| 209 | # lm_head weights belong to this module. |
| 210 | def _filter( |
| 211 | weights: Iterable[tuple[str, torch.Tensor]], |
| 212 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 213 | for name, weight in weights: |
| 214 | if any(k in name for k in ("mtp", "embed_tokens", "lm_head")): |
| 215 | yield name, weight |
| 216 | |
| 217 | loader = AutoWeightsLoader(self) |
| 218 | return loader.load_weights(_filter(weights), mapper=self.hf_to_vllm_mapper) |