| 1 | # |
| 2 | # Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved. |
| 3 | # Copyright 2023 The vLLM team. |
| 4 | # |
| 5 | # This file is a part of the vllm-ascend project. |
| 6 | # |
| 7 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 8 | # you may not use this file except in compliance with the License. |
| 9 | # You may obtain a copy of the License at |
| 10 | # |
| 11 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 12 | # |
| 13 | # Unless required by applicable law or agreed to in writing, software |
| 14 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 15 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 16 | # See the License for the specific language governing permissions and |
| 17 | # limitations under the License. |
| 18 | |
| 19 | # Adapted from |
| 20 | # https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/model_executor/models/deepseek_mtp.py |
| 21 | from collections.abc import Iterable |
| 22 | |
| 23 | import torch |
| 24 | import torch.nn as nn |
| 25 | |
| 26 | from vllm.compilation.decorators import support_torch_compile |
| 27 | from vllm.config import VllmConfig |
| 28 | |
| 29 | # SPDX-License-Identifier: Apache-2.0 |
| 30 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 31 | from vllm.model_executor.layers.fused_moe import ( |
| 32 | fused_moe_make_expert_params_mapping, |
| 33 | ) |
| 34 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 35 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 36 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 37 | VocabParallelEmbedding, |
| 38 | ) |
| 39 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 40 | from vllm.model_executor.models.deepseek_mtp import ( |
| 41 | DeepSeekMultiTokenPredictor, |
| 42 | DeepSeekMultiTokenPredictorLayer, |
| 43 | SharedHead, |
| 44 | ) |
| 45 | from vllm.model_executor.models.utils import maybe_prefix |
| 46 | from vllm.sequence import IntermediateTensors |
| 47 | |
| 48 | from .openpangu import OpenPanguDecoderLayer |
| 49 | |
| 50 | |
| 51 | class OpenPanguMultiTokenPredictorLayer(DeepSeekMultiTokenPredictorLayer): |
| 52 | def __init__(self, vllm_config: VllmConfig, prefix: str) -> None: |
| 53 | nn.Module.__init__(self) |
| 54 | |
| 55 | speculative_config = vllm_config.speculative_config |
| 56 | assert speculative_config is not None |
| 57 | config = speculative_config.draft_model_config.hf_config |
| 58 | self.config = config |
| 59 | quant_config = vllm_config.quant_config |
| 60 | |
| 61 | self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 62 | self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 63 | self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False) |
| 64 | self.shared_head = SharedHead( |
| 65 | config=config, |
| 66 | quant_config=quant_config, |
| 67 | prefix=maybe_prefix(prefix, "shared_head"), |
| 68 | ) |
| 69 | self.mtp_block = OpenPanguDecoderLayer(config, prefix, vllm_config) |
| 70 | |
| 71 | |
| 72 | class OpenPanguMultiTokenPredictor(DeepSeekMultiTokenPredictor): |
| 73 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 74 | nn.Module.__init__(self) |
| 75 | config = vllm_config.model_config.hf_config |
| 76 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 77 | self.num_mtp_layers = config.num_nextn_predict_layers |
| 78 | # to map the exact layer index from weights |
| 79 | self.layers = torch.nn.ModuleDict( |
| 80 | { |
| 81 | str(idx): OpenPanguMultiTokenPredictorLayer( |
| 82 | vllm_config, f"{prefix}.layers.{idx}" |
| 83 | ) |
| 84 | for idx in range( |
| 85 | self.mtp_start_layer_idx, |
| 86 | self.mtp_start_layer_idx + self.num_mtp_layers, |
| 87 | ) |
| 88 | } |
| 89 | ) |
| 90 | self.embed_tokens = VocabParallelEmbedding( |
| 91 | config.vocab_size, |
| 92 | config.hidden_size, |
| 93 | ) |
| 94 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 95 | |
| 96 | |
| 97 | @support_torch_compile |
| 98 | class OpenPanguMTP(nn.Module): |
| 99 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 100 | super().__init__() |
| 101 | self.config = vllm_config.model_config.hf_config |
| 102 | self.model = OpenPanguMultiTokenPredictor( |
| 103 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 104 | ) |
| 105 | |
| 106 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 107 | return self.model.embed_input_ids(input_ids) |
| 108 | |
| 109 | def forward( |
| 110 | self, |
| 111 | input_ids: torch.Tensor | None, |
| 112 | positions: torch.Tensor, |
| 113 | hidden_states: torch.Tensor, |
| 114 | intermediate_tensors: IntermediateTensors | None = None, |
| 115 | inputs_embeds: torch.Tensor | None = None, |
| 116 | spec_step_idx: int = 0, |
| 117 | ) -> torch.Tensor: |
| 118 | hidden_states = self.model( |
| 119 | input_ids, |
| 120 | positions, |
| 121 | hidden_states, |
| 122 | inputs_embeds, |
| 123 | spec_step_idx, |
| 124 | ) |
| 125 | return hidden_states |
| 126 | |
| 127 | def compute_logits( |
| 128 | self, |
| 129 | hidden_states: torch.Tensor, |
| 130 | spec_step_idx: int = 0, |
| 131 | ) -> torch.Tensor | None: |
| 132 | return self.model.compute_logits(hidden_states, spec_step_idx) |
| 133 | |
| 134 | def get_spec_layer(self, name): |
| 135 | if ( |
| 136 | "layers" in name |
| 137 | and hasattr(self.config, "num_nextn_predict_layers") |
| 138 | and self.config.num_nextn_predict_layers > 0 |
| 139 | ): |
| 140 | layer_idx = int(name.split("layers.")[-1].split(".")[0]) |
| 141 | mtp_idx = layer_idx - self.config.num_hidden_layers |
| 142 | if mtp_idx >= 0 and mtp_idx < self.config.num_nextn_predict_layers: |
| 143 | return layer_idx |
| 144 | return None |
| 145 | |
| 146 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 147 | stacked_params_mapping = [ |
| 148 | ("gate_up_proj", "gate_proj", 0), |
| 149 | ("gate_up_proj", "up_proj", 1), |
| 150 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 151 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 152 | ] |
| 153 | |
| 154 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 155 | self, |
| 156 | ckpt_gate_proj_name="gate_proj", |
| 157 | ckpt_down_proj_name="down_proj", |
| 158 | ckpt_up_proj_name="up_proj", |
| 159 | num_experts=self.config.n_routed_experts, |
| 160 | ) |
| 161 | |
| 162 | params_dict = dict(self.named_parameters()) |
| 163 | loaded_params: set[str] = set() |
| 164 | for name, loaded_weight in weights: |
| 165 | if "rotary_emb.inv_freq" in name: |
| 166 | continue |
| 167 | spec_layer = self.get_spec_layer(name) |
| 168 | if spec_layer is None: |
| 169 | continue |
| 170 | |
| 171 | name = self._rewrite_spec_layer_name(spec_layer, name) |
| 172 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 173 | # Skip non-stacked layers and experts (experts handled below). |
| 174 | if weight_name not in name: |
| 175 | continue |
| 176 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 177 | # Since we handle the experts below in expert_params_mapping, |
| 178 | # we need to skip here BEFORE we update the name, otherwise |
| 179 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 180 | # will then be updated below in expert_params_mapping |
| 181 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 182 | if ("mlp.experts." in name) and name not in params_dict: |
| 183 | continue |
| 184 | name_mapped = name.replace(weight_name, param_name) |
| 185 | |
| 186 | # QKV fusion is optional, fall back to normal |
| 187 | # weight loading if it's not enabled |
| 188 | if ( |
| 189 | param_name == "fused_qkv_a_proj" |
| 190 | ) and name_mapped not in params_dict: |
| 191 | continue |
| 192 | else: |
| 193 | name = name_mapped |
| 194 | |
| 195 | # Skip loading extra bias for GPTQ models. |
| 196 | if name.endswith(".bias") and name not in params_dict: |
| 197 | continue |
| 198 | |
| 199 | param = params_dict[name] |
| 200 | weight_loader = param.weight_loader |
| 201 | weight_loader(param, loaded_weight, shard_id) |
| 202 | break |
| 203 | else: |
| 204 | for mapping in expert_params_mapping: |
| 205 | ( |
| 206 | expert_param_name, |
| 207 | expert_weight_name, |
| 208 | expert_id, |
| 209 | expert_shard_id, |
| 210 | ) = mapping |
| 211 | if expert_weight_name not in name: |
| 212 | continue |
| 213 | name = name.replace(expert_weight_name, expert_param_name) |
| 214 | |
| 215 | param = params_dict[name] |
| 216 | weight_loader = param.weight_loader |
| 217 | weight_loader( |
| 218 | param, |
| 219 | loaded_weight, |
| 220 | name, |
| 221 | shard_id=expert_shard_id, |
| 222 | expert_id=expert_id, |
| 223 | ) |
| 224 | break |
| 225 | else: |
| 226 | # Skip loading extra bias for GPTQ models. |
| 227 | if name.endswith(".bias") and name not in params_dict: |
| 228 | continue |
| 229 | |
| 230 | if ( |
| 231 | spec_layer != self.model.mtp_start_layer_idx |
| 232 | and ".layers" not in name |
| 233 | ): |
| 234 | continue |
| 235 | |
| 236 | param = params_dict[name] |
| 237 | weight_loader = getattr( |
| 238 | param, "weight_loader", default_weight_loader |
| 239 | ) |
| 240 | weight_loader(param, loaded_weight) |
| 241 | loaded_params.add(name) |
| 242 | return loaded_params |
| 243 | |
| 244 | def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str: |
| 245 | """Rewrite the weight name to match the format of the original model. |
| 246 | Add .mtp_block for modules in transformer layer block for spec layer |
| 247 | and rename shared layer weights to be top level. |
| 248 | """ |
| 249 | spec_layer_weight_names = [ |
| 250 | "embed_tokens", |
| 251 | "enorm", |
| 252 | "hnorm", |
| 253 | "eh_proj", |
| 254 | "shared_head", |
| 255 | ] |
| 256 | shared_weight_names = ["embed_tokens"] |
| 257 | spec_layer_weight = False |
| 258 | shared_weight = False |
| 259 | for weight_name in spec_layer_weight_names: |
| 260 | if weight_name in name: |
| 261 | spec_layer_weight = True |
| 262 | if weight_name in shared_weight_names: |
| 263 | shared_weight = True |
| 264 | break |
| 265 | if not spec_layer_weight: |
| 266 | # treat rest weights as weights for transformer layer block |
| 267 | name = name.replace( |
| 268 | f"model.layers.{spec_layer}.", f"model.layers.{spec_layer}.mtp_block." |
| 269 | ) |
| 270 | elif shared_weight: |
| 271 | # treat shared weights as top level weights |
| 272 | name = name.replace(f"model.layers.{spec_layer}.", "model.") |
| 273 | return name |