| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only ExaoneMoe MTP model.""" |
| 4 | |
| 5 | from collections.abc import Iterable |
| 6 | |
| 7 | import torch |
| 8 | from torch import nn |
| 9 | |
| 10 | from vllm.compilation.decorators import support_torch_compile |
| 11 | from vllm.config import VllmConfig |
| 12 | from vllm.distributed.parallel_state import get_pp_group |
| 13 | from vllm.logger import init_logger |
| 14 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 15 | from vllm.model_executor.layers.linear import ColumnParallelLinear |
| 16 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 17 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 18 | ParallelLMHead, |
| 19 | VocabParallelEmbedding, |
| 20 | ) |
| 21 | from vllm.model_executor.models.exaone_moe import ExaoneMoeDecoderLayer |
| 22 | from vllm.sequence import IntermediateTensors |
| 23 | |
| 24 | from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix |
| 25 | |
| 26 | logger = init_logger(__name__) |
| 27 | |
| 28 | KVCache = tuple[torch.Tensor, torch.Tensor] |
| 29 | |
| 30 | |
| 31 | @support_torch_compile |
| 32 | class ExaoneMoeMultiTokenPredictor(nn.Module): |
| 33 | hf_to_vllm_mapper = WeightsMapper( |
| 34 | orig_to_new_stacked={ |
| 35 | ".q_proj": (".qkv_proj", "q"), |
| 36 | ".k_proj": (".qkv_proj", "k"), |
| 37 | ".v_proj": (".qkv_proj", "v"), |
| 38 | # Scope to dense mlp; experts are handled separately. |
| 39 | ".mlp.gate_proj": (".mlp.gate_up_proj", 0), |
| 40 | ".mlp.up_proj": (".mlp.gate_up_proj", 1), |
| 41 | } |
| 42 | ) |
| 43 | |
| 44 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 45 | super().__init__() |
| 46 | |
| 47 | model_config = vllm_config.model_config |
| 48 | quant_config = vllm_config.quant_config |
| 49 | config = model_config.hf_config |
| 50 | |
| 51 | self.config = config |
| 52 | self.vocab_size = config.vocab_size |
| 53 | self.org_vocab_size = config.vocab_size |
| 54 | |
| 55 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 56 | self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1) |
| 57 | |
| 58 | self.embed_tokens = VocabParallelEmbedding( |
| 59 | self.vocab_size, |
| 60 | config.hidden_size, |
| 61 | ) |
| 62 | |
| 63 | self.fc = ColumnParallelLinear( |
| 64 | self.config.hidden_size * 2, |
| 65 | self.config.hidden_size, |
| 66 | gather_output=True, |
| 67 | bias=False, |
| 68 | return_bias=False, |
| 69 | quant_config=quant_config, |
| 70 | prefix=f"{prefix}.fc", |
| 71 | ) |
| 72 | self.layers = nn.ModuleList( |
| 73 | ExaoneMoeDecoderLayer( |
| 74 | vllm_config.model_config.hf_config, |
| 75 | quant_config=quant_config, |
| 76 | prefix=f"{prefix}.layers.{idx}", |
| 77 | is_mtp=True, |
| 78 | ) |
| 79 | for idx in range(self.num_mtp_layers) |
| 80 | ) |
| 81 | |
| 82 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 83 | self.pre_fc_norm_hidden = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 84 | self.pre_fc_norm_embedding = RMSNorm( |
| 85 | config.hidden_size, eps=config.rms_norm_eps |
| 86 | ) |
| 87 | |
| 88 | def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 89 | return self.embed_tokens(input_ids) |
| 90 | |
| 91 | def forward( |
| 92 | self, |
| 93 | input_ids: torch.Tensor, |
| 94 | positions: torch.Tensor, |
| 95 | hidden_states: torch.Tensor, |
| 96 | intermediate_tensors: IntermediateTensors | None = None, |
| 97 | inputs_embeds: torch.Tensor | None = None, |
| 98 | spec_step_idx: int = 0, |
| 99 | ) -> torch.Tensor: |
| 100 | if get_pp_group().is_first_rank: |
| 101 | if inputs_embeds is None: |
| 102 | inputs_embeds = self.get_input_embeddings(input_ids) |
| 103 | assert hidden_states.shape[-1] == inputs_embeds.shape[-1] |
| 104 | inputs_embeds = self.pre_fc_norm_embedding(inputs_embeds) |
| 105 | hidden_states = self.pre_fc_norm_hidden(hidden_states) |
| 106 | hidden_states = torch.cat([inputs_embeds, hidden_states], dim=-1) |
| 107 | hidden_states = self.fc(hidden_states) |
| 108 | residual = None |
| 109 | else: |
| 110 | assert intermediate_tensors is not None |
| 111 | hidden_states = intermediate_tensors["hidden_states"] |
| 112 | residual = intermediate_tensors["residual"] |
| 113 | |
| 114 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 115 | hidden_states, residual = self.layers[current_step_idx]( |
| 116 | positions=positions, |
| 117 | hidden_states=hidden_states, |
| 118 | residual=residual, |
| 119 | ) |
| 120 | |
| 121 | if not get_pp_group().is_last_rank: |
| 122 | return IntermediateTensors( |
| 123 | {"hidden_states": hidden_states, "residual": residual} |
| 124 | ) |
| 125 | |
| 126 | hidden_states, _ = self.norm(hidden_states, residual) |
| 127 | return hidden_states |
| 128 | |
| 129 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 130 | loader = AutoWeightsLoader(self) |
| 131 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 132 | |
| 133 | |
| 134 | @support_torch_compile |
| 135 | class ExaoneMoeMTP(nn.Module): |
| 136 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 137 | config = vllm_config.model_config.hf_config |
| 138 | self.vllm_config = vllm_config |
| 139 | self.quant_config = vllm_config.quant_config |
| 140 | |
| 141 | super().__init__() |
| 142 | self.config = config |
| 143 | self.model = ExaoneMoeMultiTokenPredictor( |
| 144 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp") |
| 145 | ) |
| 146 | self.unpadded_vocab_size = config.vocab_size |
| 147 | self.lm_head = ParallelLMHead( |
| 148 | self.unpadded_vocab_size, |
| 149 | config.hidden_size, |
| 150 | org_num_embeddings=config.vocab_size, |
| 151 | # padding_size=DEFAULT_VOCAB_PADDING_SIZE, |
| 152 | prefix=maybe_prefix(prefix, "lm_head"), |
| 153 | ) |
| 154 | if config.tie_word_embeddings: |
| 155 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 156 | self.logits_processor = LogitsProcessor( |
| 157 | self.unpadded_vocab_size, config.vocab_size |
| 158 | ) |
| 159 | |
| 160 | def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 161 | return self.model.get_input_embeddings(input_ids) |
| 162 | |
| 163 | def forward( |
| 164 | self, |
| 165 | input_ids: torch.Tensor, |
| 166 | positions: torch.Tensor, |
| 167 | hidden_states: torch.Tensor, |
| 168 | intermediate_tensors: IntermediateTensors | None = None, |
| 169 | inputs_embeds: torch.Tensor | None = None, |
| 170 | spec_step_idx: int = 0, |
| 171 | **kwargs: object, |
| 172 | ): |
| 173 | hidden_states = self.model( |
| 174 | input_ids, |
| 175 | positions, |
| 176 | hidden_states, |
| 177 | intermediate_tensors, |
| 178 | inputs_embeds, |
| 179 | spec_step_idx, |
| 180 | ) |
| 181 | return hidden_states |
| 182 | |
| 183 | def compute_logits( |
| 184 | self, |
| 185 | hidden_states: torch.Tensor, |
| 186 | spec_step_idx: int = 0, |
| 187 | ) -> torch.Tensor | None: |
| 188 | return self.logits_processor(self.lm_head, hidden_states) |
| 189 | |
| 190 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 191 | shared_weight_names = ["embed_tokens", "lm_head"] |
| 192 | |
| 193 | def remap_weight_names(weights): |
| 194 | for name, weight in weights: |
| 195 | if name.startswith("mtp."): |
| 196 | name = name.replace("mtp.", "model.") |
| 197 | elif not any(key in name for key in shared_weight_names): |
| 198 | continue |
| 199 | yield name, weight |
| 200 | |
| 201 | loader = AutoWeightsLoader(self) |
| 202 | return loader.load_weights(remap_weight_names(weights)) |