| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only EXAONE-4_5 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.exaone4 import Exaone4DecoderLayer |
| 22 | from vllm.model_executor.models.exaone_moe_mtp import ( |
| 23 | ExaoneMoeMTP, |
| 24 | ExaoneMoeMultiTokenPredictor, |
| 25 | ) |
| 26 | from vllm.sequence import IntermediateTensors |
| 27 | |
| 28 | from .interfaces import ( |
| 29 | MultiModalEmbeddings, |
| 30 | SupportsMultiModal, |
| 31 | _require_is_multimodal, |
| 32 | ) |
| 33 | from .utils import ( |
| 34 | AutoWeightsLoader, |
| 35 | _merge_multimodal_embeddings, |
| 36 | maybe_prefix, |
| 37 | ) |
| 38 | |
| 39 | logger = init_logger(__name__) |
| 40 | |
| 41 | KVCache = tuple[torch.Tensor, torch.Tensor] |
| 42 | |
| 43 | |
| 44 | @support_torch_compile |
| 45 | class Exaone4_5MultiTokenPredictor(ExaoneMoeMultiTokenPredictor): |
| 46 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 47 | nn.Module.__init__(self) |
| 48 | |
| 49 | model_config = vllm_config.model_config |
| 50 | quant_config = vllm_config.quant_config |
| 51 | config = model_config.hf_config |
| 52 | text_config = config.text_config |
| 53 | |
| 54 | self.config = config |
| 55 | self.vocab_size = config.vocab_size |
| 56 | self.org_vocab_size = config.vocab_size |
| 57 | |
| 58 | self.mtp_start_layer_idx = text_config.num_hidden_layers |
| 59 | self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1) |
| 60 | |
| 61 | self.embed_tokens = VocabParallelEmbedding( |
| 62 | self.vocab_size, |
| 63 | text_config.hidden_size, |
| 64 | ) |
| 65 | |
| 66 | self.fc = ColumnParallelLinear( |
| 67 | text_config.hidden_size * 2, |
| 68 | text_config.hidden_size, |
| 69 | gather_output=True, |
| 70 | bias=False, |
| 71 | return_bias=False, |
| 72 | quant_config=quant_config, |
| 73 | prefix=f"{prefix}.fc", |
| 74 | ) |
| 75 | self.layers = nn.ModuleList( |
| 76 | Exaone4DecoderLayer( |
| 77 | text_config, |
| 78 | quant_config=quant_config, |
| 79 | prefix=f"{prefix}.layers.{idx}", |
| 80 | ) |
| 81 | for idx in range(self.num_mtp_layers) |
| 82 | ) |
| 83 | |
| 84 | self.norm = RMSNorm(text_config.hidden_size, eps=text_config.rms_norm_eps) |
| 85 | self.pre_fc_norm_hidden = RMSNorm( |
| 86 | text_config.hidden_size, eps=text_config.rms_norm_eps |
| 87 | ) |
| 88 | self.pre_fc_norm_embedding = RMSNorm( |
| 89 | text_config.hidden_size, eps=text_config.rms_norm_eps |
| 90 | ) |
| 91 | |
| 92 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 93 | return self.embed_tokens(input_ids) |
| 94 | |
| 95 | def forward( |
| 96 | self, |
| 97 | input_ids: torch.Tensor, |
| 98 | positions: torch.Tensor, |
| 99 | hidden_states: torch.Tensor, |
| 100 | intermediate_tensors: IntermediateTensors | None = None, |
| 101 | inputs_embeds: torch.Tensor | None = None, |
| 102 | spec_step_idx: int = 0, |
| 103 | ) -> torch.Tensor: |
| 104 | if get_pp_group().is_first_rank: |
| 105 | if inputs_embeds is None: |
| 106 | inputs_embeds = self.get_input_embeddings(input_ids) |
| 107 | assert hidden_states.shape[-1] == inputs_embeds.shape[-1] |
| 108 | inputs_embeds = self.pre_fc_norm_embedding(inputs_embeds) |
| 109 | hidden_states = self.pre_fc_norm_hidden(hidden_states) |
| 110 | hidden_states = torch.cat([inputs_embeds, hidden_states], dim=-1) |
| 111 | hidden_states = self.fc(hidden_states) |
| 112 | residual = None |
| 113 | else: |
| 114 | assert intermediate_tensors is not None |
| 115 | hidden_states = intermediate_tensors["hidden_states"] |
| 116 | residual = intermediate_tensors["residual"] |
| 117 | |
| 118 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 119 | hidden_states, residual = self.layers[current_step_idx]( |
| 120 | positions=positions, |
| 121 | hidden_states=hidden_states, |
| 122 | residual=residual, |
| 123 | ) |
| 124 | |
| 125 | if not get_pp_group().is_last_rank: |
| 126 | return IntermediateTensors( |
| 127 | {"hidden_states": hidden_states, "residual": residual} |
| 128 | ) |
| 129 | |
| 130 | hidden_states = self.norm(hidden_states) |
| 131 | return hidden_states |
| 132 | |
| 133 | |
| 134 | @support_torch_compile |
| 135 | class Exaone4_5_MTP(ExaoneMoeMTP, SupportsMultiModal): |
| 136 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 137 | config = vllm_config.model_config.hf_config |
| 138 | text_config = config.text_config |
| 139 | self.vllm_config = vllm_config |
| 140 | self.quant_config = vllm_config.quant_config |
| 141 | |
| 142 | nn.Module.__init__(self) |
| 143 | self.config = config |
| 144 | self.model = Exaone4_5MultiTokenPredictor( |
| 145 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp") |
| 146 | ) |
| 147 | self.unpadded_vocab_size = config.vocab_size |
| 148 | self.lm_head = ParallelLMHead( |
| 149 | self.unpadded_vocab_size, |
| 150 | text_config.hidden_size, |
| 151 | org_num_embeddings=config.vocab_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 embed_input_ids( |
| 161 | self, |
| 162 | input_ids: torch.Tensor, |
| 163 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 164 | *, |
| 165 | is_multimodal: torch.Tensor | None = None, |
| 166 | ) -> torch.Tensor: |
| 167 | inputs_embeds = self._embed_text_input_ids( |
| 168 | input_ids, |
| 169 | self.model.embed_input_ids, |
| 170 | is_multimodal=is_multimodal, |
| 171 | ) |
| 172 | |
| 173 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 174 | return inputs_embeds |
| 175 | |
| 176 | is_multimodal = _require_is_multimodal(is_multimodal) |
| 177 | |
| 178 | inputs_embeds = _merge_multimodal_embeddings( |
| 179 | inputs_embeds=inputs_embeds, |
| 180 | multimodal_embeddings=multimodal_embeddings, |
| 181 | is_multimodal=is_multimodal, |
| 182 | ) |
| 183 | |
| 184 | return inputs_embeds |
| 185 | |
| 186 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 187 | shared_weight_names = ["embed_tokens", "lm_head"] |
| 188 | |
| 189 | def remap_weight_names(weights): |
| 190 | for name, weight in weights: |
| 191 | if name.startswith("mtp."): |
| 192 | name = name.replace("mtp.", "model.") |
| 193 | elif any(key in name for key in shared_weight_names): |
| 194 | if "embed_tokens" in name: |
| 195 | name = name.replace("language_model.", "") |
| 196 | else: |
| 197 | continue |
| 198 | yield name, weight |
| 199 | |
| 200 | loader = AutoWeightsLoader(self) |
| 201 | return loader.load_weights(remap_weight_names(weights)) |