| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from vllm.compilation.decorators import support_torch_compile |
| 11 | from vllm.config import VllmConfig |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 14 | from vllm.model_executor.layers.linear import RowParallelLinear |
| 15 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 16 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 17 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 18 | from vllm.model_executor.models.interfaces import ( |
| 19 | MultiModalEmbeddings, |
| 20 | SupportsMultiModalEmbeddings, |
| 21 | ) |
| 22 | from vllm.model_executor.models.llama import LlamaConfig |
| 23 | from vllm.model_executor.models.mistral import ( |
| 24 | MistralDecoderLayer, |
| 25 | MistralForCausalLM, |
| 26 | MistralModel, |
| 27 | ) |
| 28 | from vllm.model_executor.models.utils import ( |
| 29 | _merge_multimodal_embeddings, |
| 30 | get_draft_quant_config, |
| 31 | maybe_prefix, |
| 32 | ) |
| 33 | |
| 34 | logger = init_logger(__name__) |
| 35 | |
| 36 | |
| 37 | if TYPE_CHECKING: |
| 38 | |
| 39 | class _EagleMistralModelBase(nn.Module): |
| 40 | pass |
| 41 | |
| 42 | class _EagleMistralForCausalLMBase(nn.Module): |
| 43 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 44 | raise NotImplementedError |
| 45 | |
| 46 | else: |
| 47 | _EagleMistralModelBase = MistralModel |
| 48 | _EagleMistralForCausalLMBase = MistralForCausalLM |
| 49 | |
| 50 | |
| 51 | class EagleMistralDecoderLayer(MistralDecoderLayer): |
| 52 | def __init__( |
| 53 | self, |
| 54 | vllm_config: VllmConfig, |
| 55 | prefix: str = "", |
| 56 | config: LlamaConfig | None = None, |
| 57 | ) -> None: |
| 58 | super().__init__(vllm_config, prefix=prefix, config=config) |
| 59 | |
| 60 | def get_quant_config(self, vllm_config: VllmConfig) -> QuantizationConfig | None: |
| 61 | return get_draft_quant_config(vllm_config) |
| 62 | |
| 63 | |
| 64 | @support_torch_compile |
| 65 | class EagleMistralModel(_EagleMistralModelBase): |
| 66 | def __init__( |
| 67 | self, |
| 68 | *, |
| 69 | vllm_config: VllmConfig, |
| 70 | prefix: str = "", |
| 71 | start_layer_id: int = 0, |
| 72 | ) -> None: |
| 73 | # Bypass MistralModel.__init__ to avoid creating duplicate attention |
| 74 | # layer entries in the global context. |
| 75 | nn.Module.__init__(self) |
| 76 | speculative_config = vllm_config.speculative_config |
| 77 | assert speculative_config is not None |
| 78 | self.config = speculative_config.draft_model_config.hf_config |
| 79 | self.vocab_size = self.config.vocab_size |
| 80 | # Get drafter's quantization config |
| 81 | self.quant_config = get_draft_quant_config(vllm_config) |
| 82 | |
| 83 | self.embed_tokens = VocabParallelEmbedding( |
| 84 | self.config.vocab_size, |
| 85 | self.config.hidden_size, |
| 86 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 87 | quant_config=self.quant_config, |
| 88 | ) |
| 89 | |
| 90 | self.layers = nn.ModuleList( |
| 91 | [ |
| 92 | EagleMistralDecoderLayer( |
| 93 | vllm_config, |
| 94 | prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"), |
| 95 | config=self.config, |
| 96 | ) |
| 97 | for i in range(self.config.num_hidden_layers) |
| 98 | ] |
| 99 | ) |
| 100 | self.fc = RowParallelLinear( |
| 101 | self.config.hidden_size * 2, |
| 102 | self.config.hidden_size, |
| 103 | bias=False, |
| 104 | input_is_parallel=False, |
| 105 | quant_config=self.quant_config, |
| 106 | prefix=maybe_prefix(prefix, "fc"), |
| 107 | return_bias=False, |
| 108 | ) |
| 109 | self.norm = RMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps) |
| 110 | |
| 111 | def forward( |
| 112 | self, |
| 113 | input_ids: torch.Tensor, |
| 114 | positions: torch.Tensor, |
| 115 | hidden_states: torch.Tensor, |
| 116 | inputs_embeds: torch.Tensor | None = None, |
| 117 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 118 | if inputs_embeds is None: |
| 119 | inputs_embeds = self.embed_input_ids(input_ids) |
| 120 | hidden_states = self.fc(torch.cat((inputs_embeds, hidden_states), dim=-1)) |
| 121 | residual = None |
| 122 | for layer in self.layers: |
| 123 | hidden_states, residual = layer( |
| 124 | positions, |
| 125 | hidden_states, |
| 126 | residual, |
| 127 | ) |
| 128 | hidden_states, _ = self.norm(hidden_states, residual) |
| 129 | return hidden_states, hidden_states |
| 130 | |
| 131 | |
| 132 | class EagleMistralForCausalLM( |
| 133 | _EagleMistralForCausalLMBase, SupportsMultiModalEmbeddings |
| 134 | ): |
| 135 | mistral_mapping = MistralForCausalLM.mistral_mapping | { |
| 136 | "eagle_linear": "model.fc", |
| 137 | } |
| 138 | |
| 139 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 140 | # Bypass MistralForCausalLM.__init__ to use the draft model config |
| 141 | # and to avoid creating an lm_head. |
| 142 | nn.Module.__init__(self) |
| 143 | speculative_config = vllm_config.speculative_config |
| 144 | assert speculative_config is not None |
| 145 | self.config = speculative_config.draft_model_config.hf_config |
| 146 | target_layer_num = vllm_config.model_config.get_num_layers( |
| 147 | vllm_config.parallel_config |
| 148 | ) |
| 149 | self.model = EagleMistralModel( |
| 150 | vllm_config=vllm_config, |
| 151 | prefix=maybe_prefix(prefix, "model"), |
| 152 | start_layer_id=target_layer_num, |
| 153 | ) |
| 154 | |
| 155 | logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 156 | self.logits_processor = LogitsProcessor( |
| 157 | self.config.vocab_size, scale=logit_scale |
| 158 | ) |
| 159 | |
| 160 | def forward( |
| 161 | self, |
| 162 | input_ids: torch.Tensor, |
| 163 | positions: torch.Tensor, |
| 164 | hidden_states: torch.Tensor, |
| 165 | inputs_embeds: torch.Tensor | None = None, |
| 166 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 167 | return self.model(input_ids, positions, hidden_states, inputs_embeds) |
| 168 | |
| 169 | def embed_input_ids( |
| 170 | self, |
| 171 | input_ids: torch.Tensor, |
| 172 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 173 | *, |
| 174 | is_multimodal: torch.Tensor | None = None, |
| 175 | ) -> torch.Tensor: |
| 176 | inputs_embeds = super().embed_input_ids(input_ids) |
| 177 | |
| 178 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 179 | return inputs_embeds |
| 180 | |
| 181 | assert is_multimodal is not None |
| 182 | |
| 183 | return _merge_multimodal_embeddings( |
| 184 | inputs_embeds=inputs_embeds, |
| 185 | multimodal_embeddings=multimodal_embeddings, |
| 186 | is_multimodal=is_multimodal, |
| 187 | ) |
| 188 | |
| 189 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 190 | # Pretend embed_tokens is loaded; the actual weight is shared |
| 191 | # from the target model at runtime by `load_eagle_model`. |
| 192 | return super().load_weights(weights) | {"model.embed_tokens.weight"} |