| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | from transformers import CohereConfig |
| 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.linear import ReplicatedLinear |
| 14 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 15 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 16 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 17 | from vllm.model_executor.models.commandr import ( |
| 18 | CohereDecoderLayer, |
| 19 | CohereForCausalLM, |
| 20 | LayerNorm, |
| 21 | ) |
| 22 | |
| 23 | from .utils import ( |
| 24 | AutoWeightsLoader, |
| 25 | WeightsMapper, |
| 26 | get_draft_quant_config, |
| 27 | maybe_prefix, |
| 28 | process_eagle_weight, |
| 29 | ) |
| 30 | |
| 31 | logger = init_logger(__name__) |
| 32 | |
| 33 | |
| 34 | class CohereEagleDecoderLayer(CohereDecoderLayer): |
| 35 | """Eagle draft variant of CohereDecoderLayer.""" |
| 36 | |
| 37 | def __init__( |
| 38 | self, |
| 39 | config: CohereConfig, |
| 40 | cache_config=None, |
| 41 | quant_config: QuantizationConfig | None = None, |
| 42 | prefix: str = "", |
| 43 | ) -> None: |
| 44 | super().__init__( |
| 45 | config, |
| 46 | cache_config=cache_config, |
| 47 | quant_config=quant_config, |
| 48 | prefix=prefix, |
| 49 | ) |
| 50 | |
| 51 | |
| 52 | @support_torch_compile |
| 53 | class CohereEagleModel(nn.Module): |
| 54 | def __init__( |
| 55 | self, |
| 56 | *, |
| 57 | vllm_config: VllmConfig, |
| 58 | prefix: str = "", |
| 59 | start_layer_id: int = 0, |
| 60 | ) -> None: |
| 61 | super().__init__() |
| 62 | assert vllm_config.speculative_config is not None |
| 63 | self.config = vllm_config.speculative_config.draft_model_config.hf_config |
| 64 | self.quant_config = get_draft_quant_config(vllm_config) |
| 65 | |
| 66 | # Cohere2-targeted EAGLE drafts inherit the target's sliding-window |
| 67 | # attention pattern. ``CohereAttention`` resolves per-layer behavior |
| 68 | # via ``config.layer_types[layer_idx]`` and the eagle layers use |
| 69 | # absolute indices (target_layer_num + i), so prepend the target's |
| 70 | # ``layer_types`` to the draft's so the lookup succeeds. |
| 71 | target_text_config = vllm_config.model_config.hf_text_config |
| 72 | if hasattr(target_text_config, "layer_types") and hasattr( |
| 73 | self.config, "layer_types" |
| 74 | ): |
| 75 | self.config.layer_types = list(target_text_config.layer_types) + list( |
| 76 | self.config.layer_types |
| 77 | ) |
| 78 | |
| 79 | self.vocab_size = self.config.vocab_size |
| 80 | self.embed_tokens = VocabParallelEmbedding( |
| 81 | self.config.vocab_size, |
| 82 | self.config.hidden_size, |
| 83 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 84 | ) |
| 85 | |
| 86 | self.layers = nn.ModuleList( |
| 87 | [ |
| 88 | CohereEagleDecoderLayer( |
| 89 | self.config, |
| 90 | cache_config=vllm_config.cache_config, |
| 91 | quant_config=self.quant_config, |
| 92 | prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"), |
| 93 | ) |
| 94 | for i in range(self.config.num_hidden_layers) |
| 95 | ] |
| 96 | ) |
| 97 | |
| 98 | # Cohere EAGLE checkpoints include a bias term on the input fusion |
| 99 | # projection (unlike LLaMA EAGLE which uses bias=False). |
| 100 | self.fc = ReplicatedLinear( |
| 101 | input_size=self.config.hidden_size * 2, |
| 102 | output_size=self.config.hidden_size, |
| 103 | bias=True, |
| 104 | params_dtype=vllm_config.model_config.dtype, |
| 105 | quant_config=self.quant_config, |
| 106 | prefix=maybe_prefix(prefix, "fc"), |
| 107 | return_bias=False, |
| 108 | ) |
| 109 | |
| 110 | # Cohere EAGLE applies an explicit final LayerNorm to the draft |
| 111 | # hidden states before they are consumed by the logits processor. |
| 112 | self.norm = LayerNorm( |
| 113 | param_shape=(self.config.hidden_size), |
| 114 | eps=self.config.layer_norm_eps, |
| 115 | ) |
| 116 | |
| 117 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 118 | return self.embed_tokens(input_ids) |
| 119 | |
| 120 | def forward( |
| 121 | self, |
| 122 | input_ids: torch.Tensor, |
| 123 | positions: torch.Tensor, |
| 124 | hidden_states: torch.Tensor, |
| 125 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 126 | input_embeds = self.embed_tokens(input_ids) |
| 127 | hidden_states = self.fc(torch.cat((input_embeds, hidden_states), dim=-1)) |
| 128 | residual = None |
| 129 | for layer in self.layers: |
| 130 | hidden_states, residual = layer( |
| 131 | positions, |
| 132 | hidden_states, |
| 133 | residual, |
| 134 | ) |
| 135 | hidden_states, _ = self.norm(hidden_states, residual) |
| 136 | return hidden_states, hidden_states |
| 137 | |
| 138 | |
| 139 | class EagleCohereForCausalLM(CohereForCausalLM): |
| 140 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 141 | nn.Module.__init__(self) |
| 142 | assert vllm_config.speculative_config is not None |
| 143 | self.config = vllm_config.speculative_config.draft_model_config.hf_config |
| 144 | # Flags checked by the speculative proposer to decide whether to share |
| 145 | # embed_tokens / lm_head with the target model. Cohere EAGLE checkpoints |
| 146 | # use tied embeddings so these weights are absent from the draft file. |
| 147 | self.has_own_embed_tokens = False |
| 148 | self.has_own_lm_head = False |
| 149 | if self.config.tie_word_embeddings: |
| 150 | self.hf_to_vllm_mapper = self.hf_to_vllm_mapper | WeightsMapper( |
| 151 | orig_to_new_prefix={"model.embed_tokens.": None} |
| 152 | ) |
| 153 | target_layer_num = vllm_config.model_config.get_num_layers( |
| 154 | vllm_config.parallel_config |
| 155 | ) |
| 156 | self.model = CohereEagleModel( |
| 157 | vllm_config=vllm_config, |
| 158 | prefix=maybe_prefix(prefix, "model"), |
| 159 | start_layer_id=target_layer_num, |
| 160 | ) |
| 161 | |
| 162 | logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 163 | self.logits_processor = LogitsProcessor( |
| 164 | self.config.vocab_size, scale=logit_scale |
| 165 | ) |
| 166 | |
| 167 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 168 | return self.model.embed_input_ids(input_ids) |
| 169 | |
| 170 | def forward( # type: ignore[override] |
| 171 | self, |
| 172 | input_ids: torch.Tensor, |
| 173 | positions: torch.Tensor, |
| 174 | hidden_states: torch.Tensor, |
| 175 | inputs_embeds: torch.Tensor | None = None, |
| 176 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 177 | if inputs_embeds is not None: |
| 178 | raise NotImplementedError( |
| 179 | f"{type(self).__name__} does not support multimodal inputs yet." |
| 180 | ) |
| 181 | return self.model(input_ids, positions, hidden_states) |
| 182 | |
| 183 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 184 | def _track_and_forward(inputs): |
| 185 | name, weight = inputs |
| 186 | process_eagle_weight(self, name) |
| 187 | return name, weight |
| 188 | |
| 189 | loader = AutoWeightsLoader(self) |
| 190 | |
| 191 | loaded_weight_names = loader.load_weights( |
| 192 | map(_track_and_forward, weights), mapper=self.hf_to_vllm_mapper |
| 193 | ) |
| 194 | |
| 195 | # Embed tokens are tied with the target model and therefore not |
| 196 | # present in the EAGLE checkpoint; mark them as loaded explicitly to |
| 197 | # avoid a spurious "weight not found" warning from the default |
| 198 | # weight loader. |
| 199 | loaded_weight_names.add("model.embed_tokens.weight") |
| 200 | return loaded_weight_names |