| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright 2025 the LLAMA4, Meta Inc., vLLM, and HuggingFace Inc. team. |
| 4 | # All rights reserved. |
| 5 | # |
| 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 | from collections.abc import Iterable |
| 20 | from typing import TYPE_CHECKING |
| 21 | |
| 22 | import torch |
| 23 | import torch.nn as nn |
| 24 | |
| 25 | from vllm.compilation.decorators import support_torch_compile |
| 26 | from vllm.config import VllmConfig |
| 27 | from vllm.logger import init_logger |
| 28 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 29 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 30 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 31 | from vllm.model_executor.layers.quantization.torchao import TorchAOConfig |
| 32 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 33 | ParallelLMHead, |
| 34 | VocabParallelEmbedding, |
| 35 | ) |
| 36 | from vllm.model_executor.models.llama4 import Llama4DecoderLayer, Llama4ForCausalLM |
| 37 | from vllm.model_executor.models.utils import extract_layer_index |
| 38 | |
| 39 | from .interfaces import SupportsMultiModal, SupportsMultiModalEmbeddings |
| 40 | from .utils import ( |
| 41 | AutoWeightsLoader, |
| 42 | WeightsMapper, |
| 43 | maybe_prefix, |
| 44 | process_eagle_weight, |
| 45 | ) |
| 46 | |
| 47 | logger = init_logger(__name__) |
| 48 | |
| 49 | |
| 50 | if TYPE_CHECKING: |
| 51 | |
| 52 | class _EagleLlama4ForCausalLMBase(nn.Module): |
| 53 | def set_moe_parameters(self) -> None: ... |
| 54 | |
| 55 | def permute_qk_weight_for_rotary( |
| 56 | self, name: str, loaded_weight: torch.Tensor |
| 57 | ) -> tuple[str, torch.Tensor]: ... |
| 58 | |
| 59 | else: |
| 60 | _EagleLlama4ForCausalLMBase = Llama4ForCausalLM |
| 61 | |
| 62 | |
| 63 | @support_torch_compile |
| 64 | class LlamaModel(nn.Module): |
| 65 | hf_to_vllm_mapper = WeightsMapper( |
| 66 | orig_to_new_prefix={"model.": ""}, |
| 67 | orig_to_new_stacked={ |
| 68 | ".q_proj": (".qkv_proj", "q"), |
| 69 | ".k_proj": (".qkv_proj", "k"), |
| 70 | ".v_proj": (".qkv_proj", "v"), |
| 71 | ".gate_proj": (".gate_up_proj", 0), |
| 72 | ".up_proj": (".gate_up_proj", 1), |
| 73 | }, |
| 74 | ) |
| 75 | |
| 76 | def __init__( |
| 77 | self, |
| 78 | *, |
| 79 | vllm_config: VllmConfig, |
| 80 | prefix: str = "", |
| 81 | start_layer_id: int = 0, |
| 82 | quant_config: QuantizationConfig | None = None, |
| 83 | ) -> None: |
| 84 | super().__init__() |
| 85 | speculative_config = vllm_config.speculative_config |
| 86 | assert speculative_config is not None |
| 87 | self.config = speculative_config.draft_model_config.hf_config |
| 88 | self.validate_and_update_config(start_layer_id, quant_config) |
| 89 | self.vocab_size = self.config.vocab_size |
| 90 | self.embed_tokens = VocabParallelEmbedding( |
| 91 | self.config.vocab_size, |
| 92 | self.config.hidden_size, |
| 93 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 94 | ) |
| 95 | |
| 96 | # Temporarily modify vllm_config.quant_config for draft model layers |
| 97 | original_quant_config = vllm_config.quant_config |
| 98 | vllm_config.quant_config = quant_config |
| 99 | try: |
| 100 | self.layers = nn.ModuleList( |
| 101 | [ |
| 102 | Llama4DecoderLayer( |
| 103 | vllm_config=vllm_config, |
| 104 | prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"), |
| 105 | config=self.config, |
| 106 | ) |
| 107 | for i in range(self.config.num_hidden_layers) |
| 108 | ] |
| 109 | ) |
| 110 | finally: |
| 111 | # Restore original quant_config |
| 112 | vllm_config.quant_config = original_quant_config |
| 113 | self.fc = torch.nn.Linear( |
| 114 | self.config.hidden_size * 2, self.config.hidden_size, bias=False |
| 115 | ) |
| 116 | self.norm = RMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps) |
| 117 | |
| 118 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 119 | return self.embed_tokens(input_ids) |
| 120 | |
| 121 | def forward( |
| 122 | self, |
| 123 | input_ids: torch.Tensor | None, |
| 124 | positions: torch.Tensor, |
| 125 | hidden_states: torch.Tensor, |
| 126 | inputs_embeds: torch.Tensor | None = None, |
| 127 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 128 | if inputs_embeds is None: |
| 129 | inputs_embeds = self.embed_input_ids(input_ids) |
| 130 | hidden_states = self.fc(torch.cat((inputs_embeds, hidden_states), dim=-1)) |
| 131 | residual = None |
| 132 | for layer in self.layers: |
| 133 | hidden_states, residual = layer( |
| 134 | positions, |
| 135 | hidden_states, |
| 136 | residual, |
| 137 | ) |
| 138 | hidden_states, _ = self.norm(hidden_states, residual) |
| 139 | return hidden_states, hidden_states |
| 140 | |
| 141 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 142 | loader = AutoWeightsLoader(self) |
| 143 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 144 | |
| 145 | def validate_and_update_config( |
| 146 | self, start_layer_id: int, quant_config: QuantizationConfig | None = None |
| 147 | ) -> None: |
| 148 | # yoco and moe is not supported by draft model yet |
| 149 | assert self.config.yoco_global_kv_layer is None |
| 150 | assert self.config.yoco_local_kv_layer is None |
| 151 | assert len(self.config.moe_layers) == 0 |
| 152 | # draft model layer index is increased by start_layer_id, |
| 153 | # so we need to pad relevant configs accordingly |
| 154 | self.config.no_rope_layers = [0] * start_layer_id + self.config.no_rope_layers |
| 155 | # currently only TorchAO quantization is supported |
| 156 | if isinstance(quant_config, TorchAOConfig): |
| 157 | |
| 158 | def pad_layer_name(layer: str) -> str: |
| 159 | layer_index = extract_layer_index(layer) |
| 160 | return layer.replace( |
| 161 | str(layer_index), str(layer_index + start_layer_id) |
| 162 | ) |
| 163 | |
| 164 | torchao_config = quant_config.torchao_config |
| 165 | torchao_config.module_fqn_to_config = { |
| 166 | pad_layer_name(layer): quantization |
| 167 | for layer, quantization in torchao_config.module_fqn_to_config.items() |
| 168 | } |
| 169 | |
| 170 | |
| 171 | class EagleLlama4ForCausalLM(_EagleLlama4ForCausalLMBase, SupportsMultiModalEmbeddings): |
| 172 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 173 | nn.Module.__init__(self) |
| 174 | speculative_config = vllm_config.speculative_config |
| 175 | assert speculative_config is not None |
| 176 | self.config = speculative_config.draft_model_config.hf_config |
| 177 | target_layer_num = vllm_config.model_config.get_num_layers( |
| 178 | vllm_config.parallel_config |
| 179 | ) |
| 180 | # draft model quantization config may differ from target model |
| 181 | quant_config = VllmConfig.get_quantization_config( |
| 182 | speculative_config.draft_model_config, vllm_config.load_config |
| 183 | ) |
| 184 | self.model = LlamaModel( |
| 185 | vllm_config=vllm_config, |
| 186 | prefix=maybe_prefix(prefix, "model"), |
| 187 | start_layer_id=target_layer_num, |
| 188 | quant_config=quant_config, |
| 189 | ) |
| 190 | logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 191 | self.logits_processor = LogitsProcessor( |
| 192 | self.config.vocab_size, scale=logit_scale |
| 193 | ) |
| 194 | |
| 195 | self.lm_head = ParallelLMHead( |
| 196 | self.config.draft_vocab_size, |
| 197 | self.config.hidden_size, |
| 198 | prefix=maybe_prefix(prefix, "lm_head"), |
| 199 | ) |
| 200 | |
| 201 | # Set MoE hyperparameters |
| 202 | self.set_moe_parameters() |
| 203 | |
| 204 | def get_language_model(self) -> torch.nn.Module: |
| 205 | return self.model |
| 206 | |
| 207 | embed_input_ids = SupportsMultiModal.embed_input_ids # type: ignore |
| 208 | |
| 209 | def forward( |
| 210 | self, |
| 211 | input_ids: torch.Tensor, |
| 212 | positions: torch.Tensor, |
| 213 | hidden_states: torch.Tensor, |
| 214 | inputs_embeds: torch.Tensor | None = None, |
| 215 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 216 | return self.model(input_ids, positions, hidden_states, inputs_embeds) |
| 217 | |
| 218 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> None: |
| 219 | def transform(inputs): |
| 220 | name, loaded_weight = inputs |
| 221 | name, weight = self.permute_qk_weight_for_rotary(name, loaded_weight) |
| 222 | if "lm_head" not in name: |
| 223 | name = "model." + name |
| 224 | process_eagle_weight(self, name) |
| 225 | return name, weight |
| 226 | |
| 227 | loader = AutoWeightsLoader(self) |
| 228 | loader.load_weights(map(transform, weights)) |