|
| 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 Optional |
| 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.distributed.parallel_state import get_pp_group |
| 28 | +from vllm.logger import init_logger |
| 29 | +from vllm.model_executor.layers.layernorm import RMSNorm |
| 30 | +from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 31 | +from vllm.model_executor.layers.quantization.base_config import ( |
| 32 | + QuantizationConfig) |
| 33 | +from vllm.model_executor.layers.quantization.torchao import TorchAOConfig |
| 34 | +from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 35 | + VocabParallelEmbedding) |
| 36 | +from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 37 | +from vllm.model_executor.models.llama4 import (Llama4DecoderLayer, |
| 38 | + Llama4ForCausalLM) |
| 39 | +from vllm.model_executor.models.utils import extract_layer_index |
| 40 | + |
| 41 | +from .utils import AutoWeightsLoader, maybe_prefix |
| 42 | + |
| 43 | +logger = init_logger(__name__) |
| 44 | + |
| 45 | + |
| 46 | +@support_torch_compile |
| 47 | +class LlamaModel(nn.Module): |
| 48 | + |
| 49 | + def __init__( |
| 50 | + self, |
| 51 | + *, |
| 52 | + vllm_config: VllmConfig, |
| 53 | + prefix: str = "", |
| 54 | + start_layer_id: int = 0, |
| 55 | + quant_config: Optional[QuantizationConfig] = None, |
| 56 | + ) -> None: |
| 57 | + super().__init__() |
| 58 | + self.config = ( |
| 59 | + vllm_config.speculative_config.draft_model_config.hf_config) |
| 60 | + self.validate_and_update_config(start_layer_id, quant_config) |
| 61 | + self.vocab_size = self.config.vocab_size |
| 62 | + self.embed_tokens = VocabParallelEmbedding( |
| 63 | + self.config.vocab_size, |
| 64 | + self.config.hidden_size, |
| 65 | + prefix=maybe_prefix(prefix, "embed_tokens"), |
| 66 | + ) |
| 67 | + |
| 68 | + self.layers = nn.ModuleList([ |
| 69 | + Llama4DecoderLayer( |
| 70 | + self.config, |
| 71 | + quant_config=quant_config, |
| 72 | + prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"), |
| 73 | + ) for i in range(self.config.num_hidden_layers) |
| 74 | + ]) |
| 75 | + self.fc = torch.nn.Linear(self.config.hidden_size * 2, |
| 76 | + self.config.hidden_size, |
| 77 | + bias=False) |
| 78 | + self.norm = RMSNorm(self.config.hidden_size, |
| 79 | + eps=self.config.rms_norm_eps) |
| 80 | + |
| 81 | + def forward( |
| 82 | + self, |
| 83 | + input_ids: Optional[torch.Tensor], |
| 84 | + positions: torch.Tensor, |
| 85 | + hidden_states: torch.Tensor, |
| 86 | + ) -> tuple[torch.Tensor, torch.Tensor]: |
| 87 | + input_embeds = self.embed_tokens(input_ids) |
| 88 | + hidden_states = self.fc( |
| 89 | + torch.cat((input_embeds, hidden_states), dim=-1)) |
| 90 | + residual = None |
| 91 | + for layer in self.layers: |
| 92 | + hidden_states, residual = layer( |
| 93 | + positions, |
| 94 | + hidden_states, |
| 95 | + residual, |
| 96 | + ) |
| 97 | + hidden_states, _ = self.norm(hidden_states, residual) |
| 98 | + return hidden_states, hidden_states |
| 99 | + |
| 100 | + def load_weights(self, weights: Iterable[tuple[str, |
| 101 | + torch.Tensor]]) -> set[str]: |
| 102 | + stacked_params_mapping = [ |
| 103 | + # (param_name, shard_name, shard_id) |
| 104 | + (".qkv_proj", ".q_proj", "q"), |
| 105 | + (".qkv_proj", ".k_proj", "k"), |
| 106 | + (".qkv_proj", ".v_proj", "v"), |
| 107 | + (".gate_up_proj", ".gate_proj", 0), |
| 108 | + (".gate_up_proj", ".up_proj", 1), |
| 109 | + ] |
| 110 | + params_dict = dict(self.named_parameters()) |
| 111 | + loaded_params: set[str] = set() |
| 112 | + for name, loaded_weight in weights: |
| 113 | + name = name.removeprefix("model.") |
| 114 | + for param_name, weight_name, shard_id in stacked_params_mapping: |
| 115 | + if weight_name not in name: |
| 116 | + continue |
| 117 | + name = name.replace(weight_name, param_name) |
| 118 | + param = params_dict[name] |
| 119 | + weight_loader = param.weight_loader |
| 120 | + weight_loader(param, loaded_weight, shard_id) |
| 121 | + break |
| 122 | + else: |
| 123 | + # if PP disabled then draft will share embed with target |
| 124 | + if get_pp_group().world_size == 1 and \ |
| 125 | + "embed_tokens." in name: |
| 126 | + continue |
| 127 | + param = params_dict[name] |
| 128 | + weight_loader = getattr(param, "weight_loader", |
| 129 | + default_weight_loader) |
| 130 | + weight_loader(param, loaded_weight) |
| 131 | + loaded_params.add(name) |
| 132 | + for name in params_dict: |
| 133 | + # if PP disabled then draft will share embed with target |
| 134 | + if get_pp_group().world_size == 1 and \ |
| 135 | + "embed_tokens." in name: |
| 136 | + continue |
| 137 | + assert name in loaded_params, f"{name} is not loaded!" |
| 138 | + return loaded_params |
| 139 | + |
| 140 | + def validate_and_update_config( |
| 141 | + self, |
| 142 | + start_layer_id: int, |
| 143 | + quant_config: Optional[QuantizationConfig] = None) -> None: |
| 144 | + # yoco and moe is not supported by draft model yet |
| 145 | + assert self.config.yoco_global_kv_layer is None |
| 146 | + assert self.config.yoco_local_kv_layer is None |
| 147 | + assert len(self.config.moe_layers) == 0 |
| 148 | + # draft model layer index is increased by start_layer_id, |
| 149 | + # so we need to pad relevant configs accordingly |
| 150 | + self.config.no_rope_layers = [ |
| 151 | + 0 |
| 152 | + ] * start_layer_id + self.config.no_rope_layers |
| 153 | + # currently only TorchAO quantization is supported |
| 154 | + if isinstance(quant_config, TorchAOConfig): |
| 155 | + |
| 156 | + def pad_layer_name(layer: str) -> str: |
| 157 | + layer_index = extract_layer_index(layer) |
| 158 | + return layer.replace(str(layer_index), |
| 159 | + str(layer_index + start_layer_id)) |
| 160 | + |
| 161 | + quant_config.torchao_config.module_fqn_to_config = { |
| 162 | + pad_layer_name(layer): quantization |
| 163 | + for layer, quantization in |
| 164 | + quant_config.torchao_config.module_fqn_to_config.items() |
| 165 | + } |
| 166 | + |
| 167 | + |
| 168 | +class EagleLlama4ForCausalLM(Llama4ForCausalLM): |
| 169 | + |
| 170 | + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 171 | + nn.Module.__init__(self) |
| 172 | + self.config = ( |
| 173 | + vllm_config.speculative_config.draft_model_config.hf_config) |
| 174 | + target_layer_num = vllm_config.model_config.get_num_layers( |
| 175 | + vllm_config.parallel_config) |
| 176 | + # draft model quantization config may differ from target model |
| 177 | + quant_config = VllmConfig.get_quantization_config( |
| 178 | + vllm_config.speculative_config.draft_model_config, |
| 179 | + vllm_config.load_config) |
| 180 | + self.model = LlamaModel(vllm_config=vllm_config, |
| 181 | + prefix="model", |
| 182 | + start_layer_id=target_layer_num, |
| 183 | + quant_config=quant_config) |
| 184 | + logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 185 | + self.logits_processor = LogitsProcessor(self.config.vocab_size, |
| 186 | + scale=logit_scale) |
| 187 | + |
| 188 | + def forward( |
| 189 | + self, |
| 190 | + input_ids: torch.Tensor, |
| 191 | + positions: torch.Tensor, |
| 192 | + hidden_states: torch.Tensor, |
| 193 | + ) -> tuple[torch.Tensor, torch.Tensor]: |
| 194 | + return self.model(input_ids, positions, hidden_states) |
| 195 | + |
| 196 | + def load_weights(self, weights: Iterable[tuple[str, |
| 197 | + torch.Tensor]]) -> None: |
| 198 | + loader = AutoWeightsLoader( |
| 199 | + self, |
| 200 | + # lm_head is tied with target model (Llama4ForCausalLM) |
| 201 | + skip_prefixes=(["lm_head."]), |
| 202 | + ) |
| 203 | + |
| 204 | + model_weights = {} |
| 205 | + weights = [ |
| 206 | + self.permute_qk_weight_for_rotary(name, loaded_weight) |
| 207 | + for name, loaded_weight in weights |
| 208 | + ] |
| 209 | + for name, loaded_weight in weights: |
| 210 | + if "lm_head" not in name: |
| 211 | + name = "model." + name |
| 212 | + model_weights[name] = loaded_weight |
| 213 | + |
| 214 | + loader.load_weights(model_weights.items()) |
0 commit comments