| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # adapted from |
| 5 | # https://github.com/deepseek-ai/DeepSeek-OCR-2/blob/main/DeepSeek-OCR2-master/DeepSeek-OCR2-vllm/deepencoderv2/qwen2_d2e.py |
| 6 | |
| 7 | # Copyright (c) Meta Platforms, Inc. and affiliates. |
| 8 | # All rights reserved. |
| 9 | |
| 10 | # This source code is licensed under the license found in the |
| 11 | # LICENSE file in the root directory of this source tree. |
| 12 | |
| 13 | from functools import lru_cache |
| 14 | |
| 15 | import torch |
| 16 | import torch.nn as nn |
| 17 | import transformers |
| 18 | from transformers.models.qwen2.modeling_qwen2 import Qwen2Model |
| 19 | |
| 20 | from vllm.model_executor.custom_op import PluggableLayer |
| 21 | |
| 22 | |
| 23 | # --8<-- [start:qwen2_decoder] |
| 24 | @PluggableLayer.register("qwen2_decoder") |
| 25 | class CustomQwen2Decoder(PluggableLayer): |
| 26 | """Qwen2 visual encoder |
| 27 | non-causal attention + causal attention |
| 28 | token_type_ids ：0=non-causal, 1=causal |
| 29 | """ |
| 30 | |
| 31 | # --8<-- [end:qwen2_decoder] |
| 32 | |
| 33 | def __init__( |
| 34 | self, |
| 35 | decoder_layer: int = 24, |
| 36 | max_position_embeddings: int = 131072, |
| 37 | hidden_dimension: int = 896, |
| 38 | num_attention_heads: int = 14, |
| 39 | num_key_value_heads: int = 2, |
| 40 | intermediate_size: int = 4864, |
| 41 | vocab_size: int = 151936, |
| 42 | attn_implementation: str = "sdpa", |
| 43 | rms_norm_eps: float = 1e-06, |
| 44 | rope_theta: float = 1000000.0, |
| 45 | attention_dropout: float = 0.0, |
| 46 | hidden_act: str = "silu", |
| 47 | initializer_range: float = 0.02, |
| 48 | ): |
| 49 | super().__init__() |
| 50 | |
| 51 | # load |
| 52 | Qwen2Config = transformers.Qwen2Config |
| 53 | |
| 54 | # config |
| 55 | config = Qwen2Config( |
| 56 | hidden_size=hidden_dimension, |
| 57 | num_hidden_layers=decoder_layer, |
| 58 | num_attention_heads=num_attention_heads, |
| 59 | num_key_value_heads=num_key_value_heads, |
| 60 | intermediate_size=intermediate_size, |
| 61 | max_position_embeddings=max_position_embeddings, |
| 62 | vocab_size=vocab_size, |
| 63 | rms_norm_eps=rms_norm_eps, |
| 64 | rope_theta=rope_theta, |
| 65 | attention_dropout=attention_dropout, |
| 66 | hidden_act=hidden_act, |
| 67 | initializer_range=initializer_range, |
| 68 | _attn_implementation=attn_implementation, # ⭐ |
| 69 | ) |
| 70 | |
| 71 | # |
| 72 | self.model = self._create_custom_model(config) |
| 73 | |
| 74 | del self.model.embed_tokens |
| 75 | |
| 76 | def _create_custom_model(self, config): |
| 77 | """Qwen2Model.""" |
| 78 | |
| 79 | class CustomQwen2ModelInner(Qwen2Model): |
| 80 | def forward( |
| 81 | self, |
| 82 | input_ids=None, |
| 83 | attention_mask=None, |
| 84 | position_ids=None, |
| 85 | past_key_values=None, |
| 86 | inputs_embeds=None, |
| 87 | token_type_ids=None, # ⭐ |
| 88 | use_cache=None, |
| 89 | output_attentions=None, |
| 90 | output_hidden_states=None, |
| 91 | return_dict=None, |
| 92 | cache_position=None, |
| 93 | ): |
| 94 | causal_mask_mapping = { |
| 95 | "full_attention": self._update_causal_mask( |
| 96 | attention_mask, |
| 97 | inputs_embeds, |
| 98 | cache_position, |
| 99 | past_key_values, |
| 100 | output_attentions, |
| 101 | ) |
| 102 | } |
| 103 | outputs = super().forward( |
| 104 | input_ids=input_ids, |
| 105 | attention_mask=causal_mask_mapping, |
| 106 | position_ids=position_ids, |
| 107 | past_key_values=past_key_values, |
| 108 | inputs_embeds=inputs_embeds, |
| 109 | use_cache=use_cache, |
| 110 | output_attentions=output_attentions, |
| 111 | output_hidden_states=output_hidden_states, |
| 112 | return_dict=return_dict, |
| 113 | cache_position=cache_position, |
| 114 | ) |
| 115 | |
| 116 | return outputs |
| 117 | |
| 118 | def _update_causal_mask( |
| 119 | self, |
| 120 | attention_mask, |
| 121 | input_tensor, |
| 122 | cache_position, |
| 123 | past_key_values, |
| 124 | output_attentions, |
| 125 | ): |
| 126 | dtype, device = input_tensor.dtype, input_tensor.device |
| 127 | min_dtype = torch.finfo(dtype).min |
| 128 | batch_size, sequence_length = ( |
| 129 | input_tensor.shape[0], |
| 130 | input_tensor.shape[1], |
| 131 | ) |
| 132 | |
| 133 | # attention mask |
| 134 | causal_mask = self._create_custom_4d_mask( |
| 135 | sequence_length=sequence_length, |
| 136 | dtype=dtype, |
| 137 | device=device, |
| 138 | batch_size=batch_size, |
| 139 | ) |
| 140 | |
| 141 | # padding mask |
| 142 | if attention_mask is not None and attention_mask.dim() == 2: |
| 143 | padding_mask = attention_mask[:, None, None, :].to(dtype=dtype) |
| 144 | padding_mask = (1.0 - padding_mask) * min_dtype |
| 145 | causal_mask = causal_mask + padding_mask |
| 146 | |
| 147 | return causal_mask |
| 148 | |
| 149 | @classmethod |
| 150 | @lru_cache(maxsize=8) |
| 151 | def compute_mask_base(cls, sequence_length, dtype, device): |
| 152 | # token_type_ids is the fixed pattern [0]*n_query + [1]*n_query, |
| 153 | # identical across the batch, so the mask depends only on |
| 154 | # sequence_length: img tokens (first half) attend to |
| 155 | # everything, txt tokens (second half) attend causally among |
| 156 | # themselves. lru_cache keeps one batch-invariant [1, 1, S, S] |
| 157 | # mask per (S, dtype, device). |
| 158 | min_dtype = torch.finfo(dtype).min |
| 159 | n_query = sequence_length // 2 |
| 160 | img = torch.arange(sequence_length, device=device) < n_query |
| 161 | txt = ~img |
| 162 | causal = torch.tril( |
| 163 | torch.ones( |
| 164 | sequence_length, |
| 165 | sequence_length, |
| 166 | dtype=torch.bool, |
| 167 | device=device, |
| 168 | ) |
| 169 | ) |
| 170 | allow = img[None, :] | (txt[:, None] & txt[None, :] & causal) |
| 171 | return torch.where( |
| 172 | allow, |
| 173 | torch.zeros((), dtype=dtype, device=device), |
| 174 | torch.full((), min_dtype, dtype=dtype, device=device), |
| 175 | )[None, None] |
| 176 | |
| 177 | def _create_custom_4d_mask( |
| 178 | self, |
| 179 | sequence_length, |
| 180 | dtype, |
| 181 | device, |
| 182 | batch_size, |
| 183 | ): |
| 184 | base = self.compute_mask_base(sequence_length, dtype, device) |
| 185 | return base.expand(batch_size, -1, -1, -1) |
| 186 | |
| 187 | return CustomQwen2ModelInner(config) |
| 188 | |
| 189 | def forward( |
| 190 | self, |
| 191 | inputs_embeds: torch.Tensor, |
| 192 | token_type_ids: torch.Tensor, |
| 193 | attention_mask: torch.Tensor = None, |
| 194 | **kwargs, |
| 195 | ): |
| 196 | """Args: |
| 197 | inputs_embeds: [batch_size, seq_len, hidden_dim] |
| 198 | token_type_ids: [batch_size, seq_len], 0=non-causal, 1=causal |
| 199 | attention_mask: [batch_size, seq_len], optional |
| 200 | |
| 201 | """ |
| 202 | return self.model( |
| 203 | inputs_embeds=inputs_embeds, |
| 204 | token_type_ids=token_type_ids, |
| 205 | attention_mask=attention_mask, |
| 206 | **kwargs, |
| 207 | ) |
| 208 | |
| 209 | |
| 210 | class Qwen2Decoder2Encoder(nn.Module): |
| 211 | """Decoder based on Multilingual BART |
| 212 | Set the initial weights and configuration with a pretrained multilingual BART model, |
| 213 | and modify the detailed configurations as a Nougat decoder |
| 214 | """ |
| 215 | |
| 216 | def __init__( |
| 217 | self, |
| 218 | decoder_layer: int, |
| 219 | hidden_dimension: int, |
| 220 | num_attention_heads: int, |
| 221 | num_key_value_heads: int, |
| 222 | intermediate_size: int, |
| 223 | ): |
| 224 | super().__init__() |
| 225 | |
| 226 | self.model = CustomQwen2Decoder( |
| 227 | decoder_layer=decoder_layer, |
| 228 | hidden_dimension=hidden_dimension, |
| 229 | num_attention_heads=num_attention_heads, |
| 230 | num_key_value_heads=num_key_value_heads, |
| 231 | intermediate_size=intermediate_size, |
| 232 | attn_implementation="sdpa", |
| 233 | ) |
| 234 | self.query_768 = nn.Embedding(144, hidden_dimension) |
| 235 | self.query_1024 = nn.Embedding(256, hidden_dimension) |
| 236 | |
| 237 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 238 | x = x.flatten(2).transpose(1, 2) |
| 239 | |
| 240 | bs, n_query, _ = x.shape |
| 241 | |
| 242 | if n_query == 144: |
| 243 | param_img = self.query_768.weight |
| 244 | elif n_query == 256: |
| 245 | param_img = self.query_1024.weight |
| 246 | |
| 247 | batch_query_imgs = param_img.unsqueeze(0).expand( |
| 248 | bs, -1, -1 |
| 249 | ) # (batch_size, num_queries, hidden_size) |
| 250 | |
| 251 | x_combined = torch.cat([x, batch_query_imgs], dim=1) |
| 252 | |
| 253 | token_type_ids = torch.cat( |
| 254 | [ |
| 255 | torch.zeros(bs, n_query, dtype=torch.long), |
| 256 | torch.ones(bs, n_query, dtype=torch.long), |
| 257 | ], |
| 258 | dim=1, |
| 259 | ) |
| 260 | |
| 261 | y = self.model(x_combined, token_type_ids)[0] |
| 262 | |
| 263 | y = y[:, n_query:, :] # causal flow query |
| 264 | |
| 265 | return y |
| 266 | |
| 267 | |
| 268 | def build_qwen2_decoder_as_encoder( |
| 269 | decoder_layer=24, |
| 270 | hidden_dimension=896, |
| 271 | num_attention_heads=14, |
| 272 | num_key_value_heads=2, |
| 273 | intermediate_size=4864, |
| 274 | ): |
| 275 | decoder_as_encoder = Qwen2Decoder2Encoder( |
| 276 | decoder_layer=decoder_layer, |
| 277 | hidden_dimension=hidden_dimension, |
| 278 | num_attention_heads=num_attention_heads, |
| 279 | num_key_value_heads=num_key_value_heads, |
| 280 | intermediate_size=intermediate_size, |
| 281 | ) |
| 282 | |
| 283 | return decoder_as_encoder |