modeling_qwen2.py
| 1 | # coding=utf-8 |
| 2 | # Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. |
| 3 | # |
| 4 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 5 | # and OPT implementations in this library. It has been modified from its |
| 6 | # original forms to accommodate minor architectural differences compared |
| 7 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 8 | # |
| 9 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 10 | # you may not use this file except in compliance with the License. |
| 11 | # You may obtain a copy of the License at |
| 12 | # |
| 13 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 14 | # |
| 15 | # Unless required by applicable law or agreed to in writing, software |
| 16 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 17 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 18 | # See the License for the specific language governing permissions and |
| 19 | # limitations under the License. |
| 20 | """ PyTorch Qwen2 model.""" |
| 21 | import inspect |
| 22 | import math |
| 23 | import copy |
| 24 | import warnings |
| 25 | from functools import partial |
| 26 | from typing import List, Optional, Tuple, Union |
| 27 | |
| 28 | import torch |
| 29 | import torch.nn.functional as F |
| 30 | import torch.utils.checkpoint |
| 31 | from torch import nn |
| 32 | from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss |
| 33 | |
| 34 | from transformers.activations import ACT2FN |
| 35 | from transformers.cache_utils import Cache, DynamicCache |
| 36 | from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa |
| 37 | from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast |
| 38 | from transformers.modeling_utils import PreTrainedModel |
| 39 | from transformers.utils import ( |
| 40 | add_start_docstrings, |
| 41 | add_start_docstrings_to_model_forward, |
| 42 | is_flash_attn_2_available, |
| 43 | is_flash_attn_greater_or_equal_2_10, |
| 44 | logging, |
| 45 | replace_return_docstrings, |
| 46 | ) |
| 47 | from .configuration_qwen2 import Qwen2Config |
| 48 | |
| 49 | if is_flash_attn_2_available(): |
| 50 | from flash_attn import flash_attn_func, flash_attn_varlen_func |
| 51 | from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa |
| 52 | |
| 53 | _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters) |
| 54 | |
| 55 | |
| 56 | logger = logging.get_logger(__name__) |
| 57 | |
| 58 | # Magi Attention Supported |
| 59 | _MAGI_AVAILABLE = False |
| 60 | try: |
| 61 | from magi_attention.functional.flex_flash_attn import flex_flash_attn_func |
| 62 | _MAGI_AVAILABLE = True |
| 63 | except ImportError: |
| 64 | flex_flash_attn_func = None |
| 65 | |
| 66 | |
| 67 | _CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta" |
| 68 | _CONFIG_FOR_DOC = "Qwen2Config" |
| 69 | |
| 70 | QWEN2_PRETRAINED_MODEL_ARCHIVE_LIST = [ |
| 71 | "Qwen/Qwen2-7B-beta", |
| 72 | # See all Qwen2 models at https://huggingface.co/models?filter=qwen2 |
| 73 | ] |
| 74 | |
| 75 | from .mask_sdpa_utils import ( |
| 76 | find_prefix_seq_length_by_pe, |
| 77 | update_causal_mask_with_pad_non_visible_2d, |
| 78 | update_causal_mask_for_one_gen_window_2d, |
| 79 | create_block_diff_mask_by_pe_4d, |
| 80 | find_pred_pos_from_input_ids |
| 81 | ) |
| 82 | |
| 83 | from .mask_magi_utils import build_magi_ranges |
| 84 | |
| 85 | # Copied from transformers.models.llama.modeling_llama._get_unpad_data |
| 86 | def _get_unpad_data(attention_mask): |
| 87 | seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) |
| 88 | indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() |
| 89 | max_seqlen_in_batch = seqlens_in_batch.max().item() |
| 90 | cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0)) |
| 91 | return ( |
| 92 | indices, |
| 93 | cu_seqlens, |
| 94 | max_seqlen_in_batch, |
| 95 | ) |
| 96 | |
| 97 | |
| 98 | # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2 |
| 99 | class Qwen2RMSNorm(nn.Module): |
| 100 | def __init__(self, hidden_size, eps=1e-6): |
| 101 | """ |
| 102 | Qwen2RMSNorm is equivalent to T5LayerNorm |
| 103 | """ |
| 104 | super().__init__() |
| 105 | self.weight = nn.Parameter(torch.ones(hidden_size)) |
| 106 | self.variance_epsilon = eps |
| 107 | |
| 108 | def forward(self, hidden_states): |
| 109 | input_dtype = hidden_states.dtype |
| 110 | hidden_states = hidden_states.to(torch.float32) |
| 111 | variance = hidden_states.pow(2).mean(-1, keepdim=True) |
| 112 | hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) |
| 113 | return self.weight * hidden_states.to(input_dtype) |
| 114 | |
| 115 | |
| 116 | # Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Qwen2 |
| 117 | class Qwen2RotaryEmbedding(nn.Module): |
| 118 | def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None): |
| 119 | super().__init__() |
| 120 | |
| 121 | self.dim = dim |
| 122 | self.max_position_embeddings = max_position_embeddings |
| 123 | self.base = base |
| 124 | inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)) |
| 125 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 126 | |
| 127 | # Build here to make `torch.jit.trace` work. |
| 128 | self._set_cos_sin_cache( |
| 129 | seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype() |
| 130 | ) |
| 131 | |
| 132 | def _set_cos_sin_cache(self, seq_len, device, dtype): |
| 133 | self.max_seq_len_cached = seq_len |
| 134 | t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype) |
| 135 | |
| 136 | freqs = torch.outer(t, self.inv_freq) |
| 137 | # Different from paper, but it uses a different permutation in order to obtain the same calculation |
| 138 | emb = torch.cat((freqs, freqs), dim=-1) |
| 139 | self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False) |
| 140 | self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False) |
| 141 | |
| 142 | def forward(self, x, seq_len=None): |
| 143 | # x: [bs, num_attention_heads, seq_len, head_size] |
| 144 | if seq_len > self.max_seq_len_cached: |
| 145 | self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype) |
| 146 | |
| 147 | return ( |
| 148 | self.cos_cached[:seq_len].to(dtype=x.dtype), |
| 149 | self.sin_cached[:seq_len].to(dtype=x.dtype), |
| 150 | ) |
| 151 | |
| 152 | |
| 153 | # Copied from transformers.models.llama.modeling_llama.rotate_half |
| 154 | def rotate_half(x): |
| 155 | """Rotates half the hidden dims of the input.""" |
| 156 | x1 = x[..., : x.shape[-1] // 2] |
| 157 | x2 = x[..., x.shape[-1] // 2 :] |
| 158 | return torch.cat((-x2, x1), dim=-1) |
| 159 | |
| 160 | |
| 161 | # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb |
| 162 | def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1): |
| 163 | """Applies Rotary Position Embedding to the query and key tensors. |
| 164 | |
| 165 | Args: |
| 166 | q (`torch.Tensor`): The query tensor. |
| 167 | k (`torch.Tensor`): The key tensor. |
| 168 | cos (`torch.Tensor`): The cosine part of the rotary embedding. |
| 169 | sin (`torch.Tensor`): The sine part of the rotary embedding. |
| 170 | position_ids (`torch.Tensor`): |
| 171 | The position indices of the tokens corresponding to the query and key tensors. For example, this can be |
| 172 | used to pass offsetted position ids when working with a KV-cache. |
| 173 | unsqueeze_dim (`int`, *optional*, defaults to 1): |
| 174 | The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and |
| 175 | sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note |
| 176 | that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and |
| 177 | k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes |
| 178 | cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have |
| 179 | the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. |
| 180 | Returns: |
| 181 | `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. |
| 182 | """ |
| 183 | cos = cos[position_ids].unsqueeze(unsqueeze_dim) |
| 184 | sin = sin[position_ids].unsqueeze(unsqueeze_dim) |
| 185 | q_embed = (q * cos) + (rotate_half(q) * sin) |
| 186 | k_embed = (k * cos) + (rotate_half(k) * sin) |
| 187 | return q_embed, k_embed |
| 188 | |
| 189 | |
| 190 | # Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2 |
| 191 | class Qwen2MLP(nn.Module): |
| 192 | def __init__(self, config): |
| 193 | super().__init__() |
| 194 | self.config = config |
| 195 | self.hidden_size = config.hidden_size |
| 196 | self.intermediate_size = config.intermediate_size |
| 197 | self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) |
| 198 | self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) |
| 199 | self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) |
| 200 | self.act_fn = ACT2FN[config.hidden_act] |
| 201 | |
| 202 | def forward(self, x): |
| 203 | return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) |
| 204 | |
| 205 | |
| 206 | # Copied from transformers.models.llama.modeling_llama.repeat_kv |
| 207 | def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: |
| 208 | """ |
| 209 | This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, |
| 210 | num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) |
| 211 | """ |
| 212 | batch, num_key_value_heads, slen, head_dim = hidden_states.shape |
| 213 | if n_rep == 1: |
| 214 | return hidden_states |
| 215 | hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) |
| 216 | return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) |
| 217 | |
| 218 | |
| 219 | class Qwen2Attention(nn.Module): |
| 220 | """ |
| 221 | Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer |
| 222 | and "Generating Long Sequences with Sparse Transformers". |
| 223 | """ |
| 224 | |
| 225 | def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None): |
| 226 | super().__init__() |
| 227 | self.config = config |
| 228 | self.layer_idx = layer_idx |
| 229 | if layer_idx is None: |
| 230 | logger.warning_once( |
| 231 | f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will " |
| 232 | "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` " |
| 233 | "when creating this class." |
| 234 | ) |
| 235 | |
| 236 | self.hidden_size = config.hidden_size |
| 237 | self.num_heads = config.num_attention_heads |
| 238 | self.head_dim = self.hidden_size // self.num_heads |
| 239 | self.num_key_value_heads = config.num_key_value_heads |
| 240 | self.num_key_value_groups = self.num_heads // self.num_key_value_heads |
| 241 | self.max_position_embeddings = config.max_position_embeddings |
| 242 | self.rope_theta = config.rope_theta |
| 243 | self.is_causal = True |
| 244 | self.attention_dropout = config.attention_dropout |
| 245 | |
| 246 | if (self.head_dim * self.num_heads) != self.hidden_size: |
| 247 | raise ValueError( |
| 248 | f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" |
| 249 | f" and `num_heads`: {self.num_heads})." |
| 250 | ) |
| 251 | self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True) |
| 252 | self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) |
| 253 | self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) |
| 254 | self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) |
| 255 | |
| 256 | self.rotary_emb = Qwen2RotaryEmbedding( |
| 257 | self.head_dim, |
| 258 | max_position_embeddings=self.max_position_embeddings, |
| 259 | base=self.rope_theta, |
| 260 | ) |
| 261 | |
| 262 | def forward( |
| 263 | self, |
| 264 | hidden_states: torch.Tensor, |
| 265 | attention_mask: Optional[torch.Tensor] = None, |
| 266 | position_ids: Optional[torch.LongTensor] = None, |
| 267 | past_key_value: Optional[Cache] = None, |
| 268 | output_attentions: bool = False, |
| 269 | use_cache: bool = False, |
| 270 | **kwargs, |
| 271 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 272 | if "padding_mask" in kwargs: |
| 273 | warnings.warn( |
| 274 | "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" |
| 275 | ) |
| 276 | bsz, q_len, _ = hidden_states.size() |
| 277 | |
| 278 | query_states = self.q_proj(hidden_states) |
| 279 | key_states = self.k_proj(hidden_states) |
| 280 | value_states = self.v_proj(hidden_states) |
| 281 | |
| 282 | query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 283 | key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 284 | value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 285 | |
| 286 | kv_seq_len = key_states.shape[-2] |
| 287 | if past_key_value is not None: |
| 288 | if self.layer_idx is None: |
| 289 | raise ValueError( |
| 290 | f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " |
| 291 | "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " |
| 292 | "with a layer index." |
| 293 | ) |
| 294 | kv_seq_len += past_key_value.get_seq_length(self.layer_idx) |
| 295 | cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) |
| 296 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) |
| 297 | |
| 298 | if past_key_value is not None: |
| 299 | cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models |
| 300 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 301 | |
| 302 | # repeat k/v heads if n_kv_heads < n_heads |
| 303 | key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 304 | value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 305 | |
| 306 | attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) |
| 307 | |
| 308 | if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len): |
| 309 | raise ValueError( |
| 310 | f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is" |
| 311 | f" {attn_weights.size()}" |
| 312 | ) |
| 313 | |
| 314 | if attention_mask is not None: |
| 315 | if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): |
| 316 | raise ValueError( |
| 317 | f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" |
| 318 | ) |
| 319 | |
| 320 | attn_weights = attn_weights + attention_mask |
| 321 | |
| 322 | # upcast attention to fp32 |
| 323 | attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) |
| 324 | attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) |
| 325 | attn_output = torch.matmul(attn_weights, value_states) |
| 326 | |
| 327 | if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): |
| 328 | raise ValueError( |
| 329 | f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" |
| 330 | f" {attn_output.size()}" |
| 331 | ) |
| 332 | |
| 333 | attn_output = attn_output.transpose(1, 2).contiguous() |
| 334 | attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) |
| 335 | |
| 336 | attn_output = self.o_proj(attn_output) |
| 337 | |
| 338 | if not output_attentions: |
| 339 | attn_weights = None |
| 340 | |
| 341 | return attn_output, attn_weights, past_key_value |
| 342 | |
| 343 | |
| 344 | class Qwen2FlashAttention2(Qwen2Attention): |
| 345 | """ |
| 346 | Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention` |
| 347 | as the weights of the module stays untouched. The only required change would be on the forward pass |
| 348 | where it needs to correctly call the public API of flash attention and deal with padding tokens |
| 349 | in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom |
| 350 | config.max_window_layers layers. |
| 351 | """ |
| 352 | |
| 353 | # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ |
| 354 | def __init__(self, *args, **kwargs): |
| 355 | super().__init__(*args, **kwargs) |
| 356 | |
| 357 | # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. |
| 358 | # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. |
| 359 | # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). |
| 360 | self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() |
| 361 | |
| 362 | def forward( |
| 363 | self, |
| 364 | hidden_states: torch.Tensor, |
| 365 | attention_mask: Optional[torch.Tensor] = None, |
| 366 | position_ids: Optional[torch.LongTensor] = None, |
| 367 | past_key_value: Optional[Cache] = None, |
| 368 | output_attentions: bool = False, |
| 369 | use_cache: bool = False, |
| 370 | **kwargs, |
| 371 | ): |
| 372 | if "padding_mask" in kwargs: |
| 373 | warnings.warn( |
| 374 | "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" |
| 375 | ) |
| 376 | |
| 377 | # overwrite attention_mask with padding_mask |
| 378 | attention_mask = kwargs.pop("padding_mask") |
| 379 | bsz, q_len, _ = hidden_states.size() |
| 380 | |
| 381 | query_states = self.q_proj(hidden_states) |
| 382 | key_states = self.k_proj(hidden_states) |
| 383 | value_states = self.v_proj(hidden_states) |
| 384 | |
| 385 | query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 386 | key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 387 | value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 388 | |
| 389 | kv_seq_len = key_states.shape[-2] |
| 390 | if past_key_value is not None: |
| 391 | if self.layer_idx is None: |
| 392 | raise ValueError( |
| 393 | f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " |
| 394 | "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " |
| 395 | "with a layer index." |
| 396 | ) |
| 397 | kv_seq_len += past_key_value.get_seq_length(self.layer_idx) |
| 398 | |
| 399 | # Because the input can be padded, the absolute sequence length depends on the max position id. |
| 400 | rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1 |
| 401 | cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len) |
| 402 | |
| 403 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) |
| 404 | |
| 405 | use_sliding_windows = ( |
| 406 | _flash_supports_window_size |
| 407 | and getattr(self.config, "sliding_window", None) is not None |
| 408 | and kv_seq_len > self.config.sliding_window |
| 409 | and self.config.use_sliding_window |
| 410 | ) |
| 411 | |
| 412 | if not _flash_supports_window_size: |
| 413 | logger.warning_once( |
| 414 | "The current flash attention version does not support sliding window attention, for a more memory efficient implementation" |
| 415 | " make sure to upgrade flash-attn library." |
| 416 | ) |
| 417 | |
| 418 | if past_key_value is not None: |
| 419 | # Activate slicing cache only if the config has a value `sliding_windows` attribute |
| 420 | cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0 |
| 421 | if ( |
| 422 | getattr(self.config, "sliding_window", None) is not None |
| 423 | and kv_seq_len > self.config.sliding_window |
| 424 | and cache_has_contents |
| 425 | ): |
| 426 | slicing_tokens = 1 - self.config.sliding_window |
| 427 | |
| 428 | past_key = past_key_value[self.layer_idx][0] |
| 429 | past_value = past_key_value[self.layer_idx][1] |
| 430 | |
| 431 | past_key = past_key[:, :, slicing_tokens:, :].contiguous() |
| 432 | past_value = past_value[:, :, slicing_tokens:, :].contiguous() |
| 433 | |
| 434 | if past_key.shape[-2] != self.config.sliding_window - 1: |
| 435 | raise ValueError( |
| 436 | f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got" |
| 437 | f" {past_key.shape}" |
| 438 | ) |
| 439 | |
| 440 | if attention_mask is not None: |
| 441 | attention_mask = attention_mask[:, slicing_tokens:] |
| 442 | attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) |
| 443 | |
| 444 | cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models |
| 445 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 446 | |
| 447 | # repeat k/v heads if n_kv_heads < n_heads |
| 448 | key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 449 | value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 450 | dropout_rate = 0.0 if not self.training else self.attention_dropout |
| 451 | |
| 452 | # In PEFT, usually we cast the layer norms in float32 for training stability reasons |
| 453 | # therefore the input hidden states gets silently casted in float32. Hence, we need |
| 454 | # cast them back in float16 just to be sure everything works as expected. |
| 455 | input_dtype = query_states.dtype |
| 456 | if input_dtype == torch.float32: |
| 457 | if torch.is_autocast_enabled(): |
| 458 | target_dtype = torch.get_autocast_gpu_dtype() |
| 459 | # Handle the case where the model is quantized |
| 460 | elif hasattr(self.config, "_pre_quantization_dtype"): |
| 461 | target_dtype = self.config._pre_quantization_dtype |
| 462 | else: |
| 463 | target_dtype = self.q_proj.weight.dtype |
| 464 | |
| 465 | logger.warning_once( |
| 466 | f"The input hidden states seems to be silently casted in float32, this might be related to" |
| 467 | f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" |
| 468 | f" {target_dtype}." |
| 469 | ) |
| 470 | |
| 471 | query_states = query_states.to(target_dtype) |
| 472 | key_states = key_states.to(target_dtype) |
| 473 | value_states = value_states.to(target_dtype) |
| 474 | |
| 475 | # Reashape to the expected shape for Flash Attention |
| 476 | query_states = query_states.transpose(1, 2) |
| 477 | key_states = key_states.transpose(1, 2) |
| 478 | value_states = value_states.transpose(1, 2) |
| 479 | |
| 480 | attn_output = self._flash_attention_forward( |
| 481 | query_states, |
| 482 | key_states, |
| 483 | value_states, |
| 484 | attention_mask, |
| 485 | q_len, |
| 486 | dropout=dropout_rate, |
| 487 | use_sliding_windows=use_sliding_windows, |
| 488 | ) |
| 489 | |
| 490 | attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() |
| 491 | attn_output = self.o_proj(attn_output) |
| 492 | |
| 493 | if not output_attentions: |
| 494 | attn_weights = None |
| 495 | |
| 496 | return attn_output, attn_weights, past_key_value |
| 497 | |
| 498 | def _flash_attention_forward( |
| 499 | self, |
| 500 | query_states, |
| 501 | key_states, |
| 502 | value_states, |
| 503 | attention_mask, |
| 504 | query_length, |
| 505 | dropout=0.0, |
| 506 | softmax_scale=None, |
| 507 | use_sliding_windows=False, |
| 508 | ): |
| 509 | """ |
| 510 | Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token |
| 511 | first unpad the input, then computes the attention scores and pad the final attention scores. |
| 512 | |
| 513 | Args: |
| 514 | query_states (`torch.Tensor`): |
| 515 | Input query states to be passed to Flash Attention API |
| 516 | key_states (`torch.Tensor`): |
| 517 | Input key states to be passed to Flash Attention API |
| 518 | value_states (`torch.Tensor`): |
| 519 | Input value states to be passed to Flash Attention API |
| 520 | attention_mask (`torch.Tensor`): |
| 521 | The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the |
| 522 | position of padding tokens and 1 for the position of non-padding tokens. |
| 523 | dropout (`int`, *optional*): |
| 524 | Attention dropout |
| 525 | softmax_scale (`float`, *optional*): |
| 526 | The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) |
| 527 | use_sliding_windows (`bool`, *optional*): |
| 528 | Whether to activate sliding window attention. |
| 529 | """ |
| 530 | if not self._flash_attn_uses_top_left_mask: |
| 531 | causal = self.is_causal |
| 532 | else: |
| 533 | # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__. |
| 534 | causal = self.is_causal and query_length != 1 |
| 535 | |
| 536 | # Decide whether to use SWA or not by layer index. |
| 537 | if use_sliding_windows and self.layer_idx >= self.config.max_window_layers: |
| 538 | use_sliding_windows = False |
| 539 | |
| 540 | # Contains at least one padding token in the sequence |
| 541 | if attention_mask is not None: |
| 542 | batch_size = query_states.shape[0] |
| 543 | query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( |
| 544 | query_states, key_states, value_states, attention_mask, query_length |
| 545 | ) |
| 546 | |
| 547 | cu_seqlens_q, cu_seqlens_k = cu_seq_lens |
| 548 | max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens |
| 549 | |
| 550 | if not use_sliding_windows: |
| 551 | attn_output_unpad = flash_attn_varlen_func( |
| 552 | query_states, |
| 553 | key_states, |
| 554 | value_states, |
| 555 | cu_seqlens_q=cu_seqlens_q, |
| 556 | cu_seqlens_k=cu_seqlens_k, |
| 557 | max_seqlen_q=max_seqlen_in_batch_q, |
| 558 | max_seqlen_k=max_seqlen_in_batch_k, |
| 559 | dropout_p=dropout, |
| 560 | softmax_scale=softmax_scale, |
| 561 | causal=causal, |
| 562 | ) |
| 563 | else: |
| 564 | attn_output_unpad = flash_attn_varlen_func( |
| 565 | query_states, |
| 566 | key_states, |
| 567 | value_states, |
| 568 | cu_seqlens_q=cu_seqlens_q, |
| 569 | cu_seqlens_k=cu_seqlens_k, |
| 570 | max_seqlen_q=max_seqlen_in_batch_q, |
| 571 | max_seqlen_k=max_seqlen_in_batch_k, |
| 572 | dropout_p=dropout, |
| 573 | softmax_scale=softmax_scale, |
| 574 | causal=causal, |
| 575 | window_size=(self.config.sliding_window, self.config.sliding_window), |
| 576 | ) |
| 577 | |
| 578 | attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) |
| 579 | else: |
| 580 | if not use_sliding_windows: |
| 581 | attn_output = flash_attn_func( |
| 582 | query_states, |
| 583 | key_states, |
| 584 | value_states, |
| 585 | dropout, |
| 586 | softmax_scale=softmax_scale, |
| 587 | causal=causal, |
| 588 | ) |
| 589 | else: |
| 590 | attn_output = flash_attn_func( |
| 591 | query_states, |
| 592 | key_states, |
| 593 | value_states, |
| 594 | dropout, |
| 595 | softmax_scale=softmax_scale, |
| 596 | causal=causal, |
| 597 | window_size=(self.config.sliding_window, self.config.sliding_window), |
| 598 | ) |
| 599 | |
| 600 | return attn_output |
| 601 | |
| 602 | # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input |
| 603 | def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): |
| 604 | batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape |
| 605 | |
| 606 | # On the first iteration we need to properly re-create the padding mask |
| 607 | # by slicing it on the proper place |
| 608 | if kv_seq_len != attention_mask.shape[-1]: |
| 609 | attention_mask_num_tokens = attention_mask.shape[-1] |
| 610 | attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :] |
| 611 | |
| 612 | indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) |
| 613 | |
| 614 | key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) |
| 615 | value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) |
| 616 | |
| 617 | if query_length == kv_seq_len: |
| 618 | query_layer = index_first_axis( |
| 619 | query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k |
| 620 | ) |
| 621 | cu_seqlens_q = cu_seqlens_k |
| 622 | max_seqlen_in_batch_q = max_seqlen_in_batch_k |
| 623 | indices_q = indices_k |
| 624 | elif query_length == 1: |
| 625 | max_seqlen_in_batch_q = 1 |
| 626 | cu_seqlens_q = torch.arange( |
| 627 | batch_size + 1, dtype=torch.int32, device=query_layer.device |
| 628 | ) # There is a memcpy here, that is very bad. |
| 629 | indices_q = cu_seqlens_q[:-1] |
| 630 | query_layer = query_layer.squeeze(1) |
| 631 | else: |
| 632 | # The -q_len: slice assumes left padding. |
| 633 | attention_mask = attention_mask[:, -query_length:] |
| 634 | query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) |
| 635 | |
| 636 | return ( |
| 637 | query_layer, |
| 638 | key_layer, |
| 639 | value_layer, |
| 640 | indices_q, |
| 641 | (cu_seqlens_q, cu_seqlens_k), |
| 642 | (max_seqlen_in_batch_q, max_seqlen_in_batch_k), |
| 643 | ) |
| 644 | |
| 645 | |
| 646 | # Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Qwen2 |
| 647 | class Qwen2SdpaAttention(Qwen2Attention): |
| 648 | """ |
| 649 | Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from |
| 650 | `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to |
| 651 | SDPA API. |
| 652 | """ |
| 653 | |
| 654 | # Adapted from Qwen2Attention.forward |
| 655 | def forward( |
| 656 | self, |
| 657 | hidden_states: torch.Tensor, |
| 658 | attention_mask: Optional[torch.Tensor] = None, |
| 659 | position_ids: Optional[torch.LongTensor] = None, |
| 660 | past_key_value: Optional[Cache] = None, |
| 661 | output_attentions: bool = False, |
| 662 | use_cache: bool = False, |
| 663 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 664 | if output_attentions: |
| 665 | # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. |
| 666 | logger.warning_once( |
| 667 | "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " |
| 668 | 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' |
| 669 | ) |
| 670 | return super().forward( |
| 671 | hidden_states=hidden_states, |
| 672 | attention_mask=attention_mask, |
| 673 | position_ids=position_ids, |
| 674 | past_key_value=past_key_value, |
| 675 | output_attentions=output_attentions, |
| 676 | use_cache=use_cache, |
| 677 | ) |
| 678 | |
| 679 | bsz, q_len, _ = hidden_states.size() |
| 680 | |
| 681 | query_states = self.q_proj(hidden_states) |
| 682 | key_states = self.k_proj(hidden_states) |
| 683 | value_states = self.v_proj(hidden_states) |
| 684 | |
| 685 | query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 686 | key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 687 | value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 688 | |
| 689 | kv_seq_len = key_states.shape[-2] |
| 690 | if past_key_value is not None: |
| 691 | kv_seq_len += past_key_value.get_seq_length(self.layer_idx) |
| 692 | cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) |
| 693 | |
| 694 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) |
| 695 | |
| 696 | if past_key_value is not None: |
| 697 | cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models |
| 698 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 699 | |
| 700 | key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 701 | value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 702 | |
| 703 | if attention_mask is not None: |
| 704 | if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): |
| 705 | raise ValueError( |
| 706 | f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" |
| 707 | ) |
| 708 | |
| 709 | # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, |
| 710 | # Reference: https://github.com/pytorch/pytorch/issues/112577. |
| 711 | if query_states.device.type == "cuda" and attention_mask is not None: |
| 712 | query_states = query_states.contiguous() |
| 713 | key_states = key_states.contiguous() |
| 714 | value_states = value_states.contiguous() |
| 715 | |
| 716 | attn_output = torch.nn.functional.scaled_dot_product_attention( |
| 717 | query_states, |
| 718 | key_states, |
| 719 | value_states, |
| 720 | attn_mask=attention_mask, |
| 721 | dropout_p=self.attention_dropout if self.training else 0.0, |
| 722 | is_causal=False, |
| 723 | ) |
| 724 | |
| 725 | attn_output = attn_output.transpose(1, 2).contiguous() |
| 726 | attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) |
| 727 | |
| 728 | attn_output = self.o_proj(attn_output) |
| 729 | |
| 730 | return attn_output, None, past_key_value |
| 731 | |
| 732 | |
| 733 | class Qwen2SdpaAttentionGqa(Qwen2Attention): |
| 734 | """ |
| 735 | Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from |
| 736 | `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to |
| 737 | SDPA API. |
| 738 | """ |
| 739 | |
| 740 | # Adapted from Qwen2Attention.forward |
| 741 | def forward( |
| 742 | self, |
| 743 | hidden_states: torch.Tensor, |
| 744 | attention_mask: Optional[torch.Tensor] = None, |
| 745 | position_ids: Optional[torch.LongTensor] = None, |
| 746 | past_key_value: Optional[Cache] = None, |
| 747 | output_attentions: bool = False, |
| 748 | use_cache: bool = False, |
| 749 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 750 | if output_attentions: |
| 751 | # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. |
| 752 | logger.warning_once( |
| 753 | "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " |
| 754 | 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' |
| 755 | ) |
| 756 | return super().forward( |
| 757 | hidden_states=hidden_states, |
| 758 | attention_mask=attention_mask, |
| 759 | position_ids=position_ids, |
| 760 | past_key_value=past_key_value, |
| 761 | output_attentions=output_attentions, |
| 762 | use_cache=use_cache, |
| 763 | ) |
| 764 | |
| 765 | bsz, q_len, _ = hidden_states.size() |
| 766 | |
| 767 | query_states = self.q_proj(hidden_states) |
| 768 | key_states = self.k_proj(hidden_states) |
| 769 | value_states = self.v_proj(hidden_states) |
| 770 | |
| 771 | query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 772 | key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 773 | value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 774 | |
| 775 | kv_seq_len = key_states.shape[-2] |
| 776 | if past_key_value is not None: |
| 777 | kv_seq_len += past_key_value.get_seq_length(self.layer_idx) |
| 778 | cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) |
| 779 | |
| 780 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) |
| 781 | |
| 782 | if past_key_value is not None: |
| 783 | cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models |
| 784 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 785 | |
| 786 | # key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 787 | # value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 788 | |
| 789 | if attention_mask is not None: |
| 790 | if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): |
| 791 | raise ValueError( |
| 792 | f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" |
| 793 | ) |
| 794 | |
| 795 | # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, |
| 796 | # Reference: https://github.com/pytorch/pytorch/issues/112577. |
| 797 | if query_states.device.type == "cuda" and attention_mask is not None: |
| 798 | query_states = query_states.contiguous() |
| 799 | key_states = key_states.contiguous() |
| 800 | value_states = value_states.contiguous() |
| 801 | |
| 802 | with torch.backends.cuda.sdp_kernel(enable_flash=True, |
| 803 | enable_math=True, |
| 804 | enable_mem_efficient=False): |
| 805 | |
| 806 | attn_output = torch.nn.functional.scaled_dot_product_attention( |
| 807 | query_states, |
| 808 | key_states, |
| 809 | value_states, |
| 810 | attn_mask=attention_mask, |
| 811 | enable_gqa=True, |
| 812 | dropout_p=self.attention_dropout if self.training else 0.0, |
| 813 | is_causal=False, |
| 814 | ) |
| 815 | |
| 816 | attn_output = attn_output.transpose(1, 2).contiguous() |
| 817 | attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) |
| 818 | |
| 819 | attn_output = self.o_proj(attn_output) |
| 820 | |
| 821 | return attn_output, None, past_key_value |
| 822 | |
| 823 | |
| 824 | class Qwen2MagiAttention(Qwen2Attention): |
| 825 | """ |
| 826 | Qwen2 attention using MagiAttention for efficient training with MTP packing support. |
| 827 | |
| 828 | MagiAttention uses range-based sparse attention patterns: |
| 829 | - q_ranges/k_ranges define which query/key ranges attend to each other |
| 830 | - attn_type_map specifies causal(1) or full(0) attention for each range pair |
| 831 | """ |
| 832 | |
| 833 | def __init__(self, *args, **kwargs): |
| 834 | super().__init__(*args, **kwargs) |
| 835 | if not _MAGI_AVAILABLE: |
| 836 | raise ImportError( |
| 837 | "magi_attention is not installed. Install with: pip install magi-attention" |
| 838 | ) |
| 839 | self.softmax_scale = self.head_dim ** -0.5 |
| 840 | |
| 841 | def forward( |
| 842 | self, |
| 843 | hidden_states: torch.Tensor, |
| 844 | attention_mask: Optional[dict] = None, # magi_plan dict |
| 845 | position_ids: Optional[torch.LongTensor] = None, |
| 846 | past_key_value: Optional[Cache] = None, |
| 847 | output_attentions: bool = False, |
| 848 | use_cache: bool = False, |
| 849 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 850 | if output_attentions: |
| 851 | raise NotImplementedError('MagiAttention does not support output_attentions=True') |
| 852 | |
| 853 | bsz, q_len, _ = hidden_states.size() |
| 854 | assert bsz == 1, "MagiAttention only supports batch_size=1 (use packing instead)" |
| 855 | |
| 856 | query_states = self.q_proj(hidden_states) |
| 857 | key_states = self.k_proj(hidden_states) |
| 858 | value_states = self.v_proj(hidden_states) |
| 859 | |
| 860 | # Magi expects [T, H, D] format (no batch dimension) |
| 861 | query_states = query_states.view(q_len, self.num_heads, self.head_dim) |
| 862 | key_states = key_states.view(q_len, self.num_key_value_heads, self.head_dim) |
| 863 | value_states = value_states.view(q_len, self.num_key_value_heads, self.head_dim) |
| 864 | |
| 865 | kv_seq_len = q_len |
| 866 | if past_key_value is not None: |
| 867 | if self.layer_idx is None: |
| 868 | raise ValueError( |
| 869 | f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " |
| 870 | "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " |
| 871 | "with a layer index." |
| 872 | ) |
| 873 | kv_seq_len += past_key_value.get_seq_length(self.layer_idx) |
| 874 | |
| 875 | cos, sin = self.rotary_emb(value_states.unsqueeze(0).transpose(1, 2), seq_len=kv_seq_len) |
| 876 | |
| 877 | # Apply RoPE: need [B, H, L, D] format for apply_rotary_pos_emb |
| 878 | q_for_rope = query_states.unsqueeze(0).transpose(1, 2) # [1, H, L, D] |
| 879 | k_for_rope = key_states.unsqueeze(0).transpose(1, 2) # [1, Hkv, L, D] |
| 880 | q_for_rope, k_for_rope = apply_rotary_pos_emb(q_for_rope, k_for_rope, cos, sin, position_ids) |
| 881 | |
| 882 | # Back to [T, H, D] |
| 883 | query_states = q_for_rope.squeeze(0).transpose(0, 1).contiguous() # [L, H, D] |
| 884 | key_states = k_for_rope.squeeze(0).transpose(0, 1).contiguous() # [L, Hkv, D] |
| 885 | |
| 886 | if past_key_value is not None: |
| 887 | cache_kwargs = {"sin": sin, "cos": cos} |
| 888 | # Note: Magi doesn't support KV cache in training, this is for potential future use |
| 889 | key_states_4d = key_states.unsqueeze(0).transpose(1, 2) |
| 890 | value_states_4d = value_states.unsqueeze(0).transpose(1, 2) |
| 891 | key_states_4d, value_states_4d = past_key_value.update( |
| 892 | key_states_4d, value_states_4d, self.layer_idx, cache_kwargs |
| 893 | ) |
| 894 | key_states = key_states_4d.squeeze(0).transpose(0, 1).contiguous() |
| 895 | value_states = value_states_4d.squeeze(0).transpose(0, 1).contiguous() |
| 896 | |
| 897 | # Run Magi Attention |
| 898 | # attention_mask is a magi_plan dict with q_ranges, k_ranges, attn_type_map, etc. |
| 899 | |
| 900 | attn_output, _ = flex_flash_attn_func( |
| 901 | query_states.contiguous(), |
| 902 | key_states.contiguous(), |
| 903 | value_states.contiguous(), |
| 904 | q_ranges=attention_mask["q_ranges"], |
| 905 | k_ranges=attention_mask["k_ranges"], |
| 906 | attn_type_map=attention_mask["attn_type_map"], |
| 907 | softmax_scale=self.softmax_scale, |
| 908 | softcap=0.0, |
| 909 | deterministic=False, |
| 910 | ) # [T, H, D] |
| 911 | |
| 912 | # Reshape to [B, L, H*D] |
| 913 | attn_output = attn_output.view(1, q_len, self.hidden_size) |
| 914 | attn_output = self.o_proj(attn_output) |
| 915 | |
| 916 | return attn_output, None, past_key_value |
| 917 | |
| 918 | |
| 919 | QWEN2_ATTENTION_CLASSES = { |
| 920 | "eager": Qwen2Attention, |
| 921 | "flash_attention_2": Qwen2FlashAttention2, |
| 922 | "sdpa": Qwen2SdpaAttention, |
| 923 | "magi": Qwen2MagiAttention, |
| 924 | } |
| 925 | |
| 926 | |
| 927 | class Qwen2DecoderLayer(nn.Module): |
| 928 | def __init__(self, config: Qwen2Config, layer_idx: int): |
| 929 | super().__init__() |
| 930 | self.hidden_size = config.hidden_size |
| 931 | |
| 932 | if config._attn_implementation == 'magi' and not _MAGI_AVAILABLE: |
| 933 | if is_flash_attn_2_available(): |
| 934 | logger.warning_once( |
| 935 | 'magi_attention not available, falling back to flash_attention_2' |
| 936 | ) |
| 937 | config._attn_implementation = 'flash_attention_2' |
| 938 | else: |
| 939 | logger.warning_once( |
| 940 | 'magi_attention not available, falling back to sdpa' |
| 941 | ) |
| 942 | config._attn_implementation = 'sdpa' |
| 943 | if config._attn_implementation == 'flash_attention_2' and not is_flash_attn_2_available(): |
| 944 | logger.warning_once( |
| 945 | 'flash_attn is not available, falling back to sdpa' |
| 946 | ) |
| 947 | config._attn_implementation = 'sdpa' |
| 948 | |
| 949 | self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx) |
| 950 | |
| 951 | self.mlp = Qwen2MLP(config) |
| 952 | self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 953 | self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 954 | |
| 955 | def forward( |
| 956 | self, |
| 957 | hidden_states: torch.Tensor, |
| 958 | attention_mask: Optional[torch.Tensor] = None, |
| 959 | position_ids: Optional[torch.LongTensor] = None, |
| 960 | past_key_value: Optional[Tuple[torch.Tensor]] = None, |
| 961 | output_attentions: Optional[bool] = False, |
| 962 | use_cache: Optional[bool] = False, |
| 963 | **kwargs, |
| 964 | ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: |
| 965 | if "padding_mask" in kwargs: |
| 966 | warnings.warn( |
| 967 | "Passing `padding_mask` is deprecated and will be removed in v4.37. " |
| 968 | "Please make sure use `attention_mask` instead.`" |
| 969 | ) |
| 970 | """ |
| 971 | Args: |
| 972 | hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` |
| 973 | attention_mask (`torch.FloatTensor`, *optional*): attention mask of size |
| 974 | `(batch, sequence_length)` where padding elements are indicated by 0. |
| 975 | output_attentions (`bool`, *optional*): |
| 976 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under |
| 977 | returned tensors for more detail. |
| 978 | use_cache (`bool`, *optional*): |
| 979 | If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding |
| 980 | (see `past_key_values`). |
| 981 | past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states |
| 982 | """ |
| 983 | |
| 984 | residual = hidden_states |
| 985 | |
| 986 | hidden_states = self.input_layernorm(hidden_states) |
| 987 | |
| 988 | # Self Attention |
| 989 | hidden_states, self_attn_weights, present_key_value = self.self_attn( |
| 990 | hidden_states=hidden_states, |
| 991 | attention_mask=attention_mask, |
| 992 | position_ids=position_ids, |
| 993 | past_key_value=past_key_value, |
| 994 | output_attentions=output_attentions, |
| 995 | use_cache=use_cache, |
| 996 | ) |
| 997 | hidden_states = residual + hidden_states |
| 998 | |
| 999 | # Fully Connected |
| 1000 | residual = hidden_states |
| 1001 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 1002 | hidden_states = self.mlp(hidden_states) |
| 1003 | hidden_states = residual + hidden_states |
| 1004 | |
| 1005 | outputs = (hidden_states,) |
| 1006 | |
| 1007 | if output_attentions: |
| 1008 | outputs += (self_attn_weights,) |
| 1009 | |
| 1010 | if use_cache: |
| 1011 | outputs += (present_key_value,) |
| 1012 | |
| 1013 | return outputs |
| 1014 | |
| 1015 | |
| 1016 | QWEN2_START_DOCSTRING = r""" |
| 1017 | This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the |
| 1018 | library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads |
| 1019 | etc.) |
| 1020 | |
| 1021 | This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. |
| 1022 | Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage |
| 1023 | and behavior. |
| 1024 | |
| 1025 | Parameters: |
| 1026 | config ([`Qwen2Config`]): |
| 1027 | Model configuration class with all the parameters of the model. Initializing with a config file does not |
| 1028 | load the weights associated with the model, only the configuration. Check out the |
| 1029 | [`~PreTrainedModel.from_pretrained`] method to load the model weights. |
| 1030 | """ |
| 1031 | |
| 1032 | |
| 1033 | @add_start_docstrings( |
| 1034 | "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.", |
| 1035 | QWEN2_START_DOCSTRING, |
| 1036 | ) |
| 1037 | class Qwen2PreTrainedModel(PreTrainedModel): |
| 1038 | config_class = Qwen2Config |
| 1039 | base_model_prefix = "model" |
| 1040 | supports_gradient_checkpointing = True |
| 1041 | _no_split_modules = ["Qwen2DecoderLayer"] |
| 1042 | _skip_keys_device_placement = "past_key_values" |
| 1043 | _supports_flash_attn_2 = True |
| 1044 | _supports_sdpa = True |
| 1045 | _supports_cache_class = True |
| 1046 | |
| 1047 | @classmethod |
| 1048 | def _autoset_attn_implementation(cls, config, *args, **kwargs): |
| 1049 | if getattr(config, '_attn_implementation', None) == 'magi': |
| 1050 | return config |
| 1051 | return super()._autoset_attn_implementation(config, *args, **kwargs) |
| 1052 | |
| 1053 | def _check_and_adjust_attn_implementation(self, attn_implementation, is_init_check=False): |
| 1054 | if attn_implementation == "magi": |
| 1055 | return "magi" |
| 1056 | return super()._check_and_adjust_attn_implementation(attn_implementation, is_init_check) |
| 1057 | |
| 1058 | def _init_weights(self, module): |
| 1059 | std = self.config.initializer_range |
| 1060 | if isinstance(module, nn.Linear): |
| 1061 | module.weight.data.normal_(mean=0.0, std=std) |
| 1062 | if module.bias is not None: |
| 1063 | module.bias.data.zero_() |
| 1064 | elif isinstance(module, nn.Embedding): |
| 1065 | module.weight.data.normal_(mean=0.0, std=std) |
| 1066 | if module.padding_idx is not None: |
| 1067 | module.weight.data[module.padding_idx].zero_() |
| 1068 | |
| 1069 | |
| 1070 | QWEN2_INPUTS_DOCSTRING = r""" |
| 1071 | Args: |
| 1072 | input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): |
| 1073 | Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide |
| 1074 | it. |
| 1075 | |
| 1076 | Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and |
| 1077 | [`PreTrainedTokenizer.__call__`] for details. |
| 1078 | |
| 1079 | [What are input IDs?](../glossary#input-ids) |
| 1080 | attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): |
| 1081 | Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: |
| 1082 | |
| 1083 | - 1 for tokens that are **not masked**, |
| 1084 | - 0 for tokens that are **masked**. |
| 1085 | |
| 1086 | [What are attention masks?](../glossary#attention-mask) |
| 1087 | |
| 1088 | Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and |
| 1089 | [`PreTrainedTokenizer.__call__`] for details. |
| 1090 | |
| 1091 | If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see |
| 1092 | `past_key_values`). |
| 1093 | |
| 1094 | If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] |
| 1095 | and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more |
| 1096 | information on the default strategy. |
| 1097 | |
| 1098 | - 1 indicates the head is **not masked**, |
| 1099 | - 0 indicates the head is **masked**. |
| 1100 | position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): |
| 1101 | Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, |
| 1102 | config.n_positions - 1]`. |
| 1103 | |
| 1104 | [What are position IDs?](../glossary#position-ids) |
| 1105 | past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): |
| 1106 | Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention |
| 1107 | blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` |
| 1108 | returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. |
| 1109 | |
| 1110 | Two formats are allowed: |
| 1111 | - a [`~cache_utils.Cache`] instance; |
| 1112 | - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of |
| 1113 | shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy |
| 1114 | cache format. |
| 1115 | |
| 1116 | The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the |
| 1117 | legacy cache format will be returned. |
| 1118 | |
| 1119 | If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't |
| 1120 | have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` |
| 1121 | of shape `(batch_size, sequence_length)`. |
| 1122 | inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): |
| 1123 | Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This |
| 1124 | is useful if you want more control over how to convert `input_ids` indices into associated vectors than the |
| 1125 | model's internal embedding lookup matrix. |
| 1126 | use_cache (`bool`, *optional*): |
| 1127 | If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see |
| 1128 | `past_key_values`). |
| 1129 | output_attentions (`bool`, *optional*): |
| 1130 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned |
| 1131 | tensors for more detail. |
| 1132 | output_hidden_states (`bool`, *optional*): |
| 1133 | Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for |
| 1134 | more detail. |
| 1135 | return_dict (`bool`, *optional*): |
| 1136 | Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. |
| 1137 | """ |
| 1138 | |
| 1139 | |
| 1140 | @add_start_docstrings( |
| 1141 | "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.", |
| 1142 | QWEN2_START_DOCSTRING, |
| 1143 | ) |
| 1144 | class Qwen2Model(Qwen2PreTrainedModel): |
| 1145 | """ |
| 1146 | Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`] |
| 1147 | |
| 1148 | Args: |
| 1149 | config: Qwen2Config |
| 1150 | """ |
| 1151 | |
| 1152 | def __init__(self, config: Qwen2Config): |
| 1153 | super().__init__(config) |
| 1154 | self.padding_idx = config.pad_token_id |
| 1155 | self.vocab_size = config.vocab_size |
| 1156 | |
| 1157 | self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) |
| 1158 | self.layers = nn.ModuleList( |
| 1159 | [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] |
| 1160 | ) |
| 1161 | self._attn_implementation = config._attn_implementation |
| 1162 | self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1163 | |
| 1164 | self.gradient_checkpointing = False |
| 1165 | # Initialize weights and apply final processing |
| 1166 | self.post_init() |
| 1167 | |
| 1168 | self.block_size = getattr(config, 'block_size', 6) |
| 1169 | self.causal_attn = getattr(config, 'causal_attn', False) |
| 1170 | self.text_mask_token_id = getattr(config, 'text_mask_token_id', 151676) |
| 1171 | |
| 1172 | |
| 1173 | def get_input_embeddings(self): |
| 1174 | return self.embed_tokens |
| 1175 | |
| 1176 | def set_input_embeddings(self, value): |
| 1177 | self.embed_tokens = value |
| 1178 | |
| 1179 | def image_processing(self, input_ids, visual_features, image_token_index): |
| 1180 | if visual_features is not None: |
| 1181 | input_embeds = self.get_input_embeddings()(input_ids) |
| 1182 | B, N, C = input_embeds.shape |
| 1183 | input_embeds = input_embeds.reshape(B * N, C) |
| 1184 | |
| 1185 | input_ids = input_ids.reshape(B * N) |
| 1186 | selected = (input_ids == image_token_index) |
| 1187 | assert selected.sum() != 0 |
| 1188 | input_embeds[selected] = visual_features.reshape(-1, C).to(input_embeds.device) |
| 1189 | input_embeds = input_embeds.reshape(B, N, C) |
| 1190 | else: |
| 1191 | input_embeds = self.get_input_embeddings()(input_ids) |
| 1192 | return input_embeds |
| 1193 | |
| 1194 | @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) |
| 1195 | def forward( |
| 1196 | self, |
| 1197 | input_ids: torch.LongTensor = None, |
| 1198 | visual_features: Optional[torch.FloatTensor] = None, |
| 1199 | image_token_index: int = None, |
| 1200 | attention_mask: Optional[torch.Tensor] = None, |
| 1201 | position_ids: Optional[torch.LongTensor] = None, |
| 1202 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 1203 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 1204 | use_cache: Optional[bool] = None, |
| 1205 | output_attentions: Optional[bool] = None, |
| 1206 | output_hidden_states: Optional[bool] = None, |
| 1207 | return_dict: Optional[bool] = None, |
| 1208 | ) -> Union[Tuple, BaseModelOutputWithPast]: |
| 1209 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions |
| 1210 | output_hidden_states = ( |
| 1211 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states |
| 1212 | ) |
| 1213 | use_cache = use_cache if use_cache is not None else self.config.use_cache |
| 1214 | |
| 1215 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 1216 | |
| 1217 | # retrieve input_ids and inputs_embeds |
| 1218 | if input_ids is not None and inputs_embeds is not None: |
| 1219 | raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") |
| 1220 | elif input_ids is not None: |
| 1221 | batch_size, seq_length = input_ids.shape |
| 1222 | elif inputs_embeds is not None: |
| 1223 | batch_size, seq_length, _ = inputs_embeds.shape |
| 1224 | else: |
| 1225 | raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") |
| 1226 | |
| 1227 | if self.gradient_checkpointing and self.training: |
| 1228 | if use_cache: |
| 1229 | logger.warning_once( |
| 1230 | "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." |
| 1231 | ) |
| 1232 | use_cache = False |
| 1233 | |
| 1234 | past_key_values_length = 0 |
| 1235 | |
| 1236 | if use_cache: |
| 1237 | use_legacy_cache = not isinstance(past_key_values, Cache) |
| 1238 | if use_legacy_cache: |
| 1239 | if past_key_values is None: |
| 1240 | past_key_values = DynamicCache() |
| 1241 | else: |
| 1242 | past_key_values = DynamicCache.from_legacy_cache(past_key_values) |
| 1243 | past_key_values_length = past_key_values.get_seq_length() |
| 1244 | |
| 1245 | if position_ids is None: |
| 1246 | device = input_ids.device if input_ids is not None else inputs_embeds.device |
| 1247 | position_ids = torch.arange( |
| 1248 | past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device |
| 1249 | ) |
| 1250 | position_ids = position_ids.unsqueeze(0).view(-1, seq_length) |
| 1251 | else: |
| 1252 | position_ids = position_ids.view(-1, seq_length).long() |
| 1253 | |
| 1254 | if inputs_embeds is None: |
| 1255 | inputs_embeds = self.image_processing(input_ids, visual_features, image_token_index) |
| 1256 | |
| 1257 | if attention_mask is not None and self._attn_implementation == "magi" and use_cache: |
| 1258 | is_padding_right = attention_mask[:, -1].sum().item() != batch_size |
| 1259 | if is_padding_right: |
| 1260 | raise ValueError( |
| 1261 | "You are attempting to perform batched generation with padding_side='right'" |
| 1262 | " this may lead to unexpected behaviour for Flash Attention version of Qwen2. Make sure to " |
| 1263 | " call `tokenizer.padding_side = 'left'` before tokenizing the input. " |
| 1264 | ) |
| 1265 | |
| 1266 | device = input_ids.device if input_ids is not None else inputs_embeds.device |
| 1267 | |
| 1268 | x0_len = find_prefix_seq_length_by_pe(position_ids).to(device=device) |
| 1269 | |
| 1270 | def _prepare_block_mask_for_inference(attention_mask): |
| 1271 | attention_mask = _prepare_4d_causal_attention_mask( |
| 1272 | attention_mask, |
| 1273 | (batch_size, seq_length), |
| 1274 | inputs_embeds, |
| 1275 | past_key_values_length, |
| 1276 | sliding_window=self.config.sliding_window, |
| 1277 | ) |
| 1278 | # switch to ar mode |
| 1279 | if seq_length == 1 or (input_ids is not None and input_ids[0][-1].item() != self.text_mask_token_id): |
| 1280 | return attention_mask |
| 1281 | |
| 1282 | |
| 1283 | if attention_mask is None or len(attention_mask.shape) != 4: |
| 1284 | return attention_mask |
| 1285 | |
| 1286 | # For SDLM, the generation window should set to bidirectional attention |
| 1287 | if use_cache: |
| 1288 | update_mask_func = partial( |
| 1289 | update_causal_mask_for_one_gen_window_2d, |
| 1290 | block_size=self.block_size, |
| 1291 | use_cache=use_cache, |
| 1292 | causal_attn=self.causal_attn, |
| 1293 | ) |
| 1294 | else: |
| 1295 | update_mask_func = partial( |
| 1296 | update_causal_mask_with_pad_non_visible_2d, |
| 1297 | block_size=self.block_size, |
| 1298 | text_mask_token_id=self.text_mask_token_id, |
| 1299 | causal_attn=self.causal_attn, |
| 1300 | ) |
| 1301 | |
| 1302 | new_attention_mask = [] |
| 1303 | for b in range(attention_mask.shape[0]): |
| 1304 | new_attention_mask.append( |
| 1305 | update_mask_func( |
| 1306 | input_ids[b], |
| 1307 | attention_mask[b][0], |
| 1308 | ).unsqueeze(0) |
| 1309 | ) |
| 1310 | return torch.stack(new_attention_mask, dim=0) |
| 1311 | |
| 1312 | def _prepare_block_mask_for_training(): |
| 1313 | block_mask, _ = create_block_diff_mask_by_pe_4d( |
| 1314 | block_size=self.block_size, |
| 1315 | x0_len_list=x0_len, |
| 1316 | position_ids=position_ids, |
| 1317 | causal_attn=self.causal_attn, |
| 1318 | ) |
| 1319 | return block_mask |
| 1320 | |
| 1321 | if self._attn_implementation == "magi": |
| 1322 | ar_decode = seq_length == 1 or (input_ids is not None and input_ids[0][-1].item() != self.text_mask_token_id) |
| 1323 | attention_mask = build_magi_ranges( |
| 1324 | kv_len=seq_length + past_key_values_length, |
| 1325 | q_len=seq_length, |
| 1326 | block_size=self.block_size, |
| 1327 | ar_decode=ar_decode, |
| 1328 | device=device |
| 1329 | ) |
| 1330 | |
| 1331 | elif self._attn_implementation == "sdpa": |
| 1332 | attention_mask = _prepare_block_mask_for_training() if self.training else _prepare_block_mask_for_inference(attention_mask) |
| 1333 | |
| 1334 | else: |
| 1335 | raise NotImplementedError(f'{self._attn_implementation=}') |
| 1336 | |
| 1337 | |
| 1338 | hidden_states = inputs_embeds |
| 1339 | |
| 1340 | # decoder layers |
| 1341 | all_hidden_states = () if output_hidden_states else None |
| 1342 | all_self_attns = () if output_attentions else None |
| 1343 | next_decoder_cache = None |
| 1344 | |
| 1345 | for decoder_layer in self.layers: |
| 1346 | if output_hidden_states: |
| 1347 | all_hidden_states += (hidden_states,) |
| 1348 | |
| 1349 | if self.gradient_checkpointing and self.training: |
| 1350 | layer_outputs = self._gradient_checkpointing_func( |
| 1351 | decoder_layer.__call__, |
| 1352 | hidden_states, |
| 1353 | attention_mask, |
| 1354 | position_ids, |
| 1355 | past_key_values, |
| 1356 | output_attentions, |
| 1357 | use_cache, |
| 1358 | ) |
| 1359 | else: |
| 1360 | layer_outputs = decoder_layer( |
| 1361 | hidden_states, |
| 1362 | attention_mask=attention_mask, |
| 1363 | position_ids=position_ids, |
| 1364 | past_key_value=past_key_values, |
| 1365 | output_attentions=output_attentions, |
| 1366 | use_cache=use_cache, |
| 1367 | ) |
| 1368 | |
| 1369 | hidden_states = layer_outputs[0] |
| 1370 | |
| 1371 | if use_cache: |
| 1372 | next_decoder_cache = layer_outputs[2 if output_attentions else 1] |
| 1373 | |
| 1374 | if output_attentions: |
| 1375 | all_self_attns += (layer_outputs[1],) |
| 1376 | |
| 1377 | hidden_states = self.norm(hidden_states) |
| 1378 | |
| 1379 | # add hidden states from the last decoder layer |
| 1380 | if output_hidden_states: |
| 1381 | all_hidden_states += (hidden_states,) |
| 1382 | |
| 1383 | next_cache = None |
| 1384 | if use_cache: |
| 1385 | next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache |
| 1386 | |
| 1387 | if not return_dict: |
| 1388 | return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) |
| 1389 | return BaseModelOutputWithPast( |
| 1390 | last_hidden_state=hidden_states, |
| 1391 | past_key_values=next_cache, |
| 1392 | hidden_states=all_hidden_states, |
| 1393 | attentions=all_self_attns, |
| 1394 | ) |
| 1395 | |
| 1396 | |
| 1397 | class Qwen2ForCausalLM(Qwen2PreTrainedModel): |
| 1398 | _tied_weights_keys = ["lm_head.weight"] |
| 1399 | |
| 1400 | def __init__(self, config): |
| 1401 | super().__init__(config) |
| 1402 | self.model = Qwen2Model(config) |
| 1403 | self.vocab_size = config.vocab_size |
| 1404 | self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) |
| 1405 | |
| 1406 | self.text_mask_token_id = getattr(config, 'text_mask_token_id', 151676) |
| 1407 | |
| 1408 | # Initialize weights and apply final processing |
| 1409 | self.post_init() |
| 1410 | |
| 1411 | |
| 1412 | def get_input_embeddings(self): |
| 1413 | return self.model.embed_tokens |
| 1414 | |
| 1415 | def set_input_embeddings(self, value): |
| 1416 | self.model.embed_tokens = value |
| 1417 | |
| 1418 | def get_output_embeddings(self): |
| 1419 | return self.lm_head |
| 1420 | |
| 1421 | def set_output_embeddings(self, new_embeddings): |
| 1422 | self.lm_head = new_embeddings |
| 1423 | |
| 1424 | def set_decoder(self, decoder): |
| 1425 | self.model = decoder |
| 1426 | |
| 1427 | def get_decoder(self): |
| 1428 | return self.model |
| 1429 | |
| 1430 | @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) |
| 1431 | @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) |
| 1432 | def forward( |
| 1433 | self, |
| 1434 | input_ids: torch.LongTensor = None, |
| 1435 | visual_features: Optional[torch.FloatTensor] = None, |
| 1436 | image_token_index: int = None, |
| 1437 | attention_mask: Optional[torch.Tensor] = None, |
| 1438 | position_ids: Optional[torch.LongTensor] = None, |
| 1439 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 1440 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 1441 | labels: Optional[torch.LongTensor] = None, |
| 1442 | use_cache: Optional[bool] = None, |
| 1443 | output_attentions: Optional[bool] = None, |
| 1444 | output_hidden_states: Optional[bool] = None, |
| 1445 | return_dict: Optional[bool] = None, |
| 1446 | ) -> Union[Tuple, CausalLMOutputWithPast]: |
| 1447 | r""" |
| 1448 | Args: |
| 1449 | labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): |
| 1450 | Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., |
| 1451 | config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored |
| 1452 | (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. |
| 1453 | |
| 1454 | Returns: |
| 1455 | |
| 1456 | Example: |
| 1457 | |
| 1458 | ```python |
| 1459 | >>> from transformers import AutoTokenizer, Qwen2ForCausalLM |
| 1460 | |
| 1461 | >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS) |
| 1462 | >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER) |
| 1463 | |
| 1464 | >>> prompt = "Hey, are you conscious? Can you talk to me?" |
| 1465 | >>> inputs = tokenizer(prompt, return_tensors="pt") |
| 1466 | |
| 1467 | >>> # Generate |
| 1468 | >>> generate_ids = model.generate(inputs.input_ids, max_length=30) |
| 1469 | >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] |
| 1470 | "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." |
| 1471 | ```""" |
| 1472 | |
| 1473 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions |
| 1474 | output_hidden_states = ( |
| 1475 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states |
| 1476 | ) |
| 1477 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 1478 | |
| 1479 | # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) |
| 1480 | outputs = self.model( |
| 1481 | input_ids=input_ids, |
| 1482 | visual_features=visual_features, |
| 1483 | image_token_index=image_token_index, |
| 1484 | attention_mask=attention_mask, |
| 1485 | position_ids=position_ids, |
| 1486 | past_key_values=past_key_values, |
| 1487 | inputs_embeds=inputs_embeds, |
| 1488 | use_cache=use_cache, |
| 1489 | output_attentions=output_attentions, |
| 1490 | output_hidden_states=output_hidden_states, |
| 1491 | return_dict=return_dict, |
| 1492 | ) |
| 1493 | |
| 1494 | hidden_states = outputs[0] |
| 1495 | logits = self.lm_head(hidden_states) |
| 1496 | logits = logits.float() |
| 1497 | |
| 1498 | loss = None |
| 1499 | if labels is not None: |
| 1500 | |
| 1501 | # Shift so that tokens < n predict n |
| 1502 | shift_logits = logits[..., :-1, :].contiguous() |
| 1503 | shift_labels = labels[..., 1:].contiguous() |
| 1504 | |
| 1505 | # Flatten the tokens |
| 1506 | loss_fct = CrossEntropyLoss() |
| 1507 | shift_logits = shift_logits.view(-1, self.config.vocab_size) |
| 1508 | |
| 1509 | shift_labels = shift_labels.view(-1) |
| 1510 | shift_labels = shift_labels.to(shift_logits.device) |
| 1511 | loss = loss_fct(shift_logits, shift_labels) |
| 1512 | |
| 1513 | pos_masks = find_pred_pos_from_input_ids(input_ids, text_mask_token_id=self.text_mask_token_id) |
| 1514 | shift_input_ids = input_ids[..., :-1].contiguous() |
| 1515 | shift_pos_masks = pos_masks[:, :-1] |
| 1516 | shift_input_ids = shift_input_ids.view(-1) |
| 1517 | max_n_future_tokens = min(4, self.model.block_size) |
| 1518 | pos_loss_list = torch.zeros(max_n_future_tokens, device=shift_logits.device) |
| 1519 | shift_pos_masks = shift_pos_masks.reshape(-1) |
| 1520 | |
| 1521 | for ix in range(max_n_future_tokens): |
| 1522 | seg_loss = F.cross_entropy( |
| 1523 | shift_logits[shift_pos_masks == ix], |
| 1524 | shift_labels[shift_pos_masks == ix], |
| 1525 | reduction='mean' |
| 1526 | ) |
| 1527 | pos_loss_list[ix] = seg_loss |
| 1528 | |
| 1529 | |
| 1530 | if not return_dict: |
| 1531 | output = (logits,) + outputs[1:] |
| 1532 | return (loss,) + output if loss is not None else output |
| 1533 | |
| 1534 | if self.training: |
| 1535 | return CausalLMOutputWithPast( |
| 1536 | loss=loss, |
| 1537 | logits=logits, |
| 1538 | past_key_values=outputs.past_key_values, |
| 1539 | hidden_states=outputs.hidden_states, |
| 1540 | attentions=outputs.attentions, |
| 1541 | ), pos_loss_list |
| 1542 | |
| 1543 | return CausalLMOutputWithPast( |
| 1544 | loss=loss, |
| 1545 | logits=logits, |
| 1546 | past_key_values=outputs.past_key_values, |
| 1547 | hidden_states=outputs.hidden_states, |
| 1548 | attentions=outputs.attentions, |
| 1549 | ) |
| 1550 | |
| 1551 | def prepare_inputs_for_generation( |
| 1552 | self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs |
| 1553 | ): |
| 1554 | # Omit tokens covered by past_key_values |
| 1555 | if past_key_values is not None: |
| 1556 | if isinstance(past_key_values, Cache): |
| 1557 | cache_length = past_key_values.get_seq_length() |
| 1558 | past_length = past_key_values.seen_tokens |
| 1559 | max_cache_length = past_key_values.get_max_length() |
| 1560 | else: |
| 1561 | cache_length = past_length = past_key_values[0][0].shape[2] |
| 1562 | max_cache_length = None |
| 1563 | |
| 1564 | # Keep only the unprocessed tokens: |
| 1565 | # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where |
| 1566 | # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as |
| 1567 | # input) |
| 1568 | if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: |
| 1569 | input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] |
| 1570 | # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard |
| 1571 | # input_ids based on the past_length. |
| 1572 | elif past_length < input_ids.shape[1]: |
| 1573 | input_ids = input_ids[:, past_length:] |
| 1574 | # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. |
| 1575 | |
| 1576 | # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. |
| 1577 | if ( |
| 1578 | max_cache_length is not None |
| 1579 | and attention_mask is not None |
| 1580 | and cache_length + input_ids.shape[1] > max_cache_length |
| 1581 | ): |
| 1582 | attention_mask = attention_mask[:, -max_cache_length:] |
| 1583 | |
| 1584 | position_ids = kwargs.get("position_ids", None) |
| 1585 | if attention_mask is not None and position_ids is None: |
| 1586 | # create position_ids on the fly for batch generation |
| 1587 | position_ids = attention_mask.long().cumsum(-1) - 1 |
| 1588 | position_ids.masked_fill_(attention_mask == 0, 1) |
| 1589 | if past_key_values: |
| 1590 | position_ids = position_ids[:, -input_ids.shape[1] :] |
| 1591 | |
| 1592 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step |
| 1593 | if inputs_embeds is not None and past_key_values is None: |
| 1594 | model_inputs = {"inputs_embeds": inputs_embeds} |
| 1595 | else: |
| 1596 | model_inputs = {"input_ids": input_ids} |
| 1597 | |
| 1598 | model_inputs.update( |
| 1599 | { |
| 1600 | "position_ids": position_ids, |
| 1601 | "past_key_values": past_key_values, |
| 1602 | "use_cache": kwargs.get("use_cache"), |
| 1603 | "attention_mask": attention_mask, |
| 1604 | } |
| 1605 | ) |
| 1606 | return model_inputs |
| 1607 | |
| 1608 | @staticmethod |
| 1609 | def _reorder_cache(past_key_values, beam_idx): |
| 1610 | reordered_past = () |
| 1611 | for layer_past in past_key_values: |
| 1612 | reordered_past += ( |
| 1613 | tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), |
| 1614 | ) |
| 1615 | return reordered_past |
| 1616 | |
| 1617 | |
| 1618 | @add_start_docstrings( |
| 1619 | """ |
| 1620 | The Qwen2 Model transformer with a sequence classification head on top (linear layer). |
| 1621 | |
| 1622 | [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models |
| 1623 | (e.g. GPT-2) do. |
| 1624 | |
| 1625 | Since it does classification on the last token, it requires to know the position of the last token. If a |
| 1626 | `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If |
| 1627 | no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the |
| 1628 | padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in |
| 1629 | each row of the batch). |
| 1630 | """, |
| 1631 | QWEN2_START_DOCSTRING, |
| 1632 | ) |
| 1633 | class Qwen2ForSequenceClassification(Qwen2PreTrainedModel): |
| 1634 | def __init__(self, config): |
| 1635 | super().__init__(config) |
| 1636 | self.num_labels = config.num_labels |
| 1637 | self.model = Qwen2Model(config) |
| 1638 | self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) |
| 1639 | |
| 1640 | # Initialize weights and apply final processing |
| 1641 | self.post_init() |
| 1642 | |
| 1643 | def get_input_embeddings(self): |
| 1644 | return self.model.embed_tokens |
| 1645 | |
| 1646 | def set_input_embeddings(self, value): |
| 1647 | self.model.embed_tokens = value |
| 1648 | |
| 1649 | @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) |
| 1650 | def forward( |
| 1651 | self, |
| 1652 | input_ids: torch.LongTensor = None, |
| 1653 | attention_mask: Optional[torch.Tensor] = None, |
| 1654 | position_ids: Optional[torch.LongTensor] = None, |
| 1655 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 1656 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 1657 | labels: Optional[torch.LongTensor] = None, |
| 1658 | use_cache: Optional[bool] = None, |
| 1659 | output_attentions: Optional[bool] = None, |
| 1660 | output_hidden_states: Optional[bool] = None, |
| 1661 | return_dict: Optional[bool] = None, |
| 1662 | ) -> Union[Tuple, SequenceClassifierOutputWithPast]: |
| 1663 | r""" |
| 1664 | labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): |
| 1665 | Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., |
| 1666 | config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If |
| 1667 | `config.num_labels > 1` a classification loss is computed (Cross-Entropy). |
| 1668 | """ |
| 1669 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 1670 | |
| 1671 | transformer_outputs = self.model( |
| 1672 | input_ids, |
| 1673 | attention_mask=attention_mask, |
| 1674 | position_ids=position_ids, |
| 1675 | past_key_values=past_key_values, |
| 1676 | inputs_embeds=inputs_embeds, |
| 1677 | use_cache=use_cache, |
| 1678 | output_attentions=output_attentions, |
| 1679 | output_hidden_states=output_hidden_states, |
| 1680 | return_dict=return_dict, |
| 1681 | ) |
| 1682 | hidden_states = transformer_outputs[0] |
| 1683 | logits = self.score(hidden_states) |
| 1684 | |
| 1685 | if input_ids is not None: |
| 1686 | batch_size = input_ids.shape[0] |
| 1687 | else: |
| 1688 | batch_size = inputs_embeds.shape[0] |
| 1689 | |
| 1690 | if self.config.pad_token_id is None and batch_size != 1: |
| 1691 | raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") |
| 1692 | if self.config.pad_token_id is None: |
| 1693 | sequence_lengths = -1 |
| 1694 | else: |
| 1695 | if input_ids is not None: |
| 1696 | # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility |
| 1697 | sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 |
| 1698 | sequence_lengths = sequence_lengths % input_ids.shape[-1] |
| 1699 | sequence_lengths = sequence_lengths.to(logits.device) |
| 1700 | else: |
| 1701 | sequence_lengths = -1 |
| 1702 | |
| 1703 | pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] |
| 1704 | |
| 1705 | loss = None |
| 1706 | if labels is not None: |
| 1707 | labels = labels.to(logits.device) |
| 1708 | if self.config.problem_type is None: |
| 1709 | if self.num_labels == 1: |
| 1710 | self.config.problem_type = "regression" |
| 1711 | elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): |
| 1712 | self.config.problem_type = "single_label_classification" |
| 1713 | else: |
| 1714 | self.config.problem_type = "multi_label_classification" |
| 1715 | |
| 1716 | if self.config.problem_type == "regression": |
| 1717 | loss_fct = MSELoss() |
| 1718 | if self.num_labels == 1: |
| 1719 | loss = loss_fct(pooled_logits.squeeze(), labels.squeeze()) |
| 1720 | else: |
| 1721 | loss = loss_fct(pooled_logits, labels) |
| 1722 | elif self.config.problem_type == "single_label_classification": |
| 1723 | loss_fct = CrossEntropyLoss() |
| 1724 | loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1)) |
| 1725 | elif self.config.problem_type == "multi_label_classification": |
| 1726 | loss_fct = BCEWithLogitsLoss() |
| 1727 | loss = loss_fct(pooled_logits, labels) |
| 1728 | if not return_dict: |
| 1729 | output = (pooled_logits,) + transformer_outputs[1:] |
| 1730 | return ((loss,) + output) if loss is not None else output |
| 1731 | |
| 1732 | return SequenceClassifierOutputWithPast( |
| 1733 | loss=loss, |
| 1734 | logits=pooled_logits, |
| 1735 | past_key_values=transformer_outputs.past_key_values, |
| 1736 | hidden_states=transformer_outputs.hidden_states, |
| 1737 | attentions=transformer_outputs.attentions, |
| 1738 | ) |
| 1739 | |