WebOrganizer/TopicClassifier-NoURL
143.6k
1# coding=utf-82# Copyright 2024 The GTE Team Authors and Alibaba Group.3# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.4#5# Licensed under the Apache License, Version 2.0 (the "License");6# you may not use this file except in compliance with the License.7# You may obtain a copy of the License at8#9# http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# distributed under the License is distributed on an "AS IS" BASIS,13# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.14# See the License for the specific language governing permissions and15# limitations under the License.16"""PyTorch NEW model."""17 18import math19from dataclasses import dataclass20from typing import List, Optional, Tuple, Union21 22import torch23import torch.utils.checkpoint24from torch import nn25 26from transformers.activations import ACT2FN27from transformers.modeling_outputs import (28 BaseModelOutput,29 BaseModelOutputWithPooling,30 MaskedLMOutput,31 MultipleChoiceModelOutput,32 QuestionAnsweringModelOutput,33 SequenceClassifierOutput,34 ModelOutput,35)36from transformers.modeling_utils import PreTrainedModel37from transformers.utils import logging38 39try:40 import xformers.ops as xops41except ImportError as e:42 xops = None43 44from .configuration import NewConfig45 46 47logger = logging.get_logger(__name__)48 49 50# Adapted from https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/bert_padding.py51# Which was adapted from https://github.com/mlcommons/training_results_v1.1/blob/main/NVIDIA/benchmarks/bert/implementations/pytorch/padding.py52class IndexFirstAxis(torch.autograd.Function):53 @staticmethod54 def forward(ctx, input, indices):55 ctx.save_for_backward(indices)56 assert input.ndim >= 257 ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]58 second_dim = other_shape.numel()59 # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.60 # return input[indices]61 # return torch.gather(62 # rearrange(input, "b ... -> b (...)"), 0, repeat(indices, "z -> z d", d=second_dim)63 # ).reshape(-1, *other_shape)64 return torch.gather(65 input.view(ctx.first_axis_dim, second_dim),66 0,67 indices.unsqueeze(-1).expand(indices.size(0), second_dim)68 ).reshape(-1, *other_shape)69 70 @staticmethod71 def backward(ctx, grad_output):72 (indices,) = ctx.saved_tensors73 assert grad_output.ndim >= 274 other_shape = grad_output.shape[1:]75 # grad_output = rearrange(grad_output, "b ... -> b (...)")76 grad_output = grad_output.view(grad_output.size(0), other_shape.numel())77 grad_input = torch.zeros(78 [ctx.first_axis_dim, grad_output.shape[1]],79 device=grad_output.device,80 dtype=grad_output.dtype,81 )82 # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.83 # grad_input[indices] = grad_output84 # grad_input.scatter_(0, repeat(indices, "z -> z d", d=grad_output.shape[1]), grad_output)85 grad_input.scatter_(86 0, indices.unsqueeze(-1).expand(indices.size(0), grad_output.size(1)), grad_output87 )88 return grad_input.reshape(ctx.first_axis_dim, *other_shape), None89 90 91index_first_axis = IndexFirstAxis.apply92 93 94def unpad_input(hidden_states, attention_mask=None, indices=None):95 """96 Arguments:97 hidden_states: (batch, seqlen, ...)98 attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.99 indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.100 Return:101 hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.102 """103 if indices is None:104 assert attention_mask is not None105 indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()106 107 # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the108 # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim109 # times larger than it needs to be, wasting memory. It's faster and more memory-efficient to110 # index with integer indices. Moreover, torch's index is a bit slower than it needs to be,111 # so we write custom forward and backward to make it a bit faster.112 hidden_states = hidden_states.view(-1, *hidden_states.shape[2:])113 return index_first_axis(hidden_states, indices)114 115 116class IndexPutFirstAxis(torch.autograd.Function):117 @staticmethod118 def forward(119 ctx,120 values: torch.Tensor,121 indices: torch.Tensor,122 first_axis_dim123 ) -> torch.Tensor:124 ctx.save_for_backward(indices)125 assert indices.ndim == 1126 assert values.ndim >= 2127 output = torch.zeros(128 first_axis_dim, *values.shape[1:], device=values.device, dtype=values.dtype129 )130 output[indices] = values131 return output132 133 @staticmethod134 def backward(ctx, grad_output: torch.Tensor) -> Tuple[torch.Tensor, None, None]:135 indices, = ctx.saved_tensors136 grad_values = grad_output[indices]137 return grad_values, None, None138 139 140index_put_first_axis = IndexPutFirstAxis.apply141 142 143def pad_input(inputs: torch.Tensor, indices: torch.Tensor, batch: int, seqlen: int) -> torch.Tensor:144 """Add padding to sequences.145 146 Arguments:147 inputs: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.148 indices: (total_nnz), `indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()`149 batch: int batch_size150 seqlen: int max sequence length151 152 Returns:153 inputs: (batch, seqlen, ...)154 """155 output = index_put_first_axis(inputs, indices, batch * seqlen)156 return output.view(batch, seqlen, *inputs.shape[1:])157 158 159def rotate_half(x):160 """Rotates half the hidden dims of the input."""161 x1 = x[..., : x.shape[-1] // 2]162 x2 = x[..., x.shape[-1] // 2 :]163 return torch.cat((-x2, x1), dim=-1)164 165 166def apply_rotary_pos_emb(q, k, cos, sin):167 """Applies Rotary Position Embedding to the query and key tensors.168 169 Args:170 q (`torch.Tensor`): The query tensor.171 k (`torch.Tensor`): The key tensor.172 cos (`torch.Tensor`): The cosine part of the rotary embedding.173 sin (`torch.Tensor`): The sine part of the rotary embedding.174 Returns:175 `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.176 """177 cos, sin = cos.to(q.dtype), sin.to(q.dtype)178 q_embed = (q * cos) + (rotate_half(q) * sin)179 k_embed = (k * cos) + (rotate_half(k) * sin)180 return q_embed, k_embed181 182 183class RotaryEmbedding(torch.nn.Module):184 def __init__(self, dim, max_position_embeddings=512, base=10000.0, device=None):185 super().__init__()186 187 self.dim = dim188 self.max_position_embeddings = max_position_embeddings189 self.base = base190 inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))191 self.register_buffer("inv_freq", inv_freq, persistent=False)192 193 # Build here to make `torch.jit.trace` work.194 self._set_cos_sin_cache(195 seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()196 )197 198 def _set_cos_sin_cache(self, seq_len, device, dtype):199 self.max_seq_len_cached = seq_len200 t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32)201 202 freqs = torch.einsum("i,j->ij", t, self.inv_freq)203 # Different from paper, but it uses a different permutation in order to obtain the same calculation204 emb = torch.cat((freqs, freqs), dim=-1)205 self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)206 self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)207 208 def forward(self, x, seq_len=None):209 # x: [bs, num_attention_heads, seq_len, head_size]210 if seq_len > self.max_seq_len_cached:211 self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)212 213 return (214 self.cos_cached[:seq_len, ...].to(dtype=x.dtype),215 self.sin_cached[:seq_len, ...].to(dtype=x.dtype),216 )217 218 219class NTKScalingRotaryEmbedding(RotaryEmbedding):220 """RotaryEmbedding extended with fixed and mixed NTK scaling. https://kexue.fm/archives/9706 """221 222 def __init__(self, dim, max_position_embeddings=512, base=10000, device=None, scaling_factor=1.0, mixed_b=None):223 self.scaling_factor = scaling_factor224 self.mixed_b = mixed_b225 super().__init__(dim, max_position_embeddings, base, device)226 max_position_embeddings = max_position_embeddings * self.scaling_factor227 self._set_cos_sin_cache(max_position_embeddings, self.inv_freq.device, torch.get_default_dtype())228 229 def _set_cos_sin_cache(self, seq_len, device, dtype):230 self.max_seq_len_cached = seq_len231 232 if seq_len > self.max_position_embeddings:233 base = self.base * (self.scaling_factor if self.mixed_b is None else 1)234 inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))235 236 if self.mixed_b is None:237 inv_freq = inv_freq / self.scaling_factor ** (2 / self.dim) # (6)238 else:239 a = torch.tensor(self.scaling_factor).log() / (self.dim / 2) ** self.mixed_b # (13)240 lambda_1_m = (a * torch.arange(1, self.dim // 2 + 1).float().to(device) ** self.mixed_b).exp() # (12)241 inv_freq = inv_freq / lambda_1_m # (10)242 243 self.register_buffer("inv_freq", inv_freq, persistent=False)244 245 t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32)246 247 freqs = torch.einsum("i,j->ij", t, self.inv_freq)248 # Different from paper, but it uses a different permutation in order to obtain the same calculation249 emb = torch.cat((freqs, freqs), dim=-1)250 self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)251 self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)252 253 254class RMSNorm(nn.Module):255 def __init__(self, hidden_size, eps=1e-6):256 """257 RMSNorm is equivalent to T5LayerNorm258 """259 super().__init__()260 self.weight = nn.Parameter(torch.ones(hidden_size))261 self.variance_epsilon = eps262 263 def forward(self, hidden_states):264 input_dtype = hidden_states.dtype265 hidden_states = hidden_states.to(torch.float32)266 variance = hidden_states.pow(2).mean(-1, keepdim=True)267 hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)268 return self.weight * hidden_states.to(input_dtype)269 270 271LAYER_NORM = {272 'layer_norm': nn.LayerNorm,273 'rms_norm': RMSNorm274}275 276 277class NewEmbeddings(nn.Module):278 """279 Embedding and Unpadding.280 """281 282 def __init__(self, config: NewConfig):283 super().__init__()284 self.padding_idx = config.pad_token_id285 self.word_embeddings = nn.Embedding(286 config.vocab_size, config.hidden_size, padding_idx=self.padding_idx287 )288 289 self.position_embedding_type = config.position_embedding_type290 if self.position_embedding_type == 'absolute':291 self.position_embeddings = nn.Embedding(292 config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx293 )294 elif self.position_embedding_type == 'rope':295 self._init_rope(config)296 else:297 raise ValueError298 299 self.type_vocab_size = config.type_vocab_size300 if self.type_vocab_size > 0:301 self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)302 303 # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load304 # any TensorFlow checkpoint file305 self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)306 self.dropout = nn.Dropout(config.hidden_dropout_prob)307 # position_ids is contiguous in memory and excluded when serialized308 self.register_buffer(309 "position_ids", torch.arange(config.max_position_embeddings), persistent=False310 )311 312 def _init_rope(self, config):313 kwargs = dict(314 dim=int(config.hidden_size / config.num_attention_heads),315 max_position_embeddings=config.max_position_embeddings,316 base=config.rope_theta317 )318 if config.rope_scaling is None:319 self.rotary_emb = RotaryEmbedding(**kwargs)320 else:321 kwargs.update(scaling_factor=config.rope_scaling["factor"])322 scaling_type = config.rope_scaling["type"]323 if scaling_type == 'ntk':324 kwargs.update(mixed_b=config.rope_scaling.get('mixed_b', None))325 self.rotary_emb = NTKScalingRotaryEmbedding(**kwargs)326 # elif scaling_type == "linear":327 # self.rotary_emb = LinearScalingRotaryEmbedding(**kwargs)328 # elif scaling_type == "dynamic":329 # self.rotary_emb = DynamicNTKScalingRotaryEmbedding(**kwargs)330 else:331 raise ValueError(f"Unknown RoPE scaling type {scaling_type}")332 333 def forward(334 self,335 unpad_inputs: bool,336 input_ids: Optional[torch.Tensor] = None,337 attention_mask: Optional[torch.Tensor] = None,338 length: Optional[List[int]] = None,339 token_type_ids: Optional[torch.Tensor] = None,340 position_ids: Optional[torch.Tensor] = None,341 inputs_embeds: Optional[torch.Tensor] = None,342 ) -> Tuple[torch.Tensor, torch.Tensor, Optional[Tuple], Optional[List[int]]]:343 """344 """345 if inputs_embeds is None:346 device, input_shape = input_ids.device, input_ids.shape347 else:348 device, input_shape = inputs_embeds.device, inputs_embeds.shape[:2]349 batch_size, seq_length = input_shape350 351 # Set attention_mask if it's None352 if attention_mask is None:353 attention_mask = torch.ones(input_shape, device=device)354 if length is not None:355 for i, l in enumerate(length):356 attention_mask[i, l:] = 0357 358 # Set attention_mask_bool for unpadding359 if unpad_inputs:360 attention_mask_bool = attention_mask.bool()361 if length is None:362 length = attention_mask.sum(-1).tolist()363 364 # Get word embeddings365 if inputs_embeds is None:366 if unpad_inputs:367 input_ids = input_ids[attention_mask_bool].unsqueeze(0)368 inputs_embeds = self.word_embeddings(input_ids)369 else:370 if unpad_inputs:371 inputs_embeds = inputs_embeds[attention_mask_bool].unsqueeze(0)372 embeddings = inputs_embeds373 374 # Set and unpad position_ids375 if position_ids is None:376 if seq_length > self.position_ids.size(0):377 self.register_buffer(378 "position_ids", torch.arange(seq_length, device=embeddings.device), persistent=False379 )380 if unpad_inputs:381 # [1, cumsum_seq_len]382 position_ids = torch.cat([self.position_ids[:l] for l in length]).unsqueeze(0)383 else:384 # [bs, seq_len]385 position_ids = self.position_ids[:seq_length].expand(batch_size, -1)386 elif unpad_inputs:387 position_ids = position_ids[attention_mask_bool].unsqueeze(0) # [1, cumsum_seq_len]388 389 # Compute rotary embedding390 if self.position_embedding_type == 'rope':391 rope_cos, rope_sin = self.rotary_emb(inputs_embeds, seq_len=seq_length)392 rope_cos = rope_cos[position_ids].unsqueeze(2) # [bs, seq_len, 1, dim]393 rope_sin = rope_sin[position_ids].unsqueeze(2) # [bs, seq_len, 1, dim]394 rope_embeds = rope_cos, rope_sin395 else:396 rope_embeds = None397 398 if self.type_vocab_size > 0:399 if token_type_ids is None:400 token_type_ids = position_ids.mul(0)401 else:402 if self.type_vocab_size < 2:403 token_type_ids.mul_(0)404 if unpad_inputs:405 token_type_ids = token_type_ids[attention_mask_bool].unsqueeze(0)406 407 token_type_embeddings = self.token_type_embeddings(token_type_ids)408 embeddings = embeddings + token_type_embeddings409 410 # BERT position411 if self.position_embedding_type == "absolute":412 position_embeddings = self.position_embeddings(position_ids)413 embeddings = embeddings + position_embeddings414 415 embeddings = self.LayerNorm(embeddings)416 embeddings = self.dropout(embeddings)417 418 return embeddings, attention_mask, rope_embeds, length419 420 421class NewAttention(nn.Module):422 def __init__(self, config: NewConfig, pack_qkv=None, use_memory_efficient_attention=None):423 super().__init__()424 self.config = config425 if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):426 raise ValueError(427 f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "428 f"heads ({config.num_attention_heads})"429 )430 431 self.hidden_size = config.hidden_size432 self.num_attention_heads = config.num_attention_heads433 self.attention_head_size = int(config.hidden_size / config.num_attention_heads)434 self.all_head_size = self.num_attention_heads * self.attention_head_size435 436 if pack_qkv is None:437 pack_qkv = config.pack_qkv438 self.pack_qkv = pack_qkv439 440 if self.pack_qkv:441 self.qkv_proj = nn.Linear(config.hidden_size, self.all_head_size * 3, bias=True)442 else:443 self.q_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)444 self.k_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)445 self.v_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)446 447 self.dropout = nn.Dropout(config.attention_probs_dropout_prob)448 self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=True)449 450 if use_memory_efficient_attention is None:451 use_memory_efficient_attention = self.config.use_memory_efficient_attention452 self.use_memory_efficient_attention = use_memory_efficient_attention453 self.memory_efficient_attention = None if xops is None else xops.memory_efficient_attention454 if self.use_memory_efficient_attention:455 assert self.memory_efficient_attention is not None, 'please install xformers'456 457 def forward(458 self,459 hidden_states: torch.Tensor,460 attention_bias: torch.FloatTensor,461 rope_embeds: Optional[Tuple[torch.FloatTensor, torch.FloatTensor]] = None,462 padding_inputs: Optional[Tuple] = None, # indices, batch, seqlen463 attention_scale: Optional[torch.FloatTensor] = None,464 head_mask: Optional[torch.FloatTensor] = None,465 output_attentions: Optional[bool] = False,466 qkv_inputs: Optional[Tuple] = None, # For RetroMAE467 ) -> Tuple[torch.Tensor, ...]:468 shape_hd = (self.num_attention_heads, self.attention_head_size)469 # qkv470 if self.pack_qkv and qkv_inputs is None:471 qkv_pack = self.qkv_proj(hidden_states).split(self.all_head_size, dim=-1)472 else:473 if qkv_inputs is None:474 qkv_inputs = (hidden_states, hidden_states, hidden_states)475 qkv_pack = [476 getattr(self, n + '_proj')(s) for s, n in zip(qkv_inputs, 'qkv')477 ]478 query_states, key_states, value_states = [t.view(t.shape[:-1] + shape_hd) for t in qkv_pack]479 480 if self.config.position_embedding_type == 'rope':481 query_states, key_states = apply_rotary_pos_emb(query_states, key_states, *rope_embeds)482 483 dtype = query_states.dtype484 485 if self.config.logn_attention_scale and attention_scale is not None:486 # https://kexue.fm/archives/8823487 query_states = query_states * attention_scale.to(dtype)488 489 if padding_inputs is not None:490 query_states = pad_input(query_states.squeeze(), *padding_inputs)491 key_states = pad_input(key_states.squeeze(), *padding_inputs)492 value_states = pad_input(value_states.squeeze(), *padding_inputs)493 494 if self.use_memory_efficient_attention:495 assert self.memory_efficient_attention is not None, "xformers is not loaded"496 assert output_attentions is False, "memory_efficient_attention do not output attentions"497 assert head_mask is None, "Not support yet"498 attention_probs = None499 if torch.is_tensor(attention_bias):500 attention_bias = attention_bias.to(dtype)501 context_layer = self.memory_efficient_attention(502 query_states,503 key_states,504 value_states,505 attn_bias=attention_bias,506 p=self.dropout.p507 )508 else:509 if output_attentions and isinstance(self, NewSdpaAttention):510 raise RuntimeError("SDPA do not output attentions")511 context_layer, attention_probs = self._attention(512 query_states, key_states, value_states, attention_bias, head_mask513 )514 515 if padding_inputs is not None:516 context_layer = unpad_input(context_layer, indices=padding_inputs[0])517 518 new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)519 context_layer = context_layer.view(new_context_layer_shape)520 521 # output proj522 attn_output = self.o_proj(context_layer)523 524 # add attentions if we output them525 outputs = (attn_output, attention_probs) if output_attentions else (attn_output,)526 return outputs527 528 def _attention(self, query_states, key_states, value_states, attention_bias, head_mask):529 """530 Args:531 q/k/v: (B, L, n_head, head_dim),532 Returns:533 attn_output: (B L, n_head, head_dim)534 """535 query_states = query_states.transpose(1, 2)536 key_states = key_states.transpose(1, 2)537 value_states = value_states.transpose(1, 2)538 # Take the dot product between "query" and "key" to get the raw attention scores.539 attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))540 541 attention_scores = attention_scores / math.sqrt(self.attention_head_size)542 if attention_bias is not None:543 # Apply the attention mask is (precomputed for all layers in BertModel forward() function)544 attention_scores = attention_scores + attention_bias545 546 # Normalize the attention scores to probabilities.547 attention_probs = nn.functional.softmax(attention_scores, dim=-1)548 549 # This is actually dropping out entire tokens to attend to, which might550 # seem a bit unusual, but is taken from the original Transformer paper.551 if self.dropout.p > 0:552 attention_probs = self.dropout(attention_probs)553 554 # Mask heads if we want to555 if head_mask is not None:556 attention_probs = attention_probs * head_mask557 558 context_layer = torch.matmul(attention_probs, value_states)559 560 context_layer = context_layer.permute(0, 2, 1, 3).contiguous()561 return context_layer, attention_probs562 563 564class NewSdpaAttention(NewAttention):565 """566 New attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from567 `NewAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to568 SDPA API.569 """570 def __init__(self, config: NewConfig, **kwargs):571 super().__init__(config, **kwargs)572 # torch.backends.cuda.enable_mem_efficient_sdp(False)573 # logger.warning(574 # "Disable memory efficient attention kernel for `NewSdpaAttention`, you can set "575 # "`use_memory_efficient_attention=True` if it expected to use."576 # )577 578 def _attention(self, query_states, key_states, value_states, attention_bias, head_mask):579 attn_output = torch.nn.functional.scaled_dot_product_attention(580 query_states.transpose(1, 2),581 key_states.transpose(1, 2),582 value_states.transpose(1, 2),583 attn_mask=attention_bias,584 dropout_p=self.dropout.p if self.training else 0.0,585 )586 attn_output = attn_output.permute(0, 2, 1, 3).contiguous()587 return attn_output, None588 589 590NEW_ATTENTION_CLASSES = {591 "eager": NewAttention,592 # "flash_attention_2": , # TODO593 "sdpa": NewSdpaAttention,594}595 596 597class NewGatedMLP(nn.Module):598 """599 GLU Variants Improve Transformer.600 """601 602 def __init__(self, config: NewConfig):603 super().__init__()604 self.intermediate_size = config.intermediate_size605 self.up_gate_proj = nn.Linear(config.hidden_size, self.intermediate_size * 2, bias=False)606 self.down_proj = nn.Linear(self.intermediate_size, config.hidden_size, bias=True)607 self.act_fn = ACT2FN[config.hidden_act]608 if config.hidden_dropout_prob > 0:609 self.hidden_dropout = nn.Dropout(config.hidden_dropout_prob)610 else:611 self.hidden_dropout = None612 613 def forward(self, hidden_states):614 up_gate = self.up_gate_proj(hidden_states)615 up_states, gate = torch.split(up_gate, self.intermediate_size, dim=-1)616 gate = self.act_fn(gate)617 gated_states = gate * up_states618 if self.hidden_dropout is not None:619 gated_states = self.hidden_dropout(gated_states)620 down_states = self.down_proj(gated_states)621 return down_states622 623 624class NewLayer(nn.Module):625 def __init__(626 self,627 config: NewConfig,628 pack_qkv=None,629 use_memory_efficient_attention=None,630 attn_implementation=None631 ):632 super().__init__()633 if attn_implementation is None:634 attn_implementation = config._attn_implementation635 if use_memory_efficient_attention is None:636 use_memory_efficient_attention = config.use_memory_efficient_attention637 if use_memory_efficient_attention:638 if attn_implementation != 'eager':639 logger.warning_once(f"Override {attn_implementation=} to 'eager' as {use_memory_efficient_attention=}")640 attn_implementation = 'eager' # Since it will be SDPA by default for torch>=2.1.1641 self.attention = NEW_ATTENTION_CLASSES[attn_implementation](642 config, pack_qkv=pack_qkv, use_memory_efficient_attention=use_memory_efficient_attention643 )644 self.mlp = NewGatedMLP(config)645 646 ln_class = LAYER_NORM[config.layer_norm_type]647 self.attn_ln = ln_class(config.hidden_size, eps=config.layer_norm_eps)648 self.mlp_ln = ln_class(config.hidden_size, eps=config.layer_norm_eps)649 650 if config.hidden_dropout_prob > 0:651 self.hidden_dropout = nn.Dropout(config.hidden_dropout_prob)652 else:653 self.hidden_dropout = None654 655 def forward(656 self,657 hidden_states: torch.Tensor,658 attention_bias: torch.FloatTensor,659 rope_embeds: Optional[Tuple[torch.FloatTensor, torch.FloatTensor]] = None,660 padding_inputs: Optional[Tuple] = None, # indices, batch, seqlen661 attention_scale: Optional[torch.FloatTensor] = None,662 subset_indices: Optional[torch.LongTensor] = None,663 head_mask: Optional[torch.FloatTensor] = None,664 output_attentions: Optional[bool] = False,665 qkv_inputs: Optional[Tuple] = None, # For RetroMAE666 ) -> Tuple[torch.Tensor, ...]:667 # Multi head self attention668 residual = hidden_states if qkv_inputs is None else qkv_inputs[0]669 attention_outputs = self.attention(670 hidden_states,671 attention_bias,672 rope_embeds,673 padding_inputs,674 attention_scale,675 head_mask,676 output_attentions=output_attentions,677 qkv_inputs=qkv_inputs,678 )679 hidden_states = attention_outputs[0]680 if self.hidden_dropout is not None:681 hidden_states = self.hidden_dropout(hidden_states)682 hidden_states = residual + hidden_states683 684 # In pretraining, after the attention of last layer, we only need the masked tokens.685 if subset_indices is not None:686 hidden_states = hidden_states[subset_indices]687 688 hidden_states = self.attn_ln(hidden_states)689 690 # Fully Connected691 residual = hidden_states692 hidden_states = self.mlp(hidden_states)693 if self.hidden_dropout is not None:694 hidden_states = self.hidden_dropout(hidden_states)695 hidden_states = residual + hidden_states696 hidden_states = self.mlp_ln(hidden_states)697 698 # add self attentions if we output attention weights699 outputs = (hidden_states,) + attention_outputs[1:]700 return outputs701 702 703class NewEncoder(nn.Module):704 def __init__(self, config):705 super().__init__()706 self.config = config707 self.layer = nn.ModuleList([NewLayer(config) for _ in range(config.num_hidden_layers)])708 self.gradient_checkpointing = False709 710 def forward(711 self,712 hidden_states: torch.Tensor,713 attention_bias: Optional[torch.FloatTensor] = None,714 rope_embeds: Optional[Tuple[torch.FloatTensor, torch.FloatTensor]] = None,715 padding_inputs: Optional[Tuple] = None, # indices, batch, seqlen716 attention_scale: Optional[torch.FloatTensor] = None,717 subset_indices: Optional[torch.LongTensor] = None,718 head_mask: Optional[torch.FloatTensor] = None,719 output_attentions: Optional[bool] = False,720 output_hidden_states: Optional[bool] = False,721 return_dict: Optional[bool] = True,722 ) -> Union[Tuple[torch.Tensor], BaseModelOutput]:723 all_hidden_states = () if output_hidden_states else None724 all_self_attentions = () if output_attentions else None725 726 for i, layer_module in enumerate(self.layer):727 if output_hidden_states:728 all_hidden_states = all_hidden_states + (hidden_states,)729 730 if i >= len(self.layer) - 1:731 layer_subset_indices = subset_indices732 else:733 layer_subset_indices = None734 735 layer_head_mask = head_mask[i] if head_mask is not None else None736 737 if self.gradient_checkpointing and self.training:738 layer_outputs = self._gradient_checkpointing_func(739 layer_module.__call__,740 hidden_states,741 attention_bias,742 rope_embeds,743 padding_inputs,744 attention_scale,745 layer_subset_indices,746 layer_head_mask,747 )748 else:749 layer_outputs = layer_module(750 hidden_states,751 attention_bias,752 rope_embeds,753 padding_inputs,754 attention_scale,755 layer_subset_indices,756 layer_head_mask,757 output_attentions,758 )759 760 hidden_states = layer_outputs[0]761 if output_attentions:762 all_self_attentions = all_self_attentions + (layer_outputs[1],)763 764 if output_hidden_states:765 all_hidden_states = all_hidden_states + (hidden_states,)766 767 if not return_dict:768 return tuple(769 v770 for v in [771 hidden_states,772 all_hidden_states,773 all_self_attentions,774 ]775 if v is not None776 )777 return BaseModelOutput(778 last_hidden_state=hidden_states,779 hidden_states=all_hidden_states,780 attentions=all_self_attentions,781 )782 783 784# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert->New785class NewPooler(nn.Module):786 def __init__(self, config):787 super().__init__()788 self.dense = nn.Linear(config.hidden_size, config.hidden_size)789 self.activation = nn.Tanh()790 791 def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:792 # We "pool" the model by simply taking the hidden state corresponding793 # to the first token.794 first_token_tensor = hidden_states[:, 0]795 pooled_output = self.dense(first_token_tensor)796 pooled_output = self.activation(pooled_output)797 return pooled_output798 799 800class NewPreTrainedModel(PreTrainedModel):801 """802 An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained803 models.804 """805 806 config_class = NewConfig807 base_model_prefix = "new"808 supports_gradient_checkpointing = True809 _supports_sdpa = True810 811 def _init_weights(self, module):812 """Initialize the weights"""813 if isinstance(module, nn.Linear):814 # Slightly different from the TF version which uses truncated_normal for initialization815 # cf https://github.com/pytorch/pytorch/pull/5617816 module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)817 if module.bias is not None:818 module.bias.data.zero_()819 elif isinstance(module, nn.Embedding):820 module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)821 if module.padding_idx is not None:822 module.weight.data[module.padding_idx].zero_()823 elif isinstance(module, nn.LayerNorm):824 module.bias.data.zero_()825 module.weight.data.fill_(1.0)826 827 828class NewModel(NewPreTrainedModel):829 """830 The bare New Model transformer outputting raw hidden-states without any specific head on top.831 """832 833 def __init__(self, config: NewConfig, add_pooling_layer=False):834 super().__init__(config)835 self.config = config836 837 self.embeddings = NewEmbeddings(config)838 self.encoder = NewEncoder(config)839 840 self.pooler = NewPooler(config) if add_pooling_layer else None841 842 # Initialize weights and apply final processing843 self.post_init()844 845 def get_input_embeddings(self):846 return self.embeddings.word_embeddings847 848 def set_input_embeddings(self, value):849 self.embeddings.word_embeddings = value850 851 def forward(852 self,853 input_ids: Optional[torch.Tensor] = None,854 attention_mask: Optional[torch.Tensor] = None,855 length: Optional[List[int]] = None,856 subset_indices: Optional[torch.LongTensor] = None,857 token_type_ids: Optional[torch.Tensor] = None,858 position_ids: Optional[torch.Tensor] = None,859 head_mask: Optional[torch.Tensor] = None,860 inputs_embeds: Optional[torch.Tensor] = None,861 output_attentions: Optional[bool] = None,862 output_hidden_states: Optional[bool] = None,863 return_dict: Optional[bool] = None,864 unpad_inputs: Optional[bool] = None,865 ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPooling]:866 r"""867 length (`list` of length `batch_size`, *optional*):868 If is `None`, return padded `last_hidden_state`.869 subset_indices ():870 pass871 unpad_inputs (`bool`, *optional*):872 pass873 """874 output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions875 output_hidden_states = (876 output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states877 )878 return_dict = return_dict if return_dict is not None else self.config.use_return_dict879 unpad_inputs = unpad_inputs if unpad_inputs is not None else self.config.unpad_inputs880 output_padded = length is None881 882 if input_ids is not None and inputs_embeds is not None:883 raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")884 elif input_ids is not None:885 self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)886 input_shape = input_ids.size()887 elif inputs_embeds is not None:888 input_shape = inputs_embeds.size()[:-1]889 else:890 raise ValueError("You have to specify either input_ids or inputs_embeds")891 892 # TODO: not used893 # # Prepare head mask if needed894 # # 1.0 in head_mask indicate we keep the head895 # # attention_probs has shape bsz x n_heads x N x N896 # # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]897 # # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]898 # head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)899 900 # Get embeddings, may unpad them901 (embedding_output, attention_mask, rope_embeds, length) = self.embeddings(902 unpad_inputs,903 input_ids=input_ids,904 attention_mask=attention_mask,905 length=length,906 token_type_ids=token_type_ids,907 position_ids=position_ids,908 inputs_embeds=inputs_embeds909 )910 911 batch_size, seq_length = input_shape912 if unpad_inputs and self.config.use_memory_efficient_attention:913 attention_bias = xops.fmha.attn_bias.BlockDiagonalMask.from_seqlens(length)914 else:915 # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]916 # ourselves in which case we just need to make it broadcastable to all heads.917 attention_bias = self.get_extended_attention_mask(attention_mask, input_shape)918 if self.config.use_memory_efficient_attention:919 # Invalid shape for attention bias: torch.Size([48, 1, 1, 512]) (expected (48, 12, 512, 512))920 attention_bias = attention_bias.expand(-1, self.config.num_attention_heads, seq_length, -1)921 922 padding_inputs = None923 if unpad_inputs and (output_padded or not self.config.use_memory_efficient_attention):924 indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()925 if not self.config.use_memory_efficient_attention:926 padding_inputs = (indices, *input_shape)927 928 attention_scale = None929 if self.config.logn_attention_scale:930 logger.warning_once("TODO: logn_attention_scale")931 # # attention scale log_512(input_len)932 # attention_scale = attention_mask.sum(1).log() / torch.tensor(self.config.max_position_embeddings).log()933 # # inference-time logn scale need clip 1934 # if self.config.logn_attention_clip1:935 # attention_scale.clip_(1)936 # attention_scale = attention_scale[:, None, None, None]937 # else:938 # attention_scale = None939 940 encoder_outputs = self.encoder(941 embedding_output,942 attention_bias=attention_bias,943 rope_embeds=rope_embeds,944 padding_inputs=padding_inputs,945 attention_scale=attention_scale,946 subset_indices=subset_indices,947 head_mask=head_mask,948 output_attentions=output_attentions,949 output_hidden_states=output_hidden_states,950 return_dict=return_dict,951 )952 sequence_output = encoder_outputs[0]953 if unpad_inputs and output_padded:954 sequence_output = pad_input(955 sequence_output.squeeze(), indices, batch_size, seq_length956 )957 958 pooled_output = self.pooler(sequence_output) if self.pooler is not None else None959 960 if not return_dict:961 return (sequence_output, pooled_output) + encoder_outputs[1:]962 963 return BaseModelOutputWithPooling(964 last_hidden_state=sequence_output,965 pooler_output=pooled_output,966 hidden_states=encoder_outputs.hidden_states,967 attentions=encoder_outputs.attentions,968 )969 970 971class NewLMPredictionHead(nn.Module):972 def __init__(self, config):973 super().__init__()974 self.dense = nn.Linear(config.hidden_size, config.hidden_size)975 self.transform_act_fn = ACT2FN[config.hidden_act]976 self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)977 978 # The output weights are the same as the input embeddings, but there is979 # an output-only bias for each token.980 self.decoder = nn.Linear(config.hidden_size, config.vocab_size)981 982 def forward(self, hidden_states):983 hidden_states = self.dense(hidden_states)984 hidden_states = self.transform_act_fn(hidden_states)985 hidden_states = self.norm(hidden_states)986 hidden_states = self.decoder(hidden_states)987 return hidden_states988 989 990class NewForMaskedLM(NewPreTrainedModel):991 _tied_weights_keys = ["lm_head.decoder.bias", "lm_head.decoder.weight"]992 993 def __init__(self, config: NewConfig):994 super().__init__(config)995 self.new = NewModel(config, add_pooling_layer=False)996 self.lm_head = NewLMPredictionHead(config)997 self.loss_fct = nn.CrossEntropyLoss()998 999 # Initialize weights and apply final processing1000 self.post_init()1001 1002 def get_output_embeddings(self):1003 return self.lm_head.decoder1004 1005 def set_output_embeddings(self, new_embeddings):1006 self.lm_head.decoder = new_embeddings1007 1008 def forward(1009 self,1010 input_ids: Optional[torch.Tensor] = None,1011 attention_mask: Optional[torch.Tensor] = None,1012 token_type_ids: Optional[torch.Tensor] = None,1013 position_ids: Optional[torch.Tensor] = None,1014 head_mask: Optional[torch.Tensor] = None,1015 inputs_embeds: Optional[torch.Tensor] = None,1016 labels: Optional[torch.Tensor] = None,1017 output_attentions: Optional[bool] = None,1018 output_hidden_states: Optional[bool] = None,1019 return_dict: Optional[bool] = None,1020 unpad_inputs: Optional[bool] = None,1021 ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:1022 r"""1023 labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):1024 Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,1025 config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the1026 loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`1027 """1028 1029 return_dict = return_dict if return_dict is not None else self.config.use_return_dict1030 1031 if labels is None or not self.new.config.unpad_inputs:1032 length = None1033 subset_indices = None1034 else:1035 length = attention_mask.sum(-1).tolist()1036 labels = labels[attention_mask.bool()].unsqueeze(0)1037 subset_indices = labels > -1001038 1039 outputs = self.new(1040 input_ids,1041 attention_mask=attention_mask,1042 length=length,1043 subset_indices=subset_indices,1044 token_type_ids=token_type_ids,1045 position_ids=position_ids,1046 head_mask=head_mask,1047 inputs_embeds=inputs_embeds,1048 output_attentions=output_attentions,1049 output_hidden_states=output_hidden_states,1050 return_dict=return_dict,1051 unpad_inputs=unpad_inputs,1052 )1053 1054 sequence_output = outputs[0]1055 prediction_scores = self.lm_head(sequence_output)1056 1057 masked_lm_loss = None1058 if labels is not None:1059 if subset_indices is None:1060 mask = attention_mask.bool()1061 prediction_scores = prediction_scores[mask]1062 labels = labels[mask]1063 else:1064 labels = labels[subset_indices]1065 masked_lm_loss = self.loss_fct(prediction_scores, labels)1066 1067 if not return_dict:1068 output = (prediction_scores,) + outputs[2:]1069 return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output1070 1071 return MaskedLMOutput(1072 loss=masked_lm_loss,1073 logits=prediction_scores,1074 hidden_states=outputs.hidden_states,1075 attentions=outputs.attentions,1076 )1077 1078 1079class NewForSequenceClassification(NewPreTrainedModel):1080 def __init__(self, config):1081 super().__init__(config)1082 self.num_labels = config.num_labels1083 self.config = config1084 1085 self.new = NewModel(config, add_pooling_layer=True)1086 classifier_dropout = (1087 config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob1088 )1089 self.dropout = nn.Dropout(classifier_dropout)1090 self.classifier = nn.Linear(config.hidden_size, config.num_labels)1091 1092 # Initialize weights and apply final processing1093 self.post_init()1094 1095 def forward(1096 self,1097 input_ids: Optional[torch.Tensor] = None,1098 attention_mask: Optional[torch.Tensor] = None,1099 token_type_ids: Optional[torch.Tensor] = None,1100 position_ids: Optional[torch.Tensor] = None,1101 head_mask: Optional[torch.Tensor] = None,1102 inputs_embeds: Optional[torch.Tensor] = None,1103 labels: Optional[torch.Tensor] = None,1104 output_attentions: Optional[bool] = None,1105 output_hidden_states: Optional[bool] = None,1106 return_dict: Optional[bool] = None,1107 unpad_inputs: Optional[bool] = None,1108 ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:1109 r"""1110 labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):1111 Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,1112 config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If1113 `config.num_labels > 1` a classification loss is computed (Cross-Entropy).1114 """1115 return_dict = return_dict if return_dict is not None else self.config.use_return_dict1116 1117 outputs = self.new(1118 input_ids,1119 attention_mask=attention_mask,1120 token_type_ids=token_type_ids,1121 position_ids=position_ids,1122 head_mask=head_mask,1123 inputs_embeds=inputs_embeds,1124 output_attentions=output_attentions,1125 output_hidden_states=output_hidden_states,1126 return_dict=return_dict,1127 unpad_inputs=unpad_inputs,1128 )1129 1130 pooled_output = outputs[1]1131 1132 pooled_output = self.dropout(pooled_output)1133 logits = self.classifier(pooled_output)1134 1135 loss = None1136 if labels is not None:1137 if self.config.problem_type is None:1138 if self.num_labels == 1:1139 self.config.problem_type = "regression"1140 elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):1141 self.config.problem_type = "single_label_classification"1142 else:1143 self.config.problem_type = "multi_label_classification"1144 1145 if self.config.problem_type == "regression":1146 loss_fct = nn.MSELoss()1147 if self.num_labels == 1:1148 loss = loss_fct(logits.squeeze(), labels.squeeze())1149 else:1150 loss = loss_fct(logits, labels)1151 elif self.config.problem_type == "single_label_classification":1152 loss_fct = nn.CrossEntropyLoss()1153 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))1154 elif self.config.problem_type == "multi_label_classification":1155 loss_fct = nn.BCEWithLogitsLoss()1156 loss = loss_fct(logits, labels)1157 1158 if not return_dict:1159 output = (logits,) + outputs[2:]1160 return ((loss,) + output) if loss is not None else output1161 1162 return SequenceClassifierOutput(1163 loss=loss,1164 logits=logits,1165 hidden_states=outputs.hidden_states,1166 attentions=outputs.attentions,1167 )1168 1169 1170class NewForMultipleChoice(NewPreTrainedModel):1171 def __init__(self, config):1172 super().__init__(config)1173 1174 self.new = NewModel(config, add_pooling_layer=True)1175 classifier_dropout = (1176 config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob1177 )1178 self.dropout = nn.Dropout(classifier_dropout)1179 self.classifier = nn.Linear(config.hidden_size, 1)1180 1181 # Initialize weights and apply final processing1182 self.post_init()1183 1184 def forward(1185 self,1186 input_ids: Optional[torch.Tensor] = None,1187 attention_mask: Optional[torch.Tensor] = None,1188 token_type_ids: Optional[torch.Tensor] = None,1189 position_ids: Optional[torch.Tensor] = None,1190 head_mask: Optional[torch.Tensor] = None,1191 inputs_embeds: Optional[torch.Tensor] = None,1192 labels: Optional[torch.Tensor] = None,1193 output_attentions: Optional[bool] = None,1194 output_hidden_states: Optional[bool] = None,1195 return_dict: Optional[bool] = None,1196 unpad_inputs: Optional[bool] = None,1197 ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:1198 r"""1199 labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):1200 Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,