Salesforce/codet5p-16b
67158
1# coding=utf-82# Copyright 2023 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.3""" PyTorch CodeT5+ 2B 6B 16B models.4The implementation is mainly based on transformers.models.codegen.modeling_codegen by adding cross-attention5and transformers.models.encoder_decoder.modeling_encoder_decoder.EncoderDecoderModel.6"""7from typing import Optional, Tuple, Union8import torch9import torch.utils.checkpoint10from torch import nn11from torch.nn import CrossEntropyLoss12 13from transformers.activations import ACT2FN14from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput, \15 BaseModelOutputWithPast, CausalLMOutputWithPast, \16 BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions17from transformers.modeling_utils import PreTrainedModel18from transformers.configuration_utils import PretrainedConfig19from transformers.utils import add_code_sample_docstrings, add_start_docstrings, logging20from .configuration_codet5p import CodeT5pConfig, CodeT5pModuleConfig21 22logger = logging.get_logger(__name__)23 24CODET5P_PRETRAINED_MODEL_ARCHIVE_LIST = [25 "Salesforce/codet5p-220m",26 "Salesforce/codet5p-770m",27 "Salesforce/codet5p-220m-py",28 "Salesforce/codet5p-770m-py",29 "Salesforce/codet5p-2b",30 "Salesforce/codet5p-6b",31 "Salesforce/codet5p-16b",32 "Salesforce/instructcodet5p-16b",33 # See all CodeT5+ models at https://huggingface.co/models?filter=codet5p34]35 36 37# Copied from transformers.models.gptj.modeling_gptj.fixed_pos_embedding38def fixed_pos_embedding(x, seq_dim=1, seq_len=None):39 dim = x.shape[-1]40 if seq_len is None:41 seq_len = x.shape[seq_dim]42 inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))43 sinusoid_inp = (44 torch.einsum("i , j -> i j", torch.arange(seq_len, dtype=torch.float), inv_freq).to(x.device).float()45 )46 return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)47 48 49# Copied from transformers.models.gptj.modeling_gptj.rotate_every_two50def rotate_every_two(x):51 x1 = x[:, :, :, ::2]52 x2 = x[:, :, :, 1::2]53 x = torch.stack((-x2, x1), dim=-1)54 return x.flatten(-2) # in einsum notation: rearrange(x, '... d j -> ... (d j)')55 56 57# Copied from transformers.models.gptj.modeling_gptj.duplicate_interleave58def duplicate_interleave(m):59 """60 A simple version of `torch.repeat_interleave` for duplicating a matrix while interleaving the copy.61 """62 dim0 = m.shape[0]63 m = m.view(-1, 1) # flatten the matrix64 m = m.repeat(1, 2) # repeat all elements into the 2nd dimension65 m = m.view(dim0, -1) # reshape into a matrix, interleaving the copy66 return m67 68 69# Copied from transformers.models.gptj.modeling_gptj.apply_rotary_pos_emb70def apply_rotary_pos_emb(x, sincos, offset=0):71 sin, cos = (duplicate_interleave(t)[None, offset: x.shape[1] + offset, None, :] for t in sincos)72 # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)73 return (x * cos) + (rotate_every_two(x) * sin)74 75 76# Adapted from transformers.models.codegen.modeling_codegen.CodeGenAttention77class CodeT5pAttention(nn.Module):78 def __init__(self, config, is_cross_attention=False, is_decoder=True):79 super().__init__()80 81 max_positions = config.max_position_embeddings82 self.register_buffer(83 "causal_mask",84 torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(85 1, 1, max_positions, max_positions86 ),87 )88 89 self.attn_dropout = nn.Dropout(config.attn_pdrop)90 self.resid_dropout = nn.Dropout(config.resid_pdrop)91 92 self.embed_dim = config.hidden_size93 self.num_attention_heads = config.num_attention_heads94 self.head_dim = self.embed_dim // self.num_attention_heads95 if self.head_dim * self.num_attention_heads != self.embed_dim:96 raise ValueError(97 f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and"98 f" `num_attention_heads`: {self.num_attention_heads})."99 )100 101 self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())102 self.is_decoder = is_decoder103 self.is_cross_attention = is_cross_attention104 if self.is_cross_attention:105 self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 2, bias=False)106 self.q_attn = nn.Linear(self.embed_dim, self.embed_dim, bias=False)107 else:108 self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)109 110 self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)111 self.rotary_dim = None112 if config.rotary_dim is not None:113 self.rotary_dim = config.rotary_dim114 115 def _split_heads(self, x, n_head, dim_head, mp_num):116 reshaped = x.reshape(x.shape[:-1] + (n_head // mp_num, dim_head))117 reshaped = reshaped.reshape(x.shape[:-2] + (-1,) + reshaped.shape[-1:])118 return reshaped119 120 def _merge_heads(self, tensor, num_attention_heads, attn_head_size):121 """122 Merges attn_head_size dim and num_attn_heads dim into n_ctx123 """124 if len(tensor.shape) == 5:125 tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()126 elif len(tensor.shape) == 4:127 tensor = tensor.permute(0, 2, 1, 3).contiguous()128 else:129 raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")130 new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)131 return tensor.view(new_shape)132 133 def _attn(134 self,135 query,136 key,137 value,138 attention_mask=None,139 head_mask=None,140 ):141 # Keep the attention weights computation in fp32 to avoid overflow issues142 query = query.to(torch.float32)143 key = key.to(torch.float32)144 145 attn_weights = torch.matmul(query, key.transpose(-1, -2))146 attn_weights = attn_weights / self.scale_attn147 148 if not self.is_cross_attention and self.is_decoder:149 # compute causal mask from causal mask buffer150 query_length, key_length = query.size(-2), key.size(-2)151 causal_mask = self.causal_mask[:, :, key_length - query_length: key_length, :key_length]152 mask_value = torch.finfo(attn_weights.dtype).min153 # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.154 # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`155 mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)156 attn_weights = torch.where(causal_mask.bool(), attn_weights, mask_value)157 158 if attention_mask is not None:159 # Apply the attention mask160 attn_weights = attn_weights + attention_mask161 162 attn_weights = nn.Softmax(dim=-1)(attn_weights)163 attn_weights = attn_weights.to(value.dtype)164 attn_weights = self.attn_dropout(attn_weights)165 166 # Mask heads if we want to167 if head_mask is not None:168 attn_weights = attn_weights * head_mask169 170 attn_output = torch.matmul(attn_weights, value)171 172 return attn_output, attn_weights173 174 def forward(175 self,176 hidden_states: Optional[torch.FloatTensor],177 attention_mask: Optional[torch.FloatTensor] = None,178 layer_past: Optional[Tuple[torch.Tensor]] = None,179 head_mask: Optional[torch.FloatTensor] = None,180 encoder_hidden_states: Optional[torch.Tensor] = None,181 encoder_attention_mask: Optional[torch.FloatTensor] = None,182 use_cache: Optional[bool] = False,183 output_attentions: Optional[bool] = False,184 ) -> Union[185 Tuple[torch.Tensor, Tuple[torch.Tensor]],186 Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],187 ]:188 189 if encoder_hidden_states is not None:190 if not hasattr(self, "q_attn"):191 raise ValueError(192 "If class is used as cross attention, the weights `q_attn` have to be defined. "193 "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."194 )195 196 mp_num = 4197 local_dim = self.head_dim * self.num_attention_heads // mp_num198 q = self.q_attn(hidden_states)199 q_split = q.reshape(q.shape[:-1] + (mp_num, -1))200 query = torch.split(q_split, local_dim, dim=-1)[0]201 202 qkv = self.qkv_proj(encoder_hidden_states)203 qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))204 value, key = torch.split(qkv_split, local_dim, dim=-1)205 206 attention_mask = encoder_attention_mask207 else:208 qkv = self.qkv_proj(hidden_states)209 mp_num = 4210 qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))211 212 local_dim = self.head_dim * self.num_attention_heads // mp_num213 query, value, key = torch.split(qkv_split, local_dim, dim=-1)214 215 query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)216 key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)217 218 value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)219 value = value.permute(0, 2, 1, 3)220 221 seq_len = key.shape[1]222 offset = 0223 224 if layer_past is not None:225 offset = layer_past[0].shape[-2]226 seq_len += offset227 228 if self.rotary_dim is not None:229 k_rot = key[:, :, :, : self.rotary_dim]230 k_pass = key[:, :, :, self.rotary_dim:]231 232 q_rot = query[:, :, :, : self.rotary_dim]233 q_pass = query[:, :, :, self.rotary_dim:]234 235 sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)236 k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)237 seq_len_q = query.shape[1]238 sincos_q = fixed_pos_embedding(q_rot, 1, seq_len=seq_len_q)239 q_rot = apply_rotary_pos_emb(q_rot, sincos_q, offset=offset)240 241 key = torch.cat([k_rot, k_pass], dim=-1)242 query = torch.cat([q_rot, q_pass], dim=-1)243 else:244 sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)245 key = apply_rotary_pos_emb(key, sincos, offset=offset)246 query = apply_rotary_pos_emb(query, sincos, offset=offset)247 248 key = key.permute(0, 2, 1, 3)249 query = query.permute(0, 2, 1, 3)250 251 if layer_past is not None:252 past_key = layer_past[0]253 past_value = layer_past[1]254 key = torch.cat((past_key, key), dim=-2)255 value = torch.cat((past_value, value), dim=-2)256 257 if use_cache is True:258 present = (key, value)259 else:260 present = None261 262 # compute self-attention: V x Softmax(QK^T)263 attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)264 265 attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)266 attn_output = self.out_proj(attn_output)267 attn_output = self.resid_dropout(attn_output)268 269 outputs = (attn_output, present)270 if output_attentions:271 outputs += (attn_weights,)272 273 return outputs # a, present, (attentions)274 275 276# Adapted from transformers.models.codegen.modeling_codegen.CodeGenMLP277class CodeT5pMLP(nn.Module):278 def __init__(self, intermediate_size, config): # in MLP: intermediate_size= 4 * embed_dim279 super().__init__()280 embed_dim = config.n_embd281 282 self.fc_in = nn.Linear(embed_dim, intermediate_size)283 self.fc_out = nn.Linear(intermediate_size, embed_dim)284 285 self.act = ACT2FN[config.activation_function]286 self.dropout = nn.Dropout(config.resid_pdrop)287 288 def forward(self, hidden_states: Optional[torch.FloatTensor]) -> torch.FloatTensor:289 hidden_states = self.fc_in(hidden_states)290 hidden_states = self.act(hidden_states)291 hidden_states = self.fc_out(hidden_states)292 hidden_states = self.dropout(hidden_states)293 return hidden_states294 295 296# Adapted from transformers.models.codegen.modeling_codegen.CodeGenBlock297class CodeT5pBlock(nn.Module):298 def __init__(self, config, layer_idx=None):299 super().__init__()300 inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd301 self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)302 303 if config.is_decoder is False:304 self.attn = CodeT5pAttention(config, is_cross_attention=False, is_decoder=False)305 else:306 self.attn = CodeT5pAttention(config)307 self.mlp = CodeT5pMLP(inner_dim, config)308 309 # Adding 1 cross-attention layer at the final decoder layer310 self.add_cross_attention_by_layer = True \311 if config.add_cross_attention and layer_idx == config.n_layer - 1 else False312 313 if config.add_cross_attention and self.add_cross_attention_by_layer:314 self.crossattention = CodeT5pAttention(config, is_cross_attention=True)315 316 def forward(317 self,318 hidden_states: Optional[torch.FloatTensor],319 layer_past: Optional[Tuple[torch.Tensor]] = None,320 attention_mask: Optional[torch.FloatTensor] = None,321 head_mask: Optional[torch.FloatTensor] = None,322 encoder_hidden_states: Optional[torch.Tensor] = None,323 encoder_attention_mask: Optional[torch.FloatTensor] = None,324 use_cache: Optional[bool] = False,325 output_attentions: Optional[bool] = False,326 ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:327 residual = hidden_states328 hidden_states = self.ln_1(hidden_states)329 attn_outputs = self.attn(330 hidden_states,331 layer_past=layer_past,332 attention_mask=attention_mask,333 head_mask=head_mask,334 use_cache=use_cache,335 output_attentions=output_attentions,336 )337 attn_output = attn_outputs[0] # output_attn: a, present, (attentions)338 outputs = attn_outputs[1:]339 feed_forward_hidden_states = self.mlp(hidden_states)340 341 if encoder_hidden_states is not None and self.add_cross_attention_by_layer:342 # add one self-attention block for cross-attention343 if not hasattr(self, "crossattention"):344 raise ValueError(345 f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "346 "cross-attention layers by setting `config.add_cross_attention=True`"347 )348 # residual = hidden_states349 # hidden_states = self.ln_cross_attn(residual)350 cross_attn_outputs = self.crossattention(351 hidden_states,352 attention_mask=attention_mask,353 head_mask=head_mask,354 encoder_hidden_states=encoder_hidden_states,355 encoder_attention_mask=encoder_attention_mask,356 output_attentions=output_attentions,357 )358 xattn_output = cross_attn_outputs[0]359 attn_output = attn_output + xattn_output360 outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights361 362 hidden_states = attn_output + feed_forward_hidden_states + residual363 364 if use_cache:365 outputs = (hidden_states,) + outputs366 else:367 outputs = (hidden_states,) + outputs[1:]368 369 return outputs # hidden_states, present, (attentions)370 371 372# Adapted from transformers.models.codegen.modeling_codegen.CodeGenPreTrainedModel373class CodeT5pPreTrainedModel(PreTrainedModel):374 """375 An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained376 models.377 """378 config_class = CodeT5pModuleConfig379 base_model_prefix = "transformer"380 supports_gradient_checkpointing = True381 _no_split_modules = ["CodeT5pBlock"]382 383 def __init__(self, *inputs, **kwargs):384 super().__init__(*inputs, **kwargs)385 386 def _init_weights(self, module):387 """Initialize the weights."""388 if isinstance(module, (nn.Linear,)):389 # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization390 # cf https://github.com/pytorch/pytorch/pull/5617391 module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)392 if module.bias is not None:393 module.bias.data.zero_()394 elif isinstance(module, nn.Embedding):395 module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)396 if module.padding_idx is not None:397 module.weight.data[module.padding_idx].zero_()398 elif isinstance(module, nn.LayerNorm):399 module.bias.data.zero_()400 module.weight.data.fill_(1.0)401 402 def _set_gradient_checkpointing(self, module, value=False):403 if isinstance(module, CodeT5pModel):404 module.gradient_checkpointing = value405 406 407# Adapted from transformers.models.codegen.modeling_codegen.CodeGenModel408class CodeT5pModel(CodeT5pPreTrainedModel):409 def __init__(self, config):410 super().__init__(config)411 412 self.embed_dim = config.n_embd413 self.vocab_size = config.vocab_size414 self.wte = nn.Embedding(config.vocab_size, self.embed_dim)415 self.drop = nn.Dropout(config.embd_pdrop)416 self.h = nn.ModuleList([CodeT5pBlock(config, idx) for idx in range(config.n_layer)])417 self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)418 self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)419 420 self.gradient_checkpointing = False421 422 # Initialize weights and apply final processing423 self.post_init()424 425 def get_input_embeddings(self):426 return self.wte427 428 def set_input_embeddings(self, new_embeddings):429 self.wte = new_embeddings430 431 def forward(432 self,433 input_ids: Optional[torch.LongTensor] = None,434 past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,435 attention_mask: Optional[torch.FloatTensor] = None,436 token_type_ids: Optional[torch.LongTensor] = None,437 position_ids: Optional[torch.LongTensor] = None,438 head_mask: Optional[torch.FloatTensor] = None,439 inputs_embeds: Optional[torch.FloatTensor] = None,440 encoder_hidden_states: Optional[torch.Tensor] = None,441 encoder_attention_mask: Optional[torch.FloatTensor] = None,442 use_cache: Optional[bool] = None,443 output_attentions: Optional[bool] = None,444 output_hidden_states: Optional[bool] = None,445 return_dict: Optional[bool] = None,446 ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:447 output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions448 output_hidden_states = (449 output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states450 )451 use_cache = use_cache if use_cache is not None else self.config.use_cache452 return_dict = return_dict if return_dict is not None else self.config.use_return_dict453 454 if input_ids is not None and inputs_embeds is not None:455 raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")456 elif input_ids is not None:457 input_shape = input_ids.size()458 input_ids = input_ids.view(-1, input_shape[-1])459 batch_size = input_ids.shape[0]460 elif inputs_embeds is not None:461 input_shape = inputs_embeds.size()[:-1]462 batch_size = inputs_embeds.shape[0]463 else:464 raise ValueError("You have to specify either input_ids or inputs_embeds")465 466 device = input_ids.device if input_ids is not None else inputs_embeds.device467 468 if token_type_ids is not None:469 token_type_ids = token_type_ids.view(-1, input_shape[-1])470 471 if position_ids is not None:472 position_ids = position_ids.view(-1, input_shape[-1])473 474 if past_key_values is None:475 past_length = 0476 past_key_values = tuple([None] * len(self.h))477 else:478 past_length = past_key_values[0][0].size(-2)479 480 if position_ids is None:481 position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)482 position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])483 484 # Attention mask.485 if attention_mask is not None:486 if batch_size <= 0:487 raise ValueError("batch_size has to be defined and > 0")488 attention_mask = attention_mask.view(batch_size, -1)489 # We create a 3D attention mask from a 2D tensor mask.490 # Sizes are [batch_size, 1, 1, to_seq_length]491 # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]492 # this attention mask is more simple than the triangular masking of causal attention493 # used in OpenAI GPT, we just need to prepare the broadcast dimension here.494 attention_mask = attention_mask[:, None, None, :]495 496 # Since attention_mask is 1.0 for positions we want to attend and 0.0 for497 # masked positions, this operation will create a tensor which is 0.0 for498 # positions we want to attend and the dtype's smallest value for masked positions.499 # Since we are adding it to the raw scores before the softmax, this is500 # effectively the same as removing these entirely.501 attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility502 attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min503 504 # If a 2D or 3D attention mask is provided for the cross-attention505 # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]506 if self.config.add_cross_attention and encoder_hidden_states is not None:507 encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()508 encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)509 if encoder_attention_mask is None:510 encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)511 encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)512 else:513 encoder_attention_mask = None514 515 # Prepare head mask if needed516 # 1.0 in head_mask indicate we keep the head517 # attention_probs has shape bsz x num_attention_heads x N x N518 # head_mask has shape n_layer x batch x num_attention_heads x N x N519 head_mask = self.get_head_mask(head_mask, self.config.n_layer)520 521 if inputs_embeds is None:522 inputs_embeds = self.wte(input_ids)523 524 hidden_states = inputs_embeds525 526 if token_type_ids is not None:527 token_type_embeds = self.wte(token_type_ids)528 hidden_states = hidden_states + token_type_embeds529 530 hidden_states = self.drop(hidden_states)531 532 output_shape = input_shape + (hidden_states.size(-1),)533 534 presents = () if use_cache else None535 all_self_attentions = () if output_attentions else None536 all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None537 all_hidden_states = () if output_hidden_states else None538 for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):539 if output_hidden_states:540 all_hidden_states = all_hidden_states + (hidden_states,)541 542 if self.gradient_checkpointing and self.training:543 if use_cache:544 logger.warning(545 "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "546 "`use_cache=False`..."547 )548 use_cache = False549 550 def create_custom_forward(module):551 def custom_forward(*inputs):552 # None for past_key_value553 return module(*inputs, use_cache, output_attentions)554 555 return custom_forward556 557 outputs = torch.utils.checkpoint.checkpoint(558 create_custom_forward(block),559 hidden_states,560 None,561 attention_mask,562 head_mask[i],563 encoder_hidden_states,564 encoder_attention_mask,565 )566 else:567 outputs = block(568 hidden_states,569 layer_past=layer_past,570 attention_mask=attention_mask,571 head_mask=head_mask[i],572 encoder_hidden_states=encoder_hidden_states,573 encoder_attention_mask=encoder_attention_mask,574 use_cache=use_cache,575 output_attentions=output_attentions,576 )577 578 hidden_states = outputs[0]579 if use_cache is True:580 presents = presents + (outputs[1],)581 582 if output_attentions:583 all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)584 if self.config.add_cross_attention and self.add_cross_attention_by_layer:585 all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)586 587 hidden_states = self.ln_f(hidden_states)588 589 hidden_states = hidden_states.view(output_shape)590 # Add last hidden state591 if output_hidden_states:592 all_hidden_states = all_hidden_states + (hidden_states,)593 594 if not return_dict:595 return tuple(596 v for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions] if597 v is not None)598 599 return BaseModelOutputWithPastAndCrossAttentions(600 last_hidden_state=hidden_states,601 past_key_values=presents,602 hidden_states=all_hidden_states,603 attentions=all_self_attentions,604 cross_attentions=all_cross_attentions,605 )606 607 608# Adapted from transformers.models.codegen.modeling_codegen.CodeGenForCausalLM609class CodeT5pForCausalLM(CodeT5pPreTrainedModel):610 _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.causal_mask"]611 612 def __init__(self, config):613 super().__init__(config)614 self.transformer = CodeT5pModel(config)615 self.lm_head = nn.Linear(config.n_embd, config.vocab_size)616 617 # Initialize weights and apply final processing618 self.post_init()619 620 def get_output_embeddings(self):621 return self.lm_head622 623 def set_output_embeddings(self, new_embeddings):624 self.lm_head = new_embeddings625 626 def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):627 token_type_ids = kwargs.get("token_type_ids", None)628 # only last token for inputs_ids if past is defined in kwargs629 if past_key_values:630 input_ids = input_ids[:, -1].unsqueeze(-1)631 if token_type_ids is not None:632 token_type_ids = token_type_ids[:, -1].unsqueeze(-1)633 634 attention_mask = kwargs.get("attention_mask", None)635 position_ids = kwargs.get("position_ids", None)636 637 if attention_mask is not None and position_ids is None:638 # create position_ids on the fly for batch generation639 position_ids = attention_mask.long().cumsum(-1) - 1640 position_ids.masked_fill_(attention_mask == 0, 1)641 if past_key_values:642 position_ids = position_ids[:, -1].unsqueeze(-1)643 else:644 position_ids = None645 return {646 "input_ids": input_ids,647 "past_key_values": past_key_values,648 "use_cache": kwargs.get("use_cache"),649 "position_ids": position_ids,650 "attention_mask": attention_mask,651 "token_type_ids": token_type_ids,652 }653 654 def forward(655 self,656 input_ids: Optional[torch.LongTensor] = None,657 past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,658 attention_mask: Optional[torch.FloatTensor] = None,659 token_type_ids: Optional[torch.LongTensor] = None,660 position_ids: Optional[torch.LongTensor] = None,661 head_mask: Optional[torch.FloatTensor] = None,662 inputs_embeds: Optional[torch.FloatTensor] = None,663 encoder_hidden_states: Optional[torch.Tensor] = None,664 encoder_attention_mask: Optional[torch.FloatTensor] = None,665 labels: Optional[torch.LongTensor] = None,666 use_cache: Optional[bool] = None,667 output_attentions: Optional[bool] = None,668 output_hidden_states: Optional[bool] = None,669 return_dict: Optional[bool] = None,670 ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:671 r"""672 labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):673 Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set674 `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`675 are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`676 """677 return_dict = return_dict if return_dict is not None else self.config.use_return_dict678 679 transformer_outputs = self.transformer(680 input_ids,681 past_key_values=past_key_values,682 attention_mask=attention_mask,683 token_type_ids=token_type_ids,684 position_ids=position_ids,685 head_mask=head_mask,686 inputs_embeds=inputs_embeds,687 encoder_hidden_states=encoder_hidden_states,688 encoder_attention_mask=encoder_attention_mask,689 use_cache=use_cache,690 output_attentions=output_attentions,691 output_hidden_states=output_hidden_states,692 return_dict=return_dict,693 )694 hidden_states = transformer_outputs[0]695 696 # make sure sampling in fp16 works correctly and697 # compute loss in fp32 to match with mesh-tf version698 # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179699 lm_logits = self.lm_head(hidden_states).to(torch.float32)700 701 loss = None702 if labels is not None:703 # Shift so that tokens < n predict n704 shift_logits = lm_logits[..., :-1, :].contiguous()705 shift_labels = labels[..., 1:].contiguous()706 # Flatten the tokens707 loss_fct = CrossEntropyLoss()708 loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))709 710 loss = loss.to(hidden_states.dtype)711 712 if not return_dict:713 output = (lm_logits,) + transformer_outputs[1:]714 return ((loss,) + output) if loss is not None else output715 716 return CausalLMOutputWithCrossAttentions(717 loss=loss,718 logits=lm_logits,719 past_key_values=transformer_outputs.past_key_values,720 hidden_states=transformer_outputs.hidden_states,721 attentions=transformer_outputs.attentions,722 cross_attentions=transformer_outputs.cross_attentions,723 )724 725 @staticmethod726 def _reorder_cache(727 past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor728 ) -> Tuple[Tuple[torch.Tensor]]:729 """730 This function is used to re-order the `past_key_values` cache if [`~PretrainedModel.beam_search`] or731 [`~PretrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct732 beam_idx at every generation step.733 """734 return tuple(735 tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)736 for layer_past in past_key_values737 )738 739 740def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):741 """742 Shift input ids one token to the right.743 """744 shifted_input_ids = input_ids.new_zeros(input_ids.shape)745 shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()746 if decoder_start_token_id is None:747 raise ValueError("Make sure to set the decoder_start_token_id attribute of the model's configuration.")748 shifted_input_ids[:, 0] = decoder_start_token_id749 750 if pad_token_id is None:751 raise ValueError("Make sure to set the pad_token_id attribute of the model's configuration.")752 # replace possible -100 values in labels by `pad_token_id`753 shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)754 755 return shifted_input_ids756 757 758# Adapted from transformers.models.encoder_decoder.modeling_encoder_decoder.EncoderDecoderModel759class CodeT5pEncoderDecoderModel(PreTrainedModel):760 config_class = CodeT5pConfig761 _no_split_modules = ["CodeT5pBlock"]762 def __init__(763 self,764 config: Optional[PretrainedConfig] = None,765 encoder: Optional[PreTrainedModel] = None,766 decoder: Optional[PreTrainedModel] = None,767 ):768 if config is None and (encoder is None or decoder is None):769 raise ValueError("Either a configuration or an encoder and a decoder has to be provided.")770 if config is None:771 config = CodeT5pConfig.from_encoder_decoder_configs(encoder.config, decoder.config)772 else:773 if not isinstance(config, self.config_class):774 raise ValueError(f"Config: {config} has to be of type {self.config_class}")775 776 if config.decoder.cross_attention_hidden_size is not None:777 if config.decoder.cross_attention_hidden_size != config.encoder.hidden_size:778 raise ValueError(779 "If `cross_attention_hidden_size` is specified in the decoder's configuration, it has to be equal"780 f" to the encoder's `hidden_size`. Got {config.decoder.cross_attention_hidden_size} for"781 f" `config.decoder.cross_attention_hidden_size` and {config.encoder.hidden_size} for"782 " `config.encoder.hidden_size`."783 )784 785 # initialize with config786 super().__init__(config)787 788 if encoder is None:789 encoder = CodeT5pModel(config.encoder)790 791 if decoder is None:792 decoder = CodeT5pForCausalLM(config.decoder)793 794 self.encoder = encoder795 self.decoder = decoder796 797 if self.encoder.config.to_dict() != self.config.encoder.to_dict():798 logger.warning(799 f"Config of the encoder: {self.encoder.__class__} is overwritten by shared encoder config:"800 f" {self.config.encoder}"801 )802 if self.decoder.config.to_dict() != self.config.decoder.to_dict():803 logger.warning(804 f"Config of the decoder: {self.decoder.__class__} is overwritten by shared decoder config:"805 f" {self.config.decoder}"806 )807 808 # make sure that the individual model's config refers to the shared config809 # so that the updates to the config will be synced810 self.encoder.config = self.config.encoder811 self.decoder.config = self.config.decoder812 813 # encoder outputs might need to be projected to different dimension for decoder814 if (815 self.encoder.config.hidden_size != self.decoder.config.hidden_size816 and self.decoder.config.cross_attention_hidden_size is None817 ):818 self.enc_to_dec_proj = nn.Linear(self.encoder.config.hidden_size, self.decoder.config.hidden_size)819 820 if self.encoder.get_output_embeddings() is not None:821 raise ValueError(822 f"The encoder {self.encoder} should not have a LM Head. Please use a model without LM Head"823 )824 # tie encoder, decoder weights if config set accordingly825 self.tie_weights()826 827 def tie_weights(self):828 # tie encoder & decoder if needed829 if self.config.tie_encoder_decoder:830 # tie encoder and decoder base model831 decoder_base_model_prefix = self.decoder.base_model_prefix832 self._tie_encoder_decoder_weights(833 self.encoder, self.decoder._modules[decoder_base_model_prefix], self.decoder.base_model_prefix834 )835 836 def get_encoder(self):837 return self.encoder838 839 def get_decoder(self):840 return self.decoder841 842 def get_input_embeddings(self):843 return self.encoder.get_input_embeddings()844 845 def get_output_embeddings(self):846 return self.decoder.get_output_embeddings()847 848 def set_output_embeddings(self, new_embeddings):849 return self.decoder.set_output_embeddings(new_embeddings)850 851 @classmethod852 def from_pretrained(cls, *args, **kwargs):853 # At the moment fast initialization is not supported for composite models854 if kwargs.get("_fast_init", False):855 logger.warning(856 "Fast initialization is currently not supported for EncoderDecoderModel. "857 "Falling back to slow initialization..."858 )859 kwargs["_fast_init"] = False860 return super().from_pretrained(*args, **kwargs)861 862 def forward(863 self,864 input_ids: Optional[torch.LongTensor] = None,865 attention_mask: Optional[torch.FloatTensor] = None,866 decoder_input_ids: Optional[torch.LongTensor] = None,867 decoder_attention_mask: Optional[torch.BoolTensor] = None,868 encoder_outputs: Optional[Tuple[torch.FloatTensor]] = None,869 past_key_values: Tuple[Tuple[torch.FloatTensor]] = None,870 inputs_embeds: Optional[torch.FloatTensor] = None,871 decoder_inputs_embeds: Optional[torch.FloatTensor] = None,872 labels: Optional[torch.LongTensor] = None,873 use_cache: Optional[bool] = None,874 output_attentions: Optional[bool] = None,875 output_hidden_states: Optional[bool] = None,876 return_dict: Optional[bool] = None,877 **kwargs,878 ) -> Union[Tuple, Seq2SeqLMOutput]:879 return_dict = return_dict if return_dict is not None else self.config.use_return_dict880 881 kwargs_encoder = {argument: value for argument, value in kwargs.items() if not argument.startswith("decoder_")}882 883 kwargs_decoder = {884 argument[len("decoder_"):]: value for argument, value in kwargs.items() if argument.startswith("decoder_")885 }886 887 if encoder_outputs is None:888 encoder_outputs = self.encoder(889 input_ids=input_ids,890 attention_mask=attention_mask,891 inputs_embeds=inputs_embeds,892 output_attentions=output_attentions,893 output_hidden_states=output_hidden_states,894 return_dict=return_dict,895 **kwargs_encoder,896 )897 elif isinstance(encoder_outputs, tuple):898 encoder_outputs = BaseModelOutput(*encoder_outputs)899 900 encoder_hidden_states = encoder_outputs[0]901 902 # optionally project encoder_hidden_states903 if (904 self.encoder.config.hidden_size != self.decoder.config.hidden_size905 and self.decoder.config.cross_attention_hidden_size is None906 ):907 encoder_hidden_states = self.enc_to_dec_proj(encoder_hidden_states)908 909 if (labels is not None) and (decoder_input_ids is None and decoder_inputs_embeds is None):910 decoder_input_ids = shift_tokens_right(911 labels, self.config.pad_token_id, self.config.decoder_start_token_id912 )913 914 # Decode915 decoder_outputs = self.decoder(916 input_ids=decoder_input_ids,917 attention_mask=decoder_attention_mask,918 encoder_hidden_states=encoder_hidden_states,919 encoder_attention_mask=attention_mask,920 inputs_embeds=decoder_inputs_embeds,921 output_attentions=output_attentions,922 output_hidden_states=output_hidden_states,923 use_cache=use_cache,924 past_key_values=past_key_values,925 return_dict=return_dict,926 **kwargs_decoder,927 )928 929 # Compute loss independent from decoder (as some shift the logits inside them)930 loss = None931 if labels is not None:932 # warnings.warn(DEPRECATION_WARNING, FutureWarning)933 logits = decoder_outputs.logits if return_dict else decoder_outputs[0]934 loss_fct = CrossEntropyLoss()935 loss = loss_fct(logits.reshape(-1, self.decoder.config.vocab_size), labels.view(-1))936 937 if not return_dict:938 if loss is not None:939 return (loss,) + decoder_outputs + encoder_outputs940 else:941 return decoder_outputs + encoder_outputs942 943 return Seq2SeqLMOutput(944 loss=loss,945 logits=decoder_outputs.logits,946 past_key_values=decoder_outputs.past_key_values,947 decoder_hidden_states=decoder_outputs.hidden_states,948 decoder_attentions=decoder_outputs.attentions,949 cross_attentions=decoder_outputs.cross_attentions,950 encoder_last_hidden_state=encoder_outputs.last_hidden_state,951 encoder_hidden_states=encoder_outputs.hidden_states,952 encoder_attentions=encoder_outputs.attentions,953 )954 955 def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):956 return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)957 958 def prepare_inputs_for_generation(959 self, input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs960 ):961 decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids, past=past)962 decoder_attention_mask = decoder_inputs["attention_mask"] if "attention_mask" in decoder_inputs else None963 input_dict = {964 "attention_mask": attention_mask,965 "decoder_attention_mask": decoder_attention_mask,966 "decoder_input_ids": decoder_inputs["input_ids"],967 "encoder_outputs": encoder_outputs,968 "past_key_values": decoder_inputs["past_key_values"],969 "use_cache": use_cache,970 }971 return input_dict972 973 def resize_token_embeddings(self, *args, **kwargs):974 raise NotImplementedError(975 "Resizing the embedding layers via the EncoderDecoderModel directly is not supported. Please use the"976 " respective methods of the wrapped objects (model.encoder.resize_token_embeddings(...) or"977 " model.decoder.resize_token_embeddings(...))"978 )979 980 def _reorder_cache(self, past, beam_idx):981 # apply decoder cache reordering here982 return self.decoder._reorder_cache(past, beam_idx)983