fix sequence length in santacoder and introduce new model type
#23
by
mayank-mishra
- opened
- config.json +1 -1
- modeling_gpt2_mq.py +200 -23
config.json
CHANGED
@@ -14,7 +14,7 @@
|
|
14 |
"eos_token_id": 50256,
|
15 |
"initializer_range": 0.02,
|
16 |
"layer_norm_epsilon": 1e-05,
|
17 |
-
"model_type": "
|
18 |
"n_embd": 2048,
|
19 |
"n_head": 16,
|
20 |
"n_inner": 8192,
|
|
|
14 |
"eos_token_id": 50256,
|
15 |
"initializer_range": 0.02,
|
16 |
"layer_norm_epsilon": 1e-05,
|
17 |
+
"model_type": "santacoder",
|
18 |
"n_embd": 2048,
|
19 |
"n_head": 16,
|
20 |
"n_inner": 8192,
|
modeling_gpt2_mq.py
CHANGED
@@ -1,39 +1,21 @@
|
|
1 |
"""PyTorch OpenAI GPT-2 model modified with MultiQuery attention"""
|
2 |
|
3 |
|
4 |
-
import math
|
5 |
-
import os
|
6 |
-
from dataclasses import dataclass
|
7 |
from typing import Optional, Tuple, Union
|
8 |
|
9 |
import torch
|
10 |
import torch.utils.checkpoint
|
11 |
from torch import nn
|
12 |
from torch.cuda.amp import autocast
|
13 |
-
|
14 |
-
|
15 |
-
from transformers.activations import ACT2FN
|
16 |
-
from transformers.modeling_outputs import (
|
17 |
-
BaseModelOutputWithPastAndCrossAttentions,
|
18 |
-
CausalLMOutputWithCrossAttentions,
|
19 |
-
SequenceClassifierOutputWithPast,
|
20 |
-
TokenClassifierOutput,
|
21 |
-
)
|
22 |
-
from transformers.modeling_utils import PreTrainedModel, SequenceSummary
|
23 |
from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
|
24 |
|
25 |
-
from transformers.utils import
|
26 |
-
ModelOutput,
|
27 |
-
add_code_sample_docstrings,
|
28 |
-
add_start_docstrings,
|
29 |
-
add_start_docstrings_to_model_forward,
|
30 |
-
logging,
|
31 |
-
replace_return_docstrings,
|
32 |
-
)
|
33 |
-
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
|
34 |
from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2Block, GPT2PreTrainedModel, GPT2LMHeadModel
|
35 |
-
from .configuration_gpt2_mq import GPT2CustomConfig, MULTI_QUERY
|
36 |
|
|
|
37 |
|
38 |
|
39 |
class GPT2MQAttention(nn.Module):
|
@@ -329,6 +311,201 @@ class GPT2CustomModel(GPT2Model):
|
|
329 |
# Initialize weights and apply final processing
|
330 |
self.post_init()
|
331 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
|
333 |
class GPT2LMHeadCustomModel(GPT2LMHeadModel):
|
334 |
config_class = GPT2CustomConfig
|
|
|
1 |
"""PyTorch OpenAI GPT-2 model modified with MultiQuery attention"""
|
2 |
|
3 |
|
|
|
|
|
|
|
4 |
from typing import Optional, Tuple, Union
|
5 |
|
6 |
import torch
|
7 |
import torch.utils.checkpoint
|
8 |
from torch import nn
|
9 |
from torch.cuda.amp import autocast
|
10 |
+
|
11 |
+
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
|
13 |
|
14 |
+
from transformers.utils import logging
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2Block, GPT2PreTrainedModel, GPT2LMHeadModel
|
16 |
+
from .configuration_gpt2_mq import GPT2CustomConfig, MULTI_QUERY
|
17 |
|
18 |
+
logger = logging.get_logger(__name__)
|
19 |
|
20 |
|
21 |
class GPT2MQAttention(nn.Module):
|
|
|
311 |
# Initialize weights and apply final processing
|
312 |
self.post_init()
|
313 |
|
314 |
+
def forward(
|
315 |
+
self,
|
316 |
+
input_ids: Optional[torch.LongTensor] = None,
|
317 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
318 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
319 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
320 |
+
position_ids: Optional[torch.LongTensor] = None,
|
321 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
322 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
323 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
324 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
325 |
+
use_cache: Optional[bool] = None,
|
326 |
+
output_attentions: Optional[bool] = None,
|
327 |
+
output_hidden_states: Optional[bool] = None,
|
328 |
+
return_dict: Optional[bool] = None,
|
329 |
+
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
330 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
331 |
+
output_hidden_states = (
|
332 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
333 |
+
)
|
334 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
335 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
336 |
+
|
337 |
+
if input_ids is not None and inputs_embeds is not None:
|
338 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
339 |
+
elif input_ids is not None:
|
340 |
+
input_shape = input_ids.size()
|
341 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
342 |
+
batch_size = input_ids.shape[0]
|
343 |
+
elif inputs_embeds is not None:
|
344 |
+
input_shape = inputs_embeds.size()[:-1]
|
345 |
+
batch_size = inputs_embeds.shape[0]
|
346 |
+
else:
|
347 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
348 |
+
|
349 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
350 |
+
|
351 |
+
if token_type_ids is not None:
|
352 |
+
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
353 |
+
if position_ids is not None:
|
354 |
+
position_ids = position_ids.view(-1, input_shape[-1])
|
355 |
+
|
356 |
+
if past_key_values is None:
|
357 |
+
past_length = 0
|
358 |
+
past_key_values = tuple([None] * len(self.h))
|
359 |
+
else:
|
360 |
+
# this is different from GPT2
|
361 |
+
past_length = past_key_values[0][0].size(-1)
|
362 |
+
if position_ids is None:
|
363 |
+
position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
|
364 |
+
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
|
365 |
+
|
366 |
+
# GPT2Attention mask.
|
367 |
+
if attention_mask is not None:
|
368 |
+
if batch_size <= 0:
|
369 |
+
raise ValueError("batch_size has to be defined and > 0")
|
370 |
+
attention_mask = attention_mask.view(batch_size, -1)
|
371 |
+
# We create a 3D attention mask from a 2D tensor mask.
|
372 |
+
# Sizes are [batch_size, 1, 1, to_seq_length]
|
373 |
+
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
|
374 |
+
# this attention mask is more simple than the triangular masking of causal attention
|
375 |
+
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
|
376 |
+
attention_mask = attention_mask[:, None, None, :]
|
377 |
+
|
378 |
+
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
|
379 |
+
# masked positions, this operation will create a tensor which is 0.0 for
|
380 |
+
# positions we want to attend and the dtype's smallest value for masked positions.
|
381 |
+
# Since we are adding it to the raw scores before the softmax, this is
|
382 |
+
# effectively the same as removing these entirely.
|
383 |
+
attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
|
384 |
+
attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
|
385 |
+
|
386 |
+
# If a 2D or 3D attention mask is provided for the cross-attention
|
387 |
+
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
388 |
+
if self.config.add_cross_attention and encoder_hidden_states is not None:
|
389 |
+
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
|
390 |
+
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
|
391 |
+
if encoder_attention_mask is None:
|
392 |
+
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
|
393 |
+
encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
|
394 |
+
else:
|
395 |
+
encoder_attention_mask = None
|
396 |
+
|
397 |
+
# Prepare head mask if needed
|
398 |
+
# 1.0 in head_mask indicate we keep the head
|
399 |
+
# attention_probs has shape bsz x n_heads x N x N
|
400 |
+
# head_mask has shape n_layer x batch x n_heads x N x N
|
401 |
+
head_mask = self.get_head_mask(head_mask, self.config.n_layer)
|
402 |
+
|
403 |
+
if inputs_embeds is None:
|
404 |
+
inputs_embeds = self.wte(input_ids)
|
405 |
+
position_embeds = self.wpe(position_ids)
|
406 |
+
hidden_states = inputs_embeds + position_embeds
|
407 |
+
|
408 |
+
if token_type_ids is not None:
|
409 |
+
token_type_embeds = self.wte(token_type_ids)
|
410 |
+
hidden_states = hidden_states + token_type_embeds
|
411 |
+
|
412 |
+
hidden_states = self.drop(hidden_states)
|
413 |
+
|
414 |
+
output_shape = input_shape + (hidden_states.size(-1),)
|
415 |
+
|
416 |
+
presents = () if use_cache else None
|
417 |
+
all_self_attentions = () if output_attentions else None
|
418 |
+
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
419 |
+
all_hidden_states = () if output_hidden_states else None
|
420 |
+
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
421 |
+
|
422 |
+
# Model parallel
|
423 |
+
if self.model_parallel:
|
424 |
+
torch.cuda.set_device(hidden_states.device)
|
425 |
+
# Ensure layer_past is on same device as hidden_states (might not be correct)
|
426 |
+
if layer_past is not None:
|
427 |
+
layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
|
428 |
+
# Ensure that attention_mask is always on the same device as hidden_states
|
429 |
+
if attention_mask is not None:
|
430 |
+
attention_mask = attention_mask.to(hidden_states.device)
|
431 |
+
if isinstance(head_mask, torch.Tensor):
|
432 |
+
head_mask = head_mask.to(hidden_states.device)
|
433 |
+
if output_hidden_states:
|
434 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
435 |
+
|
436 |
+
if self.gradient_checkpointing and self.training:
|
437 |
+
|
438 |
+
if use_cache:
|
439 |
+
logger.warning(
|
440 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
441 |
+
)
|
442 |
+
use_cache = False
|
443 |
+
|
444 |
+
def create_custom_forward(module):
|
445 |
+
def custom_forward(*inputs):
|
446 |
+
# None for past_key_value
|
447 |
+
return module(*inputs, use_cache, output_attentions)
|
448 |
+
|
449 |
+
return custom_forward
|
450 |
+
|
451 |
+
outputs = torch.utils.checkpoint.checkpoint(
|
452 |
+
create_custom_forward(block),
|
453 |
+
hidden_states,
|
454 |
+
None,
|
455 |
+
attention_mask,
|
456 |
+
head_mask[i],
|
457 |
+
encoder_hidden_states,
|
458 |
+
encoder_attention_mask,
|
459 |
+
)
|
460 |
+
else:
|
461 |
+
outputs = block(
|
462 |
+
hidden_states,
|
463 |
+
layer_past=layer_past,
|
464 |
+
attention_mask=attention_mask,
|
465 |
+
head_mask=head_mask[i],
|
466 |
+
encoder_hidden_states=encoder_hidden_states,
|
467 |
+
encoder_attention_mask=encoder_attention_mask,
|
468 |
+
use_cache=use_cache,
|
469 |
+
output_attentions=output_attentions,
|
470 |
+
)
|
471 |
+
|
472 |
+
hidden_states = outputs[0]
|
473 |
+
if use_cache is True:
|
474 |
+
presents = presents + (outputs[1],)
|
475 |
+
|
476 |
+
if output_attentions:
|
477 |
+
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
|
478 |
+
if self.config.add_cross_attention:
|
479 |
+
all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
|
480 |
+
|
481 |
+
# Model Parallel: If it's the last layer for that device, put things on the next device
|
482 |
+
if self.model_parallel:
|
483 |
+
for k, v in self.device_map.items():
|
484 |
+
if i == v[-1] and "cuda:" + str(k) != self.last_device:
|
485 |
+
hidden_states = hidden_states.to("cuda:" + str(k + 1))
|
486 |
+
|
487 |
+
hidden_states = self.ln_f(hidden_states)
|
488 |
+
|
489 |
+
hidden_states = hidden_states.view(output_shape)
|
490 |
+
# Add last hidden state
|
491 |
+
if output_hidden_states:
|
492 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
493 |
+
|
494 |
+
if not return_dict:
|
495 |
+
return tuple(
|
496 |
+
v
|
497 |
+
for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
|
498 |
+
if v is not None
|
499 |
+
)
|
500 |
+
|
501 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
502 |
+
last_hidden_state=hidden_states,
|
503 |
+
past_key_values=presents,
|
504 |
+
hidden_states=all_hidden_states,
|
505 |
+
attentions=all_self_attentions,
|
506 |
+
cross_attentions=all_cross_attentions,
|
507 |
+
)
|
508 |
+
|
509 |
|
510 |
class GPT2LMHeadCustomModel(GPT2LMHeadModel):
|
511 |
config_class = GPT2CustomConfig
|