Modify to original GLM-4-9B code
Browse files- modeling_chatglm.py +298 -132
modeling_chatglm.py
CHANGED
@@ -1,42 +1,39 @@
|
|
1 |
""" PyTorch ChatGLM model. """
|
2 |
|
3 |
import math
|
4 |
-
import copy
|
5 |
-
import warnings
|
6 |
-
import re
|
7 |
import sys
|
8 |
-
|
9 |
import torch
|
10 |
import torch.utils.checkpoint
|
11 |
import torch.nn.functional as F
|
12 |
from torch import nn
|
13 |
-
from torch.nn import CrossEntropyLoss, LayerNorm
|
14 |
from torch.nn.utils import skip_init
|
15 |
-
from typing import Optional, Tuple, Union, List,
|
16 |
|
17 |
from transformers.modeling_outputs import (
|
18 |
BaseModelOutputWithPast,
|
19 |
CausalLMOutputWithPast,
|
|
|
20 |
)
|
21 |
from transformers.modeling_utils import PreTrainedModel
|
22 |
-
from transformers.utils import logging
|
23 |
from transformers.generation.logits_process import LogitsProcessor
|
24 |
-
from transformers.generation.utils import
|
25 |
|
26 |
from .configuration_chatglm import ChatGLMConfig
|
27 |
-
|
28 |
try:
|
29 |
-
from
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
from flash_attn.
|
34 |
-
|
35 |
-
|
36 |
|
37 |
# flags required to enable jit fusion kernels
|
38 |
|
39 |
-
if sys.platform != 'darwin':
|
40 |
torch._C._jit_set_profiling_mode(False)
|
41 |
torch._C._jit_set_profiling_executor(False)
|
42 |
torch._C._jit_override_can_fuse_on_cpu(True)
|
@@ -44,13 +41,9 @@ if sys.platform != 'darwin':
|
|
44 |
|
45 |
logger = logging.get_logger(__name__)
|
46 |
|
47 |
-
_CHECKPOINT_FOR_DOC = "THUDM/
|
48 |
-
_CONFIG_FOR_DOC = "
|
49 |
|
50 |
-
CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
51 |
-
"THUDM/chatglm2-6b",
|
52 |
-
# See all ChatGLM models at https://huggingface.co/models?filter=chatglm
|
53 |
-
]
|
54 |
|
55 |
def default_init(cls, *args, **kwargs):
|
56 |
return cls(*args, **kwargs)
|
@@ -60,22 +53,21 @@ class InvalidScoreLogitsProcessor(LogitsProcessor):
|
|
60 |
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
61 |
if torch.isnan(scores).any() or torch.isinf(scores).any():
|
62 |
scores.zero_()
|
63 |
-
scores[...,
|
64 |
return scores
|
65 |
|
|
|
66 |
def split_tensor_along_last_dim(
|
67 |
tensor: torch.Tensor,
|
68 |
num_partitions: int,
|
69 |
contiguous_split_chunks: bool = False,
|
70 |
) -> List[torch.Tensor]:
|
71 |
"""Split a tensor along its last dimension.
|
72 |
-
|
73 |
Arguments:
|
74 |
tensor: input tensor.
|
75 |
num_partitions: number of partitions to split the tensor
|
76 |
contiguous_split_chunks: If True, make each chunk contiguous
|
77 |
in memory.
|
78 |
-
|
79 |
Returns:
|
80 |
A list of Tensors
|
81 |
"""
|
@@ -104,13 +96,11 @@ class RotaryEmbedding(nn.Module):
|
|
104 |
self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
|
105 |
):
|
106 |
"""Enhanced Transformer with Rotary Position Embedding.
|
107 |
-
|
108 |
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
109 |
transformers/rope/__init__.py. MIT License:
|
110 |
https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
|
111 |
"""
|
112 |
# $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
|
113 |
-
|
114 |
base = base * self.rope_ratio
|
115 |
theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem))
|
116 |
|
@@ -135,14 +125,14 @@ class RotaryEmbedding(nn.Module):
|
|
135 |
|
136 |
@torch.jit.script
|
137 |
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
|
138 |
-
# x: [
|
139 |
-
|
140 |
rot_dim = rope_cache.shape[-2] * 2
|
141 |
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
142 |
# truncate to support variable sizes
|
143 |
-
rope_cache = rope_cache[:sq]
|
144 |
-
xshaped = x.reshape(
|
145 |
-
rope_cache = rope_cache.view(
|
146 |
x_out2 = torch.stack(
|
147 |
[
|
148 |
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
|
@@ -171,12 +161,13 @@ class RMSNorm(torch.nn.Module):
|
|
171 |
class CoreAttention(torch.nn.Module):
|
172 |
def __init__(self, config: ChatGLMConfig, layer_number):
|
173 |
super(CoreAttention, self).__init__()
|
174 |
-
|
175 |
self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
|
176 |
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
177 |
if self.apply_query_key_layer_scaling:
|
178 |
self.attention_softmax_in_fp32 = True
|
179 |
self.layer_number = max(1, layer_number)
|
|
|
180 |
|
181 |
projection_size = config.kv_channels * config.num_attention_heads
|
182 |
|
@@ -185,43 +176,213 @@ class CoreAttention(torch.nn.Module):
|
|
185 |
self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
|
186 |
self.num_attention_heads_per_partition = config.num_attention_heads
|
187 |
|
|
|
188 |
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
189 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
190 |
|
191 |
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
#
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
self.attention_dropout,
|
214 |
-
softmax_scale=1.0 / self.norm_factor, causal=is_causal
|
215 |
)
|
216 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
218 |
context_layer = context_layer.reshape(*new_context_layer_shape)
|
219 |
return context_layer
|
220 |
|
221 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
class SelfAttention(torch.nn.Module):
|
223 |
"""Parallel self-attention layer abstract class.
|
224 |
-
|
225 |
Self-attention layer takes input with size [s, b, h]
|
226 |
and returns output of the same size.
|
227 |
"""
|
@@ -248,7 +409,7 @@ class SelfAttention(torch.nn.Module):
|
|
248 |
device=device, **_config_to_kwargs(config)
|
249 |
)
|
250 |
|
251 |
-
self.core_attention =
|
252 |
|
253 |
# Output.
|
254 |
self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
|
@@ -272,7 +433,7 @@ class SelfAttention(torch.nn.Module):
|
|
272 |
def forward(
|
273 |
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
|
274 |
):
|
275 |
-
# hidden_states: [
|
276 |
|
277 |
# =================================================
|
278 |
# Pre-allocate memory for key-values for inference.
|
@@ -281,7 +442,7 @@ class SelfAttention(torch.nn.Module):
|
|
281 |
# Query, Key, and Value
|
282 |
# =====================
|
283 |
|
284 |
-
# Attention heads [
|
285 |
mixed_x_layer = self.query_key_value(hidden_states)
|
286 |
|
287 |
if self.multi_query_attention:
|
@@ -309,39 +470,45 @@ class SelfAttention(torch.nn.Module):
|
|
309 |
3 * self.hidden_size_per_attention_head)
|
310 |
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
|
311 |
|
312 |
-
# [
|
313 |
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
|
314 |
|
|
|
|
|
|
|
315 |
# apply relative positional encoding (rotary embedding)
|
316 |
if rotary_pos_emb is not None:
|
317 |
query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
|
318 |
key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
|
319 |
|
320 |
# adjust key and value for inference
|
|
|
|
|
|
|
|
|
321 |
if use_cache:
|
322 |
-
if kv_cache is
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
else:
|
328 |
kv_cache = None
|
329 |
-
|
330 |
-
|
331 |
if self.multi_query_attention:
|
332 |
-
key_layer = key_layer.unsqueeze(
|
333 |
key_layer = key_layer.expand(
|
334 |
-
-1, -1,
|
335 |
)
|
336 |
key_layer = key_layer.contiguous().view(
|
337 |
-
key_layer.size()[:
|
338 |
)
|
339 |
-
value_layer = value_layer.unsqueeze(
|
340 |
value_layer = value_layer.expand(
|
341 |
-
-1, -1,
|
342 |
)
|
343 |
value_layer = value_layer.contiguous().view(
|
344 |
-
value_layer.size()[:
|
345 |
)
|
346 |
|
347 |
# ==================================
|
@@ -368,7 +535,6 @@ def _config_to_kwargs(args):
|
|
368 |
|
369 |
class MLP(torch.nn.Module):
|
370 |
"""MLP.
|
371 |
-
|
372 |
MLP will take the input with h hidden state, project it to 4*h
|
373 |
hidden dimension, perform nonlinear transformation, and project the
|
374 |
state back into h hidden dimension.
|
@@ -414,7 +580,6 @@ class MLP(torch.nn.Module):
|
|
414 |
|
415 |
class GLMBlock(torch.nn.Module):
|
416 |
"""A single transformer layer.
|
417 |
-
|
418 |
Transformer layer takes input with size [s, b, h] and returns an
|
419 |
output of the same size.
|
420 |
"""
|
@@ -525,9 +690,9 @@ class GLMTransformer(torch.nn.Module):
|
|
525 |
presents = () if use_cache else None
|
526 |
if self.gradient_checkpointing and self.training:
|
527 |
if use_cache:
|
528 |
-
|
529 |
-
|
530 |
-
|
531 |
use_cache = False
|
532 |
|
533 |
all_self_attentions = None
|
@@ -557,7 +722,15 @@ class GLMTransformer(torch.nn.Module):
|
|
557 |
)
|
558 |
hidden_states, kv_cache = layer_ret
|
559 |
if use_cache:
|
560 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
561 |
|
562 |
if output_hidden_states:
|
563 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
@@ -580,18 +753,24 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
|
|
580 |
config_class = ChatGLMConfig
|
581 |
base_model_prefix = "transformer"
|
582 |
_no_split_modules = ["GLMBlock"]
|
|
|
|
|
583 |
|
584 |
def _init_weights(self, module: nn.Module):
|
585 |
"""Initialize the weights."""
|
586 |
return
|
587 |
|
588 |
def get_masks(self, input_ids, past_key_values, padding_mask=None):
|
|
|
|
|
|
|
|
|
589 |
batch_size, seq_length = input_ids.shape
|
590 |
full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
|
591 |
full_attention_mask.tril_()
|
592 |
past_length = 0
|
593 |
if past_key_values:
|
594 |
-
past_length = past_key_values[0][0].shape[
|
595 |
if past_length:
|
596 |
full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
|
597 |
device=input_ids.device), full_attention_mask), dim=-1)
|
@@ -608,11 +787,6 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
|
|
608 |
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
609 |
return position_ids
|
610 |
|
611 |
-
def _set_gradient_checkpointing(self, module, value=False):
|
612 |
-
if isinstance(module, GLMTransformer):
|
613 |
-
module.gradient_checkpointing = value
|
614 |
-
|
615 |
-
|
616 |
class Embedding(torch.nn.Module):
|
617 |
"""Language model embeddings."""
|
618 |
|
@@ -633,8 +807,6 @@ class Embedding(torch.nn.Module):
|
|
633 |
# Embeddings.
|
634 |
words_embeddings = self.word_embeddings(input_ids)
|
635 |
embeddings = words_embeddings
|
636 |
-
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
|
637 |
-
embeddings = embeddings.transpose(0, 1).contiguous()
|
638 |
# If the input flag for fp32 residual connection is set, convert for float.
|
639 |
if self.fp32_residual_connection:
|
640 |
embeddings = embeddings.float()
|
@@ -652,6 +824,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
652 |
if device is not None:
|
653 |
init_kwargs["device"] = device
|
654 |
self.embedding = init_method(Embedding, config, **init_kwargs)
|
|
|
|
|
|
|
655 |
|
656 |
# Rotary positional embeddings
|
657 |
self.seq_length = config.seq_length
|
@@ -659,7 +834,8 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
659 |
config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
|
660 |
)
|
661 |
|
662 |
-
self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, rope_ratio=config.rope_ratio,
|
|
|
663 |
device=device, dtype=config.torch_dtype)
|
664 |
self.encoder = init_method(GLMTransformer, config, **init_kwargs)
|
665 |
self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
|
@@ -668,6 +844,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
668 |
def get_input_embeddings(self):
|
669 |
return self.embedding.word_embeddings
|
670 |
|
|
|
|
|
|
|
671 |
def forward(
|
672 |
self,
|
673 |
input_ids,
|
@@ -677,6 +856,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
677 |
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
678 |
inputs_embeds: Optional[torch.Tensor] = None,
|
679 |
use_cache: Optional[bool] = None,
|
|
|
680 |
output_hidden_states: Optional[bool] = None,
|
681 |
return_dict: Optional[bool] = None,
|
682 |
):
|
@@ -691,9 +871,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
691 |
if inputs_embeds is None:
|
692 |
inputs_embeds = self.embedding(input_ids)
|
693 |
|
694 |
-
|
695 |
-
|
696 |
-
|
697 |
|
698 |
# Rotary positional embeddings
|
699 |
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
|
@@ -701,13 +881,18 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
701 |
rotary_pos_emb = rotary_pos_emb[position_ids]
|
702 |
else:
|
703 |
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
|
704 |
-
rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
|
705 |
|
706 |
# Run encoder.
|
707 |
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
|
708 |
-
inputs_embeds,
|
709 |
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
|
710 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
711 |
|
712 |
if not return_dict:
|
713 |
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
|
@@ -727,7 +912,6 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
727 |
self.max_sequence_length = config.max_length
|
728 |
self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
|
729 |
self.config = config
|
730 |
-
self.pack_loss = False
|
731 |
|
732 |
def _update_model_kwargs_for_generation(
|
733 |
self,
|
@@ -764,6 +948,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
764 |
past_key_values: Optional[torch.Tensor] = None,
|
765 |
attention_mask: Optional[torch.Tensor] = None,
|
766 |
position_ids: Optional[torch.Tensor] = None,
|
|
|
767 |
is_first_forward: bool = True,
|
768 |
**kwargs
|
769 |
) -> dict:
|
@@ -771,14 +956,16 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
771 |
if position_ids is None:
|
772 |
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
773 |
if not is_first_forward:
|
774 |
-
|
775 |
-
|
|
|
776 |
return {
|
777 |
"input_ids": input_ids,
|
778 |
"past_key_values": past_key_values,
|
779 |
"position_ids": position_ids,
|
780 |
"attention_mask": attention_mask,
|
781 |
-
"return_last_logit": True
|
|
|
782 |
}
|
783 |
|
784 |
def forward(
|
@@ -788,7 +975,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
788 |
attention_mask: Optional[torch.Tensor] = None,
|
789 |
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
790 |
inputs_embeds: Optional[torch.Tensor] = None,
|
791 |
-
labels: Optional[
|
792 |
use_cache: Optional[bool] = None,
|
793 |
output_attentions: Optional[bool] = None,
|
794 |
output_hidden_states: Optional[bool] = None,
|
@@ -811,30 +998,19 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
811 |
|
812 |
hidden_states = transformer_outputs[0]
|
813 |
if return_last_logit:
|
814 |
-
hidden_states = hidden_states[-1:]
|
815 |
lm_logits = self.transformer.output_layer(hidden_states)
|
816 |
-
lm_logits = lm_logits.transpose(0, 1).contiguous()
|
817 |
|
818 |
loss = None
|
819 |
if labels is not None:
|
820 |
lm_logits = lm_logits.to(torch.float32)
|
|
|
821 |
# Shift so that tokens < n predict n
|
822 |
shift_logits = lm_logits[..., :-1, :].contiguous()
|
823 |
-
if isinstance(labels, tuple) or isinstance(labels, list):
|
824 |
-
labels, weights = labels
|
825 |
shift_labels = labels[..., 1:].contiguous()
|
826 |
-
|
827 |
-
|
828 |
-
|
829 |
-
loss *= weights
|
830 |
-
# if self.pack_loss:
|
831 |
-
# shift_weights = weights[..., 1:].contiguous()
|
832 |
-
# loss_fct = CrossEntropyLoss(ignore_index=-100, reduction='none')
|
833 |
-
# loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
834 |
-
# loss = (loss * shift_weights).sum()
|
835 |
-
else:
|
836 |
-
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
837 |
-
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
838 |
|
839 |
lm_logits = lm_logits.to(hidden_states.dtype)
|
840 |
loss = loss.to(hidden_states.dtype)
|
@@ -859,33 +1035,24 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
859 |
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
860 |
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
861 |
beam_idx at every generation step.
|
862 |
-
|
863 |
Output shares the same memory storage as `past`.
|
864 |
"""
|
865 |
return tuple(
|
866 |
(
|
867 |
-
layer_past[0].index_select(
|
868 |
-
layer_past[1].index_select(
|
869 |
)
|
870 |
for layer_past in past
|
871 |
)
|
872 |
|
873 |
-
def process_response(self, response):
|
874 |
-
response = response.strip()
|
875 |
-
response = response.replace("[[训练时间]]", "2023年")
|
876 |
-
return response
|
877 |
-
|
878 |
@torch.inference_mode()
|
879 |
def chat(self, tokenizer, query: str, history: List[Dict] = None, role: str = "user",
|
880 |
-
max_length: int = 8192, num_beams=1, do_sample=True, top_p=0.8, temperature=0.8,
|
881 |
**kwargs):
|
882 |
if history is None:
|
883 |
history = []
|
884 |
-
if logits_processor is None:
|
885 |
-
logits_processor = LogitsProcessorList()
|
886 |
-
logits_processor.append(InvalidScoreLogitsProcessor())
|
887 |
gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
|
888 |
-
"temperature": temperature,
|
889 |
inputs = tokenizer.build_chat_input(query, history=history, role=role)
|
890 |
inputs = inputs.to(self.device)
|
891 |
eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"),
|
@@ -894,5 +1061,4 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
894 |
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1]
|
895 |
response = tokenizer.decode(outputs)
|
896 |
history.append({"role": role, "content": query})
|
897 |
-
response = self.process_response(response)
|
898 |
return response, history
|
|
|
1 |
""" PyTorch ChatGLM model. """
|
2 |
|
3 |
import math
|
|
|
|
|
|
|
4 |
import sys
|
|
|
5 |
import torch
|
6 |
import torch.utils.checkpoint
|
7 |
import torch.nn.functional as F
|
8 |
from torch import nn
|
9 |
+
from torch.nn import CrossEntropyLoss, LayerNorm, MSELoss, BCEWithLogitsLoss
|
10 |
from torch.nn.utils import skip_init
|
11 |
+
from typing import Optional, Tuple, Union, List, Dict, Any
|
12 |
|
13 |
from transformers.modeling_outputs import (
|
14 |
BaseModelOutputWithPast,
|
15 |
CausalLMOutputWithPast,
|
16 |
+
SequenceClassifierOutputWithPast,
|
17 |
)
|
18 |
from transformers.modeling_utils import PreTrainedModel
|
19 |
+
from transformers.utils import logging, is_torch_npu_available
|
20 |
from transformers.generation.logits_process import LogitsProcessor
|
21 |
+
from transformers.generation.utils import ModelOutput
|
22 |
|
23 |
from .configuration_chatglm import ChatGLMConfig
|
24 |
+
|
25 |
try:
|
26 |
+
from transformers.utils import is_flash_attn_greater_or_equal_2_10, is_flash_attn_2_available
|
27 |
+
|
28 |
+
if is_flash_attn_2_available():
|
29 |
+
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
30 |
+
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
31 |
+
except:
|
32 |
+
pass
|
33 |
|
34 |
# flags required to enable jit fusion kernels
|
35 |
|
36 |
+
if sys.platform != 'darwin' and not is_torch_npu_available():
|
37 |
torch._C._jit_set_profiling_mode(False)
|
38 |
torch._C._jit_set_profiling_executor(False)
|
39 |
torch._C._jit_override_can_fuse_on_cpu(True)
|
|
|
41 |
|
42 |
logger = logging.get_logger(__name__)
|
43 |
|
44 |
+
_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM"
|
45 |
+
_CONFIG_FOR_DOC = "ChatGLMConfig"
|
46 |
|
|
|
|
|
|
|
|
|
47 |
|
48 |
def default_init(cls, *args, **kwargs):
|
49 |
return cls(*args, **kwargs)
|
|
|
53 |
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
54 |
if torch.isnan(scores).any() or torch.isinf(scores).any():
|
55 |
scores.zero_()
|
56 |
+
scores[..., 198] = 5e4
|
57 |
return scores
|
58 |
|
59 |
+
|
60 |
def split_tensor_along_last_dim(
|
61 |
tensor: torch.Tensor,
|
62 |
num_partitions: int,
|
63 |
contiguous_split_chunks: bool = False,
|
64 |
) -> List[torch.Tensor]:
|
65 |
"""Split a tensor along its last dimension.
|
|
|
66 |
Arguments:
|
67 |
tensor: input tensor.
|
68 |
num_partitions: number of partitions to split the tensor
|
69 |
contiguous_split_chunks: If True, make each chunk contiguous
|
70 |
in memory.
|
|
|
71 |
Returns:
|
72 |
A list of Tensors
|
73 |
"""
|
|
|
96 |
self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
|
97 |
):
|
98 |
"""Enhanced Transformer with Rotary Position Embedding.
|
|
|
99 |
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
100 |
transformers/rope/__init__.py. MIT License:
|
101 |
https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
|
102 |
"""
|
103 |
# $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
|
|
|
104 |
base = base * self.rope_ratio
|
105 |
theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem))
|
106 |
|
|
|
125 |
|
126 |
@torch.jit.script
|
127 |
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
|
128 |
+
# x: [b, np, sq, hn]
|
129 |
+
b, np, sq, hn = x.size(0), x.size(1), x.size(2), x.size(3)
|
130 |
rot_dim = rope_cache.shape[-2] * 2
|
131 |
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
132 |
# truncate to support variable sizes
|
133 |
+
rope_cache = rope_cache[:, :sq]
|
134 |
+
xshaped = x.reshape(b, np, sq, rot_dim // 2, 2)
|
135 |
+
rope_cache = rope_cache.view(-1, 1, sq, xshaped.size(3), 2)
|
136 |
x_out2 = torch.stack(
|
137 |
[
|
138 |
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
|
|
|
161 |
class CoreAttention(torch.nn.Module):
|
162 |
def __init__(self, config: ChatGLMConfig, layer_number):
|
163 |
super(CoreAttention, self).__init__()
|
164 |
+
self.config = config
|
165 |
self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
|
166 |
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
167 |
if self.apply_query_key_layer_scaling:
|
168 |
self.attention_softmax_in_fp32 = True
|
169 |
self.layer_number = max(1, layer_number)
|
170 |
+
self.is_causal = True
|
171 |
|
172 |
projection_size = config.kv_channels * config.num_attention_heads
|
173 |
|
|
|
176 |
self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
|
177 |
self.num_attention_heads_per_partition = config.num_attention_heads
|
178 |
|
179 |
+
coeff = None
|
180 |
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
181 |
+
if self.apply_query_key_layer_scaling:
|
182 |
+
coeff = self.layer_number
|
183 |
+
self.norm_factor *= coeff
|
184 |
+
self.coeff = coeff
|
185 |
+
|
186 |
+
self.attention_dropout = torch.nn.Dropout(config.attention_dropout)
|
187 |
|
188 |
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
189 |
+
# [b, np, sq, sk]
|
190 |
+
output_size = (query_layer.size(0), query_layer.size(1), query_layer.size(2), key_layer.size(2))
|
191 |
+
|
192 |
+
# [b, np, sq, hn] -> [b * np, sq, hn]
|
193 |
+
query_layer = query_layer.view(output_size[0] * output_size[1], output_size[2], -1)
|
194 |
+
# [b, np, sk, hn] -> [b * np, sk, hn]
|
195 |
+
key_layer = key_layer.view(output_size[0] * output_size[1], output_size[3], -1)
|
196 |
+
|
197 |
+
# preallocting input tensor: [b * np, sq, sk]
|
198 |
+
matmul_input_buffer = torch.empty(
|
199 |
+
output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
|
200 |
+
device=query_layer.device
|
201 |
+
)
|
202 |
+
|
203 |
+
# Raw attention scores. [b * np, sq, sk]
|
204 |
+
matmul_result = torch.baddbmm(
|
205 |
+
matmul_input_buffer,
|
206 |
+
query_layer, # [b * np, sq, hn]
|
207 |
+
key_layer.transpose(1, 2), # [b * np, hn, sk]
|
208 |
+
beta=0.0,
|
209 |
+
alpha=(1.0 / self.norm_factor),
|
|
|
|
|
210 |
)
|
211 |
+
|
212 |
+
# change view to [b, np, sq, sk]
|
213 |
+
attention_scores = matmul_result.view(*output_size)
|
214 |
+
|
215 |
+
# ===========================
|
216 |
+
# Attention probs and dropout
|
217 |
+
# ===========================
|
218 |
+
|
219 |
+
# attention scores and attention mask [b, np, sq, sk]
|
220 |
+
if self.attention_softmax_in_fp32:
|
221 |
+
attention_scores = attention_scores.float()
|
222 |
+
if self.coeff is not None:
|
223 |
+
attention_scores = attention_scores * self.coeff
|
224 |
+
if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
|
225 |
+
attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
|
226 |
+
device=attention_scores.device, dtype=torch.bool)
|
227 |
+
attention_mask.tril_()
|
228 |
+
attention_mask = ~attention_mask
|
229 |
+
if attention_mask is not None:
|
230 |
+
attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
|
231 |
+
attention_probs = F.softmax(attention_scores, dim=-1)
|
232 |
+
attention_probs = attention_probs.type_as(value_layer)
|
233 |
+
|
234 |
+
# This is actually dropping out entire tokens to attend to, which might
|
235 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
236 |
+
attention_probs = self.attention_dropout(attention_probs)
|
237 |
+
|
238 |
+
# query layer shape: [b * np, sq, hn]
|
239 |
+
# value layer shape: [b, np, sk, hn]
|
240 |
+
# attention shape: [b, np, sq, sk]
|
241 |
+
# context layer shape: [b, np, sq, hn]
|
242 |
+
output_size = (value_layer.size(0), value_layer.size(1), query_layer.size(1), value_layer.size(3))
|
243 |
+
# change view [b * np, sk, hn]
|
244 |
+
value_layer = value_layer.view(output_size[0] * output_size[1], value_layer.size(2), -1)
|
245 |
+
# change view [b * np, sq, sk]
|
246 |
+
attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
|
247 |
+
# matmul: [b * np, sq, hn]
|
248 |
+
context_layer = torch.bmm(attention_probs, value_layer)
|
249 |
+
# change view [b, np, sq, hn]
|
250 |
+
context_layer = context_layer.view(*output_size)
|
251 |
+
# [b, np, sq, hn] --> [b, sq, np, hn]
|
252 |
+
context_layer = context_layer.transpose(1, 2).contiguous()
|
253 |
+
# [b, sq, np, hn] --> [b, sq, hp]
|
254 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
255 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
256 |
+
|
257 |
+
return context_layer
|
258 |
+
|
259 |
+
|
260 |
+
class SdpaAttention(CoreAttention):
|
261 |
+
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
262 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
263 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
264 |
+
is_causal=True,
|
265 |
+
dropout_p=self.config.attention_dropout if self.training else 0.0)
|
266 |
+
else:
|
267 |
+
if attention_mask is not None:
|
268 |
+
attention_mask = ~attention_mask
|
269 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
270 |
+
attention_mask,
|
271 |
+
dropout_p=self.config.attention_dropout if self.training else 0.0)
|
272 |
+
context_layer = context_layer.transpose(1, 2).contiguous()
|
273 |
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
274 |
context_layer = context_layer.reshape(*new_context_layer_shape)
|
275 |
return context_layer
|
276 |
|
277 |
|
278 |
+
def _get_unpad_data(attention_mask):
|
279 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
280 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
281 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
282 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
|
283 |
+
return (
|
284 |
+
indices,
|
285 |
+
cu_seqlens,
|
286 |
+
max_seqlen_in_batch,
|
287 |
+
)
|
288 |
+
|
289 |
+
|
290 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2
|
291 |
+
class FlashAttention2(CoreAttention):
|
292 |
+
def __init__(self, *args, **kwargs):
|
293 |
+
super().__init__(*args, **kwargs)
|
294 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
295 |
+
|
296 |
+
def forward(self, query_states, key_states, value_states, attention_mask):
|
297 |
+
query_states = query_states.transpose(1, 2)
|
298 |
+
key_states = key_states.transpose(1, 2)
|
299 |
+
value_states = value_states.transpose(1, 2)
|
300 |
+
batch_size, query_length = query_states.shape[:2]
|
301 |
+
if not self._flash_attn_uses_top_left_mask:
|
302 |
+
causal = self.is_causal
|
303 |
+
else:
|
304 |
+
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
|
305 |
+
causal = self.is_causal and query_length != 1
|
306 |
+
dropout = self.config.attention_dropout if self.training else 0.0
|
307 |
+
# Contains at least one padding token in the sequence
|
308 |
+
if attention_mask is not None:
|
309 |
+
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
310 |
+
query_states, key_states, value_states, attention_mask, query_length
|
311 |
+
)
|
312 |
+
|
313 |
+
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
314 |
+
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
315 |
+
|
316 |
+
attn_output_unpad = flash_attn_varlen_func(
|
317 |
+
query_states,
|
318 |
+
key_states,
|
319 |
+
value_states,
|
320 |
+
cu_seqlens_q=cu_seqlens_q,
|
321 |
+
cu_seqlens_k=cu_seqlens_k,
|
322 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
323 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
324 |
+
dropout_p=dropout,
|
325 |
+
softmax_scale=None,
|
326 |
+
causal=causal,
|
327 |
+
)
|
328 |
+
|
329 |
+
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
330 |
+
else:
|
331 |
+
attn_output = flash_attn_func(
|
332 |
+
query_states, key_states, value_states, dropout, softmax_scale=None, causal=causal
|
333 |
+
)
|
334 |
+
attn_output = attn_output.reshape(batch_size, query_length, self.hidden_size_per_partition).contiguous()
|
335 |
+
return attn_output
|
336 |
+
|
337 |
+
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
338 |
+
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
339 |
+
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
340 |
+
|
341 |
+
key_layer = index_first_axis(
|
342 |
+
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
343 |
+
)
|
344 |
+
value_layer = index_first_axis(
|
345 |
+
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
346 |
+
)
|
347 |
+
if query_length == kv_seq_len:
|
348 |
+
query_layer = index_first_axis(
|
349 |
+
query_layer.reshape(batch_size * kv_seq_len, self.num_attention_heads_per_partition, head_dim),
|
350 |
+
indices_k
|
351 |
+
)
|
352 |
+
cu_seqlens_q = cu_seqlens_k
|
353 |
+
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
354 |
+
indices_q = indices_k
|
355 |
+
elif query_length == 1:
|
356 |
+
max_seqlen_in_batch_q = 1
|
357 |
+
cu_seqlens_q = torch.arange(
|
358 |
+
batch_size + 1, dtype=torch.int32, device=query_layer.device
|
359 |
+
) # There is a memcpy here, that is very bad.
|
360 |
+
indices_q = cu_seqlens_q[:-1]
|
361 |
+
query_layer = query_layer.squeeze(1)
|
362 |
+
else:
|
363 |
+
# The -q_len: slice assumes left padding.
|
364 |
+
attention_mask = attention_mask[:, -query_length:]
|
365 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
|
366 |
+
|
367 |
+
return (
|
368 |
+
query_layer,
|
369 |
+
key_layer,
|
370 |
+
value_layer,
|
371 |
+
indices_q,
|
372 |
+
(cu_seqlens_q, cu_seqlens_k),
|
373 |
+
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
374 |
+
)
|
375 |
+
|
376 |
+
|
377 |
+
CORE_ATTENTION_CLASSES = {
|
378 |
+
"eager": CoreAttention,
|
379 |
+
"sdpa": SdpaAttention,
|
380 |
+
"flash_attention_2": FlashAttention2
|
381 |
+
}
|
382 |
+
|
383 |
+
|
384 |
class SelfAttention(torch.nn.Module):
|
385 |
"""Parallel self-attention layer abstract class.
|
|
|
386 |
Self-attention layer takes input with size [s, b, h]
|
387 |
and returns output of the same size.
|
388 |
"""
|
|
|
409 |
device=device, **_config_to_kwargs(config)
|
410 |
)
|
411 |
|
412 |
+
self.core_attention = CORE_ATTENTION_CLASSES[config._attn_implementation](config, self.layer_number)
|
413 |
|
414 |
# Output.
|
415 |
self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
|
|
|
433 |
def forward(
|
434 |
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
|
435 |
):
|
436 |
+
# hidden_states: [b, sq, h]
|
437 |
|
438 |
# =================================================
|
439 |
# Pre-allocate memory for key-values for inference.
|
|
|
442 |
# Query, Key, and Value
|
443 |
# =====================
|
444 |
|
445 |
+
# Attention heads [b, sq, h] --> [b, sq, (np * 3 * hn)]
|
446 |
mixed_x_layer = self.query_key_value(hidden_states)
|
447 |
|
448 |
if self.multi_query_attention:
|
|
|
470 |
3 * self.hidden_size_per_attention_head)
|
471 |
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
|
472 |
|
473 |
+
# [b, sq, np, 3 * hn] --> 3 [b, sq, np, hn]
|
474 |
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
|
475 |
|
476 |
+
# [b, sq, np, hn] -> [b, np, sq, hn]
|
477 |
+
query_layer, key_layer, value_layer = [k.transpose(1, 2) for k in [query_layer, key_layer, value_layer]]
|
478 |
+
|
479 |
# apply relative positional encoding (rotary embedding)
|
480 |
if rotary_pos_emb is not None:
|
481 |
query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
|
482 |
key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
|
483 |
|
484 |
# adjust key and value for inference
|
485 |
+
if kv_cache is not None:
|
486 |
+
cache_k, cache_v = kv_cache
|
487 |
+
key_layer = torch.cat((cache_k, key_layer), dim=2)
|
488 |
+
value_layer = torch.cat((cache_v, value_layer), dim=2)
|
489 |
if use_cache:
|
490 |
+
if kv_cache is None:
|
491 |
+
kv_cache = torch.cat((key_layer.unsqueeze(0).unsqueeze(0), value_layer.unsqueeze(0).unsqueeze(0)),
|
492 |
+
dim=1)
|
493 |
+
else:
|
494 |
+
kv_cache = (key_layer, value_layer)
|
495 |
else:
|
496 |
kv_cache = None
|
497 |
+
|
|
|
498 |
if self.multi_query_attention:
|
499 |
+
key_layer = key_layer.unsqueeze(2)
|
500 |
key_layer = key_layer.expand(
|
501 |
+
-1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1, -1
|
502 |
)
|
503 |
key_layer = key_layer.contiguous().view(
|
504 |
+
key_layer.size()[:1] + (self.num_attention_heads_per_partition,) + key_layer.size()[3:]
|
505 |
)
|
506 |
+
value_layer = value_layer.unsqueeze(2)
|
507 |
value_layer = value_layer.expand(
|
508 |
+
-1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1, -1
|
509 |
)
|
510 |
value_layer = value_layer.contiguous().view(
|
511 |
+
value_layer.size()[:1] + (self.num_attention_heads_per_partition,) + value_layer.size()[3:]
|
512 |
)
|
513 |
|
514 |
# ==================================
|
|
|
535 |
|
536 |
class MLP(torch.nn.Module):
|
537 |
"""MLP.
|
|
|
538 |
MLP will take the input with h hidden state, project it to 4*h
|
539 |
hidden dimension, perform nonlinear transformation, and project the
|
540 |
state back into h hidden dimension.
|
|
|
580 |
|
581 |
class GLMBlock(torch.nn.Module):
|
582 |
"""A single transformer layer.
|
|
|
583 |
Transformer layer takes input with size [s, b, h] and returns an
|
584 |
output of the same size.
|
585 |
"""
|
|
|
690 |
presents = () if use_cache else None
|
691 |
if self.gradient_checkpointing and self.training:
|
692 |
if use_cache:
|
693 |
+
logger.warning_once(
|
694 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
695 |
+
)
|
696 |
use_cache = False
|
697 |
|
698 |
all_self_attentions = None
|
|
|
722 |
)
|
723 |
hidden_states, kv_cache = layer_ret
|
724 |
if use_cache:
|
725 |
+
# token by token decoding, use tuple format
|
726 |
+
if kv_caches[0] is not None:
|
727 |
+
presents = presents + (kv_cache,)
|
728 |
+
# prefilling in decoding, use tensor format to save cuda memory
|
729 |
+
else:
|
730 |
+
if len(presents) == 0:
|
731 |
+
presents = kv_cache
|
732 |
+
else:
|
733 |
+
presents = torch.cat((presents, kv_cache.to(presents.device)), dim=0)
|
734 |
|
735 |
if output_hidden_states:
|
736 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
753 |
config_class = ChatGLMConfig
|
754 |
base_model_prefix = "transformer"
|
755 |
_no_split_modules = ["GLMBlock"]
|
756 |
+
_supports_flash_attn_2 = True
|
757 |
+
_supports_sdpa = True
|
758 |
|
759 |
def _init_weights(self, module: nn.Module):
|
760 |
"""Initialize the weights."""
|
761 |
return
|
762 |
|
763 |
def get_masks(self, input_ids, past_key_values, padding_mask=None):
|
764 |
+
if self.config._attn_implementation == "flash_attention_2":
|
765 |
+
if padding_mask is not None and not padding_mask.all():
|
766 |
+
return padding_mask
|
767 |
+
return None
|
768 |
batch_size, seq_length = input_ids.shape
|
769 |
full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
|
770 |
full_attention_mask.tril_()
|
771 |
past_length = 0
|
772 |
if past_key_values:
|
773 |
+
past_length = past_key_values[0][0].shape[2]
|
774 |
if past_length:
|
775 |
full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
|
776 |
device=input_ids.device), full_attention_mask), dim=-1)
|
|
|
787 |
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
788 |
return position_ids
|
789 |
|
|
|
|
|
|
|
|
|
|
|
790 |
class Embedding(torch.nn.Module):
|
791 |
"""Language model embeddings."""
|
792 |
|
|
|
807 |
# Embeddings.
|
808 |
words_embeddings = self.word_embeddings(input_ids)
|
809 |
embeddings = words_embeddings
|
|
|
|
|
810 |
# If the input flag for fp32 residual connection is set, convert for float.
|
811 |
if self.fp32_residual_connection:
|
812 |
embeddings = embeddings.float()
|
|
|
824 |
if device is not None:
|
825 |
init_kwargs["device"] = device
|
826 |
self.embedding = init_method(Embedding, config, **init_kwargs)
|
827 |
+
self.num_layers = config.num_layers
|
828 |
+
self.multi_query_group_num = config.multi_query_group_num
|
829 |
+
self.kv_channels = config.kv_channels
|
830 |
|
831 |
# Rotary positional embeddings
|
832 |
self.seq_length = config.seq_length
|
|
|
834 |
config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
|
835 |
)
|
836 |
|
837 |
+
self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, rope_ratio=config.rope_ratio,
|
838 |
+
original_impl=config.original_rope,
|
839 |
device=device, dtype=config.torch_dtype)
|
840 |
self.encoder = init_method(GLMTransformer, config, **init_kwargs)
|
841 |
self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
|
|
|
844 |
def get_input_embeddings(self):
|
845 |
return self.embedding.word_embeddings
|
846 |
|
847 |
+
def set_input_embeddings(self, value):
|
848 |
+
self.embedding.word_embeddings = value
|
849 |
+
|
850 |
def forward(
|
851 |
self,
|
852 |
input_ids,
|
|
|
856 |
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
857 |
inputs_embeds: Optional[torch.Tensor] = None,
|
858 |
use_cache: Optional[bool] = None,
|
859 |
+
output_attentions: Optional[bool] = None,
|
860 |
output_hidden_states: Optional[bool] = None,
|
861 |
return_dict: Optional[bool] = None,
|
862 |
):
|
|
|
871 |
if inputs_embeds is None:
|
872 |
inputs_embeds = self.embedding(input_ids)
|
873 |
|
874 |
+
if full_attention_mask is None:
|
875 |
+
if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
|
876 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
|
877 |
|
878 |
# Rotary positional embeddings
|
879 |
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
|
|
|
881 |
rotary_pos_emb = rotary_pos_emb[position_ids]
|
882 |
else:
|
883 |
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
|
|
|
884 |
|
885 |
# Run encoder.
|
886 |
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
|
887 |
+
inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
|
888 |
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
|
889 |
)
|
890 |
+
if presents is not None and type(presents) is torch.Tensor:
|
891 |
+
presents = presents.split(1, dim=0)
|
892 |
+
presents = list(presents)
|
893 |
+
presents = [list(x.squeeze(0).split(1, dim=0)) for x in presents]
|
894 |
+
presents = [tuple([x.squeeze(0) for x in y]) for y in presents]
|
895 |
+
presents = tuple(presents)
|
896 |
|
897 |
if not return_dict:
|
898 |
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
|
|
|
912 |
self.max_sequence_length = config.max_length
|
913 |
self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
|
914 |
self.config = config
|
|
|
915 |
|
916 |
def _update_model_kwargs_for_generation(
|
917 |
self,
|
|
|
948 |
past_key_values: Optional[torch.Tensor] = None,
|
949 |
attention_mask: Optional[torch.Tensor] = None,
|
950 |
position_ids: Optional[torch.Tensor] = None,
|
951 |
+
use_cache: Optional[bool] = None,
|
952 |
is_first_forward: bool = True,
|
953 |
**kwargs
|
954 |
) -> dict:
|
|
|
956 |
if position_ids is None:
|
957 |
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
958 |
if not is_first_forward:
|
959 |
+
if past_key_values is not None:
|
960 |
+
position_ids = position_ids[..., -1:]
|
961 |
+
input_ids = input_ids[:, -1:]
|
962 |
return {
|
963 |
"input_ids": input_ids,
|
964 |
"past_key_values": past_key_values,
|
965 |
"position_ids": position_ids,
|
966 |
"attention_mask": attention_mask,
|
967 |
+
"return_last_logit": True,
|
968 |
+
"use_cache": use_cache
|
969 |
}
|
970 |
|
971 |
def forward(
|
|
|
975 |
attention_mask: Optional[torch.Tensor] = None,
|
976 |
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
977 |
inputs_embeds: Optional[torch.Tensor] = None,
|
978 |
+
labels: Optional[torch.Tensor] = None,
|
979 |
use_cache: Optional[bool] = None,
|
980 |
output_attentions: Optional[bool] = None,
|
981 |
output_hidden_states: Optional[bool] = None,
|
|
|
998 |
|
999 |
hidden_states = transformer_outputs[0]
|
1000 |
if return_last_logit:
|
1001 |
+
hidden_states = hidden_states[:, -1:]
|
1002 |
lm_logits = self.transformer.output_layer(hidden_states)
|
|
|
1003 |
|
1004 |
loss = None
|
1005 |
if labels is not None:
|
1006 |
lm_logits = lm_logits.to(torch.float32)
|
1007 |
+
|
1008 |
# Shift so that tokens < n predict n
|
1009 |
shift_logits = lm_logits[..., :-1, :].contiguous()
|
|
|
|
|
1010 |
shift_labels = labels[..., 1:].contiguous()
|
1011 |
+
# Flatten the tokens
|
1012 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
1013 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1014 |
|
1015 |
lm_logits = lm_logits.to(hidden_states.dtype)
|
1016 |
loss = loss.to(hidden_states.dtype)
|
|
|
1035 |
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
1036 |
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
1037 |
beam_idx at every generation step.
|
|
|
1038 |
Output shares the same memory storage as `past`.
|
1039 |
"""
|
1040 |
return tuple(
|
1041 |
(
|
1042 |
+
layer_past[0].index_select(0, beam_idx.to(layer_past[0].device)),
|
1043 |
+
layer_past[1].index_select(0, beam_idx.to(layer_past[1].device)),
|
1044 |
)
|
1045 |
for layer_past in past
|
1046 |
)
|
1047 |
|
|
|
|
|
|
|
|
|
|
|
1048 |
@torch.inference_mode()
|
1049 |
def chat(self, tokenizer, query: str, history: List[Dict] = None, role: str = "user",
|
1050 |
+
max_length: int = 8192, num_beams=1, do_sample=True, top_p=0.8, temperature=0.8,
|
1051 |
**kwargs):
|
1052 |
if history is None:
|
1053 |
history = []
|
|
|
|
|
|
|
1054 |
gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
|
1055 |
+
"temperature": temperature, **kwargs}
|
1056 |
inputs = tokenizer.build_chat_input(query, history=history, role=role)
|
1057 |
inputs = inputs.to(self.device)
|
1058 |
eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"),
|
|
|
1061 |
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1]
|
1062 |
response = tokenizer.decode(outputs)
|
1063 |
history.append({"role": role, "content": query})
|
|
|
1064 |
return response, history
|