Upload model
Browse files- config.json +59 -0
- generation_config.json +12 -0
- model.safetensors +3 -0
- modelling_walsh.py +949 -0
config.json
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "/home/dinalt/ai_assets/models/walsh",
|
3 |
+
"activation_args": {},
|
4 |
+
"activation_cls": "torch.nn.GELU",
|
5 |
+
"architectures": [
|
6 |
+
"HFCausalModel"
|
7 |
+
],
|
8 |
+
"attention_args": {
|
9 |
+
"beta": 0.25,
|
10 |
+
"dropout": 0.1
|
11 |
+
},
|
12 |
+
"attention_cls": ".CausalSelfAttention",
|
13 |
+
"auto_map": {
|
14 |
+
"AutoConfig": "modelling_walsh.Config",
|
15 |
+
"AutoModelForCausalLM": "modelling_walsh.HFCausalModel"
|
16 |
+
},
|
17 |
+
"d_embed": 2048,
|
18 |
+
"dim_feedforward": 8192,
|
19 |
+
"dropout": 0.1,
|
20 |
+
"embdding_cls": "torch.nn.Embedding",
|
21 |
+
"embedding_args": {},
|
22 |
+
"feedforward_args": {
|
23 |
+
"beta": 0.25,
|
24 |
+
"bias": true
|
25 |
+
},
|
26 |
+
"feedforward_cls": ".FeedforwardLayer",
|
27 |
+
"head_args": {},
|
28 |
+
"head_cls": ".Transformer",
|
29 |
+
"init_gain": 1.0,
|
30 |
+
"layer_args": {
|
31 |
+
"alpha": 2.828427124746
|
32 |
+
},
|
33 |
+
"layer_cls": ".DeepnetLayer",
|
34 |
+
"layer_stack_args": {},
|
35 |
+
"layer_stack_cls": ".TransformerLayerStack",
|
36 |
+
"loss_function": ".causal_loss",
|
37 |
+
"max_sequence_length": 16384,
|
38 |
+
"model_type": "walsh-causal-v1",
|
39 |
+
"norm_args": {
|
40 |
+
"normalized_shape": 2084
|
41 |
+
},
|
42 |
+
"norm_cls": "torch.nn.LayerNorm",
|
43 |
+
"num_attention_heads": 32,
|
44 |
+
"num_hidden_layers": 32,
|
45 |
+
"output_proj_args": {},
|
46 |
+
"output_proj_cls": "torch.nn.Linear",
|
47 |
+
"pad_index": null,
|
48 |
+
"positional_encoder_args": {
|
49 |
+
"d_embed": 2048,
|
50 |
+
"gain": 0.3333,
|
51 |
+
"max_seq": 16384
|
52 |
+
},
|
53 |
+
"positional_encoder_cls": ".RSWalshPositionalEncoder",
|
54 |
+
"torch_dtype": "bfloat16",
|
55 |
+
"transformer_args": {},
|
56 |
+
"transformer_cls": ".Transformer",
|
57 |
+
"transformers_version": "4.37.2",
|
58 |
+
"vocab_size": 32000
|
59 |
+
}
|
generation_config.json
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"do_sample": true,
|
3 |
+
"eos_token_id": 3,
|
4 |
+
"max_new_tokens": 512,
|
5 |
+
"pad_token_id": 0,
|
6 |
+
"repetition_penalty": 1.01,
|
7 |
+
"temperature": 0.87,
|
8 |
+
"top_k": 85,
|
9 |
+
"top_p": 0.99,
|
10 |
+
"transformers_version": "4.37.2",
|
11 |
+
"typical_p": 0.68
|
12 |
+
}
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ae5d14280ec61dc8be1912d8f99655dd403f957eac816199b536a716fc09d928
|
3 |
+
size 3485189432
|
modelling_walsh.py
ADDED
@@ -0,0 +1,949 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# See: https://huggingface.co/docs/transformers/custom_models
|
2 |
+
from typing import Optional, Tuple, Union
|
3 |
+
import math
|
4 |
+
import copy
|
5 |
+
import sys
|
6 |
+
from importlib import import_module
|
7 |
+
|
8 |
+
import torch
|
9 |
+
from torch import nn, Tensor
|
10 |
+
import torch.nn.init as init
|
11 |
+
from torch.nn import functional as F
|
12 |
+
from transformers.modeling_outputs import CausalLMOutput
|
13 |
+
from transformers import (
|
14 |
+
PreTrainedModel,
|
15 |
+
PretrainedConfig,
|
16 |
+
AutoConfig,
|
17 |
+
AutoModel,
|
18 |
+
AutoModelForCausalLM,
|
19 |
+
)
|
20 |
+
|
21 |
+
from transformers.utils import (
|
22 |
+
is_flash_attn_2_available,
|
23 |
+
is_flash_attn_greater_or_equal_2_10,
|
24 |
+
)
|
25 |
+
|
26 |
+
if is_flash_attn_2_available():
|
27 |
+
from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
|
28 |
+
|
29 |
+
# The model type string to bind.
|
30 |
+
model_type = "walsh-causal-v1"
|
31 |
+
|
32 |
+
class Config(PretrainedConfig):
|
33 |
+
model_type = model_type
|
34 |
+
|
35 |
+
attribute_map = {
|
36 |
+
"hidden_size": "d_embed",
|
37 |
+
}
|
38 |
+
|
39 |
+
def __init__(
|
40 |
+
# All of these MUST have defaults, even if unused.
|
41 |
+
self,
|
42 |
+
vocab_size=16000,
|
43 |
+
pad_index=None,
|
44 |
+
hidden_size=1024,
|
45 |
+
num_attention_heads=8,
|
46 |
+
num_hidden_layers=6,
|
47 |
+
max_sequence_length=2048,
|
48 |
+
dim_feedforward = 4096,
|
49 |
+
dropout=0.1,
|
50 |
+
loss_function = "causal_loss",
|
51 |
+
|
52 |
+
# Default class to use for each of these components.
|
53 |
+
positional_encoder_cls='.PositionalEncoder',
|
54 |
+
attention_cls='.CausalSelfAttention',
|
55 |
+
activation_cls='torch.nn.ReLU',
|
56 |
+
feedforward_cls='.FeedforwardLayer',
|
57 |
+
layer_stack_cls='.TransformerLayerStack',
|
58 |
+
layer_cls='.PostLayerNorm',
|
59 |
+
transformer_cls='.Transformer',
|
60 |
+
norm_cls='torch.nn.LayerNorm',
|
61 |
+
embdding_cls='torch.nn.Embedding',
|
62 |
+
output_proj_cls='torch.nn.Linear',
|
63 |
+
|
64 |
+
positional_encoder_args={
|
65 |
+
'd_model': 1024,
|
66 |
+
'max_seq_len': 2048,
|
67 |
+
},
|
68 |
+
|
69 |
+
# Arg groups, passed to factory classes above.
|
70 |
+
transformer_args=dict(),
|
71 |
+
attention_args=dict(),
|
72 |
+
feedforward_args=dict(),
|
73 |
+
activation_args=dict(),
|
74 |
+
norm_args={
|
75 |
+
'normalized_shape': 1024,
|
76 |
+
},
|
77 |
+
layer_stack_args=dict(),
|
78 |
+
layer_args=dict(),
|
79 |
+
embedding_args=dict(),
|
80 |
+
output_proj_args=dict(),
|
81 |
+
|
82 |
+
**kwargs,
|
83 |
+
):
|
84 |
+
self.vocab_size = vocab_size
|
85 |
+
self.pad_index = pad_index
|
86 |
+
self.hidden_size = hidden_size
|
87 |
+
self.num_attention_heads = num_attention_heads
|
88 |
+
self.num_hidden_layers = num_hidden_layers
|
89 |
+
self.max_sequence_length = max_sequence_length
|
90 |
+
self.loss_function = loss_function
|
91 |
+
|
92 |
+
self.dim_feedforward = dim_feedforward
|
93 |
+
self.dropout = dropout
|
94 |
+
|
95 |
+
self.positional_encoder_cls = positional_encoder_cls
|
96 |
+
self.attention_cls = attention_cls
|
97 |
+
self.activation_cls = activation_cls
|
98 |
+
self.feedforward_cls = feedforward_cls
|
99 |
+
self.layer_stack_cls = layer_stack_cls
|
100 |
+
self.layer_cls = layer_cls
|
101 |
+
self.transformer_cls = transformer_cls
|
102 |
+
self.norm_cls = norm_cls
|
103 |
+
self.embdding_cls = embdding_cls
|
104 |
+
self.output_proj_cls = output_proj_cls
|
105 |
+
|
106 |
+
self.positional_encoder_args = positional_encoder_args
|
107 |
+
self.transformer_args = transformer_args
|
108 |
+
self.attention_args = attention_args
|
109 |
+
self.feedforward_args = feedforward_args
|
110 |
+
self.activation_args = activation_args
|
111 |
+
self.norm_args = norm_args
|
112 |
+
self.layer_stack_args = layer_stack_args
|
113 |
+
self.layer_args = layer_args
|
114 |
+
self.embedding_args = embedding_args
|
115 |
+
self.output_proj_args = output_proj_args
|
116 |
+
|
117 |
+
super().__init__(**kwargs)
|
118 |
+
|
119 |
+
def causal_loss(logits: Tensor, labels: Tensor, input_ids: Tensor, ignore_index=-100) -> Tensor:
|
120 |
+
"""
|
121 |
+
Compute and return the loss using logits and labels.
|
122 |
+
"""
|
123 |
+
# Shift so that tokens < n predict n
|
124 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
125 |
+
shift_labels = labels[..., 1:].contiguous()
|
126 |
+
|
127 |
+
loss = torch.nn.functional.cross_entropy(
|
128 |
+
shift_logits.view(-1, shift_logits.size(-1)),
|
129 |
+
shift_labels.view(-1),
|
130 |
+
ignore_index=ignore_index,
|
131 |
+
reduction='mean',
|
132 |
+
)
|
133 |
+
|
134 |
+
return loss.nan_to_num()
|
135 |
+
|
136 |
+
# Learning to Break the Loop: Analyzing and Mitigating Repetitions for Neural Text Generation
|
137 |
+
# https://arxiv.org/abs/2206.02369
|
138 |
+
def ditto_loss(logits: Tensor, labels: Tensor, input_ids: Tensor) -> Tensor:
|
139 |
+
batch_size, seq_len, vocab_size = logits.shape
|
140 |
+
rep_reduce_gamma = 0.5
|
141 |
+
ditto_weight = 1.0e5
|
142 |
+
|
143 |
+
probs = torch.softmax(logits, dim=-1)
|
144 |
+
total_loss = None
|
145 |
+
for i in range(batch_size):
|
146 |
+
context_len = labels[i, 0].item()
|
147 |
+
sentence_len = labels[i, 1].item()
|
148 |
+
n_repeats = labels[i, 2].item()
|
149 |
+
|
150 |
+
# For readability
|
151 |
+
context_end = context_len
|
152 |
+
sentence_start = context_len
|
153 |
+
sentence_end = sentence_start + sentence_len
|
154 |
+
target_start = sentence_end
|
155 |
+
|
156 |
+
# Get causal loss for context tokens
|
157 |
+
causal_ids = input_ids[i:i+1, :context_end]
|
158 |
+
c_loss = causal_loss(
|
159 |
+
logits=logits[i:i+1, :context_end],
|
160 |
+
labels=causal_ids,
|
161 |
+
input_ids=causal_ids
|
162 |
+
)
|
163 |
+
|
164 |
+
# Slice out target probabilities
|
165 |
+
target_probs = probs[i , target_start:, :]
|
166 |
+
|
167 |
+
# Slice out first instance of repeated sentence, detach is (prevents back-prop), repeat in N times,
|
168 |
+
# and trim to length of target_probs.
|
169 |
+
baseline_probs = probs[i, sentence_start:sentence_end, :].detach().repeat(n_repeats, 1)[:target_probs.size(0), :]
|
170 |
+
|
171 |
+
# Compute DITTO loss.
|
172 |
+
one_minus_probs = torch.clamp((1.0 - torch.abs((target_probs - baseline_probs * rep_reduce_gamma))), min=1e-20)
|
173 |
+
r_loss = -torch.log(one_minus_probs).mean() * ditto_weight
|
174 |
+
|
175 |
+
# Combine repitition and causal loss
|
176 |
+
loss = c_loss + r_loss
|
177 |
+
|
178 |
+
# Add this to the total
|
179 |
+
if total_loss is None:
|
180 |
+
total_loss = loss
|
181 |
+
else:
|
182 |
+
total_loss += loss
|
183 |
+
|
184 |
+
return total_loss / batch_size
|
185 |
+
|
186 |
+
# Dynamically lookup class name and return factory for class.
|
187 |
+
def get_dynamic_class(name):
|
188 |
+
try:
|
189 |
+
module_path, class_name = name.rsplit('.', 1)
|
190 |
+
if module_path == "":
|
191 |
+
return getattr(sys.modules[__name__], class_name)
|
192 |
+
module = import_module(module_path)
|
193 |
+
return getattr(module, class_name)
|
194 |
+
except (ImportError, AttributeError) as e:
|
195 |
+
raise ImportError(name)
|
196 |
+
|
197 |
+
# An easily extensible dynamic transformer class
|
198 |
+
# Many variations can be specified entirely in the configuration, without touching this code.
|
199 |
+
class HFCausalModel(PreTrainedModel):
|
200 |
+
config_class = Config
|
201 |
+
model_type = 'Transformer'
|
202 |
+
supports_gradient_checkpointing = True
|
203 |
+
# Presently needs to be manually set to match transformer layer class...
|
204 |
+
_no_split_modules = ["DeepNetLayer"]
|
205 |
+
_supports_flash_attn_2 = True
|
206 |
+
_supports_sdpa = True
|
207 |
+
|
208 |
+
def __init__(self, config):
|
209 |
+
super().__init__(config)
|
210 |
+
|
211 |
+
self.d_model = config.hidden_size
|
212 |
+
self.transformer_head = self._make_transformer(config)
|
213 |
+
self.loss_function = get_dynamic_class(config.loss_function)
|
214 |
+
self.gradient_checkpointing = False
|
215 |
+
self.post_init()
|
216 |
+
|
217 |
+
def forward(
|
218 |
+
self,
|
219 |
+
input_ids: Optional[torch.LongTensor] = None,
|
220 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
221 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
222 |
+
position_ids: Optional[torch.LongTensor] = None,
|
223 |
+
labels: Optional[torch.LongTensor] = None,
|
224 |
+
output_attentions: Optional[bool] = None,
|
225 |
+
output_hidden_states: Optional[bool] = None,
|
226 |
+
return_dict: Optional[bool] = None,
|
227 |
+
**kwargs,
|
228 |
+
) -> (Tensor, dict[str, Tensor]):
|
229 |
+
|
230 |
+
if self.gradient_checkpointing and self.training:
|
231 |
+
gradient_checkpointing_func = self._gradient_checkpointing_func
|
232 |
+
else:
|
233 |
+
gradient_checkpointing_func = None
|
234 |
+
|
235 |
+
logits, attentions = self.transformer_head(
|
236 |
+
input_ids=input_ids,
|
237 |
+
need_weights=output_attentions,
|
238 |
+
gradient_checkpointing_func=gradient_checkpointing_func,
|
239 |
+
)
|
240 |
+
|
241 |
+
# Compute loss.
|
242 |
+
if labels is not None:
|
243 |
+
loss = self.loss_function(logits=logits, labels=labels, input_ids=input_ids)
|
244 |
+
else:
|
245 |
+
loss = None
|
246 |
+
|
247 |
+
return CausalLMOutput(loss=loss, logits=logits, attentions=attentions)
|
248 |
+
|
249 |
+
# Needed for generate() method.
|
250 |
+
def prepare_inputs_for_generation(self, input_ids, **kwargs):
|
251 |
+
attention_mask = kwargs.get("attention_mask", None)
|
252 |
+
model_inputs = {
|
253 |
+
"input_ids": input_ids,
|
254 |
+
"attention_mask": attention_mask,
|
255 |
+
}
|
256 |
+
return model_inputs
|
257 |
+
|
258 |
+
def _make_embedding(self, config):
|
259 |
+
embedding_cls = get_dynamic_class(config.embdding_cls)
|
260 |
+
return embedding_cls(config.vocab_size, self.d_model, config.pad_index, **config.embedding_args)
|
261 |
+
|
262 |
+
def _make_pos_encoder(self, config):
|
263 |
+
pos_enc_cls = get_dynamic_class(config.positional_encoder_cls)
|
264 |
+
return pos_enc_cls(**config.positional_encoder_args)
|
265 |
+
|
266 |
+
def _make_output_projection(self, config):
|
267 |
+
output_proj_cls = get_dynamic_class(config.output_proj_cls)
|
268 |
+
return output_proj_cls(self.d_model, config.vocab_size, **config.output_proj_args)
|
269 |
+
|
270 |
+
def _make_dropout(self, config):
|
271 |
+
return nn.Dropout(config.dropout)
|
272 |
+
|
273 |
+
def _make_activation(self, config):
|
274 |
+
activation_cls = get_dynamic_class(config.activation_cls)
|
275 |
+
return activation_cls(**config.activation_args)
|
276 |
+
|
277 |
+
def _make_norm(self, config):
|
278 |
+
norm_cls = get_dynamic_class(config.norm_cls)
|
279 |
+
return norm_cls(self.d_model)
|
280 |
+
|
281 |
+
def _make_self_attention(self, config):
|
282 |
+
attention_cls = get_dynamic_class(config.attention_cls)
|
283 |
+
# Map HF _attn_implementation to attn_type
|
284 |
+
match config._attn_implementation:
|
285 |
+
case "flash_attention_2":
|
286 |
+
if is_flash_attn_2_available():
|
287 |
+
if not is_flash_attn_greater_or_equal_2_10():
|
288 |
+
raise Exception("flash_attn_2 >= 2.10 is required")
|
289 |
+
attn_type = "flash2"
|
290 |
+
else:
|
291 |
+
attn_type = "torch"
|
292 |
+
case "sdpa":
|
293 |
+
attn_type = "torch"
|
294 |
+
case "eager":
|
295 |
+
attn_type = "native"
|
296 |
+
case _:
|
297 |
+
raise Exception(f"Unimplemented attention type '{config._attn_implementation}'")
|
298 |
+
return attention_cls(
|
299 |
+
d_model=self.d_model,
|
300 |
+
num_heads=config.num_attention_heads,
|
301 |
+
attn_type=attn_type,
|
302 |
+
**config.attention_args,
|
303 |
+
)
|
304 |
+
|
305 |
+
def _make_feedforward(self, config):
|
306 |
+
feedforward_cls = get_dynamic_class(config.feedforward_cls)
|
307 |
+
return feedforward_cls(
|
308 |
+
d_model=self.d_model,
|
309 |
+
feedforward_dim=config.dim_feedforward,
|
310 |
+
dropout=config.dropout,
|
311 |
+
activation=self._make_activation(config),
|
312 |
+
**config.feedforward_args,
|
313 |
+
)
|
314 |
+
|
315 |
+
def _make_layer(self, config):
|
316 |
+
layer_cls = get_dynamic_class(config.layer_cls)
|
317 |
+
return layer_cls(
|
318 |
+
d_model=self.d_model,
|
319 |
+
dropout=self._make_dropout(config),
|
320 |
+
attention=self._make_self_attention(config),
|
321 |
+
feedforward=self._make_feedforward(config),
|
322 |
+
norm1=self._make_norm(config),
|
323 |
+
norm2=self._make_norm(config),
|
324 |
+
**config.layer_args,
|
325 |
+
)
|
326 |
+
|
327 |
+
def _make_layer_stack(self, config):
|
328 |
+
layer_stack_cls = get_dynamic_class(config.layer_stack_cls)
|
329 |
+
return layer_stack_cls(
|
330 |
+
layers=nn.ModuleList([
|
331 |
+
self._make_layer(config) for _ in range(config.num_hidden_layers)
|
332 |
+
]),
|
333 |
+
**config.layer_stack_args,
|
334 |
+
)
|
335 |
+
|
336 |
+
def _make_transformer(self, config):
|
337 |
+
transformer_cls = get_dynamic_class(config.transformer_cls)
|
338 |
+
return transformer_cls(
|
339 |
+
d_model=self.d_model,
|
340 |
+
embedding=self._make_embedding(config),
|
341 |
+
positional_encoder=self._make_pos_encoder(config),
|
342 |
+
layer_stack=self._make_layer_stack(config),
|
343 |
+
output_projection=self._make_output_projection(config),
|
344 |
+
**config.transformer_args,
|
345 |
+
)
|
346 |
+
|
347 |
+
@torch.no_grad()
|
348 |
+
def _init_weights(self, module):
|
349 |
+
pass
|
350 |
+
|
351 |
+
# Register model type and configuration
|
352 |
+
AutoConfig.register(model_type, Config)
|
353 |
+
AutoModelForCausalLM.register(Config, HFCausalModel)
|
354 |
+
|
355 |
+
# A generic container class for standard transformer components.
|
356 |
+
class Transformer(nn.Module):
|
357 |
+
def __init__(self, d_model, embedding, positional_encoder, layer_stack, output_projection, **kwargs):
|
358 |
+
super().__init__()
|
359 |
+
self.embedding = embedding
|
360 |
+
self.positional_encoder = positional_encoder
|
361 |
+
self.layer_stack = layer_stack
|
362 |
+
self.output_projection = output_projection
|
363 |
+
self.d_model = d_model
|
364 |
+
self.sqrt_d_model = d_model**0.5
|
365 |
+
self.reset_parameters()
|
366 |
+
|
367 |
+
def forward(self, input_ids, need_weights, gradient_checkpointing_func):
|
368 |
+
x = self.positional_encoder(self.embedding(input_ids) * self.sqrt_d_model)
|
369 |
+
|
370 |
+
x, attentions = self.layer_stack(
|
371 |
+
x,
|
372 |
+
need_weights,
|
373 |
+
gradient_checkpointing_func,
|
374 |
+
)
|
375 |
+
|
376 |
+
# Translate output embedding ot logits.
|
377 |
+
logits = self.output_projection(x)
|
378 |
+
return logits, attentions
|
379 |
+
|
380 |
+
def reset_parameters(self):
|
381 |
+
init.xavier_uniform_(self.output_projection.weight)
|
382 |
+
init.constant_(self.output_projection.bias, 0.)
|
383 |
+
init.normal_(self.embedding.weight, std=self.d_model**-0.5)
|
384 |
+
|
385 |
+
# A vanilla positional encoder
|
386 |
+
class PositionalEncoder(nn.Module):
|
387 |
+
def __init__(self, d_embed, max_seq):
|
388 |
+
super().__init__()
|
389 |
+
self.d_embed = d_embed
|
390 |
+
self.max_seq = max_seq
|
391 |
+
|
392 |
+
weight = torch.zeros(max_seq, d_embed)
|
393 |
+
position = torch.arange(0, max_seq, dtype=torch.float).unsqueeze(1)
|
394 |
+
div_term = torch.exp(torch.arange(0, d_embed, 2).float() * (-math.log(10000.0) / d_embed))
|
395 |
+
weight[:, 0::2] = torch.sin(position * div_term)
|
396 |
+
weight[:, 1::2] = torch.cos(position * div_term)
|
397 |
+
weight = weight.unsqueeze(0)
|
398 |
+
self.register_buffer('weight', weight)
|
399 |
+
|
400 |
+
def forward(self, x):
|
401 |
+
seq_len = x.size(-2)
|
402 |
+
return x + self.weight[:, :seq_len]
|
403 |
+
|
404 |
+
# Converts a torch array of integers into their equivalent binary codes.
|
405 |
+
def binary_tensor(x, bits):
|
406 |
+
mask = 2**torch.arange(bits).to(x.device, x.dtype)
|
407 |
+
return x.unsqueeze(-1).bitwise_and(mask).ne(0).byte()
|
408 |
+
|
409 |
+
def hadamard_walsh_matrix(k: int):
|
410 |
+
# k: The dimension of the matrix is 2^k
|
411 |
+
assert k > 0
|
412 |
+
|
413 |
+
# Start with Hadamard H2^1 matrix.
|
414 |
+
h1 = torch.tensor([[1, 1], [1, -1]], dtype=torch.float)
|
415 |
+
|
416 |
+
# The series of matrices can be computed by recurisvely applying the Kronecker product,
|
417 |
+
# starting with h1.
|
418 |
+
#
|
419 |
+
# This will produce the series of Hadamard-Wlash matrices in natural order.
|
420 |
+
w = h1
|
421 |
+
for _ in range(k-1):
|
422 |
+
w = torch.kron(h1, w)
|
423 |
+
|
424 |
+
return w
|
425 |
+
|
426 |
+
# This positional encoder adds absolute binary positions to the embedding, encoded via
|
427 |
+
# Hadamard-Walsh matrix.
|
428 |
+
# See: https://en.wikipedia.org/wiki/Hadamard_code
|
429 |
+
# Each bit in the binary code word is encoded via a row the Hadamard-Walsh matrix, with a
|
430 |
+
# 1 being encoded by the presense of the row and a 0 by its absence. While training, the base
|
431 |
+
# sequence offset is randomly selected, which appears to allow the model to generalize to
|
432 |
+
# sequences longer than it was trained on. This is similar to what is described here:
|
433 |
+
# https://arxiv.org/pdf/2305.16843.pdf
|
434 |
+
# I have tried this approach and found that my approach works better for generalization.
|
435 |
+
#
|
436 |
+
# Note: Without random shifting, the early performance of this encoder is exceptionally good.
|
437 |
+
# The drawback is that the model can't generalize to longer sequences than it was trained on
|
438 |
+
# and can't easily accomidate additonal bits later in the training process.
|
439 |
+
class RSWalshPositionalEncoder(nn.Module):
|
440 |
+
def __init__(self, d_embed, max_seq, gain=0.333):
|
441 |
+
super().__init__()
|
442 |
+
self.max_seq = max_seq
|
443 |
+
self.d_embed = d_embed
|
444 |
+
|
445 |
+
# Hadamard-Walsh k, where the dimension of the matrix is 2^k
|
446 |
+
k = math.ceil(math.log2(d_embed))
|
447 |
+
|
448 |
+
# The number of bits required to encode max_seq
|
449 |
+
bits = math.ceil(math.log2(max_seq))
|
450 |
+
|
451 |
+
# Gain controls the weight given to the encodings.
|
452 |
+
# When a trainable parameter, the value appears to settle at around 0.333.
|
453 |
+
self.gain = gain
|
454 |
+
|
455 |
+
assert bits <= d_embed, "max_seq exceeds n-bits available for d_embed"
|
456 |
+
|
457 |
+
# Generate sequential binary codes for absolute positionals.
|
458 |
+
# The implementation originally used Grey codes, which where successive symbols
|
459 |
+
# differ by by only one bit. See: https://en.wikipedia.org/wiki/Gray_code
|
460 |
+
# This, along with a few other coding schemes were tested, with a simple
|
461 |
+
# binary code having the best performance.
|
462 |
+
binary_code = binary_tensor(torch.arange(0, max_seq, 1), bits)
|
463 |
+
self.register_buffer('binary_code', binary_code, persistent=False)
|
464 |
+
|
465 |
+
# Each bit is encoded via a row of a Hadamard-Walsh matrix.
|
466 |
+
# We slice off the unused rows and columns -- ideally, d_embed should be
|
467 |
+
# the same dimension as the matrix.
|
468 |
+
walsh = hadamard_walsh_matrix(k)[:bits,:d_embed] * self.gain
|
469 |
+
|
470 |
+
# This alternative appears superior to the original.
|
471 |
+
# If starting from scratch, this use this.
|
472 |
+
# walsh = (hadamard_walsh_matrix(k)[:bits,:d_embed] -0.5) * self.gain
|
473 |
+
self.register_buffer('walsh', walsh, persistent=False)
|
474 |
+
|
475 |
+
def forward(self, x):
|
476 |
+
seq_len = x.size(-2)
|
477 |
+
|
478 |
+
# Get sequence of binary codes...
|
479 |
+
# We use a random base offset when training.
|
480 |
+
# This results in slower initial gains, but appears to allow the model to generalize to
|
481 |
+
# the value of max_seq, even if never trained with sequences of this length. I also have
|
482 |
+
# a suspicion that this has a regularizing effect on training, similar to dropout. Models with
|
483 |
+
# random base offset shifting, despite slower initial improvement, appear to perform better in the long-run.
|
484 |
+
# TODO: Setup a controlled experiment to test this hypothesis.
|
485 |
+
if self.training:
|
486 |
+
shift = torch.randint(self.max_seq - seq_len + 1, (1,)).item()
|
487 |
+
seq = self.binary_code[shift:seq_len + shift,:]
|
488 |
+
|
489 |
+
# Disable shifting when not training. This does not appear to change the evaluation loss, but
|
490 |
+
# it does makes predictions easier to analyse when the attention weights are not shifting with each step.
|
491 |
+
else:
|
492 |
+
seq = self.binary_code[:seq_len,:]
|
493 |
+
|
494 |
+
# For reasons I have yet to identify, when the model is running in Textgenwebui, the matrix appears
|
495 |
+
# to evade conversion to bfloat16, despite everything else having been converted.
|
496 |
+
# This is a work-around for this.
|
497 |
+
self.walsh = self.walsh.to(dtype=x.dtype)
|
498 |
+
|
499 |
+
# Encode binary sequence with Hadamard-Walsh codes and apply to embeddings.
|
500 |
+
# If nothing else, the Walsh encodings make the positional information exceptionally
|
501 |
+
# robust with respect to dropout and other adversities. They can still be easily detected
|
502 |
+
# at the final layer.
|
503 |
+
return x + (seq.to(dtype=x.dtype) @ self.walsh)
|
504 |
+
|
505 |
+
# A generic stack of transformer layers.
|
506 |
+
class TransformerLayerStack(nn.Module):
|
507 |
+
def __init__(self, layers):
|
508 |
+
super().__init__()
|
509 |
+
self.layers = layers
|
510 |
+
|
511 |
+
def forward(self, x, need_weights, gradient_checkpointing_func=None):
|
512 |
+
attentions = []
|
513 |
+
for layer in self.layers:
|
514 |
+
if gradient_checkpointing_func is not None:
|
515 |
+
x, attention_weights = gradient_checkpointing_func(
|
516 |
+
layer.__call__,
|
517 |
+
x,
|
518 |
+
need_weights,
|
519 |
+
use_reentrant=False
|
520 |
+
)
|
521 |
+
else:
|
522 |
+
x, attention_weights = layer(x, need_weights=need_weights)
|
523 |
+
if need_weights:
|
524 |
+
attentions.append(attention_weights)
|
525 |
+
|
526 |
+
return x, attentions
|
527 |
+
|
528 |
+
# DeepNet: Scaling Transformers to 1,000 Layers
|
529 |
+
# https://arxiv.org/abs/2203.00555
|
530 |
+
class DeepnetLayer(nn.Module):
|
531 |
+
def __init__(
|
532 |
+
self,
|
533 |
+
d_model,
|
534 |
+
attention,
|
535 |
+
feedforward,
|
536 |
+
norm1,
|
537 |
+
norm2,
|
538 |
+
dropout,
|
539 |
+
alpha=1.0,
|
540 |
+
):
|
541 |
+
super().__init__()
|
542 |
+
self.d_model = d_model
|
543 |
+
self.attention = attention
|
544 |
+
self.feedforward = feedforward
|
545 |
+
self.norm1 = norm1
|
546 |
+
self.norm2 = norm2
|
547 |
+
self.dropout = dropout
|
548 |
+
# Deepnet alpha
|
549 |
+
self.alpha = alpha
|
550 |
+
|
551 |
+
def forward(self, x, need_weights=False):
|
552 |
+
# Keep input as residual
|
553 |
+
residual = x * self.alpha
|
554 |
+
|
555 |
+
# Compute attention
|
556 |
+
x, attention_weights = self.attention(x, need_weights)
|
557 |
+
|
558 |
+
# Add attention with residual and normalize.
|
559 |
+
x = self.norm1(residual + self.dropout(x))
|
560 |
+
|
561 |
+
# Keep output as next residual.
|
562 |
+
residual = x * self.alpha
|
563 |
+
|
564 |
+
# Pass through feedforward network.
|
565 |
+
x = self.feedforward(x)
|
566 |
+
|
567 |
+
# Combine residual and ff output, then normalize again.
|
568 |
+
x = self.norm2(residual + self.dropout(x))
|
569 |
+
|
570 |
+
return x, attention_weights
|
571 |
+
|
572 |
+
# A vanilla MLP transfomer layer.
|
573 |
+
class FeedforwardLayer(nn.Module):
|
574 |
+
def __init__(
|
575 |
+
self,
|
576 |
+
d_model: int,
|
577 |
+
feedforward_dim: int,
|
578 |
+
dropout,
|
579 |
+
activation=nn.ReLU(),
|
580 |
+
beta=1.0,
|
581 |
+
bias=True,
|
582 |
+
):
|
583 |
+
super().__init__()
|
584 |
+
self.d_model = d_model
|
585 |
+
self.beta = beta
|
586 |
+
self.activation = activation
|
587 |
+
self.linear1 = nn.Linear(d_model, feedforward_dim, bias=bias)
|
588 |
+
self.linear2 = nn.Linear(feedforward_dim, d_model, bias=bias)
|
589 |
+
self.dropout = nn.Dropout(dropout)
|
590 |
+
self.reset_parameters()
|
591 |
+
|
592 |
+
def forward(self, x):
|
593 |
+
return self.linear2(self.dropout(self.activation(self.linear1(x))))
|
594 |
+
|
595 |
+
def reset_parameters(self):
|
596 |
+
init.xavier_uniform_(self.linear1.weight, gain=self.beta)
|
597 |
+
init.xavier_uniform_(self.linear2.weight, gain=self.beta)
|
598 |
+
init.constant_(self.linear1.bias, 0.)
|
599 |
+
init.constant_(self.linear2.bias, 0.)
|
600 |
+
|
601 |
+
# GLU Variants Improve Transformer
|
602 |
+
# https://arxiv.org/pdf/2002.05202v1.pdf
|
603 |
+
class SwiGLUFeedforwardLayer(nn.Module):
|
604 |
+
def __init__(
|
605 |
+
self,
|
606 |
+
d_model,
|
607 |
+
d_feedforward,
|
608 |
+
beta=1.0,
|
609 |
+
dropout=0.1
|
610 |
+
):
|
611 |
+
super().__init__()
|
612 |
+
self.d_model = d_model
|
613 |
+
self.d_feedforward = d_feedforward
|
614 |
+
self.beta = 1.0
|
615 |
+
|
616 |
+
self.linear1 = nn.Linear(self.d_model, self.d_feedforward * 2, bias=False)
|
617 |
+
self.linear2 = nn.Linear(self.d_feedforward, self.d_model, bias=False)
|
618 |
+
self.dropout = nn.Dropout(dropout)
|
619 |
+
self.reset_parameters()
|
620 |
+
|
621 |
+
def forward(self, x):
|
622 |
+
x, gate = self.linear1(x).chunk(2, dim=-1)
|
623 |
+
x = x * F.silu(gate)
|
624 |
+
x = self.dropout(x)
|
625 |
+
x = self.linear2(x)
|
626 |
+
return x
|
627 |
+
|
628 |
+
def reset_parameters(self):
|
629 |
+
# Deepnet initialization
|
630 |
+
# https://arxiv.org/pdf/2203.00555.pdf
|
631 |
+
w, g = self.linear1.weight.chunk(2, dim=0)
|
632 |
+
init.xavier_uniform_(w, gain=self.beta)
|
633 |
+
init.xavier_uniform_(g, gain=self.beta)
|
634 |
+
init.xavier_uniform_(self.linear2.weight, gain=self.beta)
|
635 |
+
|
636 |
+
class CausalSelfAttention(nn.Module):
|
637 |
+
def __init__(
|
638 |
+
self,
|
639 |
+
d_model,
|
640 |
+
num_heads,
|
641 |
+
# values:
|
642 |
+
# native: Use local impementation; slowest option; good for debugging; useful when experimenting with non-standard stuff.
|
643 |
+
# torch: Use pytorch "scaled_dot_product_attention()"; faster; generally good compatibility; does not support returning attn weights.
|
644 |
+
# flash2: Use Flash-Attention2 implementation; fastest; limited to int16 and bfloat16 types; least memory usage.
|
645 |
+
attn_type,
|
646 |
+
beta=1.0,
|
647 |
+
dropout=0.1,
|
648 |
+
):
|
649 |
+
super().__init__()
|
650 |
+
self.d_model = d_model
|
651 |
+
self.num_heads = num_heads
|
652 |
+
self.beta = beta
|
653 |
+
self.attn_type = attn_type
|
654 |
+
|
655 |
+
assert d_model % num_heads == 0, "d_model must be evenly divisible by num_heads"
|
656 |
+
|
657 |
+
# The dimension of each head.
|
658 |
+
self.d_head = d_model // num_heads
|
659 |
+
|
660 |
+
# We scale the attention scores by the inverse-square-root of the head dimension
|
661 |
+
# this shifts the temerature of softmax.
|
662 |
+
self.dot_product_scale = 1.0 / math.sqrt(self.d_head)
|
663 |
+
|
664 |
+
self.in_proj = nn.Linear(self.d_model, 3 * self.d_model, bias=True)
|
665 |
+
self.output_linear = nn.Linear(self.d_model, self.d_model, bias=True)
|
666 |
+
|
667 |
+
self.dropout = nn.Dropout(dropout)
|
668 |
+
self.reset_parameters()
|
669 |
+
|
670 |
+
def extra_repr(self) -> str:
|
671 |
+
return f'd_model={self.d_model}, num_heads={self.num_heads}, beta={self.beta}, attn_type={self.attn_type}, dropout={self.dropout}'
|
672 |
+
|
673 |
+
def reset_parameters(self):
|
674 |
+
# Deepnet initialization
|
675 |
+
# https://arxiv.org/pdf/2203.00555.pdf
|
676 |
+
q, k, v = self.in_proj.weight.chunk(3)
|
677 |
+
init.xavier_uniform_(q, gain=1.0)
|
678 |
+
init.xavier_uniform_(k, gain=1.0)
|
679 |
+
init.xavier_uniform_(v, gain=self.beta)
|
680 |
+
init.xavier_uniform_(self.output_linear.weight, gain=self.beta)
|
681 |
+
init.constant_(self.in_proj.bias, 0.)
|
682 |
+
init.constant_(self.output_linear.bias, 0.)
|
683 |
+
|
684 |
+
def project_input(self, qkv):
|
685 |
+
proj = self.in_proj(qkv)
|
686 |
+
return proj.chunk(chunks=3, dim=-1)
|
687 |
+
|
688 |
+
def forward(self, qkv, need_weights):
|
689 |
+
if self.attn_type == "flash2":
|
690 |
+
return self.flash2_forward(qkv)
|
691 |
+
|
692 |
+
# qkv: (batch_size, seq_len, d_embed)
|
693 |
+
batch_size, seq_len, d_embed = qkv.shape
|
694 |
+
|
695 |
+
# Feed the inputs through the K, Q, V matrices.
|
696 |
+
query, key, value = self.project_input(qkv)
|
697 |
+
|
698 |
+
# Split projections into multiple heads and swap position of sequence / heads dimension
|
699 |
+
query = query.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
700 |
+
key = key.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
701 |
+
value = value.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
702 |
+
|
703 |
+
# Default to returning empty attention weights.
|
704 |
+
attention_weights = None
|
705 |
+
|
706 |
+
if self.attn_type == "torch":
|
707 |
+
# This context manager can be used to force which implementation to use.
|
708 |
+
#with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
709 |
+
attended_values = F.scaled_dot_product_attention(
|
710 |
+
query,
|
711 |
+
key,
|
712 |
+
value,
|
713 |
+
attn_mask=None,
|
714 |
+
dropout_p=self.dropout.p if self.training else 0.0,
|
715 |
+
is_causal=True,
|
716 |
+
scale=self.dot_product_scale
|
717 |
+
)
|
718 |
+
# "native" scaled-dot-product attention implementation.
|
719 |
+
else:
|
720 |
+
# Compute attention scores
|
721 |
+
scores = torch.matmul(query, key.transpose(-2, -1)) * self.dot_product_scale
|
722 |
+
|
723 |
+
# Mask future positions from the past
|
724 |
+
scores.masked_fill_(
|
725 |
+
torch.tril(
|
726 |
+
torch.ones(seq_len, seq_len, dtype=torch.bool, device=qkv.device),
|
727 |
+
diagonal=0,
|
728 |
+
).logical_not(),
|
729 |
+
float('-inf'),
|
730 |
+
)
|
731 |
+
|
732 |
+
# Calculate the attention weights; avoid NANs that might emerge from zeros in softmax's denominator
|
733 |
+
attention_weights = self.dropout(torch.softmax(scores, dim=-1).clamp(min=1e-10))
|
734 |
+
del scores
|
735 |
+
|
736 |
+
# Use the attention weights to get a weighted combination of value vectors
|
737 |
+
attended_values = torch.matmul(attention_weights, value)
|
738 |
+
if not need_weights:
|
739 |
+
del attention_weights
|
740 |
+
attention_weights = None
|
741 |
+
|
742 |
+
# Concatenate attention heads and project to original embedding size using the output linear layer
|
743 |
+
attended_values = attended_values.transpose(1, 2).contiguous().view(batch_size, seq_len, d_embed)
|
744 |
+
|
745 |
+
# Project the concatenated output through the output matrix.
|
746 |
+
attended_values = self.output_linear(attended_values)
|
747 |
+
return attended_values, attention_weights
|
748 |
+
|
749 |
+
def flash2_forward(self, qkv):
|
750 |
+
batch_size, seq_len, d_embed = qkv.shape
|
751 |
+
|
752 |
+
# Feed the inputs through the K, Q, V matrices.
|
753 |
+
# query : (batch_size, seq_len, d_model)
|
754 |
+
# qkv : (batch_size, seq_len, 3, num_heads, d_kq)
|
755 |
+
qkv = self.in_proj(qkv).unflatten(
|
756 |
+
-1,
|
757 |
+
(3, self.num_heads, self.d_head)
|
758 |
+
)
|
759 |
+
|
760 |
+
attended_values = flash_attn_qkvpacked_func(
|
761 |
+
qkv.bfloat16(),
|
762 |
+
dropout_p=self.dropout.p if self.training else 0.0,
|
763 |
+
softmax_scale=self.dot_product_scale,
|
764 |
+
causal=True,
|
765 |
+
)
|
766 |
+
# attended_values: (batch_size, seqlen, nheads, headdim)
|
767 |
+
|
768 |
+
# Concatentate heads back into d_embed
|
769 |
+
attended_values = attended_values.view(batch_size, seq_len, d_embed)
|
770 |
+
|
771 |
+
# Project the concatenated output through the output matrix.
|
772 |
+
attended_values = self.output_linear(attended_values)
|
773 |
+
return attended_values, None
|
774 |
+
|
775 |
+
# Attention layer with ALiBi relative positional encoding
|
776 |
+
# TRAIN SHORT, TEST LONG: ATTENTION WITH LINEAR BIASES ENABLES INPUT LENGTH EXTRAPOLATION
|
777 |
+
# https://arxiv.org/pdf/2108.12409.pdf
|
778 |
+
def alibi_biases(query_len, key_len, device='cpu'):
|
779 |
+
x = torch.arange(key_len, device=device)[None, :]
|
780 |
+
y = torch.arange(query_len, device=device)[:, None]
|
781 |
+
return x - y
|
782 |
+
|
783 |
+
class CausalAlibiAttention(nn.Module):
|
784 |
+
def __init__(
|
785 |
+
self,
|
786 |
+
d_model,
|
787 |
+
num_heads,
|
788 |
+
beta=1.0,
|
789 |
+
dropout=0.1,
|
790 |
+
# values:
|
791 |
+
# native: Use local impementation; slowest option; good for debugging; useful when experimenting with non-standard stuff.
|
792 |
+
# torch: Use pytorch "scaled_dot_product_attention()"; faster; generally good compatibility; does not support returning attn weights.
|
793 |
+
# flash2: Use Flash-Attention2 implementation; fastest; limited to int16 and bfloat16 types; can't train Alibi weights; least memory usage.
|
794 |
+
# Note: You can perform initial training with "torch," then switch to "flash2," after the Alibi weights have settled.
|
795 |
+
window_size=None,
|
796 |
+
attn_type="native",
|
797 |
+
freeze_alibi=True,
|
798 |
+
):
|
799 |
+
super().__init__()
|
800 |
+
self.d_model = d_model
|
801 |
+
self.num_heads = num_heads
|
802 |
+
self.beta = beta
|
803 |
+
self.attn_type = attn_type
|
804 |
+
|
805 |
+
assert d_model % num_heads == 0, "d_model must be evenly divisible by num_heads"
|
806 |
+
|
807 |
+
# The dimension of each head.
|
808 |
+
self.d_head = d_model // num_heads
|
809 |
+
|
810 |
+
# We scale the attention scores by the inverse-square-root of the head dimension
|
811 |
+
# this shifts the temerature of softmax.
|
812 |
+
self.dot_product_scale = 1.0 / math.sqrt(self.d_head)
|
813 |
+
|
814 |
+
self.in_proj = nn.Parameter(torch.empty(3 * self.d_model, self.d_model))
|
815 |
+
self.output_linear = nn.Linear(self.d_model, self.d_model, bias=False)
|
816 |
+
|
817 |
+
if window_size is not None:
|
818 |
+
self.window_size=(window_size, -1)
|
819 |
+
else:
|
820 |
+
self.window_size = (-1, -1)
|
821 |
+
|
822 |
+
self.dropout = nn.Dropout(dropout)
|
823 |
+
|
824 |
+
# This generates the original slope distribution from the paper.
|
825 |
+
# Observations with trainable slopes suggest that the high half of the slopes shift
|
826 |
+
# towards / past 1.0 and the low half approach zero or even go slightly negative.
|
827 |
+
# alibi_slopes = 1.0 / torch.logspace(1, 8, self.num_heads, base=2, dtype=torch.float)
|
828 |
+
|
829 |
+
# These appear to work better, as initial values, in practice.
|
830 |
+
alibi_slopes = 1.0 / torch.logspace(0, 7, self.num_heads, base=2, dtype=torch.float)
|
831 |
+
|
832 |
+
# If not trainable, it can improve performance somewhat if the low half are set to zero. Apparently
|
833 |
+
# making roughly half of the slopes position-agnostic is somehow closer to optimal?
|
834 |
+
# alibi_slopes.masked_fill_(torch.where(torch.arange(0, self.num_heads) >= (self.num_heads / 2), True, False), 0)
|
835 |
+
|
836 |
+
self.alibi_slopes = nn.Parameter(alibi_slopes)
|
837 |
+
|
838 |
+
# Optionally, allow/disallow training of ALiBi slopes.
|
839 |
+
self.alibi_slopes.requires_grad = (not freeze_alibi)
|
840 |
+
self.reset_parameters()
|
841 |
+
|
842 |
+
def extra_repr(self) -> str:
|
843 |
+
return f'd_model={self.d_model}, num_heads={self.num_heads}, beta={self.beta}, attn_type={self.attn_type}, window_size={self.window_size}, dropout={self.dropout}'
|
844 |
+
|
845 |
+
def reset_parameters(self):
|
846 |
+
# Deepnet initialization
|
847 |
+
# https://arxiv.org/pdf/2203.00555.pdf
|
848 |
+
|
849 |
+
q, k, v = self.in_proj.chunk(3)
|
850 |
+
init.xavier_uniform_(q, gain=1.0)
|
851 |
+
init.xavier_uniform_(k, gain=1.0)
|
852 |
+
init.xavier_uniform_(v, gain=self.beta)
|
853 |
+
init.xavier_uniform_(self.output_linear.weight, gain=self.beta)
|
854 |
+
|
855 |
+
def project_input(self, qkv):
|
856 |
+
proj = F.linear(qkv, self.in_proj)
|
857 |
+
return proj.chunk(chunks=3, dim=-1)
|
858 |
+
|
859 |
+
def forward(self, qkv, need_weights):
|
860 |
+
if self.attn_type == "flash2":
|
861 |
+
return self.flash2_forward(qkv)
|
862 |
+
|
863 |
+
# qkv: (batch_size, seq_len, d_embed)
|
864 |
+
batch_size, seq_len, d_embed = qkv.shape
|
865 |
+
|
866 |
+
# Feed the inputs through the K, Q, V matrices.
|
867 |
+
query, key, value = self.project_input(qkv)
|
868 |
+
|
869 |
+
# Split projections into multiple heads and swap position of sequence / heads dimension
|
870 |
+
query = query.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
871 |
+
key = key.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
872 |
+
value = value.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
873 |
+
|
874 |
+
# Apply Alibi relative positional biases.
|
875 |
+
attn_bias = alibi_biases(seq_len, seq_len, device=query.device) * self.alibi_slopes.view(-1, 1, 1)
|
876 |
+
|
877 |
+
# Mask future positions from the past
|
878 |
+
causal_mask = torch.tril(torch.ones(seq_len, seq_len, dtype=torch.bool, device=qkv.device), diagonal=0)
|
879 |
+
attn_bias.masked_fill_(causal_mask.logical_not(), float('-inf'))
|
880 |
+
del causal_mask
|
881 |
+
|
882 |
+
# Default to returning empty attention weights.
|
883 |
+
attention_weights = None
|
884 |
+
|
885 |
+
if self.attn_type == "torch":
|
886 |
+
# This context manager can be used to force which implementation to use.
|
887 |
+
#with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
888 |
+
attended_values = F.scaled_dot_product_attention(
|
889 |
+
query,
|
890 |
+
key,
|
891 |
+
value,
|
892 |
+
attn_mask=attn_bias.to(dtype=query.dtype),
|
893 |
+
dropout_p=self.dropout.p if self.training else 0.0,
|
894 |
+
is_causal=False,
|
895 |
+
scale=self.dot_product_scale
|
896 |
+
)
|
897 |
+
# "native" scaled-dot-product attention implementation.
|
898 |
+
else:
|
899 |
+
# Compute attention scores
|
900 |
+
scores = torch.matmul(query, key.transpose(-2, -1)) * self.dot_product_scale
|
901 |
+
|
902 |
+
# Adjust scores with attn_mask
|
903 |
+
scores += attn_bias
|
904 |
+
|
905 |
+
# Calculate the attention weights; avoid NANs that might emerge from zeros in softmax's denominator
|
906 |
+
attention_weights = self.dropout(torch.softmax(scores, dim=-1).clamp(min=1e-10))
|
907 |
+
|
908 |
+
# Use the attention weights to get a weighted combination of value vectors
|
909 |
+
attended_values = torch.matmul(attention_weights, value)
|
910 |
+
if not need_weights:
|
911 |
+
attention_weights = None
|
912 |
+
|
913 |
+
# Concatenate attention heads and project to original embedding size using the output linear layer
|
914 |
+
attended_values = attended_values.transpose(1, 2).contiguous().view(batch_size, seq_len, d_embed)
|
915 |
+
|
916 |
+
# Project the concatenated output through the output matrix.
|
917 |
+
attended_values = self.output_linear(attended_values)
|
918 |
+
return attended_values, attention_weights
|
919 |
+
|
920 |
+
def flash2_forward(self, qkv):
|
921 |
+
batch_size, seq_len, d_embed = qkv.shape
|
922 |
+
|
923 |
+
# Feed the inputs through the K, Q, V matrices.
|
924 |
+
# query : (batch_size, seq_len, d_model)
|
925 |
+
# qkv : (batch_size, seq_len, 3, num_heads, d_kq)
|
926 |
+
qkv = F.linear(
|
927 |
+
qkv,
|
928 |
+
self.in_proj,
|
929 |
+
).unflatten(
|
930 |
+
-1,
|
931 |
+
(3, self.num_heads, self.d_head)
|
932 |
+
)
|
933 |
+
|
934 |
+
attended_values = flash_attn_qkvpacked_func(
|
935 |
+
qkv.bfloat16(),
|
936 |
+
dropout_p=self.dropout.p if self.training else 0.0,
|
937 |
+
softmax_scale=self.dot_product_scale,
|
938 |
+
causal=True,
|
939 |
+
window_size=self.window_size,
|
940 |
+
alibi_slopes=self.alibi_slopes.float(),
|
941 |
+
).to(dtype=qkv.dtype)
|
942 |
+
# attended_values: (batch_size, seqlen, nheads, headdim)
|
943 |
+
|
944 |
+
# Concatentate heads back into d_embed
|
945 |
+
attended_values = attended_values.view(batch_size, seq_len, d_embed)
|
946 |
+
|
947 |
+
# Project the concatenated output through the output matrix.
|
948 |
+
attended_values = self.output_linear(attended_values)
|
949 |
+
return attended_values, None
|