reymondzzzz
commited on
Commit
•
1799ef3
1
Parent(s):
c4f295d
Upload GPTRefactForCausalLM
Browse files- config.json +4 -0
- configuration_gpt_refact.py +62 -0
- modeling_gpt_refact.py +651 -0
config.json
CHANGED
@@ -5,6 +5,10 @@
|
|
5 |
],
|
6 |
"attention_softmax_in_fp32": true,
|
7 |
"attn_pdrop": 0.1,
|
|
|
|
|
|
|
|
|
8 |
"bos_token_id": 50256,
|
9 |
"do_sample": true,
|
10 |
"embd_pdrop": 0.1,
|
|
|
5 |
],
|
6 |
"attention_softmax_in_fp32": true,
|
7 |
"attn_pdrop": 0.1,
|
8 |
+
"auto_map": {
|
9 |
+
"AutoConfig": "configuration_gpt_refact.GPTRefactConfig",
|
10 |
+
"AutoModelForCausalLM": "modeling_gpt_refact.GPTRefactForCausalLM"
|
11 |
+
},
|
12 |
"bos_token_id": 50256,
|
13 |
"do_sample": true,
|
14 |
"embd_pdrop": 0.1,
|
configuration_gpt_refact.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers.configuration_utils import PretrainedConfig
|
2 |
+
from transformers.utils import logging
|
3 |
+
|
4 |
+
|
5 |
+
logger = logging.get_logger(__name__)
|
6 |
+
|
7 |
+
|
8 |
+
class GPTRefactConfig(PretrainedConfig):
|
9 |
+
model_type = "gpt_refact"
|
10 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
11 |
+
attribute_map = {
|
12 |
+
"hidden_size": "n_embd",
|
13 |
+
"max_position_embeddings": "n_positions",
|
14 |
+
"num_attention_heads": "n_head",
|
15 |
+
"num_hidden_layers": "n_layer",
|
16 |
+
}
|
17 |
+
|
18 |
+
def __init__(
|
19 |
+
self,
|
20 |
+
vocab_size=50257,
|
21 |
+
n_positions=1024,
|
22 |
+
n_embd=768,
|
23 |
+
n_layer=12,
|
24 |
+
n_head=12,
|
25 |
+
n_inner=None,
|
26 |
+
activation_function="gelu_new",
|
27 |
+
resid_pdrop=0.1,
|
28 |
+
embd_pdrop=0.1,
|
29 |
+
attn_pdrop=0.1,
|
30 |
+
layer_norm_epsilon=1e-5,
|
31 |
+
initializer_range=0.02,
|
32 |
+
scale_attn_weights=True,
|
33 |
+
use_cache=True,
|
34 |
+
bos_token_id=50256,
|
35 |
+
eos_token_id=50256,
|
36 |
+
attention_softmax_in_fp32=True,
|
37 |
+
scale_attention_softmax_in_fp32=True,
|
38 |
+
multi_query=True,
|
39 |
+
**kwargs,
|
40 |
+
):
|
41 |
+
self.vocab_size = vocab_size
|
42 |
+
self.n_positions = n_positions
|
43 |
+
self.n_embd = n_embd
|
44 |
+
self.n_layer = n_layer
|
45 |
+
self.n_head = n_head
|
46 |
+
self.n_inner = n_inner
|
47 |
+
self.activation_function = activation_function
|
48 |
+
self.resid_pdrop = resid_pdrop
|
49 |
+
self.embd_pdrop = embd_pdrop
|
50 |
+
self.attn_pdrop = attn_pdrop
|
51 |
+
self.layer_norm_epsilon = layer_norm_epsilon
|
52 |
+
self.initializer_range = initializer_range
|
53 |
+
self.scale_attn_weights = scale_attn_weights
|
54 |
+
self.use_cache = use_cache
|
55 |
+
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
|
56 |
+
self.scale_attention_softmax_in_fp32 = scale_attention_softmax_in_fp32
|
57 |
+
self.multi_query = multi_query
|
58 |
+
|
59 |
+
self.bos_token_id = bos_token_id
|
60 |
+
self.eos_token_id = eos_token_id
|
61 |
+
|
62 |
+
super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
|
modeling_gpt_refact.py
ADDED
@@ -0,0 +1,651 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import List, Optional, Tuple, Union
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.utils.checkpoint
|
6 |
+
from torch import nn
|
7 |
+
import torch.nn.functional as F
|
8 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
9 |
+
|
10 |
+
from transformers.activations import ACT2FN
|
11 |
+
from transformers.modeling_outputs import (
|
12 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
13 |
+
CausalLMOutputWithCrossAttentions,
|
14 |
+
SequenceClassifierOutputWithPast,
|
15 |
+
TokenClassifierOutput,
|
16 |
+
)
|
17 |
+
from transformers.modeling_utils import PreTrainedModel
|
18 |
+
from transformers.utils import (
|
19 |
+
add_code_sample_docstrings,
|
20 |
+
add_start_docstrings,
|
21 |
+
add_start_docstrings_to_model_forward,
|
22 |
+
logging,
|
23 |
+
)
|
24 |
+
from hf.configuration_gpt_refact import GPTRefactConfig
|
25 |
+
|
26 |
+
|
27 |
+
logger = logging.get_logger(__name__)
|
28 |
+
|
29 |
+
|
30 |
+
# Fused kernels
|
31 |
+
# Use separate functions for each case because conditionals prevent kernel fusion.
|
32 |
+
# TODO: Could have better fused kernels depending on scaling, dropout and head mask.
|
33 |
+
# Is it doable without writing 32 functions?
|
34 |
+
@torch.jit.script
|
35 |
+
def upcast_masked_softmax(
|
36 |
+
x: torch.Tensor, mask: torch.Tensor, mask_value: torch.Tensor, scale: float, softmax_dtype: torch.dtype
|
37 |
+
):
|
38 |
+
input_dtype = x.dtype
|
39 |
+
x = x.to(softmax_dtype) * scale
|
40 |
+
x = torch.where(mask, x, mask_value)
|
41 |
+
x = torch.nn.functional.softmax(x, dim=-1).to(input_dtype)
|
42 |
+
return x
|
43 |
+
|
44 |
+
|
45 |
+
@torch.jit.script
|
46 |
+
def upcast_softmax(x: torch.Tensor, scale: float, softmax_dtype: torch.dtype):
|
47 |
+
input_dtype = x.dtype
|
48 |
+
x = x.to(softmax_dtype) * scale
|
49 |
+
x = torch.nn.functional.softmax(x, dim=-1).to(input_dtype)
|
50 |
+
return x
|
51 |
+
|
52 |
+
|
53 |
+
@torch.jit.script
|
54 |
+
def masked_softmax(x: torch.Tensor, mask: torch.Tensor, mask_value: torch.Tensor):
|
55 |
+
x = torch.where(mask, x, mask_value)
|
56 |
+
x = torch.nn.functional.softmax(x, dim=-1)
|
57 |
+
return x
|
58 |
+
|
59 |
+
def _get_slopes(attn_heads: int, dev: str) -> torch.Tensor:
|
60 |
+
"""
|
61 |
+
## Get head-specific slope $m$ for each head
|
62 |
+
* `n_heads` is the number of heads in the attention layer $n$
|
63 |
+
The slope for first head is
|
64 |
+
$$\frac{1}{2^{\frac{8}{n}}} = 2^{-\frac{8}{n}}$$
|
65 |
+
The slopes for the rest of the heads are in a geometric series with a ratio same as above.
|
66 |
+
For instance when the number of heads is $8$ the slopes are
|
67 |
+
$$\frac{1}{2^1}, \frac{1}{2^2}, \dots, \frac{1}{2^8}$$
|
68 |
+
"""
|
69 |
+
|
70 |
+
# Get the closest power of 2 to `n_heads`.
|
71 |
+
# If `n_heads` is not a power of 2, then we first calculate slopes to the closest (smaller) power of 2,
|
72 |
+
# and then add the remaining slopes.
|
73 |
+
n = 2 ** math.floor(math.log2(attn_heads))
|
74 |
+
# $2^{-\frac{8}{n}}$
|
75 |
+
m_0 = 2.0 ** (-8.0 / n)
|
76 |
+
# $2^{-1\frac{8}{n}}, 2^{-2 \frac{8}{n}}, 2^{-3 \frac{8}{n}}, \dots$
|
77 |
+
m = torch.pow(m_0, torch.arange(1, 1 + n, device=dev))
|
78 |
+
|
79 |
+
# If `n_heads` is not a power of 2, then we add the remaining slopes.
|
80 |
+
# We calculate the remaining slopes for $n * 2$ (avoiding slopes added previously).
|
81 |
+
# And pick the slopes upto `n_heads`.
|
82 |
+
if n < attn_heads:
|
83 |
+
# $2^{-\frac{8}{2n}}$
|
84 |
+
m_hat_0 = 2.0 ** (-4.0 / n)
|
85 |
+
# $2^{-1\frac{8}{2n}}, 2^{-3 \frac{8}{2n}}, 2^{-5 \frac{8}{2n}}, \dots$
|
86 |
+
# Note that we take steps by $2$ to avoid slopes added previously.
|
87 |
+
m_hat = torch.pow(m_hat_0, torch.arange(1, 1 + 2 * (attn_heads - n), 2, device=dev))
|
88 |
+
# Concatenate the slopes with the remaining slopes.
|
89 |
+
m = torch.cat([m, m_hat])
|
90 |
+
|
91 |
+
return m
|
92 |
+
|
93 |
+
|
94 |
+
def get_alibi_biases(
|
95 |
+
B: int,
|
96 |
+
T: int,
|
97 |
+
attn_heads: int,
|
98 |
+
dev: str,
|
99 |
+
dtype,
|
100 |
+
causal: bool = True) -> torch.Tensor:
|
101 |
+
"""
|
102 |
+
## Calculate the attention biases matrix
|
103 |
+
* `n_heads` is the number of heads in the attention layer
|
104 |
+
* `mask` is the attention mask of shape `[seq_len_q, seq_len_k]`
|
105 |
+
This returns a matrix of shape `[seq_len_q, seq_len_k, n_heads, ]` with ALiBi attention biases.
|
106 |
+
"""
|
107 |
+
|
108 |
+
# Get slopes $m$ for each head
|
109 |
+
if causal:
|
110 |
+
mask = (torch.triu(torch.ones((T, T), device=dev)) == 1).transpose(0, 1)
|
111 |
+
else:
|
112 |
+
mask = torch.ones((T, T), device=dev, dtype=torch.bool)
|
113 |
+
|
114 |
+
m = _get_slopes(attn_heads, dev)
|
115 |
+
|
116 |
+
# Calculate distances $[0, 1, \dots, N]$
|
117 |
+
# Here we calculate the distances using the mask.
|
118 |
+
#
|
119 |
+
# Since it's causal mask we can just use $[0, 1, \dots, N]$ too.
|
120 |
+
# `distance = torch.arange(mask.shape[1], dtype=torch.long, device=mask.device)[None, :]`
|
121 |
+
distance = mask.cumsum(dim=-1)
|
122 |
+
|
123 |
+
# Multiply them pair-wise to get the AliBi bias matrix
|
124 |
+
biases = distance[:, :, None] * m[None, None, :]
|
125 |
+
biases = biases.permute(2, 0, 1)[None, :, :T, :T]
|
126 |
+
biases = biases.repeat(B, 1, 1, 1)
|
127 |
+
return biases.to(dtype).contiguous()
|
128 |
+
|
129 |
+
class Attention(nn.Module):
|
130 |
+
def __init__(self, config, layer_idx=None):
|
131 |
+
super().__init__()
|
132 |
+
self.mask_value = None
|
133 |
+
|
134 |
+
self.multi_query = config.multi_query
|
135 |
+
self.embed_dim = config.hidden_size
|
136 |
+
self.num_heads = config.num_attention_heads
|
137 |
+
self.head_dim = self.embed_dim // self.num_heads
|
138 |
+
self.kv_attn_heads = 1
|
139 |
+
|
140 |
+
self.scale = self.head_dim ** -0.5
|
141 |
+
|
142 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
143 |
+
raise ValueError(
|
144 |
+
f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
145 |
+
f" {self.num_heads})."
|
146 |
+
)
|
147 |
+
|
148 |
+
self.layer_idx = layer_idx
|
149 |
+
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
150 |
+
self.scale_attention_softmax_in_fp32 = (
|
151 |
+
config.scale_attention_softmax_in_fp32 and config.attention_softmax_in_fp32
|
152 |
+
)
|
153 |
+
|
154 |
+
self.q = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
|
155 |
+
self.k = nn.Linear(self.embed_dim, self.head_dim, bias=False)
|
156 |
+
self.v = nn.Linear(self.embed_dim, self.head_dim, bias=False)
|
157 |
+
self.c_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
|
158 |
+
|
159 |
+
def _attn(self, query, key, value, attention_mask=None, alibi=None):
|
160 |
+
dtype = query.dtype
|
161 |
+
softmax_dtype = torch.float32 if self.attention_softmax_in_fp32 else dtype
|
162 |
+
upcast = dtype != softmax_dtype
|
163 |
+
unscale = self.layer_idx + 1 if self.scale_attention_softmax_in_fp32 and upcast else 1
|
164 |
+
|
165 |
+
# MQA models: (batch_size, query_length, num_heads * head_dim)
|
166 |
+
# MHA models: (batch_size, num_heads, query_length, head_dim)
|
167 |
+
attn_weights = alibi + torch.matmul(query * self.scale, key)
|
168 |
+
|
169 |
+
if upcast:
|
170 |
+
# Use a fused kernel to prevent a large overhead from casting and scaling.
|
171 |
+
# Sub-optimal when the key length is not a multiple of 8.
|
172 |
+
if attention_mask is None:
|
173 |
+
attn_weights = upcast_softmax(attn_weights, unscale, softmax_dtype)
|
174 |
+
else:
|
175 |
+
mask_value = self._get_mask_value(attn_weights.device, softmax_dtype)
|
176 |
+
attn_weights = upcast_masked_softmax(attn_weights, attention_mask, mask_value, unscale, softmax_dtype)
|
177 |
+
else:
|
178 |
+
if attention_mask is not None:
|
179 |
+
|
180 |
+
# The fused kernel is very slow when the key length is not a multiple of 8, so we skip fusion.
|
181 |
+
attn_weights = torch.masked_fill(attn_weights, attention_mask, -10000)
|
182 |
+
|
183 |
+
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
|
184 |
+
|
185 |
+
attn_output = torch.matmul(attn_weights, value)
|
186 |
+
|
187 |
+
return attn_output, attn_weights
|
188 |
+
|
189 |
+
def _split_heads(self, tensor):
|
190 |
+
new_shape = tensor.shape[:-1] + (self.num_heads, self.head_dim)
|
191 |
+
tensor = tensor.view(new_shape)
|
192 |
+
return tensor.permute(0, 2, 1, 3)
|
193 |
+
|
194 |
+
def forward(
|
195 |
+
self,
|
196 |
+
hidden_states: torch.Tensor,
|
197 |
+
layer_past: Optional[torch.Tensor] = None,
|
198 |
+
attention_mask: Optional[torch.Tensor] = None,
|
199 |
+
alibi: Optional[torch.Tensor] = None,
|
200 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
201 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
202 |
+
use_cache: Optional[bool] = False,
|
203 |
+
output_attentions: Optional[bool] = False,
|
204 |
+
) -> Union[
|
205 |
+
Tuple[torch.Tensor, Optional[torch.Tensor]],
|
206 |
+
Tuple[torch.Tensor, Optional[torch.Tensor], Tuple[torch.Tensor, ...]],
|
207 |
+
]:
|
208 |
+
b, t, _ = hidden_states.shape
|
209 |
+
query = self.q(hidden_states)
|
210 |
+
key = self.k(hidden_states)
|
211 |
+
value = self.v(hidden_states)
|
212 |
+
query = self._split_heads(query)
|
213 |
+
key = key.view(b, t, self.kv_attn_heads, self.head_dim).permute(0, 2, 1, 3)
|
214 |
+
value = value.view(b, t, self.kv_attn_heads, self.head_dim).permute(0, 2, 1, 3)
|
215 |
+
|
216 |
+
if layer_past is not None:
|
217 |
+
past_key, past_value = layer_past
|
218 |
+
key = torch.cat((past_key, key), dim=-2)
|
219 |
+
value = torch.cat((past_value, value), dim=-2)
|
220 |
+
|
221 |
+
if use_cache is True:
|
222 |
+
present = (key, value)
|
223 |
+
else:
|
224 |
+
present = None
|
225 |
+
|
226 |
+
attn_output, attn_weights = self._attn(query, key.transpose(-1, -2), value, attention_mask, alibi)
|
227 |
+
|
228 |
+
attn_output = attn_output.transpose(1, 2).reshape(hidden_states.shape)
|
229 |
+
attn_output = self.c_proj(attn_output)
|
230 |
+
|
231 |
+
outputs = (attn_output, present)
|
232 |
+
if output_attentions:
|
233 |
+
outputs += (attn_weights,)
|
234 |
+
|
235 |
+
return outputs # a, present, (attentions)
|
236 |
+
|
237 |
+
|
238 |
+
class MLP(nn.Module):
|
239 |
+
def __init__(self, intermediate_size, config, multiple_of: int = 256):
|
240 |
+
super().__init__()
|
241 |
+
embed_dim = config.hidden_size
|
242 |
+
hidden_dim = intermediate_size
|
243 |
+
hidden_dim = int(2 * hidden_dim / 3)
|
244 |
+
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
|
245 |
+
self.linear_1 = nn.Linear(embed_dim, hidden_dim, bias=False)
|
246 |
+
self.linear_3 = nn.Linear(embed_dim, hidden_dim, bias=False)
|
247 |
+
self.c_proj = nn.Linear(hidden_dim, embed_dim, bias=False)
|
248 |
+
|
249 |
+
def forward(self, x: Optional[Tuple[torch.Tensor]]) -> torch.Tensor:
|
250 |
+
x1 = F.silu(self.linear_1(x))
|
251 |
+
x2 = self.linear_3(x)
|
252 |
+
x = self.c_proj(x1 * x2)
|
253 |
+
return x
|
254 |
+
|
255 |
+
|
256 |
+
class LayerNormNoBias(nn.Module):
|
257 |
+
|
258 |
+
def __init__(self, shape: int, eps: float = 1e-5):
|
259 |
+
super().__init__()
|
260 |
+
self.shape = (shape,)
|
261 |
+
self.eps = eps
|
262 |
+
self.weight = nn.Parameter(torch.empty(self.shape))
|
263 |
+
|
264 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
265 |
+
return F.layer_norm(x, self.shape, self.weight, None, self.eps)
|
266 |
+
|
267 |
+
class GPTRefactBlock(nn.Module):
|
268 |
+
def __init__(self, config, layer_idx=None):
|
269 |
+
super().__init__()
|
270 |
+
hidden_size = config.hidden_size
|
271 |
+
self.inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
|
272 |
+
|
273 |
+
self.ln_1 = LayerNormNoBias(hidden_size, eps=config.layer_norm_epsilon)
|
274 |
+
self.attn = Attention(config, layer_idx=layer_idx)
|
275 |
+
self.ln_2 = LayerNormNoBias(hidden_size, eps=config.layer_norm_epsilon)
|
276 |
+
|
277 |
+
self.mlp = MLP(self.inner_dim, config)
|
278 |
+
|
279 |
+
def forward(
|
280 |
+
self,
|
281 |
+
hidden_states: Optional[Tuple[torch.Tensor]],
|
282 |
+
layer_past: Optional[torch.Tensor] = None,
|
283 |
+
attention_mask: Optional[torch.Tensor] = None,
|
284 |
+
alibi: Optional[torch.Tensor] = None,
|
285 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
286 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
287 |
+
use_cache: Optional[bool] = False,
|
288 |
+
output_attentions: Optional[bool] = False,
|
289 |
+
) -> Union[
|
290 |
+
Tuple[torch.Tensor], Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
|
291 |
+
]:
|
292 |
+
hidden_states_norm = self.ln_1(hidden_states)
|
293 |
+
attn_outputs = self.attn(
|
294 |
+
hidden_states_norm,
|
295 |
+
layer_past=layer_past,
|
296 |
+
attention_mask=attention_mask,
|
297 |
+
alibi=alibi,
|
298 |
+
use_cache=use_cache,
|
299 |
+
output_attentions=output_attentions,
|
300 |
+
)
|
301 |
+
attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
|
302 |
+
outputs = attn_outputs[1:]
|
303 |
+
# residual connection
|
304 |
+
mix = attn_output + hidden_states
|
305 |
+
|
306 |
+
norm_mix = self.ln_2(mix)
|
307 |
+
feed_forward_hidden_states = self.mlp(norm_mix)
|
308 |
+
# residual connection
|
309 |
+
hidden_states = mix + feed_forward_hidden_states
|
310 |
+
|
311 |
+
if use_cache:
|
312 |
+
outputs = (hidden_states,) + outputs
|
313 |
+
else:
|
314 |
+
outputs = (hidden_states,) + outputs[1:]
|
315 |
+
|
316 |
+
return outputs # hidden_states, present, (attentions, cross_attentions)
|
317 |
+
|
318 |
+
|
319 |
+
class GPTRefactPreTrainedModel(PreTrainedModel):
|
320 |
+
"""
|
321 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
322 |
+
models.
|
323 |
+
"""
|
324 |
+
|
325 |
+
config_class = GPTRefactConfig
|
326 |
+
base_model_prefix = "transformer"
|
327 |
+
supports_gradient_checkpointing = True
|
328 |
+
_no_split_modules = ["GPTRefactBlock"]
|
329 |
+
_skip_keys_device_placement = "past_key_values"
|
330 |
+
|
331 |
+
def __init__(self, *inputs, **kwargs):
|
332 |
+
super().__init__(*inputs, **kwargs)
|
333 |
+
|
334 |
+
def _init_weights(self, module):
|
335 |
+
"""Initialize the weights."""
|
336 |
+
if isinstance(module, (MLP, Attention)):
|
337 |
+
# Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
|
338 |
+
# > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
|
339 |
+
# > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
|
340 |
+
# > -- GPT-2 :: https://openai.com/blog/better-language-models/
|
341 |
+
#
|
342 |
+
# Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
|
343 |
+
module.c_proj.weight.data.normal_(
|
344 |
+
mean=0.0, std=(self.config.initializer_range / math.sqrt(2 * self.config.n_layer))
|
345 |
+
)
|
346 |
+
module.c_proj._is_hf_initialized = True
|
347 |
+
elif isinstance(module, nn.Linear):
|
348 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
349 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
350 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
351 |
+
if module.bias is not None:
|
352 |
+
module.bias.data.zero_()
|
353 |
+
elif isinstance(module, nn.Embedding):
|
354 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
355 |
+
if module.padding_idx is not None:
|
356 |
+
module.weight.data[module.padding_idx].zero_()
|
357 |
+
elif isinstance(module, nn.LayerNorm):
|
358 |
+
module.bias.data.zero_()
|
359 |
+
module.weight.data.fill_(1.0)
|
360 |
+
|
361 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
362 |
+
if isinstance(module, GPTRefactModel):
|
363 |
+
module.gradient_checkpointing = value
|
364 |
+
|
365 |
+
|
366 |
+
class GPTRefactModel(GPTRefactPreTrainedModel):
|
367 |
+
def __init__(self, config):
|
368 |
+
super().__init__(config)
|
369 |
+
self.embed_dim = config.hidden_size
|
370 |
+
self.num_heads = config.num_attention_heads
|
371 |
+
|
372 |
+
self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
|
373 |
+
|
374 |
+
self.h = nn.ModuleList([GPTRefactBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
|
375 |
+
|
376 |
+
max_positions = config.max_position_embeddings
|
377 |
+
self.register_buffer(
|
378 |
+
"bias", torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)), persistent=False
|
379 |
+
)
|
380 |
+
|
381 |
+
self.gradient_checkpointing = False
|
382 |
+
|
383 |
+
# Initialize weights and apply final processing
|
384 |
+
self.post_init()
|
385 |
+
|
386 |
+
@staticmethod
|
387 |
+
def _make_mask(seq_len: int, past_key_values_length: int):
|
388 |
+
# prompt
|
389 |
+
if past_key_values_length == 0:
|
390 |
+
mask = torch.ones((seq_len, seq_len + past_key_values_length), dtype=torch.bool)
|
391 |
+
mask = torch.triu(mask, 1)
|
392 |
+
else:
|
393 |
+
mask = torch.zeros((seq_len, seq_len + past_key_values_length), dtype=torch.bool)
|
394 |
+
return mask
|
395 |
+
|
396 |
+
def forward(
|
397 |
+
self,
|
398 |
+
input_ids: Optional[torch.Tensor] = None,
|
399 |
+
past_key_values: Optional[List[torch.Tensor]] = None,
|
400 |
+
attention_mask: Optional[torch.Tensor] = None,
|
401 |
+
token_type_ids: Optional[torch.Tensor] = None,
|
402 |
+
position_ids: Optional[torch.Tensor] = None,
|
403 |
+
head_mask: Optional[torch.Tensor] = None,
|
404 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
405 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
406 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
407 |
+
use_cache: Optional[bool] = None,
|
408 |
+
output_attentions: Optional[bool] = None,
|
409 |
+
output_hidden_states: Optional[bool] = None,
|
410 |
+
return_dict: Optional[bool] = None,
|
411 |
+
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
412 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
413 |
+
output_hidden_states = (
|
414 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
415 |
+
)
|
416 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
417 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
418 |
+
|
419 |
+
if input_ids is not None and inputs_embeds is not None:
|
420 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
421 |
+
elif input_ids is not None:
|
422 |
+
input_shape = input_ids.size()
|
423 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
424 |
+
batch_size = input_ids.shape[0]
|
425 |
+
elif inputs_embeds is not None:
|
426 |
+
input_shape = inputs_embeds.size()[:-1]
|
427 |
+
batch_size = inputs_embeds.shape[0]
|
428 |
+
else:
|
429 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
430 |
+
|
431 |
+
if batch_size <= 0:
|
432 |
+
raise ValueError("batch_size has to be defined and > 0")
|
433 |
+
|
434 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
435 |
+
|
436 |
+
if token_type_ids is not None:
|
437 |
+
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
438 |
+
if position_ids is not None:
|
439 |
+
position_ids = position_ids.view(-1, input_shape[-1])
|
440 |
+
|
441 |
+
if past_key_values is None:
|
442 |
+
past_length = 0
|
443 |
+
past_key_values = tuple([None] * len(self.h))
|
444 |
+
else:
|
445 |
+
past_length = past_key_values[0][0].size(-2)
|
446 |
+
|
447 |
+
if attention_mask is not None and len(attention_mask.shape) == 2 and position_ids is None:
|
448 |
+
# create position_ids on the fly for batch generation
|
449 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
450 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
451 |
+
if past_length > 0:
|
452 |
+
position_ids = position_ids[:, past_length : input_shape[-1] + past_length :]
|
453 |
+
elif position_ids is None:
|
454 |
+
position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
|
455 |
+
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
|
456 |
+
|
457 |
+
# Self-attention mask.
|
458 |
+
query_length = input_shape[-1]
|
459 |
+
|
460 |
+
seq_length_with_past = past_length + query_length
|
461 |
+
if attention_mask is None:
|
462 |
+
attention_mask = self._make_mask(query_length, past_length).to(device)
|
463 |
+
else:
|
464 |
+
attention_mask = attention_mask.to(device)
|
465 |
+
|
466 |
+
hidden_states = self.wte(input_ids) if inputs_embeds is None else inputs_embeds
|
467 |
+
|
468 |
+
alibi = get_alibi_biases(hidden_states.shape[0], seq_length_with_past,
|
469 |
+
self.num_heads, device, self.wte.weight.dtype)[:, :, -query_length:, :]
|
470 |
+
|
471 |
+
if token_type_ids is not None:
|
472 |
+
token_type_embeds = self.wte(token_type_ids)
|
473 |
+
hidden_states = hidden_states + token_type_embeds
|
474 |
+
|
475 |
+
output_shape = input_shape + (hidden_states.size(-1),)
|
476 |
+
|
477 |
+
presents = [] if use_cache else None
|
478 |
+
all_self_attentions = () if output_attentions else None
|
479 |
+
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
480 |
+
all_hidden_states = () if output_hidden_states else None
|
481 |
+
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
482 |
+
if output_hidden_states:
|
483 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
484 |
+
|
485 |
+
if self.gradient_checkpointing and self.training:
|
486 |
+
|
487 |
+
def create_custom_forward(module):
|
488 |
+
def custom_forward(*inputs):
|
489 |
+
# None for past_key_value
|
490 |
+
return module(*inputs, use_cache, output_attentions)
|
491 |
+
|
492 |
+
return custom_forward
|
493 |
+
|
494 |
+
outputs = torch.utils.checkpoint.checkpoint(
|
495 |
+
create_custom_forward(block),
|
496 |
+
hidden_states,
|
497 |
+
None,
|
498 |
+
attention_mask,
|
499 |
+
head_mask[i],
|
500 |
+
encoder_hidden_states,
|
501 |
+
encoder_attention_mask,
|
502 |
+
)
|
503 |
+
else:
|
504 |
+
outputs = block(
|
505 |
+
hidden_states,
|
506 |
+
layer_past=layer_past,
|
507 |
+
attention_mask=attention_mask,
|
508 |
+
alibi=alibi,
|
509 |
+
encoder_hidden_states=encoder_hidden_states,
|
510 |
+
encoder_attention_mask=encoder_attention_mask,
|
511 |
+
use_cache=use_cache,
|
512 |
+
output_attentions=output_attentions,
|
513 |
+
)
|
514 |
+
|
515 |
+
hidden_states = outputs[0]
|
516 |
+
if use_cache:
|
517 |
+
presents.append(outputs[1])
|
518 |
+
|
519 |
+
if output_attentions:
|
520 |
+
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
|
521 |
+
if self.config.add_cross_attention:
|
522 |
+
all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
|
523 |
+
|
524 |
+
hidden_states = hidden_states.view(output_shape)
|
525 |
+
# Add last hidden state
|
526 |
+
if output_hidden_states:
|
527 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
528 |
+
|
529 |
+
if not return_dict:
|
530 |
+
return tuple(
|
531 |
+
v
|
532 |
+
for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
|
533 |
+
if v is not None
|
534 |
+
)
|
535 |
+
|
536 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
537 |
+
last_hidden_state=hidden_states,
|
538 |
+
past_key_values=presents,
|
539 |
+
hidden_states=all_hidden_states,
|
540 |
+
attentions=all_self_attentions,
|
541 |
+
cross_attentions=all_cross_attentions,
|
542 |
+
)
|
543 |
+
|
544 |
+
class GPTRefactForCausalLM(GPTRefactPreTrainedModel):
|
545 |
+
_tied_weights_keys = ["lm_head.weight"]
|
546 |
+
|
547 |
+
def __init__(self, config):
|
548 |
+
super().__init__(config)
|
549 |
+
self.transformer = GPTRefactModel(config)
|
550 |
+
self.ln_f = nn.LayerNorm(self.transformer.embed_dim, eps=config.layer_norm_epsilon)
|
551 |
+
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
552 |
+
|
553 |
+
# Initialize weights and apply final processing
|
554 |
+
self.post_init()
|
555 |
+
|
556 |
+
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
|
557 |
+
|
558 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
559 |
+
if inputs_embeds is not None and past_key_values is None:
|
560 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
561 |
+
else:
|
562 |
+
if past_key_values is not None:
|
563 |
+
model_inputs = {"input_ids": input_ids[..., -1:]}
|
564 |
+
else:
|
565 |
+
model_inputs = {"input_ids": input_ids}
|
566 |
+
|
567 |
+
model_inputs.update(
|
568 |
+
{
|
569 |
+
"past_key_values": past_key_values,
|
570 |
+
"use_cache": kwargs.get("use_cache"),
|
571 |
+
}
|
572 |
+
)
|
573 |
+
return model_inputs
|
574 |
+
|
575 |
+
def forward(
|
576 |
+
self,
|
577 |
+
input_ids: Optional[torch.Tensor] = None,
|
578 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
579 |
+
attention_mask: Optional[torch.Tensor] = None,
|
580 |
+
token_type_ids: Optional[torch.Tensor] = None,
|
581 |
+
position_ids: Optional[torch.Tensor] = None,
|
582 |
+
head_mask: Optional[torch.Tensor] = None,
|
583 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
584 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
585 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
586 |
+
labels: Optional[torch.Tensor] = None,
|
587 |
+
use_cache: Optional[bool] = None,
|
588 |
+
output_attentions: Optional[bool] = None,
|
589 |
+
output_hidden_states: Optional[bool] = None,
|
590 |
+
return_dict: Optional[bool] = None,
|
591 |
+
) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
|
592 |
+
r"""
|
593 |
+
labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
594 |
+
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
595 |
+
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
|
596 |
+
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
|
597 |
+
"""
|
598 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
599 |
+
|
600 |
+
transformer_outputs = self.transformer(
|
601 |
+
input_ids,
|
602 |
+
past_key_values=past_key_values,
|
603 |
+
attention_mask=attention_mask,
|
604 |
+
token_type_ids=token_type_ids,
|
605 |
+
position_ids=position_ids,
|
606 |
+
head_mask=head_mask,
|
607 |
+
inputs_embeds=inputs_embeds,
|
608 |
+
encoder_hidden_states=encoder_hidden_states,
|
609 |
+
encoder_attention_mask=encoder_attention_mask,
|
610 |
+
use_cache=use_cache,
|
611 |
+
output_attentions=output_attentions,
|
612 |
+
output_hidden_states=output_hidden_states,
|
613 |
+
return_dict=return_dict,
|
614 |
+
)
|
615 |
+
hidden_states = transformer_outputs[0]
|
616 |
+
|
617 |
+
x = self.ln_f(hidden_states)
|
618 |
+
lm_logits = self.lm_head(x)
|
619 |
+
|
620 |
+
loss = None
|
621 |
+
if labels is not None:
|
622 |
+
# Shift so that tokens < n predict n
|
623 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
624 |
+
shift_labels = labels[..., 1:].contiguous().to(shift_logits.device)
|
625 |
+
# Flatten the tokens
|
626 |
+
loss_fct = CrossEntropyLoss()
|
627 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
628 |
+
|
629 |
+
if not return_dict:
|
630 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
631 |
+
return ((loss,) + output) if loss is not None else output
|
632 |
+
|
633 |
+
return CausalLMOutputWithCrossAttentions(
|
634 |
+
loss=loss,
|
635 |
+
logits=lm_logits,
|
636 |
+
past_key_values=transformer_outputs.past_key_values,
|
637 |
+
hidden_states=transformer_outputs.hidden_states,
|
638 |
+
attentions=transformer_outputs.attentions,
|
639 |
+
cross_attentions=transformer_outputs.cross_attentions,
|
640 |
+
)
|
641 |
+
|
642 |
+
@staticmethod
|
643 |
+
def _reorder_cache(
|
644 |
+
past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
645 |
+
) -> Tuple[Tuple[torch.Tensor]]:
|
646 |
+
"""
|
647 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
648 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
649 |
+
beam_idx at every generation step.
|
650 |
+
"""
|
651 |
+
return tuple(layer_past.index_select(0, beam_idx.to(layer_past.device)) for layer_past in past_key_values)
|