Update config and add decoding code (buggy)
Browse filesIt loads all weights, and does a forward pass, but logits have NaNs
Need to find the bug
- config.json +5 -2
- decoder_only_t5/__init__.py +0 -0
- decoder_only_t5/config.py +11 -0
- decoder_only_t5/modeling.py +620 -0
config.json
CHANGED
@@ -23,8 +23,11 @@
|
|
23 |
"relative_attention_max_distance": 128,
|
24 |
"relative_attention_num_buckets": 32,
|
25 |
"task_specific_params": {},
|
26 |
-
"tie_word_embeddings":
|
27 |
"transformers_version": "4.23.1",
|
28 |
"use_cache": true,
|
29 |
-
"vocab_size": 256512
|
|
|
|
|
|
|
30 |
}
|
|
|
23 |
"relative_attention_max_distance": 128,
|
24 |
"relative_attention_num_buckets": 32,
|
25 |
"task_specific_params": {},
|
26 |
+
"tie_word_embeddings": true,
|
27 |
"transformers_version": "4.23.1",
|
28 |
"use_cache": true,
|
29 |
+
"vocab_size": 256512,
|
30 |
+
"parallel_layers": true,
|
31 |
+
"has_relative_attention_bias": false,
|
32 |
+
"multi_query_attention": true
|
33 |
}
|
decoder_only_t5/__init__.py
ADDED
File without changes
|
decoder_only_t5/config.py
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers.models.t5.configuration_t5 import T5Config
|
2 |
+
|
3 |
+
|
4 |
+
class DecoderOnlyT5Config(T5Config):
|
5 |
+
is_decoder_only = True
|
6 |
+
# whether to call attention and mlp in parallel.
|
7 |
+
# https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/architectures/t5/t5_architecture.py#L384
|
8 |
+
parallel_layers = True
|
9 |
+
has_relative_attention_bias = False
|
10 |
+
# https://arxiv.org/abs/1911.02150
|
11 |
+
multi_query_attention = True
|
decoder_only_t5/modeling.py
ADDED
@@ -0,0 +1,620 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
from typing import Optional, Tuple, Union
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from torch import nn
|
6 |
+
from torch.nn import CrossEntropyLoss
|
7 |
+
from transformers.models.t5 import modeling_t5
|
8 |
+
from transformers.modeling_outputs import Seq2SeqLMOutput
|
9 |
+
from transformers.utils import (
|
10 |
+
add_start_docstrings_to_model_forward,
|
11 |
+
logging,
|
12 |
+
replace_return_docstrings,
|
13 |
+
)
|
14 |
+
|
15 |
+
from decoder_only_t5.config import DecoderOnlyT5Config
|
16 |
+
|
17 |
+
|
18 |
+
logger = logging.get_logger(__name__)
|
19 |
+
_CONFIG_FOR_DOC = "DecoderOnlyT5Config"
|
20 |
+
|
21 |
+
|
22 |
+
class DecoderOnlyT5LayerFF(modeling_t5.T5LayerFF):
|
23 |
+
def __init__(self, config: DecoderOnlyT5Config):
|
24 |
+
super(modeling_t5.T5LayerFF, self).__init__()
|
25 |
+
if config.is_gated_act:
|
26 |
+
self.DenseReluDense = modeling_t5.T5DenseGatedActDense(config)
|
27 |
+
else:
|
28 |
+
self.DenseReluDense = modeling_t5.T5DenseActDense(config)
|
29 |
+
|
30 |
+
if not config.parallel_layers:
|
31 |
+
self.layer_norm = modeling_t5.T5LayerNorm(
|
32 |
+
config.d_model, eps=config.layer_norm_epsilon
|
33 |
+
)
|
34 |
+
else:
|
35 |
+
self.layer_norm = nn.Identity()
|
36 |
+
self.dropout = nn.Dropout(config.dropout_rate)
|
37 |
+
|
38 |
+
|
39 |
+
# https://github.com/huggingface/transformers/blob/7ee995fd9c692761c4601ddbffa2ac2ec9f27b0b/src/transformers/models/llama/modeling_llama.py#L263
|
40 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
41 |
+
"""
|
42 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
43 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
44 |
+
"""
|
45 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
46 |
+
if n_rep == 1:
|
47 |
+
return hidden_states
|
48 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(
|
49 |
+
batch, num_key_value_heads, n_rep, slen, head_dim
|
50 |
+
)
|
51 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
52 |
+
|
53 |
+
|
54 |
+
class DecoderOnlyT5Attention(modeling_t5.T5Attention):
|
55 |
+
"""
|
56 |
+
Supports both multi-head and multi-query attention.
|
57 |
+
https://arxiv.org/abs/1911.02150
|
58 |
+
https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/components/attention/dense_attention.py#L292
|
59 |
+
"""
|
60 |
+
|
61 |
+
def __init__(self, config: DecoderOnlyT5Config, has_relative_attention_bias=False):
|
62 |
+
super(modeling_t5.T5Attention, self).__init__()
|
63 |
+
self.is_decoder = config.is_decoder
|
64 |
+
self.has_relative_attention_bias = has_relative_attention_bias
|
65 |
+
self.relative_attention_num_buckets = config.relative_attention_num_buckets
|
66 |
+
self.relative_attention_max_distance = config.relative_attention_max_distance
|
67 |
+
self.d_model = config.d_model
|
68 |
+
self.key_value_proj_dim = config.d_kv
|
69 |
+
self.n_heads = config.num_heads
|
70 |
+
self.n_kv_heads = 1 if config.multi_query_attention else self.n_heads
|
71 |
+
self.n_kv_groups = self.n_heads // self.n_kv_heads
|
72 |
+
self.dropout = config.dropout_rate
|
73 |
+
self.inner_dim = self.n_heads * self.key_value_proj_dim
|
74 |
+
self.kv_inner_dim = self.n_kv_heads * self.key_value_proj_dim
|
75 |
+
|
76 |
+
# Mesh TensorFlow initialization to avoid scaling before softmax
|
77 |
+
|
78 |
+
self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
|
79 |
+
self.k = nn.Linear(self.d_model, self.kv_inner_dim, bias=False)
|
80 |
+
self.v = nn.Linear(self.d_model, self.kv_inner_dim, bias=False)
|
81 |
+
self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
|
82 |
+
|
83 |
+
if self.has_relative_attention_bias:
|
84 |
+
self.relative_attention_bias = nn.Embedding(
|
85 |
+
self.relative_attention_num_buckets, self.n_heads
|
86 |
+
)
|
87 |
+
self.pruned_heads = set()
|
88 |
+
self.gradient_checkpointing = False
|
89 |
+
|
90 |
+
def forward(
|
91 |
+
self,
|
92 |
+
hidden_states,
|
93 |
+
mask=None,
|
94 |
+
key_value_states=None,
|
95 |
+
position_bias=None,
|
96 |
+
past_key_value=None,
|
97 |
+
layer_head_mask=None,
|
98 |
+
query_length=None,
|
99 |
+
use_cache=False,
|
100 |
+
output_attentions=False,
|
101 |
+
):
|
102 |
+
"""
|
103 |
+
Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).
|
104 |
+
"""
|
105 |
+
# Input is (batch_size, seq_length, dim)
|
106 |
+
# Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)
|
107 |
+
# past_key_value[0] is (batch_size, n_kv_heads, q_len - 1, dim_per_head)
|
108 |
+
batch_size, seq_length = hidden_states.shape[:2]
|
109 |
+
|
110 |
+
real_seq_length = seq_length
|
111 |
+
|
112 |
+
if past_key_value is not None:
|
113 |
+
if len(past_key_value) != 2:
|
114 |
+
raise ValueError(
|
115 |
+
f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
|
116 |
+
)
|
117 |
+
real_seq_length += (
|
118 |
+
past_key_value[0].shape[2] if query_length is None else query_length
|
119 |
+
)
|
120 |
+
|
121 |
+
key_length = (
|
122 |
+
real_seq_length if key_value_states is None else key_value_states.shape[1]
|
123 |
+
)
|
124 |
+
|
125 |
+
def shape(states, n_heads):
|
126 |
+
"""projection"""
|
127 |
+
return states.view(
|
128 |
+
batch_size, -1, n_heads, self.key_value_proj_dim
|
129 |
+
).transpose(1, 2)
|
130 |
+
|
131 |
+
def unshape(states):
|
132 |
+
"""reshape"""
|
133 |
+
return (
|
134 |
+
states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)
|
135 |
+
)
|
136 |
+
|
137 |
+
def project(hidden_states, proj_layer, key_value_states, past_key_value):
|
138 |
+
"""projects hidden states correctly to key/query states"""
|
139 |
+
if key_value_states is None:
|
140 |
+
# self-attn
|
141 |
+
# (batch_size, n_kv_heads, seq_length, dim_per_head)
|
142 |
+
hidden_states = shape(proj_layer(hidden_states), self.n_kv_heads)
|
143 |
+
elif past_key_value is None:
|
144 |
+
# cross-attn
|
145 |
+
# (batch_size, n_kv_heads, seq_length, dim_per_head)
|
146 |
+
hidden_states = shape(proj_layer(key_value_states), self.n_kv_heads)
|
147 |
+
|
148 |
+
if past_key_value is not None:
|
149 |
+
if key_value_states is None:
|
150 |
+
# self-attn
|
151 |
+
# (batch_size, n_kv_heads, key_length, dim_per_head)
|
152 |
+
hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
|
153 |
+
elif past_key_value.shape[2] != key_value_states.shape[1]:
|
154 |
+
# checking that the `sequence_length` of the `past_key_value` is the same as
|
155 |
+
# the provided `key_value_states` to support prefix tuning
|
156 |
+
# cross-attn
|
157 |
+
# (batch_size, n_kv_heads, seq_length, dim_per_head)
|
158 |
+
hidden_states = shape(proj_layer(key_value_states), self.n_kv_heads)
|
159 |
+
else:
|
160 |
+
# cross-attn
|
161 |
+
hidden_states = past_key_value
|
162 |
+
return hidden_states
|
163 |
+
|
164 |
+
# get query states
|
165 |
+
query_states = shape(
|
166 |
+
self.q(hidden_states), self.n_heads
|
167 |
+
) # (batch_size, n_heads, seq_length, dim_per_head)
|
168 |
+
|
169 |
+
# get key/value states
|
170 |
+
key_states = project(
|
171 |
+
hidden_states,
|
172 |
+
self.k,
|
173 |
+
key_value_states,
|
174 |
+
past_key_value[0] if past_key_value is not None else None,
|
175 |
+
)
|
176 |
+
value_states = project(
|
177 |
+
hidden_states,
|
178 |
+
self.v,
|
179 |
+
key_value_states,
|
180 |
+
past_key_value[1] if past_key_value is not None else None,
|
181 |
+
)
|
182 |
+
|
183 |
+
# compute scores
|
184 |
+
scores = torch.matmul(
|
185 |
+
query_states, repeat_kv(key_states, self.n_kv_groups).transpose(3, 2)
|
186 |
+
) # equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
|
187 |
+
|
188 |
+
if position_bias is None:
|
189 |
+
if not self.has_relative_attention_bias:
|
190 |
+
position_bias = torch.zeros(
|
191 |
+
(1, self.n_heads, real_seq_length, key_length),
|
192 |
+
device=scores.device,
|
193 |
+
dtype=scores.dtype,
|
194 |
+
)
|
195 |
+
if self.gradient_checkpointing and self.training:
|
196 |
+
position_bias.requires_grad = True
|
197 |
+
else:
|
198 |
+
position_bias = self.compute_bias(
|
199 |
+
real_seq_length, key_length, device=scores.device
|
200 |
+
)
|
201 |
+
|
202 |
+
# if key and values are already calculated
|
203 |
+
# we want only the last query position bias
|
204 |
+
if past_key_value is not None:
|
205 |
+
position_bias = position_bias[:, :, -hidden_states.size(1) :, :]
|
206 |
+
|
207 |
+
if mask is not None:
|
208 |
+
position_bias = (
|
209 |
+
position_bias + mask
|
210 |
+
) # (batch_size, n_heads, seq_length, key_length)
|
211 |
+
|
212 |
+
if self.pruned_heads:
|
213 |
+
mask = torch.ones(position_bias.shape[1])
|
214 |
+
mask[list(self.pruned_heads)] = 0
|
215 |
+
position_bias_masked = position_bias[:, mask.bool()]
|
216 |
+
else:
|
217 |
+
position_bias_masked = position_bias
|
218 |
+
|
219 |
+
scores += position_bias_masked
|
220 |
+
attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(
|
221 |
+
scores
|
222 |
+
) # (batch_size, n_heads, seq_length, key_length)
|
223 |
+
attn_weights = nn.functional.dropout(
|
224 |
+
attn_weights, p=self.dropout, training=self.training
|
225 |
+
) # (batch_size, n_heads, seq_length, key_length)
|
226 |
+
|
227 |
+
# Mask heads if we want to
|
228 |
+
if layer_head_mask is not None:
|
229 |
+
attn_weights = attn_weights * layer_head_mask
|
230 |
+
|
231 |
+
attn_output = unshape(
|
232 |
+
torch.matmul(attn_weights, value_states)
|
233 |
+
) # (batch_size, seq_length, dim)
|
234 |
+
attn_output = self.o(attn_output)
|
235 |
+
|
236 |
+
present_key_value_state = (
|
237 |
+
(key_states, value_states) if (self.is_decoder and use_cache) else None
|
238 |
+
)
|
239 |
+
outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)
|
240 |
+
|
241 |
+
if output_attentions:
|
242 |
+
outputs = outputs + (attn_weights,)
|
243 |
+
return outputs
|
244 |
+
|
245 |
+
|
246 |
+
class DecoderOnlyT5LayerSelfAttention(modeling_t5.T5LayerSelfAttention):
|
247 |
+
def __init__(self, config, has_relative_attention_bias=False):
|
248 |
+
super(modeling_t5.T5LayerSelfAttention, self).__init__()
|
249 |
+
self.SelfAttention = DecoderOnlyT5Attention(
|
250 |
+
config, has_relative_attention_bias=has_relative_attention_bias
|
251 |
+
)
|
252 |
+
self.layer_norm = modeling_t5.T5LayerNorm(
|
253 |
+
config.d_model, eps=config.layer_norm_epsilon
|
254 |
+
)
|
255 |
+
self.dropout = nn.Dropout(config.dropout_rate)
|
256 |
+
self.parallel_layers = config.parallel_layers
|
257 |
+
|
258 |
+
def forward(
|
259 |
+
self,
|
260 |
+
hidden_states,
|
261 |
+
attention_mask=None,
|
262 |
+
position_bias=None,
|
263 |
+
layer_head_mask=None,
|
264 |
+
past_key_value=None,
|
265 |
+
use_cache=False,
|
266 |
+
output_attentions=False,
|
267 |
+
):
|
268 |
+
if not self.parallel_layers:
|
269 |
+
x = self.layer_norm(hidden_states)
|
270 |
+
else:
|
271 |
+
x = hidden_states
|
272 |
+
attention_output = self.SelfAttention(
|
273 |
+
x,
|
274 |
+
mask=attention_mask,
|
275 |
+
position_bias=position_bias,
|
276 |
+
layer_head_mask=layer_head_mask,
|
277 |
+
past_key_value=past_key_value,
|
278 |
+
use_cache=use_cache,
|
279 |
+
output_attentions=output_attentions,
|
280 |
+
)
|
281 |
+
if not self.parallel_layers:
|
282 |
+
# When parallel_layers is True, the residual connection is applied
|
283 |
+
# in the decoder block instead of here.
|
284 |
+
hidden_states = hidden_states + self.dropout(attention_output[0])
|
285 |
+
else:
|
286 |
+
hidden_states = attention_output[0]
|
287 |
+
outputs = (hidden_states,) + attention_output[
|
288 |
+
1:
|
289 |
+
] # add attentions if we output them
|
290 |
+
return outputs
|
291 |
+
|
292 |
+
|
293 |
+
class DecoderOnlyT5Block(modeling_t5.T5Block):
|
294 |
+
def __init__(self, config, has_relative_attention_bias=False):
|
295 |
+
super(modeling_t5.T5Block, self).__init__()
|
296 |
+
self.is_decoder = config.is_decoder
|
297 |
+
self.is_decoder_only = config.is_decoder_only
|
298 |
+
self.layer = nn.ModuleList()
|
299 |
+
self.layer.append(
|
300 |
+
DecoderOnlyT5LayerSelfAttention(
|
301 |
+
config, has_relative_attention_bias=has_relative_attention_bias
|
302 |
+
)
|
303 |
+
)
|
304 |
+
if self.is_decoder:
|
305 |
+
if config.is_decoder_only:
|
306 |
+
self.layer.append(nn.Identity())
|
307 |
+
else:
|
308 |
+
self.layer.append(modeling_t5.T5LayerCrossAttention(config))
|
309 |
+
self.parallel_layers = config.parallel_layers
|
310 |
+
self.layer.append(DecoderOnlyT5LayerFF(config))
|
311 |
+
|
312 |
+
def forward(
|
313 |
+
self,
|
314 |
+
hidden_states,
|
315 |
+
attention_mask=None,
|
316 |
+
position_bias=None,
|
317 |
+
encoder_hidden_states=None,
|
318 |
+
encoder_attention_mask=None,
|
319 |
+
encoder_decoder_position_bias=None,
|
320 |
+
layer_head_mask=None,
|
321 |
+
cross_attn_layer_head_mask=None,
|
322 |
+
past_key_value=None,
|
323 |
+
use_cache=False,
|
324 |
+
output_attentions=False,
|
325 |
+
return_dict=True,
|
326 |
+
):
|
327 |
+
if past_key_value is not None:
|
328 |
+
if not self.is_decoder:
|
329 |
+
logger.warning(
|
330 |
+
"`past_key_values` is passed to the encoder. Please make sure this is intended."
|
331 |
+
)
|
332 |
+
expected_num_past_key_values = 2 if encoder_hidden_states is None else 4
|
333 |
+
|
334 |
+
if len(past_key_value) != expected_num_past_key_values:
|
335 |
+
raise ValueError(
|
336 |
+
f"There should be {expected_num_past_key_values} past states. "
|
337 |
+
f"{'2 (past / key) for cross attention. ' if expected_num_past_key_values == 4 else ''}"
|
338 |
+
f"Got {len(past_key_value)} past key / value states"
|
339 |
+
)
|
340 |
+
|
341 |
+
self_attn_past_key_value = past_key_value[:2]
|
342 |
+
cross_attn_past_key_value = past_key_value[2:]
|
343 |
+
else:
|
344 |
+
self_attn_past_key_value, cross_attn_past_key_value = None, None
|
345 |
+
|
346 |
+
ff_layer = self.layer[-1]
|
347 |
+
if self.parallel_layers:
|
348 |
+
x = self.layer[0].layer_norm(hidden_states)
|
349 |
+
ff_output = ff_layer(hidden_states)
|
350 |
+
else:
|
351 |
+
x = hidden_states
|
352 |
+
|
353 |
+
self_attention_outputs = self.layer[0](
|
354 |
+
x,
|
355 |
+
attention_mask=attention_mask,
|
356 |
+
position_bias=position_bias,
|
357 |
+
layer_head_mask=layer_head_mask,
|
358 |
+
past_key_value=self_attn_past_key_value,
|
359 |
+
use_cache=use_cache,
|
360 |
+
output_attentions=output_attentions,
|
361 |
+
)
|
362 |
+
x, present_key_value_state = self_attention_outputs[:2]
|
363 |
+
attention_outputs = self_attention_outputs[
|
364 |
+
2:
|
365 |
+
] # Keep self-attention outputs and relative position weights
|
366 |
+
|
367 |
+
# clamp inf values to enable fp16 training
|
368 |
+
if x.dtype == torch.float16:
|
369 |
+
clamp_value = torch.where(
|
370 |
+
torch.isinf(x).any(),
|
371 |
+
torch.finfo(x.dtype).max - 1000,
|
372 |
+
torch.finfo(x.dtype).max,
|
373 |
+
)
|
374 |
+
x = torch.clamp(x, min=-clamp_value, max=clamp_value)
|
375 |
+
|
376 |
+
do_cross_attention = (
|
377 |
+
self.is_decoder
|
378 |
+
and not self.is_decoder_only
|
379 |
+
and encoder_hidden_states is not None
|
380 |
+
)
|
381 |
+
if do_cross_attention:
|
382 |
+
# the actual query length is unknown for cross attention
|
383 |
+
# if using past key value states. Need to inject it here
|
384 |
+
if present_key_value_state is not None:
|
385 |
+
query_length = present_key_value_state[0].shape[2]
|
386 |
+
else:
|
387 |
+
query_length = None
|
388 |
+
|
389 |
+
cross_attention_outputs = self.layer[1](
|
390 |
+
x,
|
391 |
+
key_value_states=encoder_hidden_states,
|
392 |
+
attention_mask=encoder_attention_mask,
|
393 |
+
position_bias=encoder_decoder_position_bias,
|
394 |
+
layer_head_mask=cross_attn_layer_head_mask,
|
395 |
+
past_key_value=cross_attn_past_key_value,
|
396 |
+
query_length=query_length,
|
397 |
+
use_cache=use_cache,
|
398 |
+
output_attentions=output_attentions,
|
399 |
+
)
|
400 |
+
x = cross_attention_outputs[0]
|
401 |
+
|
402 |
+
# clamp inf values to enable fp16 training
|
403 |
+
if x.dtype == torch.float16:
|
404 |
+
clamp_value = torch.where(
|
405 |
+
torch.isinf(x).any(),
|
406 |
+
torch.finfo(x.dtype).max - 1000,
|
407 |
+
torch.finfo(x.dtype).max,
|
408 |
+
)
|
409 |
+
x = torch.clamp(x, min=-clamp_value, max=clamp_value)
|
410 |
+
|
411 |
+
# Combine self attn and cross attn key value states
|
412 |
+
if present_key_value_state is not None:
|
413 |
+
present_key_value_state = (
|
414 |
+
present_key_value_state + cross_attention_outputs[1]
|
415 |
+
)
|
416 |
+
|
417 |
+
# Keep cross-attention outputs and relative position weights
|
418 |
+
attention_outputs = attention_outputs + cross_attention_outputs[2:]
|
419 |
+
|
420 |
+
if self.parallel_layers:
|
421 |
+
# https://github.com/google/flaxformer/blob/ea17eb012a1d340ddff017b7a534c2162aaec34c/flaxformer/architectures/t5/t5_architecture.py#L295
|
422 |
+
hidden_states = x + ff_output
|
423 |
+
hidden_states *= 2**-0.5
|
424 |
+
hidden_states = hidden_states + self.layer[0].dropout(hidden_states)
|
425 |
+
else:
|
426 |
+
hidden_states = ff_layer(x)
|
427 |
+
|
428 |
+
# clamp inf values to enable fp16 training
|
429 |
+
if hidden_states.dtype == torch.float16:
|
430 |
+
clamp_value = torch.where(
|
431 |
+
torch.isinf(hidden_states).any(),
|
432 |
+
torch.finfo(hidden_states.dtype).max - 1000,
|
433 |
+
torch.finfo(hidden_states.dtype).max,
|
434 |
+
)
|
435 |
+
hidden_states = torch.clamp(
|
436 |
+
hidden_states, min=-clamp_value, max=clamp_value
|
437 |
+
)
|
438 |
+
|
439 |
+
outputs = (hidden_states,)
|
440 |
+
|
441 |
+
if use_cache:
|
442 |
+
outputs = outputs + (present_key_value_state,) + attention_outputs
|
443 |
+
else:
|
444 |
+
outputs = outputs + attention_outputs
|
445 |
+
|
446 |
+
return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
|
447 |
+
|
448 |
+
|
449 |
+
class DecoderOnlyT5Stack(modeling_t5.T5Stack):
|
450 |
+
def __init__(self, config, embed_tokens=None):
|
451 |
+
super(modeling_t5.T5Stack, self).__init__(config)
|
452 |
+
|
453 |
+
self.embed_tokens = embed_tokens
|
454 |
+
self.is_decoder = config.is_decoder
|
455 |
+
|
456 |
+
self.block = nn.ModuleList(
|
457 |
+
[
|
458 |
+
DecoderOnlyT5Block(
|
459 |
+
config,
|
460 |
+
has_relative_attention_bias=(
|
461 |
+
config.has_relative_attention_bias and bool(i == 0)
|
462 |
+
),
|
463 |
+
)
|
464 |
+
for i in range(config.num_layers)
|
465 |
+
]
|
466 |
+
)
|
467 |
+
if not config.parallel_layers:
|
468 |
+
self.final_layer_norm = modeling_t5.T5LayerNorm(
|
469 |
+
config.d_model, eps=config.layer_norm_epsilon
|
470 |
+
)
|
471 |
+
else:
|
472 |
+
self.final_layer_norm = nn.Identity()
|
473 |
+
self.dropout = nn.Dropout(config.dropout_rate)
|
474 |
+
|
475 |
+
# Initialize weights and apply final processing
|
476 |
+
self.post_init()
|
477 |
+
# Model parallel
|
478 |
+
self.model_parallel = False
|
479 |
+
self.device_map = None
|
480 |
+
self.gradient_checkpointing = False
|
481 |
+
|
482 |
+
|
483 |
+
class DecoderOnlyT5Model(modeling_t5.T5ForConditionalGeneration):
|
484 |
+
def __init__(self, config: DecoderOnlyT5Config):
|
485 |
+
super(modeling_t5.T5ForConditionalGeneration, self).__init__(config)
|
486 |
+
self.model_dim = config.d_model
|
487 |
+
|
488 |
+
self.shared = nn.Embedding(config.vocab_size, config.d_model)
|
489 |
+
assert (
|
490 |
+
self.config.num_layers == 0
|
491 |
+
), "Decoder only model cannot have encoder layers"
|
492 |
+
self.encoder = None
|
493 |
+
|
494 |
+
decoder_config = copy.deepcopy(config)
|
495 |
+
decoder_config.is_decoder = True
|
496 |
+
decoder_config.is_encoder_decoder = False
|
497 |
+
decoder_config.num_layers = config.num_decoder_layers
|
498 |
+
self.decoder = DecoderOnlyT5Stack(decoder_config, self.shared)
|
499 |
+
|
500 |
+
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
|
501 |
+
|
502 |
+
# Initialize weights and apply final processing
|
503 |
+
self.post_init()
|
504 |
+
|
505 |
+
# Model parallel
|
506 |
+
self.model_parallel = False
|
507 |
+
self.device_map = None
|
508 |
+
|
509 |
+
@add_start_docstrings_to_model_forward(modeling_t5.T5_INPUTS_DOCSTRING)
|
510 |
+
@replace_return_docstrings(
|
511 |
+
output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
|
512 |
+
)
|
513 |
+
def forward(
|
514 |
+
self,
|
515 |
+
_input_ids: Optional[torch.LongTensor] = None,
|
516 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
517 |
+
decoder_input_ids: Optional[torch.LongTensor] = None,
|
518 |
+
decoder_attention_mask: Optional[torch.BoolTensor] = None,
|
519 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
520 |
+
decoder_head_mask: Optional[torch.FloatTensor] = None,
|
521 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
522 |
+
encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
523 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
524 |
+
_inputs_embeds: Optional[torch.FloatTensor] = None,
|
525 |
+
decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
|
526 |
+
labels: Optional[torch.LongTensor] = None,
|
527 |
+
use_cache: Optional[bool] = None,
|
528 |
+
output_attentions: Optional[bool] = None,
|
529 |
+
output_hidden_states: Optional[bool] = None,
|
530 |
+
return_dict: Optional[bool] = None,
|
531 |
+
) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
|
532 |
+
r"""
|
533 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
534 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
|
535 |
+
config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
|
536 |
+
labels in `[0, ..., config.vocab_size]`
|
537 |
+
|
538 |
+
Returns:
|
539 |
+
|
540 |
+
Examples:
|
541 |
+
|
542 |
+
```"""
|
543 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
544 |
+
return_dict = (
|
545 |
+
return_dict if return_dict is not None else self.config.use_return_dict
|
546 |
+
)
|
547 |
+
|
548 |
+
if self.model_parallel:
|
549 |
+
torch.cuda.set_device(self.decoder.first_device)
|
550 |
+
|
551 |
+
if (
|
552 |
+
labels is not None
|
553 |
+
and decoder_input_ids is None
|
554 |
+
and decoder_inputs_embeds is None
|
555 |
+
):
|
556 |
+
# get decoder inputs from shifting lm labels to the right
|
557 |
+
decoder_input_ids = self._shift_right(labels)
|
558 |
+
|
559 |
+
# Set device for model parallelism
|
560 |
+
if self.model_parallel:
|
561 |
+
torch.cuda.set_device(self.decoder.first_device)
|
562 |
+
if decoder_input_ids is not None:
|
563 |
+
decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
|
564 |
+
if attention_mask is not None:
|
565 |
+
attention_mask = attention_mask.to(self.decoder.first_device)
|
566 |
+
if decoder_attention_mask is not None:
|
567 |
+
decoder_attention_mask = decoder_attention_mask.to(
|
568 |
+
self.decoder.first_device
|
569 |
+
)
|
570 |
+
|
571 |
+
# Decode
|
572 |
+
decoder_outputs = self.decoder(
|
573 |
+
input_ids=decoder_input_ids,
|
574 |
+
attention_mask=decoder_attention_mask,
|
575 |
+
inputs_embeds=decoder_inputs_embeds,
|
576 |
+
past_key_values=past_key_values,
|
577 |
+
# encoder_hidden_states=hidden_states,
|
578 |
+
encoder_attention_mask=attention_mask,
|
579 |
+
head_mask=decoder_head_mask,
|
580 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
581 |
+
use_cache=use_cache,
|
582 |
+
output_attentions=output_attentions,
|
583 |
+
output_hidden_states=output_hidden_states,
|
584 |
+
return_dict=return_dict,
|
585 |
+
)
|
586 |
+
|
587 |
+
sequence_output = decoder_outputs[0]
|
588 |
+
|
589 |
+
# Set device for model parallelism
|
590 |
+
if self.model_parallel:
|
591 |
+
torch.cuda.set_device(self.decoder.first_device)
|
592 |
+
self.lm_head = self.lm_head.to(self.decoder.first_device)
|
593 |
+
sequence_output = sequence_output.to(self.lm_head.weight.device)
|
594 |
+
|
595 |
+
if self.config.tie_word_embeddings:
|
596 |
+
# Rescale output before projecting on vocab
|
597 |
+
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
|
598 |
+
sequence_output = sequence_output * (self.model_dim**-0.5)
|
599 |
+
|
600 |
+
lm_logits = self.lm_head(sequence_output)
|
601 |
+
|
602 |
+
loss = None
|
603 |
+
if labels is not None:
|
604 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
605 |
+
# move labels to correct device to enable PP
|
606 |
+
labels = labels.to(lm_logits.device)
|
607 |
+
loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
|
608 |
+
# TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
|
609 |
+
|
610 |
+
if not return_dict:
|
611 |
+
output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
|
612 |
+
return ((loss,) + output) if loss is not None else output
|
613 |
+
|
614 |
+
return Seq2SeqLMOutput(
|
615 |
+
loss=loss,
|
616 |
+
logits=lm_logits,
|
617 |
+
past_key_values=decoder_outputs.past_key_values,
|
618 |
+
decoder_hidden_states=decoder_outputs.hidden_states,
|
619 |
+
decoder_attentions=decoder_outputs.attentions,
|
620 |
+
)
|