Spaces:
Sleeping
Sleeping
Update ag4masses/alphageometry/models.py
Browse files- ag4masses/alphageometry/models.py +178 -178
ag4masses/alphageometry/models.py
CHANGED
@@ -1,178 +1,178 @@
|
|
1 |
-
# Copyright 2023 DeepMind Technologies Limited
|
2 |
-
#
|
3 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
-
# you may not use this file except in compliance with the License.
|
5 |
-
# You may obtain a copy of the License at
|
6 |
-
#
|
7 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
-
#
|
9 |
-
# Unless required by applicable law or agreed to in writing, software
|
10 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
-
# See the License for the specific language governing permissions and
|
13 |
-
# limitations under the License.
|
14 |
-
# ==============================================================================
|
15 |
-
|
16 |
-
"""Transformer language model generate mode."""
|
17 |
-
|
18 |
-
from typing import Any, Tuple
|
19 |
-
import beam_search
|
20 |
-
import decoder_stack
|
21 |
-
import gin
|
22 |
-
import jax
|
23 |
-
import jax.numpy as jnp
|
24 |
-
from
|
25 |
-
|
26 |
-
|
27 |
-
@gin.configurable
|
28 |
-
class DecoderOnlyLanguageModelGenerate(models.DecoderOnlyLanguageModel):
|
29 |
-
"""Decoder only language modeling in inference mode."""
|
30 |
-
|
31 |
-
decoder_factory = decoder_stack.DecoderStackGenerate
|
32 |
-
|
33 |
-
num_heads: int = gin.REQUIRED
|
34 |
-
head_size: int = gin.REQUIRED
|
35 |
-
|
36 |
-
def get_fake_input(self) -> dict[str, Any]:
|
37 |
-
fake_input_dict = super().get_fake_input()
|
38 |
-
b = self.task_config.batch_size
|
39 |
-
n = self.num_heads
|
40 |
-
h = self.head_size
|
41 |
-
fake_input_dict.update({
|
42 |
-
'dstate': tuple(
|
43 |
-
[{
|
44 |
-
'current_index': jnp.array([0] * b, dtype=jnp.int32),
|
45 |
-
'keys': jnp.zeros((b, 2048, n, h), dtype=jnp.bfloat16),
|
46 |
-
'values': jnp.zeros((b, 2048, n, h), dtype=jnp.bfloat16),
|
47 |
-
'recurrent_kvq': None,
|
48 |
-
'relative_position_bias': jnp.zeros(
|
49 |
-
(b, n, 1, 1024), dtype=jnp.bfloat16
|
50 |
-
),
|
51 |
-
}]
|
52 |
-
* 12
|
53 |
-
),
|
54 |
-
'eos': jnp.zeros([1024], dtype=jnp.bfloat16),
|
55 |
-
'mask': jnp.ones([1024], dtype=jnp.bfloat16),
|
56 |
-
'length': 1,
|
57 |
-
'temperature': 1.0,
|
58 |
-
})
|
59 |
-
return fake_input_dict
|
60 |
-
|
61 |
-
def __call__(self, inputs: ...) -> tuple[Any, dict[str, Any]]:
|
62 |
-
# Make sure this code is not used on untested cases.
|
63 |
-
if self.mode not in ['init', 'beam_search']:
|
64 |
-
raise ValueError(f'{type(self)} cannot do mode {self.mode}')
|
65 |
-
if self.decoder.supports_generate():
|
66 |
-
raise ValueError(f'{type(self)}.decoder cannot supports_generate()')
|
67 |
-
|
68 |
-
self.decoder(
|
69 |
-
input_tokens=inputs['targets'][:, 0:1],
|
70 |
-
target_tokens=None,
|
71 |
-
start_of_sequence=inputs['start_of_sequence'],
|
72 |
-
)
|
73 |
-
|
74 |
-
b = inputs['targets'].shape[0]
|
75 |
-
no_start_of_seq = jnp.array([False] * b, dtype=jnp.bool_)
|
76 |
-
|
77 |
-
# This fn is used in both beam_search or topk_sampling.
|
78 |
-
def tokens_to_logits_fn(
|
79 |
-
input_token: jnp.ndarray, dstate: tuple[dict[str, jnp.ndarray], ...]
|
80 |
-
) -> tuple[jnp.ndarray, tuple[dict[str, jnp.ndarray], ...]]:
|
81 |
-
(logits, dstate, _) = self.decoder(
|
82 |
-
input_tokens=input_token,
|
83 |
-
target_tokens=None,
|
84 |
-
start_of_sequence=no_start_of_seq,
|
85 |
-
decoder_state=dstate,
|
86 |
-
)
|
87 |
-
return logits[:, -1, :], dstate
|
88 |
-
|
89 |
-
last_token = jax.lax.dynamic_slice_in_dim(
|
90 |
-
inputs['targets'], inputs['length'] - 1, 1, axis=1
|
91 |
-
)
|
92 |
-
|
93 |
-
# last token is used to seed beam_search
|
94 |
-
inputs['targets'] = inputs['targets'][:, 0:-1]
|
95 |
-
dstate = jax.lax.cond(
|
96 |
-
inputs['start_of_sequence'][0],
|
97 |
-
lambda: self.generate(inputs)[0],
|
98 |
-
lambda: inputs['dstate'],
|
99 |
-
)
|
100 |
-
|
101 |
-
# Then we run beam search, init with last_token & dstate.
|
102 |
-
finished_seqs, finished_scores, dstate = beam_search.beam_search_flat(
|
103 |
-
last_token,
|
104 |
-
dstate,
|
105 |
-
tokens_to_logits_fn,
|
106 |
-
max_decode_len=512,
|
107 |
-
eos=inputs['eos'].reshape((1, 1, -1)),
|
108 |
-
mask=inputs['mask'].reshape((1, 1, -1)),
|
109 |
-
)
|
110 |
-
|
111 |
-
return 0.0, {
|
112 |
-
'finished_seqs': finished_seqs,
|
113 |
-
'finished_scores': finished_scores,
|
114 |
-
'dstate': dstate,
|
115 |
-
}
|
116 |
-
|
117 |
-
def generate(
|
118 |
-
self, inputs: ...
|
119 |
-
) -> tuple[tuple[dict[str, jnp.ndarray, ...], ...], jnp.ndarray]:
|
120 |
-
"""Generate an output sequence.
|
121 |
-
|
122 |
-
Args:
|
123 |
-
inputs: the same as argument to _call_.
|
124 |
-
|
125 |
-
Returns:
|
126 |
-
An array of generated tokens of shape (batch_size, sequence_length).
|
127 |
-
"""
|
128 |
-
input_tokens = inputs['targets'] # [b,seq_len]
|
129 |
-
start_of_sequence = inputs['start_of_sequence'] # [b]
|
130 |
-
target_tokens = jnp.pad(input_tokens[:, 1:], [(0, 0), (0, 1)])
|
131 |
-
batch_size = target_tokens.shape[0]
|
132 |
-
|
133 |
-
# Assuming all sequences start at the same time.
|
134 |
-
start0 = inputs['start_of_sequence'][0]
|
135 |
-
dstate = jax.lax.cond(
|
136 |
-
start0,
|
137 |
-
lambda: self.decoder.init_decoder_state_vanilla( # pylint: disable=g-long-lambda
|
138 |
-
1024, start_of_sequence
|
139 |
-
),
|
140 |
-
lambda: inputs['dstate'],
|
141 |
-
)
|
142 |
-
|
143 |
-
first_token = input_tokens[:, 0:1]
|
144 |
-
no_start_of_seq = jnp.array([False] * batch_size, dtype=jnp.bool_)
|
145 |
-
temperature = 1
|
146 |
-
if 'temperature' in inputs:
|
147 |
-
temperature = inputs['temperature']
|
148 |
-
|
149 |
-
num_steps = inputs['length']
|
150 |
-
if self.mode == 'beam_search':
|
151 |
-
num_steps -= 1
|
152 |
-
|
153 |
-
def cond_fn(scan_state) -> jnp.bool_:
|
154 |
-
_, _, i, _ = scan_state
|
155 |
-
return i < num_steps
|
156 |
-
|
157 |
-
def loop_fn(scan_state: Any) -> Tuple[Any, Any, Any, Any]:
|
158 |
-
(dstate, input_token, i, _) = scan_state
|
159 |
-
|
160 |
-
(logits, dstate, _) = self.decoder(
|
161 |
-
input_tokens=input_token,
|
162 |
-
target_tokens=None,
|
163 |
-
start_of_sequence=no_start_of_seq,
|
164 |
-
decoder_state=dstate,
|
165 |
-
)
|
166 |
-
|
167 |
-
logits = logits / temperature
|
168 |
-
output_token = jax.lax.dynamic_slice_in_dim(target_tokens, i, 1, axis=1)
|
169 |
-
|
170 |
-
return (dstate, output_token, i + 1, logits)
|
171 |
-
|
172 |
-
# Scan over the sequence length.
|
173 |
-
dummy_logits = jnp.zeros((batch_size, 1, 1024))
|
174 |
-
initial_scan_state = (dstate, first_token, 0, dummy_logits)
|
175 |
-
dstate, _, _, logits = jax.lax.while_loop(
|
176 |
-
cond_fn, loop_fn, initial_scan_state
|
177 |
-
)
|
178 |
-
return dstate, logits
|
|
|
1 |
+
# Copyright 2023 DeepMind Technologies Limited
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
# ==============================================================================
|
15 |
+
|
16 |
+
"""Transformer language model generate mode."""
|
17 |
+
|
18 |
+
from typing import Any, Tuple
|
19 |
+
import beam_search
|
20 |
+
import decoder_stack
|
21 |
+
import gin
|
22 |
+
import jax
|
23 |
+
import jax.numpy as jnp
|
24 |
+
from aglib.meliad.transformer import models
|
25 |
+
|
26 |
+
|
27 |
+
@gin.configurable
|
28 |
+
class DecoderOnlyLanguageModelGenerate(models.DecoderOnlyLanguageModel):
|
29 |
+
"""Decoder only language modeling in inference mode."""
|
30 |
+
|
31 |
+
decoder_factory = decoder_stack.DecoderStackGenerate
|
32 |
+
|
33 |
+
num_heads: int = gin.REQUIRED
|
34 |
+
head_size: int = gin.REQUIRED
|
35 |
+
|
36 |
+
def get_fake_input(self) -> dict[str, Any]:
|
37 |
+
fake_input_dict = super().get_fake_input()
|
38 |
+
b = self.task_config.batch_size
|
39 |
+
n = self.num_heads
|
40 |
+
h = self.head_size
|
41 |
+
fake_input_dict.update({
|
42 |
+
'dstate': tuple(
|
43 |
+
[{
|
44 |
+
'current_index': jnp.array([0] * b, dtype=jnp.int32),
|
45 |
+
'keys': jnp.zeros((b, 2048, n, h), dtype=jnp.bfloat16),
|
46 |
+
'values': jnp.zeros((b, 2048, n, h), dtype=jnp.bfloat16),
|
47 |
+
'recurrent_kvq': None,
|
48 |
+
'relative_position_bias': jnp.zeros(
|
49 |
+
(b, n, 1, 1024), dtype=jnp.bfloat16
|
50 |
+
),
|
51 |
+
}]
|
52 |
+
* 12
|
53 |
+
),
|
54 |
+
'eos': jnp.zeros([1024], dtype=jnp.bfloat16),
|
55 |
+
'mask': jnp.ones([1024], dtype=jnp.bfloat16),
|
56 |
+
'length': 1,
|
57 |
+
'temperature': 1.0,
|
58 |
+
})
|
59 |
+
return fake_input_dict
|
60 |
+
|
61 |
+
def __call__(self, inputs: ...) -> tuple[Any, dict[str, Any]]:
|
62 |
+
# Make sure this code is not used on untested cases.
|
63 |
+
if self.mode not in ['init', 'beam_search']:
|
64 |
+
raise ValueError(f'{type(self)} cannot do mode {self.mode}')
|
65 |
+
if self.decoder.supports_generate():
|
66 |
+
raise ValueError(f'{type(self)}.decoder cannot supports_generate()')
|
67 |
+
|
68 |
+
self.decoder(
|
69 |
+
input_tokens=inputs['targets'][:, 0:1],
|
70 |
+
target_tokens=None,
|
71 |
+
start_of_sequence=inputs['start_of_sequence'],
|
72 |
+
)
|
73 |
+
|
74 |
+
b = inputs['targets'].shape[0]
|
75 |
+
no_start_of_seq = jnp.array([False] * b, dtype=jnp.bool_)
|
76 |
+
|
77 |
+
# This fn is used in both beam_search or topk_sampling.
|
78 |
+
def tokens_to_logits_fn(
|
79 |
+
input_token: jnp.ndarray, dstate: tuple[dict[str, jnp.ndarray], ...]
|
80 |
+
) -> tuple[jnp.ndarray, tuple[dict[str, jnp.ndarray], ...]]:
|
81 |
+
(logits, dstate, _) = self.decoder(
|
82 |
+
input_tokens=input_token,
|
83 |
+
target_tokens=None,
|
84 |
+
start_of_sequence=no_start_of_seq,
|
85 |
+
decoder_state=dstate,
|
86 |
+
)
|
87 |
+
return logits[:, -1, :], dstate
|
88 |
+
|
89 |
+
last_token = jax.lax.dynamic_slice_in_dim(
|
90 |
+
inputs['targets'], inputs['length'] - 1, 1, axis=1
|
91 |
+
)
|
92 |
+
|
93 |
+
# last token is used to seed beam_search
|
94 |
+
inputs['targets'] = inputs['targets'][:, 0:-1]
|
95 |
+
dstate = jax.lax.cond(
|
96 |
+
inputs['start_of_sequence'][0],
|
97 |
+
lambda: self.generate(inputs)[0],
|
98 |
+
lambda: inputs['dstate'],
|
99 |
+
)
|
100 |
+
|
101 |
+
# Then we run beam search, init with last_token & dstate.
|
102 |
+
finished_seqs, finished_scores, dstate = beam_search.beam_search_flat(
|
103 |
+
last_token,
|
104 |
+
dstate,
|
105 |
+
tokens_to_logits_fn,
|
106 |
+
max_decode_len=512,
|
107 |
+
eos=inputs['eos'].reshape((1, 1, -1)),
|
108 |
+
mask=inputs['mask'].reshape((1, 1, -1)),
|
109 |
+
)
|
110 |
+
|
111 |
+
return 0.0, {
|
112 |
+
'finished_seqs': finished_seqs,
|
113 |
+
'finished_scores': finished_scores,
|
114 |
+
'dstate': dstate,
|
115 |
+
}
|
116 |
+
|
117 |
+
def generate(
|
118 |
+
self, inputs: ...
|
119 |
+
) -> tuple[tuple[dict[str, jnp.ndarray, ...], ...], jnp.ndarray]:
|
120 |
+
"""Generate an output sequence.
|
121 |
+
|
122 |
+
Args:
|
123 |
+
inputs: the same as argument to _call_.
|
124 |
+
|
125 |
+
Returns:
|
126 |
+
An array of generated tokens of shape (batch_size, sequence_length).
|
127 |
+
"""
|
128 |
+
input_tokens = inputs['targets'] # [b,seq_len]
|
129 |
+
start_of_sequence = inputs['start_of_sequence'] # [b]
|
130 |
+
target_tokens = jnp.pad(input_tokens[:, 1:], [(0, 0), (0, 1)])
|
131 |
+
batch_size = target_tokens.shape[0]
|
132 |
+
|
133 |
+
# Assuming all sequences start at the same time.
|
134 |
+
start0 = inputs['start_of_sequence'][0]
|
135 |
+
dstate = jax.lax.cond(
|
136 |
+
start0,
|
137 |
+
lambda: self.decoder.init_decoder_state_vanilla( # pylint: disable=g-long-lambda
|
138 |
+
1024, start_of_sequence
|
139 |
+
),
|
140 |
+
lambda: inputs['dstate'],
|
141 |
+
)
|
142 |
+
|
143 |
+
first_token = input_tokens[:, 0:1]
|
144 |
+
no_start_of_seq = jnp.array([False] * batch_size, dtype=jnp.bool_)
|
145 |
+
temperature = 1
|
146 |
+
if 'temperature' in inputs:
|
147 |
+
temperature = inputs['temperature']
|
148 |
+
|
149 |
+
num_steps = inputs['length']
|
150 |
+
if self.mode == 'beam_search':
|
151 |
+
num_steps -= 1
|
152 |
+
|
153 |
+
def cond_fn(scan_state) -> jnp.bool_:
|
154 |
+
_, _, i, _ = scan_state
|
155 |
+
return i < num_steps
|
156 |
+
|
157 |
+
def loop_fn(scan_state: Any) -> Tuple[Any, Any, Any, Any]:
|
158 |
+
(dstate, input_token, i, _) = scan_state
|
159 |
+
|
160 |
+
(logits, dstate, _) = self.decoder(
|
161 |
+
input_tokens=input_token,
|
162 |
+
target_tokens=None,
|
163 |
+
start_of_sequence=no_start_of_seq,
|
164 |
+
decoder_state=dstate,
|
165 |
+
)
|
166 |
+
|
167 |
+
logits = logits / temperature
|
168 |
+
output_token = jax.lax.dynamic_slice_in_dim(target_tokens, i, 1, axis=1)
|
169 |
+
|
170 |
+
return (dstate, output_token, i + 1, logits)
|
171 |
+
|
172 |
+
# Scan over the sequence length.
|
173 |
+
dummy_logits = jnp.zeros((batch_size, 1, 1024))
|
174 |
+
initial_scan_state = (dstate, first_token, 0, dummy_logits)
|
175 |
+
dstate, _, _, logits = jax.lax.while_loop(
|
176 |
+
cond_fn, loop_fn, initial_scan_state
|
177 |
+
)
|
178 |
+
return dstate, logits
|