Upload ONNX export script
Browse files
export.py
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn.functional import scaled_dot_product_attention
|
6 |
+
|
7 |
+
from transformers import (
|
8 |
+
PreTrainedModel,
|
9 |
+
PretrainedConfig,
|
10 |
+
)
|
11 |
+
from transformers.modeling_outputs import BaseModelOutput
|
12 |
+
|
13 |
+
from xformers.ops import SwiGLU
|
14 |
+
|
15 |
+
|
16 |
+
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
|
17 |
+
"""
|
18 |
+
Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
|
19 |
+
|
20 |
+
This function calculates a frequency tensor with complex exponentials using the given dimension 'dim'
|
21 |
+
and the end index 'end'. The 'theta' parameter scales the frequencies.
|
22 |
+
The returned tensor contains complex values in complex64 data type.
|
23 |
+
|
24 |
+
Adapted from https://github.com/facebookresearch/llama/blob/main/llama/model.py.
|
25 |
+
|
26 |
+
Args:
|
27 |
+
dim (int): Dimension of the frequency tensor.
|
28 |
+
end (int): End index for precomputing frequencies.
|
29 |
+
theta (float, optional): Scaling factor for frequency computation. Defaults to 10000.0.
|
30 |
+
|
31 |
+
Returns:
|
32 |
+
torch.Tensor: Precomputed frequency tensor with complex exponentials.
|
33 |
+
"""
|
34 |
+
|
35 |
+
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
36 |
+
t = torch.arange(end, device=freqs.device)
|
37 |
+
freqs = torch.outer(t, freqs).float()
|
38 |
+
return torch.polar(torch.ones_like(freqs), freqs)
|
39 |
+
|
40 |
+
|
41 |
+
def apply_rotary_emb_real(
|
42 |
+
xq: torch.Tensor,
|
43 |
+
xk: torch.Tensor,
|
44 |
+
freqs_cis: Tuple[torch.Tensor, torch.Tensor],
|
45 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
46 |
+
"""
|
47 |
+
Pure-real rotary embeddings.
|
48 |
+
|
49 |
+
xq, xk: (B, seq, n_heads, dim)
|
50 |
+
freqs_cis: (cos, sin), each of shape (B, seq, dim/2)
|
51 |
+
"""
|
52 |
+
cos, sin = freqs_cis
|
53 |
+
# make (B, seq, 1, dim/2) so they broadcast to (B, seq, n_heads, dim/2)
|
54 |
+
cos = cos.unsqueeze(2)
|
55 |
+
sin = sin.unsqueeze(2)
|
56 |
+
|
57 |
+
# split even/odd dims
|
58 |
+
xq_even = xq[..., 0::2]
|
59 |
+
xq_odd = xq[..., 1::2]
|
60 |
+
xk_even = xk[..., 0::2]
|
61 |
+
xk_odd = xk[..., 1::2]
|
62 |
+
|
63 |
+
# apply the rotation formula:
|
64 |
+
q_rot_even = xq_even * cos - xq_odd * sin
|
65 |
+
q_rot_odd = xq_even * sin + xq_odd * cos
|
66 |
+
k_rot_even = xk_even * cos - xk_odd * sin
|
67 |
+
k_rot_odd = xk_even * sin + xk_odd * cos
|
68 |
+
|
69 |
+
# interleave even/odd back into last dim
|
70 |
+
xq_rot = torch.stack([q_rot_even, q_rot_odd], dim=-1).flatten(-2)
|
71 |
+
xk_rot = torch.stack([k_rot_even, k_rot_odd], dim=-1).flatten(-2)
|
72 |
+
|
73 |
+
return xq_rot.type_as(xq), xk_rot.type_as(xk)
|
74 |
+
|
75 |
+
|
76 |
+
class NeoBERTConfig(PretrainedConfig):
|
77 |
+
model_type = "neobert"
|
78 |
+
|
79 |
+
# All config parameters must have a default value.
|
80 |
+
def __init__(
|
81 |
+
self,
|
82 |
+
hidden_size: int = 768,
|
83 |
+
num_hidden_layers: int = 28,
|
84 |
+
num_attention_heads: int = 12,
|
85 |
+
intermediate_size: int = 3072,
|
86 |
+
embedding_init_range: float = 0.02,
|
87 |
+
decoder_init_range: float = 0.02,
|
88 |
+
norm_eps: float = 1e-06,
|
89 |
+
vocab_size: int = 30522,
|
90 |
+
pad_token_id: int = 0,
|
91 |
+
max_length: int = 1024,
|
92 |
+
**kwargs,
|
93 |
+
):
|
94 |
+
super().__init__(**kwargs)
|
95 |
+
|
96 |
+
self.hidden_size = hidden_size
|
97 |
+
self.num_hidden_layers = num_hidden_layers
|
98 |
+
self.num_attention_heads = num_attention_heads
|
99 |
+
if hidden_size % num_attention_heads != 0:
|
100 |
+
raise ValueError("Hidden size must be divisible by the number of heads.")
|
101 |
+
self.dim_head = hidden_size // num_attention_heads
|
102 |
+
self.intermediate_size = intermediate_size
|
103 |
+
self.embedding_init_range = embedding_init_range
|
104 |
+
self.decoder_init_range = decoder_init_range
|
105 |
+
self.norm_eps = norm_eps
|
106 |
+
self.vocab_size = vocab_size
|
107 |
+
self.pad_token_id = pad_token_id
|
108 |
+
self.max_length = max_length
|
109 |
+
self.kwargs = kwargs
|
110 |
+
|
111 |
+
|
112 |
+
class EncoderBlock(nn.Module):
|
113 |
+
"""Transformer encoder block."""
|
114 |
+
|
115 |
+
def __init__(self, config: NeoBERTConfig):
|
116 |
+
super().__init__()
|
117 |
+
|
118 |
+
self.config = config
|
119 |
+
|
120 |
+
# Attention
|
121 |
+
self.qkv = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size * 3, bias=False)
|
122 |
+
self.wo = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size, bias=False)
|
123 |
+
|
124 |
+
# Feedforward network
|
125 |
+
multiple_of = 8
|
126 |
+
intermediate_size = int(2 * config.intermediate_size / 3)
|
127 |
+
intermediate_size = multiple_of * ((intermediate_size + multiple_of - 1) // multiple_of)
|
128 |
+
self.ffn = SwiGLU(config.hidden_size, intermediate_size, config.hidden_size, bias=False)
|
129 |
+
|
130 |
+
# Layer norms
|
131 |
+
self.attention_norm = nn.RMSNorm(config.hidden_size, config.norm_eps)
|
132 |
+
self.ffn_norm = nn.RMSNorm(config.hidden_size, config.norm_eps)
|
133 |
+
|
134 |
+
def forward(
|
135 |
+
self,
|
136 |
+
x: torch.Tensor,
|
137 |
+
attention_mask: torch.Tensor,
|
138 |
+
freqs_cis: Tuple[torch.Tensor, torch.Tensor],
|
139 |
+
output_attentions: bool,
|
140 |
+
):
|
141 |
+
# Attention
|
142 |
+
attn_output, attn_weights = self._att_block(
|
143 |
+
self.attention_norm(x), attention_mask, freqs_cis, output_attentions,
|
144 |
+
)
|
145 |
+
|
146 |
+
# Residual
|
147 |
+
x = x + attn_output
|
148 |
+
|
149 |
+
# Feed-forward
|
150 |
+
x = x + self.ffn(self.ffn_norm(x))
|
151 |
+
|
152 |
+
return x, attn_weights
|
153 |
+
|
154 |
+
def _att_block(
|
155 |
+
self,
|
156 |
+
x: torch.Tensor,
|
157 |
+
attention_mask: torch.Tensor,
|
158 |
+
freqs_cis: Tuple[torch.Tensor, torch.Tensor],
|
159 |
+
output_attentions: bool,
|
160 |
+
):
|
161 |
+
batch_size, seq_len, _ = x.shape
|
162 |
+
|
163 |
+
xq, xk, xv = self.qkv(x).view(batch_size, seq_len, self.config.num_attention_heads, self.config.dim_head * 3).chunk(3, axis=-1)
|
164 |
+
|
165 |
+
xq, xk = apply_rotary_emb_real(xq, xk, freqs_cis)
|
166 |
+
|
167 |
+
# Attn block
|
168 |
+
attn_weights = None
|
169 |
+
|
170 |
+
# Eager attention if attention weights are needed in the output
|
171 |
+
if output_attentions:
|
172 |
+
attn_weights = xq.permute(0, 2, 1, 3) @ xk.permute(0, 2, 3, 1) / (xq.size(-1) ** 0.5)
|
173 |
+
if attention_mask is not None:
|
174 |
+
attn_weights = attn_weights * attention_mask
|
175 |
+
attn_weights = attn_weights.softmax(-1)
|
176 |
+
attn = attn_weights @ xv.permute(0, 2, 1, 3)
|
177 |
+
attn = attn.transpose(1, 2)
|
178 |
+
# Fall back to SDPA otherwise
|
179 |
+
else:
|
180 |
+
attn = scaled_dot_product_attention(
|
181 |
+
query=xq.transpose(1, 2),
|
182 |
+
key=xk.transpose(1, 2),
|
183 |
+
value=xv.transpose(1, 2),
|
184 |
+
attn_mask=attention_mask.bool(),
|
185 |
+
dropout_p=0,
|
186 |
+
).transpose(1, 2)
|
187 |
+
|
188 |
+
return self.wo(attn.reshape(batch_size, seq_len, self.config.num_attention_heads * self.config.dim_head)), attn_weights
|
189 |
+
|
190 |
+
|
191 |
+
class NeoBERTPreTrainedModel(PreTrainedModel):
|
192 |
+
config_class = NeoBERTConfig
|
193 |
+
base_model_prefix = "model"
|
194 |
+
_supports_cache_class = True
|
195 |
+
|
196 |
+
def _init_weights(self, module):
|
197 |
+
if isinstance(module, nn.Linear):
|
198 |
+
module.weight.data.uniform_(-self.config.decoder_init_range, self.config.decoder_init_range)
|
199 |
+
elif isinstance(module, nn.Embedding):
|
200 |
+
module.weight.data.uniform_(-self.config.embedding_init_range, self.config.embedding_init_range)
|
201 |
+
|
202 |
+
|
203 |
+
class NeoBERT(NeoBERTPreTrainedModel):
|
204 |
+
config_class = NeoBERTConfig
|
205 |
+
|
206 |
+
def __init__(self, config: NeoBERTConfig):
|
207 |
+
super().__init__(config)
|
208 |
+
|
209 |
+
self.config = config
|
210 |
+
|
211 |
+
self.encoder = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
|
212 |
+
|
213 |
+
# Ensures freqs_cis is moved to the same devices as the model. Non-persistent buffers are not saved in the state_dict.
|
214 |
+
freqs_cis = precompute_freqs_cis(config.hidden_size // config.num_attention_heads, config.max_length)
|
215 |
+
self.register_buffer("freqs_cos", freqs_cis.real, persistent=False)
|
216 |
+
self.register_buffer("freqs_sin", freqs_cis.imag, persistent=False)
|
217 |
+
|
218 |
+
self.transformer_encoder = nn.ModuleList()
|
219 |
+
for _ in range(config.num_hidden_layers):
|
220 |
+
self.transformer_encoder.append(EncoderBlock(config))
|
221 |
+
|
222 |
+
self.layer_norm = nn.RMSNorm(config.hidden_size, config.norm_eps)
|
223 |
+
|
224 |
+
# Initialize weights and apply final processing
|
225 |
+
self.post_init()
|
226 |
+
|
227 |
+
def forward(
|
228 |
+
self,
|
229 |
+
input_ids: Optional[torch.Tensor] = None,
|
230 |
+
attention_mask: torch.Tensor = None,
|
231 |
+
position_ids: torch.Tensor = None,
|
232 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
233 |
+
output_hidden_states: bool = False,
|
234 |
+
output_attentions: bool = False,
|
235 |
+
**kwargs,
|
236 |
+
):
|
237 |
+
# Initialize
|
238 |
+
hidden_states, attentions = [], []
|
239 |
+
|
240 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
241 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
242 |
+
|
243 |
+
# Expand and repeat: (Batch, Length) -> (Batch, Heads, Length, Length)
|
244 |
+
if attention_mask is not None:
|
245 |
+
attention_mask = attention_mask.unsqueeze(1).unsqueeze(1).repeat(1, self.config.num_attention_heads, attention_mask.size(-1), 1)
|
246 |
+
|
247 |
+
# RoPE
|
248 |
+
freqs_cos = (
|
249 |
+
self.freqs_cos[position_ids]
|
250 |
+
if position_ids is not None
|
251 |
+
else self.freqs_cos[: (input_ids if input_ids is not None else inputs_embeds).shape[1]].unsqueeze(0)
|
252 |
+
)
|
253 |
+
freqs_sin = (
|
254 |
+
self.freqs_sin[position_ids]
|
255 |
+
if position_ids is not None
|
256 |
+
else self.freqs_sin[: (input_ids if input_ids is not None else inputs_embeds).shape[1]].unsqueeze(0)
|
257 |
+
)
|
258 |
+
|
259 |
+
# Embedding
|
260 |
+
x = self.encoder(input_ids) if input_ids is not None else inputs_embeds
|
261 |
+
|
262 |
+
# Transformer encoder
|
263 |
+
for layer in self.transformer_encoder:
|
264 |
+
x, attn = layer(x, attention_mask, (freqs_cos, freqs_sin), output_attentions)
|
265 |
+
if output_hidden_states:
|
266 |
+
hidden_states.append(x)
|
267 |
+
if output_attentions:
|
268 |
+
attentions.append(attn)
|
269 |
+
|
270 |
+
# Final normalization layer
|
271 |
+
x = self.layer_norm(x)
|
272 |
+
|
273 |
+
# Return the output of the last hidden layer
|
274 |
+
return BaseModelOutput(
|
275 |
+
last_hidden_state=x,
|
276 |
+
hidden_states=hidden_states if output_hidden_states else None,
|
277 |
+
attentions=attentions if output_attentions else None,
|
278 |
+
)
|
279 |
+
|
280 |
+
if __name__ == "__main__":
|
281 |
+
from transformers import AutoTokenizer
|
282 |
+
|
283 |
+
model_name = "chandar-lab/NeoBERT"
|
284 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
285 |
+
model = NeoBERT.from_pretrained(model_name)
|
286 |
+
|
287 |
+
# Tokenize input text
|
288 |
+
text = [
|
289 |
+
"NeoBERT is the most efficient model of its kind!",
|
290 |
+
"This is really cool",
|
291 |
+
]
|
292 |
+
inputs = tokenizer(text, padding=True, return_tensors="pt")
|
293 |
+
|
294 |
+
# Generate embeddings
|
295 |
+
with torch.no_grad():
|
296 |
+
pytorch_outputs = model(**inputs)
|
297 |
+
|
298 |
+
# Export to ONNX
|
299 |
+
torch.onnx.export(
|
300 |
+
model,
|
301 |
+
(inputs['input_ids'], inputs['attention_mask']),
|
302 |
+
f="model.onnx",
|
303 |
+
export_params=True,
|
304 |
+
opset_version=20,
|
305 |
+
do_constant_folding=True,
|
306 |
+
input_names = ['input_ids', 'attention_mask'],
|
307 |
+
output_names = ['last_hidden_state'],
|
308 |
+
dynamic_axes = {
|
309 |
+
'input_ids': {0: 'batch_size', 1: 'sequence_length'},
|
310 |
+
'attention_mask': {0: 'batch_size', 1: 'sequence_length'},
|
311 |
+
'last_hidden_state': {0: 'batch_size', 1: 'sequence_length'},
|
312 |
+
},
|
313 |
+
dynamo=True,
|
314 |
+
)
|
315 |
+
|
316 |
+
# Validate
|
317 |
+
import onnxruntime as ort
|
318 |
+
ort_session = ort.InferenceSession("model.onnx")
|
319 |
+
ort_inputs = {
|
320 |
+
"input_ids": inputs['input_ids'].numpy(),
|
321 |
+
"attention_mask": inputs['attention_mask'].numpy(),
|
322 |
+
}
|
323 |
+
ort_outputs = ort_session.run(None, ort_inputs)
|
324 |
+
|
325 |
+
assert (pytorch_outputs.last_hidden_state.numpy() - ort_outputs[0]).max() < 1e-3
|