add custom onnx export config
#46
by
ltcs15
- opened
- configuration_chatglm.py +310 -0
configuration_chatglm.py
CHANGED
@@ -1,8 +1,15 @@
|
|
1 |
""" ChatGLM model configuration """
|
|
|
|
|
|
|
|
|
2 |
|
3 |
from transformers.configuration_utils import PretrainedConfig
|
4 |
from transformers.utils import logging
|
5 |
|
|
|
|
|
|
|
6 |
logger = logging.get_logger(__name__)
|
7 |
|
8 |
|
@@ -101,3 +108,306 @@ class ChatGLMConfig(PretrainedConfig):
|
|
101 |
eos_token_id=eos_token_id,
|
102 |
**kwargs
|
103 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
""" ChatGLM model configuration """
|
2 |
+
import torch
|
3 |
+
|
4 |
+
from collections import OrderedDict
|
5 |
+
from typing import List, Mapping, Optional, Any
|
6 |
|
7 |
from transformers.configuration_utils import PretrainedConfig
|
8 |
from transformers.utils import logging
|
9 |
|
10 |
+
from transformers.onnx import OnnxConfigWithPast, PatchingSpec
|
11 |
+
from transformers import PreTrainedTokenizer, TensorType, is_torch_available
|
12 |
+
|
13 |
logger = logging.get_logger(__name__)
|
14 |
|
15 |
|
|
|
108 |
eos_token_id=eos_token_id,
|
109 |
**kwargs
|
110 |
)
|
111 |
+
|
112 |
+
|
113 |
+
class ChatGLMOnnxConfig(OnnxConfigWithPast):
|
114 |
+
r"""
|
115 |
+
This class is the custom configuration of a ChatGLMModel needed in exporting model to ONNX.
|
116 |
+
Currently this need to pre-fix several model struct in modeling_chatglm.py
|
117 |
+
|
118 |
+
Also there is still a TODO list of current ChatGLMOnnxConfig:
|
119 |
+
1. add support for batch_size > 1
|
120 |
+
2. add support for use_past
|
121 |
+
|
122 |
+
in modeling_chatglm.py and its attention_fn function,we need to change several view into
|
123 |
+
torch tensor action since reshape param may get frozen into constant in onnx model.
|
124 |
+
here is the code:
|
125 |
+
```python
|
126 |
+
>>> def attention_fn(
|
127 |
+
>>> self,
|
128 |
+
>>> query_layer,
|
129 |
+
>>> key_layer,
|
130 |
+
>>> value_layer,
|
131 |
+
>>> attention_mask,
|
132 |
+
>>> hidden_size_per_partition,
|
133 |
+
>>> layer_id,
|
134 |
+
>>> layer_past=None,
|
135 |
+
>>> scaling_attention_score=True,
|
136 |
+
>>> use_cache=False,
|
137 |
+
>>> ):
|
138 |
+
>>> if layer_past is not None:
|
139 |
+
>>> past_key, past_value = layer_past[0], layer_past[1]
|
140 |
+
>>> key_layer = torch.cat((past_key, key_layer), dim=0)
|
141 |
+
>>> value_layer = torch.cat((past_value, value_layer), dim=0)
|
142 |
+
>>>
|
143 |
+
>>> # seqlen, batch, num_attention_heads, hidden_size_per_attention_head
|
144 |
+
>>> seq_len, b, nh, hidden_size = key_layer.shape
|
145 |
+
>>>
|
146 |
+
>>> if use_cache:
|
147 |
+
>>> present = (key_layer, value_layer)
|
148 |
+
>>> else:
|
149 |
+
>>> present = None
|
150 |
+
>>>
|
151 |
+
>>> query_key_layer_scaling_coeff = float(layer_id + 1)
|
152 |
+
>>> if scaling_attention_score:
|
153 |
+
>>> query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff)
|
154 |
+
>>>
|
155 |
+
>>> # ===================================
|
156 |
+
>>> # Raw attention scores. [b, np, s, s]
|
157 |
+
>>> # ===================================
|
158 |
+
>>>
|
159 |
+
>>> # [b, np, sq, sk]
|
160 |
+
>>> # # output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
|
161 |
+
>>>
|
162 |
+
>>> # [sq, b, np, hn] -> [sq, b * np, hn]
|
163 |
+
>>> # query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
|
164 |
+
>>> query_layer = query_layer.flatten(start_dim=1, end_dim=2)
|
165 |
+
>>>
|
166 |
+
>>> # [sk, b, np, hn] -> [sk, b * np, hn]
|
167 |
+
>>> # key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
|
168 |
+
>>> key_layer = key_layer.flatten(start_dim=1, end_dim=2)
|
169 |
+
>>>
|
170 |
+
>>> matmul_result = torch.zeros(
|
171 |
+
>>> 1, 1, 1,
|
172 |
+
>>> dtype=query_layer.dtype,
|
173 |
+
>>> device=query_layer.device,
|
174 |
+
>>> )
|
175 |
+
>>>
|
176 |
+
>>> matmul_result = torch.baddbmm(
|
177 |
+
>>> matmul_result,
|
178 |
+
>>> query_layer.transpose(0, 1), # [b * np, sq, hn]
|
179 |
+
>>> key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
|
180 |
+
>>> beta=0.0,
|
181 |
+
>>> alpha=1.0,
|
182 |
+
>>> )
|
183 |
+
>>>
|
184 |
+
>>> # [b * np, sq, sk] -> [b, np, sq, sk]
|
185 |
+
>>> # attention_scores = matmul_result.view(*output_size)
|
186 |
+
>>> attention_scores = matmul_result.unsqueeze(0)
|
187 |
+
>>>
|
188 |
+
>>> if self.scale_mask_softmax:
|
189 |
+
>>> self.scale_mask_softmax.scale = query_key_layer_scaling_coeff
|
190 |
+
>>> attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous())
|
191 |
+
>>> else:
|
192 |
+
>>> # if not (attention_mask == 0).all():
|
193 |
+
>>> # # if auto-regressive, skip
|
194 |
+
>>> attention_scores.masked_fill_(attention_mask, -10000.0)
|
195 |
+
>>> dtype = attention_scores.dtype
|
196 |
+
>>> attention_scores = attention_scores.float()
|
197 |
+
>>> attention_scores = attention_scores * query_key_layer_scaling_coeff
|
198 |
+
>>>
|
199 |
+
>>> attention_probs = F.softmax(attention_scores, dim=-1)
|
200 |
+
>>>
|
201 |
+
>>> attention_probs = attention_probs.type(dtype)
|
202 |
+
>>>
|
203 |
+
>>> # =========================
|
204 |
+
>>> # Context layer. [sq, b, hp]
|
205 |
+
>>> # =========================
|
206 |
+
>>>
|
207 |
+
>>> # value_layer -> context layer.
|
208 |
+
>>> # [sk, b, np, hn] --> [b, np, sq, hn]
|
209 |
+
>>>
|
210 |
+
>>> # context layer shape: [b, np, sq, hn]
|
211 |
+
>>> # output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
|
212 |
+
>>>
|
213 |
+
>>> # change view [sk, b * np, hn]
|
214 |
+
>>> # value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
|
215 |
+
>>> value_layer = value_layer.flatten(start_dim=1, end_dim=2)
|
216 |
+
>>>
|
217 |
+
>>> # change view [b * np, sq, sk]
|
218 |
+
>>> # attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
|
219 |
+
>>> attention_probs = attention_probs.flatten(start_dim=0, end_dim=1)
|
220 |
+
>>>
|
221 |
+
>>> # matmul: [b * np, sq, hn]
|
222 |
+
>>> context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
|
223 |
+
>>>
|
224 |
+
>>> # change view [b, np, sq, hn]
|
225 |
+
>>> # context_layer = context_layer.reshape(b, np, sq, hidden_size)
|
226 |
+
>>> context_layer = context_layer.unsqueeze(0)
|
227 |
+
>>>
|
228 |
+
>>> # [b, np, sq, hn] --> [sq, b, np, hn]
|
229 |
+
>>> context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
|
230 |
+
>>>
|
231 |
+
>>> # [sq, b, np, hn] --> [sq, b, hp]
|
232 |
+
>>> # new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,)
|
233 |
+
>>> # context_layer = context_layer.view(*new_context_layer_shape)
|
234 |
+
>>> context_layer = context_layer.flatten(start_dim=2)
|
235 |
+
>>>
|
236 |
+
>>> outputs = (context_layer, present, attention_probs)
|
237 |
+
>>>
|
238 |
+
>>> return outputs
|
239 |
+
'''
|
240 |
+
mainly aviod using view with dynamic size
|
241 |
+
|
242 |
+
after change the modeling_chatglm.py, you can simply use following code to export and test the onnx model
|
243 |
+
```python
|
244 |
+
>>> from pathlib import Path
|
245 |
+
>>> from transformers import AutoTokenizer, AutoModel
|
246 |
+
>>> from transformers.onnx import export, validate_model_outputs
|
247 |
+
>>>
|
248 |
+
>>> # load model
|
249 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
|
250 |
+
>>> pt_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
|
251 |
+
>>> pt_model = pt_model.float() # only tested in CPU for now
|
252 |
+
>>> pt_model.eval()
|
253 |
+
>>> # define path for saving onnx model
|
254 |
+
>>> onnx_path = Path(f"model/chatglm-6b.onnx")
|
255 |
+
>>> onnx_path.parent.mkdir(exist_ok=True)
|
256 |
+
>>> # convert model to onnx
|
257 |
+
>>> onnx_config_chatglm = ChatGLMOnnxConfig(pt_model.config, task="causal-lm")
|
258 |
+
>>> onnx_inputs, onnx_outputs = export(tokenizer, pt_model,
|
259 |
+
>>> onnx_config_chatglm, onnx_config_chatglm.default_onnx_opset,
|
260 |
+
>>> onnx_path)
|
261 |
+
>>> # test onnx model
|
262 |
+
>>> validate_model_outputs(onnx_config_chatglm, tokenizer, pt_model, onnx_path, onnx_outputs, atol=1e-4)
|
263 |
+
```
|
264 |
+
"""
|
265 |
+
# TODO support dynamic batch size
|
266 |
+
default_fixed_batch = 1
|
267 |
+
|
268 |
+
def __init__(
|
269 |
+
self,
|
270 |
+
config: PretrainedConfig,
|
271 |
+
task: str = "default",
|
272 |
+
patching_specs: List[PatchingSpec] = None,
|
273 |
+
use_past: bool = False,
|
274 |
+
):
|
275 |
+
super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)
|
276 |
+
|
277 |
+
@property
|
278 |
+
def inputs(self) -> Mapping[str, Mapping[int, str]]:
|
279 |
+
common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})
|
280 |
+
if self.use_past:
|
281 |
+
# TODO support use_past
|
282 |
+
# self.fill_with_past_key_values_(common_inputs, direction="inputs")
|
283 |
+
# common_inputs["attention_mask"] = \
|
284 |
+
# {0: "batch", 1: "past_sequence + sequence", 2: "past_sequence + sequence"}
|
285 |
+
raise NotImplementedError('position_ids do not support past_key_values yet.')
|
286 |
+
else:
|
287 |
+
# remind the order
|
288 |
+
common_inputs["position_ids"] = {0: "batch", 2: "sequence"}
|
289 |
+
common_inputs["attention_mask"] = {0: "batch", 2: "sequence", 3: "sequence"}
|
290 |
+
|
291 |
+
return common_inputs
|
292 |
+
|
293 |
+
@property
|
294 |
+
def num_layers(self) -> int:
|
295 |
+
return self._config.n_layer
|
296 |
+
|
297 |
+
@property
|
298 |
+
def num_attention_heads(self) -> int:
|
299 |
+
return self._config.n_head
|
300 |
+
|
301 |
+
def get_masks(self, input_ids, device=None):
|
302 |
+
"""
|
303 |
+
reference from modeling_chatglm.get_masks
|
304 |
+
"""
|
305 |
+
batch_size, seq_length = input_ids.shape
|
306 |
+
context_lengths = [seq.tolist().index(self._config.bos_token_id) for seq in input_ids]
|
307 |
+
if device:
|
308 |
+
attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
|
309 |
+
else:
|
310 |
+
attention_mask = torch.ones((batch_size, seq_length, seq_length), device=input_ids.device)
|
311 |
+
attention_mask.tril_()
|
312 |
+
for i, context_length in enumerate(context_lengths):
|
313 |
+
attention_mask[i, :, :context_length] = 1
|
314 |
+
attention_mask.unsqueeze_(1)
|
315 |
+
attention_mask = (attention_mask < 0.5).bool()
|
316 |
+
|
317 |
+
# print("attention_mask", attention_mask.shape)
|
318 |
+
return attention_mask
|
319 |
+
|
320 |
+
def get_position_ids(self, input_ids, mask_positions, device=None, use_gmasks=None):
|
321 |
+
batch_size, seq_length = input_ids.shape
|
322 |
+
if device is None:
|
323 |
+
device = input_ids.device
|
324 |
+
if use_gmasks is None:
|
325 |
+
use_gmasks = [False] * batch_size
|
326 |
+
context_lengths = [seq.tolist().index(self._config.bos_token_id) for seq in input_ids]
|
327 |
+
if self._config.position_encoding_2d:
|
328 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
329 |
+
for i, context_length in enumerate(context_lengths):
|
330 |
+
position_ids[i, context_length:] = mask_positions[i]
|
331 |
+
block_position_ids = [torch.cat((
|
332 |
+
torch.zeros(context_length, dtype=torch.long, device=device),
|
333 |
+
torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
|
334 |
+
)) for context_length in context_lengths]
|
335 |
+
block_position_ids = torch.stack(block_position_ids, dim=0)
|
336 |
+
position_ids = torch.stack((position_ids, block_position_ids), dim=1)
|
337 |
+
else:
|
338 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
339 |
+
for i, context_length in enumerate(context_lengths):
|
340 |
+
if not use_gmasks[i]:
|
341 |
+
position_ids[context_length:] = mask_positions[i]
|
342 |
+
|
343 |
+
# print("position_ids", position_ids.shape)
|
344 |
+
return position_ids
|
345 |
+
|
346 |
+
def generate_dummy_inputs(
|
347 |
+
self,
|
348 |
+
tokenizer: PreTrainedTokenizer,
|
349 |
+
batch_size: int = default_fixed_batch,
|
350 |
+
seq_length: int = -1,
|
351 |
+
is_pair: bool = False,
|
352 |
+
framework: Optional[TensorType] = None,
|
353 |
+
) -> Mapping[str, Any]:
|
354 |
+
common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(
|
355 |
+
tokenizer, batch_size=self.default_fixed_batch, seq_length=seq_length, is_pair=is_pair, framework=framework
|
356 |
+
)
|
357 |
+
# check if the mode is using fixed batch size
|
358 |
+
if batch_size != self.default_fixed_batch:
|
359 |
+
logger.warning('batch size is not fixed, force change into fixed batch size: %d.'
|
360 |
+
% self.default_fixed_batch)
|
361 |
+
|
362 |
+
# We need to order the input in the way they appears in the forward()
|
363 |
+
ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
|
364 |
+
|
365 |
+
# Need to add the past_keys
|
366 |
+
if self.use_past:
|
367 |
+
if not is_torch_available():
|
368 |
+
raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
|
369 |
+
else:
|
370 |
+
# TODO support use_past
|
371 |
+
# import torch
|
372 |
+
#
|
373 |
+
# batch, seqlen = common_inputs["input_ids"].shape
|
374 |
+
# # Not using the same length for past_key_values
|
375 |
+
# past_key_values_length = seqlen + 2
|
376 |
+
# past_shape = (
|
377 |
+
# batch,
|
378 |
+
# self.num_attention_heads,
|
379 |
+
# past_key_values_length,
|
380 |
+
# self._config.hidden_size // self.num_attention_heads,
|
381 |
+
# )
|
382 |
+
# ordered_inputs["past_key_values"] = [
|
383 |
+
# (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)
|
384 |
+
# ]
|
385 |
+
raise NotImplementedError('position_ids do not support past_key_values yet.')
|
386 |
+
|
387 |
+
# Need to add the attention_mask manually
|
388 |
+
# 1. add attention_mask
|
389 |
+
ordered_inputs["attention_mask"] = self.get_masks(common_inputs["input_ids"])
|
390 |
+
# 2. add position_ids
|
391 |
+
MASK, gMASK = self._config.mask_token_id, self._config.gmask_token_id
|
392 |
+
seqs = common_inputs["input_ids"].tolist()
|
393 |
+
mask_positions, use_gmasks = [], []
|
394 |
+
for seq in seqs:
|
395 |
+
mask_token = gMASK if gMASK in seq else MASK
|
396 |
+
use_gmask = mask_token == gMASK
|
397 |
+
mask_positions.append(seq.index(mask_token))
|
398 |
+
use_gmasks.append(use_gmask)
|
399 |
+
ordered_inputs["position_ids"] = self.get_position_ids(common_inputs["input_ids"],
|
400 |
+
mask_positions, use_gmasks=use_gmasks)
|
401 |
+
|
402 |
+
if self.use_past:
|
403 |
+
# mask_dtype = ordered_inputs["attention_mask"].dtype
|
404 |
+
# ordered_inputs["attention_mask"] = torch.cat(
|
405 |
+
# [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
|
406 |
+
# )
|
407 |
+
raise NotImplementedError('position_ids do not support past_key_values yet.')
|
408 |
+
|
409 |
+
return ordered_inputs
|
410 |
+
|
411 |
+
@property
|
412 |
+
def default_onnx_opset(self) -> int:
|
413 |
+
return 13
|