Create infer_onnx_hf.py
Browse files- onnx/infer_onnx_hf.py +147 -0
onnx/infer_onnx_hf.py
ADDED
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import onnxruntime
|
2 |
+
import numpy as np
|
3 |
+
from transformers import AutoTokenizer
|
4 |
+
import time
|
5 |
+
import os
|
6 |
+
from huggingface_hub import hf_hub_download
|
7 |
+
|
8 |
+
model_name = "skypro1111/mbart-large-50-verbalization"
|
9 |
+
|
10 |
+
|
11 |
+
def download_model_from_hf(repo_id=model_name, model_dir="onnx_hf"):
|
12 |
+
"""Download ONNX models from HuggingFace Hub."""
|
13 |
+
os.makedirs(model_dir, exist_ok=True)
|
14 |
+
|
15 |
+
files = ["onnx/encoder_model.onnx", "onnx/decoder_model.onnx", "onnx/decoder_model.onnx_data"]
|
16 |
+
|
17 |
+
for file in files:
|
18 |
+
hf_hub_download(
|
19 |
+
repo_id=repo_id,
|
20 |
+
filename=file,
|
21 |
+
local_dir=model_dir
|
22 |
+
)
|
23 |
+
|
24 |
+
return files
|
25 |
+
|
26 |
+
def create_onnx_session(model_path, use_gpu=True):
|
27 |
+
"""Create an ONNX inference session."""
|
28 |
+
# Session options
|
29 |
+
session_options = onnxruntime.SessionOptions()
|
30 |
+
session_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
|
31 |
+
session_options.enable_mem_pattern = True
|
32 |
+
session_options.enable_mem_reuse = True
|
33 |
+
session_options.intra_op_num_threads = 8
|
34 |
+
session_options.log_severity_level = 1
|
35 |
+
|
36 |
+
cuda_provider_options = {
|
37 |
+
'device_id': 0,
|
38 |
+
'arena_extend_strategy': 'kSameAsRequested',
|
39 |
+
'gpu_mem_limit': 0, # 0 means no limit
|
40 |
+
'cudnn_conv_algo_search': 'DEFAULT',
|
41 |
+
'do_copy_in_default_stream': True,
|
42 |
+
}
|
43 |
+
|
44 |
+
print(f"Available providers: {onnxruntime.get_available_providers()}")
|
45 |
+
if use_gpu and 'CUDAExecutionProvider' in onnxruntime.get_available_providers():
|
46 |
+
providers = [('CUDAExecutionProvider', cuda_provider_options)]
|
47 |
+
print("Using CUDA for inference")
|
48 |
+
else:
|
49 |
+
providers = ['CPUExecutionProvider']
|
50 |
+
print("Using CPU for inference")
|
51 |
+
|
52 |
+
session = onnxruntime.InferenceSession(
|
53 |
+
model_path,
|
54 |
+
providers=providers,
|
55 |
+
sess_options=session_options
|
56 |
+
)
|
57 |
+
|
58 |
+
return session
|
59 |
+
|
60 |
+
def generate_text(text, tokenizer, encoder_session, decoder_session, max_length=128):
|
61 |
+
"""Generate text for a single input."""
|
62 |
+
# Prepare input
|
63 |
+
inputs = tokenizer(text, return_tensors="np", padding=True, truncation=True, max_length=512)
|
64 |
+
input_ids = inputs["input_ids"].astype(np.int64)
|
65 |
+
attention_mask = inputs["attention_mask"].astype(np.int64)
|
66 |
+
|
67 |
+
# Run encoder
|
68 |
+
encoder_outputs = encoder_session.run(
|
69 |
+
output_names=["last_hidden_state"],
|
70 |
+
input_feed={
|
71 |
+
"input_ids": input_ids,
|
72 |
+
"attention_mask": attention_mask,
|
73 |
+
}
|
74 |
+
)[0]
|
75 |
+
|
76 |
+
# Initialize decoder input
|
77 |
+
decoder_input_ids = np.array([[tokenizer.pad_token_id]], dtype=np.int64)
|
78 |
+
|
79 |
+
# Generate sequence
|
80 |
+
for _ in range(max_length):
|
81 |
+
# Run decoder
|
82 |
+
decoder_outputs = decoder_session.run(
|
83 |
+
output_names=["logits"],
|
84 |
+
input_feed={
|
85 |
+
"input_ids": decoder_input_ids,
|
86 |
+
"encoder_hidden_states": encoder_outputs,
|
87 |
+
"encoder_attention_mask": attention_mask,
|
88 |
+
}
|
89 |
+
)[0]
|
90 |
+
|
91 |
+
# Get next token
|
92 |
+
next_token = decoder_outputs[:, -1:].argmax(axis=-1)
|
93 |
+
decoder_input_ids = np.concatenate([decoder_input_ids, next_token], axis=-1)
|
94 |
+
|
95 |
+
# Check if sequence is complete
|
96 |
+
if tokenizer.eos_token_id in decoder_input_ids[0]:
|
97 |
+
break
|
98 |
+
|
99 |
+
# Decode sequence
|
100 |
+
output_text = tokenizer.decode(decoder_input_ids[0], skip_special_tokens=True)
|
101 |
+
return output_text
|
102 |
+
|
103 |
+
def main():
|
104 |
+
# Print available providers
|
105 |
+
print("Available providers:", onnxruntime.get_available_providers())
|
106 |
+
|
107 |
+
# Download models from HuggingFace
|
108 |
+
print("\nDownloading models from HuggingFace...")
|
109 |
+
encoder_path, decoder_path, _ = download_model_from_hf()
|
110 |
+
|
111 |
+
# Load tokenizer and models
|
112 |
+
print("\nLoading tokenizer...")
|
113 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
114 |
+
tokenizer.src_lang = "uk_UA"
|
115 |
+
tokenizer.tgt_lang = "uk_UA"
|
116 |
+
|
117 |
+
# Create ONNX sessions
|
118 |
+
print("\nLoading encoder...")
|
119 |
+
encoder_session = create_onnx_session(encoder_path)
|
120 |
+
print("\nLoading decoder...")
|
121 |
+
decoder_session = create_onnx_session(decoder_path)
|
122 |
+
|
123 |
+
# Test examples
|
124 |
+
test_inputs = [
|
125 |
+
"мій телефон 0979456822",
|
126 |
+
"квартира площею 11 тис кв м.",
|
127 |
+
"Пропонували хабар у 1 млрд грн.",
|
128 |
+
"1 2 3 4 5 6 7 8 9 10.",
|
129 |
+
"Крім того, парламентарій володіє шістьма ділянками землі (дві площею 25000 кв м, дві по 15000 кв м та дві по 10000 кв м) розташованими в Сосновій Балці Луганської області.",
|
130 |
+
"Підписуючи цей документ у 2003 році, голови Росії та України мали намір зміцнити співпрацю та сприяти розширенню двосторонніх відносин.",
|
131 |
+
"Очікується, що цей застосунок буде запущено 22.08.2025.",
|
132 |
+
"За інформацією від Державної служби з надзвичайних ситуацій станом на 7 ранку 15 липня.",
|
133 |
+
]
|
134 |
+
|
135 |
+
print("\nWarming up...")
|
136 |
+
_ = generate_text(test_inputs[0], tokenizer, encoder_session, decoder_session)
|
137 |
+
|
138 |
+
print("\nRunning inference...")
|
139 |
+
for text in test_inputs:
|
140 |
+
print(f"\nInput: {text}")
|
141 |
+
t = time.time()
|
142 |
+
output = generate_text(text, tokenizer, encoder_session, decoder_session)
|
143 |
+
print(f"Output: {output}")
|
144 |
+
print(f"Time: {time.time() - t:.2f} seconds")
|
145 |
+
|
146 |
+
if __name__ == "__main__":
|
147 |
+
main()
|