1from transformers import AutoConfig, AutoTokenizer
2import onnxruntime
3import numpy as np
4
5# 1. Load config, processor, and model
6model_id = "./path/to/model/"
7config = AutoConfig.from_pretrained(model_id)
8tokenizer = AutoTokenizer.from_pretrained(model_id)
9model_path = f"{model_id}/onnx/model.onnx"
10decoder_session = onnxruntime.InferenceSession(model_path)
11
12## Set config values
13num_key_value_heads = config.num_key_value_heads
14head_dim = config.hidden_size // config.num_attention_heads
15num_hidden_layers = config.num_hidden_layers
16eos_token_id = config.eos_token_id
17
18# 2. Prepare inputs
19messages = [{"role": "user", "content": "Explica en español qué significa la palabra japonesa 'ikigai' y da un ejemplo práctico."}]
20inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="np")
21input_ids = inputs['input_ids']
22attention_mask = inputs['attention_mask']
23batch_size = input_ids.shape[0]
24past_key_values = {
25 f'past_key_values.{layer}.{kv}': np.zeros([batch_size, num_key_value_heads, 0, head_dim], dtype=np.float32)
26 for layer in range(num_hidden_layers)
27 for kv in ('key', 'value')
28}
29
30# 3. Generation loop
31max_new_tokens = 1024
32generated_tokens = np.array([[]], dtype=np.int64)
33for i in range(max_new_tokens):
34 logits, *present_key_values = decoder_session.run(None, dict(
35 input_ids=input_ids,
36 attention_mask=attention_mask,
37 **past_key_values,
38 ))
39
40 ## Update values for next generation loop
41 input_ids = logits[:, -1].argmax(-1, keepdims=True)
42 attention_mask = np.concatenate([attention_mask, np.ones_like(input_ids, dtype=np.int64)], axis=-1)
43 for j, key in enumerate(past_key_values):
44 past_key_values[key] = present_key_values[j]
45
46 generated_tokens = np.concatenate([generated_tokens, input_ids], axis=-1)
47 if np.isin(input_ids, eos_token_id).any():
48 break
49
50 ## (Optional) Streaming
51 print(tokenizer.decode(input_ids[0]), end='', flush=True)
52print()
53
54# 4. Output result
55print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0])
