Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -5,7 +5,13 @@ import torch
|
|
5 |
# 加载模型和分词器
|
6 |
model_name = "defog/llama-3-sqlcoder-8b" # 使用更新的模型以提高性能
|
7 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
8 |
-
model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
def generate_sql(user_question, instructions, create_table_statements):
|
11 |
prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
|
@@ -20,11 +26,24 @@ The following SQL query best answers the question `{user_question}`:
|
|
20 |
```sql
|
21 |
"""
|
22 |
|
23 |
-
inputs = tokenizer(prompt, return_tensors="pt").to(
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
|
30 |
question = f"What are our top 3 products by revenue in the New York region?"
|
|
|
5 |
# 加载模型和分词器
|
6 |
model_name = "defog/llama-3-sqlcoder-8b" # 使用更新的模型以提高性能
|
7 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
8 |
+
model = AutoModelForCausalLM.from_pretrained(
|
9 |
+
model_name,
|
10 |
+
trust_remote_code=True,
|
11 |
+
torch_dtype=torch.float16,
|
12 |
+
device_map="auto",
|
13 |
+
use_cache=True,
|
14 |
+
)
|
15 |
|
16 |
def generate_sql(user_question, instructions, create_table_statements):
|
17 |
prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
|
|
|
26 |
```sql
|
27 |
"""
|
28 |
|
29 |
+
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
|
30 |
+
generated_ids = model.generate(
|
31 |
+
**inputs,
|
32 |
+
num_return_sequences=1,
|
33 |
+
eos_token_id=tokenizer.eos_token_id,
|
34 |
+
pad_token_id=tokenizer.eos_token_id,
|
35 |
+
max_new_tokens=400,
|
36 |
+
do_sample=False,
|
37 |
+
num_beams=1,
|
38 |
+
)
|
39 |
+
outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
40 |
+
|
41 |
+
torch.cuda.empty_cache()
|
42 |
+
torch.cuda.synchronize()
|
43 |
+
# empty cache so that you do generate more results w/o memory crashing
|
44 |
+
# particularly important on Colab – memory management is much more straightforward
|
45 |
+
# when running on an inference service
|
46 |
+
return sqlparse.format(outputs[0].split("[SQL]")[-1], reindent=True)
|
47 |
|
48 |
|
49 |
question = f"What are our top 3 products by revenue in the New York region?"
|