Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -90,32 +90,34 @@ def generate_response(model_name, system_instruction, user_input):
|
|
90 |
prompt = f"""### Instruction:
|
91 |
{system_instruction}
|
92 |
Remember to ALWAYS format your response as valid JSON.
|
93 |
-
|
94 |
### Input:
|
95 |
{user_input}
|
96 |
-
|
97 |
### Response:
|
98 |
{{""" # Note the opening curly brace to hint JSON response
|
99 |
|
100 |
-
inputs = model_manager.current_tokenizer([prompt], return_tensors="pt").to(model_manager.device)
|
101 |
-
|
102 |
-
# Generation configuration optimized for JSON output
|
103 |
-
meta_config = {
|
104 |
-
"do_sample": False,
|
105 |
-
"temperature": 0.0,
|
106 |
-
"max_new_tokens": 512,
|
107 |
-
"repetition_penalty": 1.1,
|
108 |
-
"use_cache": True,
|
109 |
-
"pad_token_id": model_manager.current_tokenizer.eos_token_id,
|
110 |
-
"eos_token_id": model_manager.current_tokenizer.eos_token_id
|
111 |
-
}
|
112 |
-
generation_config = GenerationConfig(**meta_config)
|
113 |
-
|
114 |
-
# Generate response
|
115 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
with torch.no_grad():
|
117 |
outputs = model_manager.current_model.generate(
|
118 |
-
|
|
|
119 |
generation_config=generation_config
|
120 |
)
|
121 |
decoded_output = model_manager.current_tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
|
|
90 |
prompt = f"""### Instruction:
|
91 |
{system_instruction}
|
92 |
Remember to ALWAYS format your response as valid JSON.
|
|
|
93 |
### Input:
|
94 |
{user_input}
|
|
|
95 |
### Response:
|
96 |
{{""" # Note the opening curly brace to hint JSON response
|
97 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
try:
|
99 |
+
# Ensure inputs are on the correct device
|
100 |
+
inputs = model_manager.current_tokenizer([prompt], return_tensors="pt")
|
101 |
+
# Move input_ids and attention_mask to the same device as the model
|
102 |
+
inputs = {k: v.to(model_manager.device) for k, v in inputs.items()}
|
103 |
+
|
104 |
+
# Generation configuration optimized for JSON output
|
105 |
+
meta_config = {
|
106 |
+
"do_sample": False,
|
107 |
+
"temperature": 0.0,
|
108 |
+
"max_new_tokens": 512,
|
109 |
+
"repetition_penalty": 1.2,
|
110 |
+
"use_cache": True,
|
111 |
+
"pad_token_id": model_manager.current_tokenizer.eos_token_id,
|
112 |
+
"eos_token_id": model_manager.current_tokenizer.eos_token_id
|
113 |
+
}
|
114 |
+
generation_config = GenerationConfig(**meta_config)
|
115 |
+
|
116 |
+
# Generate response
|
117 |
with torch.no_grad():
|
118 |
outputs = model_manager.current_model.generate(
|
119 |
+
input_ids=inputs['input_ids'],
|
120 |
+
attention_mask=inputs['attention_mask'],
|
121 |
generation_config=generation_config
|
122 |
)
|
123 |
decoded_output = model_manager.current_tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|