Spaces:
Paused
Paused
rodrigomasini
commited on
Commit
•
9fc5e2e
1
Parent(s):
dacf75f
Update app_v4.py
Browse files
app_v4.py
CHANGED
@@ -63,10 +63,11 @@ if model_loaded:
|
|
63 |
gpu_memory_after = get_gpu_memory()
|
64 |
st.write(f"GPU Memory Info after loading the model: {gpu_memory_after}")
|
65 |
|
66 |
-
|
67 |
-
col1
|
68 |
-
|
69 |
-
|
|
|
70 |
|
71 |
# Generate button
|
72 |
if st.button("Generate the prompt"):
|
@@ -74,15 +75,17 @@ if model_loaded:
|
|
74 |
prompt_template = f'USER: {user_input}\nASSISTANT:'
|
75 |
inputs = tokenizer(prompt_template, return_tensors='pt', max_length=512, truncation=True, padding='max_length')
|
76 |
inputs = inputs.to(device) # Move inputs to the same device as model
|
77 |
-
|
|
|
78 |
with torch.inference_mode():
|
79 |
-
output = model.generate(**inputs, max_new_tokens=max_token)
|
80 |
-
|
81 |
-
#
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
86 |
except RuntimeError as e:
|
87 |
if 'CUDA out of memory' in str(e):
|
88 |
st.error("CUDA out of memory during generation. Try reducing the input length or restarting the app.")
|
@@ -93,4 +96,8 @@ if model_loaded:
|
|
93 |
# Log the error and re-raise it
|
94 |
with open('error_log.txt', 'a') as f:
|
95 |
f.write(traceback.format_exc())
|
96 |
-
raise e
|
|
|
|
|
|
|
|
|
|
63 |
gpu_memory_after = get_gpu_memory()
|
64 |
st.write(f"GPU Memory Info after loading the model: {gpu_memory_after}")
|
65 |
|
66 |
+
col1, col2 = st.columns(2)
|
67 |
+
with col1:
|
68 |
+
user_input = st.text_input("Input a phrase")
|
69 |
+
with col2:
|
70 |
+
max_token = st.number_input(label="Select max number of generated tokens", min_value=1, max_value=1024, value=50, step=5)
|
71 |
|
72 |
# Generate button
|
73 |
if st.button("Generate the prompt"):
|
|
|
75 |
prompt_template = f'USER: {user_input}\nASSISTANT:'
|
76 |
inputs = tokenizer(prompt_template, return_tensors='pt', max_length=512, truncation=True, padding='max_length')
|
77 |
inputs = inputs.to(device) # Move inputs to the same device as model
|
78 |
+
|
79 |
+
# Generate text using torch.inference_mode for better performance during inference
|
80 |
with torch.inference_mode():
|
81 |
+
output = model.generate(**inputs, max_new_tokens=max_token, num_return_sequences=2)
|
82 |
+
|
83 |
+
# Display generated texts
|
84 |
+
for i in range(2): # Loop through the number of return sequences
|
85 |
+
output_ids_cut = output[i, inputs["input_ids"].shape[1]:]
|
86 |
+
generated_text = tokenizer.decode(output_ids_cut, skip_special_tokens=True)
|
87 |
+
st.markdown(f"**Generated Text {i+1}:**\n{generated_text}")
|
88 |
+
|
89 |
except RuntimeError as e:
|
90 |
if 'CUDA out of memory' in str(e):
|
91 |
st.error("CUDA out of memory during generation. Try reducing the input length or restarting the app.")
|
|
|
96 |
# Log the error and re-raise it
|
97 |
with open('error_log.txt', 'a') as f:
|
98 |
f.write(traceback.format_exc())
|
99 |
+
raise e
|
100 |
+
|
101 |
+
# Display GPU memory information after generation
|
102 |
+
gpu_memory_after_generation = get_gpu_memory()
|
103 |
+
st.write(f"GPU Memory Info after generation: {gpu_memory_after_generation}")
|