Update app.py
Browse files
app.py
CHANGED
|
@@ -14,7 +14,6 @@ def prune_to_single_layer(model_id: str):
|
|
| 14 |
status_lines = []
|
| 15 |
status_lines.append(f"Loading base model: {model_id}")
|
| 16 |
|
| 17 |
-
# Initialize as None to prevent UnboundLocalError in the finally block
|
| 18 |
model = None
|
| 19 |
tokenizer = None
|
| 20 |
|
|
@@ -65,8 +64,8 @@ def prune_to_single_layer(model_id: str):
|
|
| 65 |
outputs = model.generate(
|
| 66 |
**inputs.to(model.device),
|
| 67 |
max_new_tokens=40,
|
| 68 |
-
do_sample=False,
|
| 69 |
-
pad_token_id=tokenizer.eos_token_id
|
| 70 |
)
|
| 71 |
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 72 |
status_lines.append("\nQuick generation test (should be at least semi-coherent):")
|
|
@@ -86,7 +85,6 @@ def prune_to_single_layer(model_id: str):
|
|
| 86 |
return "\n".join(status_lines) + f"\n\nβ Failed: {err_msg}"
|
| 87 |
|
| 88 |
finally:
|
| 89 |
-
# Safely try to free memory even on failure
|
| 90 |
if model is not None:
|
| 91 |
del model
|
| 92 |
if tokenizer is not None:
|
|
@@ -98,7 +96,8 @@ def prune_to_single_layer(model_id: str):
|
|
| 98 |
# ββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 99 |
CSS = """.gradio-container { max-width: 780px !important; }"""
|
| 100 |
|
| 101 |
-
|
|
|
|
| 102 |
gr.Markdown("""
|
| 103 |
# Single-Layer Pruner (test version)
|
| 104 |
Loads a small model β keeps **only the last layer** β shows result + quick generation test.
|
|
@@ -117,11 +116,11 @@ with gr.Blocks(title="Minimal Single-Layer Pruner", css=CSS, theme=gr.themes.Def
|
|
| 117 |
value="Qwen/Qwen2.5-0.5B-Instruct"
|
| 118 |
)
|
| 119 |
|
|
|
|
| 120 |
status = gr.Textbox(
|
| 121 |
label="Pruning log",
|
| 122 |
lines=18,
|
| 123 |
-
interactive=False
|
| 124 |
-
show_copy_button=True
|
| 125 |
)
|
| 126 |
|
| 127 |
btn = gr.Button("Prune to 1 layer β Test", variant="primary")
|
|
@@ -140,4 +139,5 @@ with gr.Blocks(title="Minimal Single-Layer Pruner", css=CSS, theme=gr.themes.Def
|
|
| 140 |
""")
|
| 141 |
|
| 142 |
if __name__ == "__main__":
|
| 143 |
-
|
|
|
|
|
|
| 14 |
status_lines = []
|
| 15 |
status_lines.append(f"Loading base model: {model_id}")
|
| 16 |
|
|
|
|
| 17 |
model = None
|
| 18 |
tokenizer = None
|
| 19 |
|
|
|
|
| 64 |
outputs = model.generate(
|
| 65 |
**inputs.to(model.device),
|
| 66 |
max_new_tokens=40,
|
| 67 |
+
do_sample=False,
|
| 68 |
+
pad_token_id=tokenizer.eos_token_id
|
| 69 |
)
|
| 70 |
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 71 |
status_lines.append("\nQuick generation test (should be at least semi-coherent):")
|
|
|
|
| 85 |
return "\n".join(status_lines) + f"\n\nβ Failed: {err_msg}"
|
| 86 |
|
| 87 |
finally:
|
|
|
|
| 88 |
if model is not None:
|
| 89 |
del model
|
| 90 |
if tokenizer is not None:
|
|
|
|
| 96 |
# ββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 97 |
CSS = """.gradio-container { max-width: 780px !important; }"""
|
| 98 |
|
| 99 |
+
# FIX 1: Removed css and theme from Blocks()
|
| 100 |
+
with gr.Blocks(title="Minimal Single-Layer Pruner") as demo:
|
| 101 |
gr.Markdown("""
|
| 102 |
# Single-Layer Pruner (test version)
|
| 103 |
Loads a small model β keeps **only the last layer** β shows result + quick generation test.
|
|
|
|
| 116 |
value="Qwen/Qwen2.5-0.5B-Instruct"
|
| 117 |
)
|
| 118 |
|
| 119 |
+
# FIX 2: Removed show_copy_button=True
|
| 120 |
status = gr.Textbox(
|
| 121 |
label="Pruning log",
|
| 122 |
lines=18,
|
| 123 |
+
interactive=False
|
|
|
|
| 124 |
)
|
| 125 |
|
| 126 |
btn = gr.Button("Prune to 1 layer β Test", variant="primary")
|
|
|
|
| 139 |
""")
|
| 140 |
|
| 141 |
if __name__ == "__main__":
|
| 142 |
+
# FIX 3: Moved css and theme into launch()
|
| 143 |
+
demo.launch(css=CSS, theme=gr.themes.Default())
|