Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import pandas as pd
|
| 3 |
+
import torch
|
| 4 |
+
|
| 5 |
+
from core.model_loader import load_model, SUPPORTED_MODELS
|
| 6 |
+
from core.profiler import ActivationProfiler
|
| 7 |
+
from core.stress import stress_layer
|
| 8 |
+
from core.sensitivity import compute_si, assign_policy
|
| 9 |
+
from core.reevaluator import reevaluate
|
| 10 |
+
from core.visualization import create_heatmap
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
def run_cortex(model_choice, calibration_text):
|
| 14 |
+
|
| 15 |
+
model, tokenizer = load_model(model_choice)
|
| 16 |
+
|
| 17 |
+
texts = calibration_text.split("\n")
|
| 18 |
+
inputs = tokenizer(texts, return_tensors="pt", padding=True)
|
| 19 |
+
|
| 20 |
+
profiler = ActivationProfiler(model)
|
| 21 |
+
profiler.register()
|
| 22 |
+
|
| 23 |
+
with torch.no_grad():
|
| 24 |
+
model(**inputs)
|
| 25 |
+
|
| 26 |
+
profiler.remove()
|
| 27 |
+
stats = profiler.get()
|
| 28 |
+
|
| 29 |
+
stress_results = {}
|
| 30 |
+
|
| 31 |
+
for name in stats:
|
| 32 |
+
mse_val, kl_val = stress_layer(model, name, inputs, bits=8)
|
| 33 |
+
stress_results[name] = (mse_val, kl_val)
|
| 34 |
+
|
| 35 |
+
scores = compute_si(stats, stress_results)
|
| 36 |
+
policy = assign_policy(scores)
|
| 37 |
+
|
| 38 |
+
policy, degradation = reevaluate(model, tokenizer, texts, policy)
|
| 39 |
+
|
| 40 |
+
heatmap = create_heatmap(scores)
|
| 41 |
+
|
| 42 |
+
df = pd.DataFrame({
|
| 43 |
+
"Layer": list(scores.keys()),
|
| 44 |
+
"Sensitivity_Index": list(scores.values()),
|
| 45 |
+
"Policy": [policy[k] for k in scores]
|
| 46 |
+
})
|
| 47 |
+
|
| 48 |
+
csv_path = "layer_metrics.csv"
|
| 49 |
+
df.to_csv(csv_path, index=False)
|
| 50 |
+
|
| 51 |
+
return policy, heatmap, degradation, csv_path
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
demo = gr.Interface(
|
| 55 |
+
fn=run_cortex,
|
| 56 |
+
inputs=[
|
| 57 |
+
gr.Dropdown(list(SUPPORTED_MODELS.keys()), label="Select Model"),
|
| 58 |
+
gr.Textbox(label="Calibration Text (One sentence per line)", lines=8)
|
| 59 |
+
],
|
| 60 |
+
outputs=[
|
| 61 |
+
gr.JSON(label="Precision Policy"),
|
| 62 |
+
gr.Image(label="Critical Layer Map"),
|
| 63 |
+
gr.Number(label="Final Model Degradation (MSE)"),
|
| 64 |
+
gr.File(label="Download CSV Metrics")
|
| 65 |
+
],
|
| 66 |
+
title="Cortex in the Loop — Research Edition",
|
| 67 |
+
description="Runtime Sensitivity Analyzer with Batch Calibration and CSV Export"
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|