fix: replace app.py with file-based upload (no heredoc f-string issue)

app.py

@@ -1,4 +1,4 @@
-"""FOG Ablation — ZeroGPU runner for stress tasks."""
+"""FOG Ablation - ZeroGPU runner for stress tasks."""
 from __future__ import annotations
 
 import json
@@ -9,7 +9,6 @@ import time
 import gradio as gr
 import spaces
 
-# Ensure src/ is importable
 sys.path.insert(0, os.path.dirname(__file__))
 
 
@@ -26,47 +25,35 @@ def _gpu_decorator(fn):
 
 
 @_gpu_decorator
-def run_fog_task(task_name: str, seeds_str: str, n_epochs: int, n_train: int, batch_size: int) -> str:
-    """Run one FOG stress task across all 3 models and given seeds."""
+def run_fog_task(task_name, seeds_str, n_epochs, n_train, batch_size):
     import torch
     from src.fog.config import BASELINE_MED, MOTIF_MED, UNIFORM_MED
     from src.fog.train import run_experiment
 
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     seeds = [int(s.strip()) for s in seeds_str.split(",")]
-
     configs = [
         ("baseline", BASELINE_MED),
         ("uniform_small", UNIFORM_MED),
         ("motif", MOTIF_MED),
     ]
-
     results = []
     for seed in seeds:
         for model_type, cfg in configs:
             result = run_experiment(
-                task_name=task_name,
-                cfg=cfg,
-                model_type=model_type,
-                n_epochs=n_epochs,
-                batch_size=batch_size,
-                lr=3e-4,
-                device=device,
-                seed=seed,
-                n_train=n_train,
-                n_eval=500,
+                task_name=task_name, cfg=cfg, model_type=model_type,
+                n_epochs=int(n_epochs), batch_size=int(batch_size), lr=3e-4,
+                device=device, seed=seed, n_train=int(n_train), n_eval=500,
             )
             results.append(result)
             print(f"  {model_type}/{task_name} seed={seed}: "
                   f"acc={result['final_accuracy']:.4f} em={result['final_exact_match']:.4f} "
                   f"time={result['elapsed_s']}s")
-
     return json.dumps(results, indent=2)
 
 
 @_gpu_decorator
-def run_all_tasks(seeds_str: str, n_epochs: int, n_train: int, batch_size: int) -> str:
-    """Run ALL 4 stress tasks sequentially."""
+def run_all_tasks(seeds_str, n_epochs, n_train, batch_size):
     import torch
     from src.fog.config import BASELINE_MED, MOTIF_MED, UNIFORM_MED
     from src.fog.train import run_experiment
@@ -74,55 +61,42 @@ def run_all_tasks(seeds_str: str, n_epochs: int, n_train: int, batch_size: int)
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     seeds = [int(s.strip()) for s in seeds_str.split(",")]
     tasks = ["conditional", "intersection", "compose_add", "multihop"]
-
     configs = [
         ("baseline", BASELINE_MED),
         ("uniform_small", UNIFORM_MED),
         ("motif", MOTIF_MED),
     ]
-
     all_results = []
     for task_name in tasks:
-        print(f"
-{'='*50} {task_name} {'='*50}")
+        sep = "=" * 50
+        print(sep + " " + task_name + " " + sep)
         for seed in seeds:
             for model_type, cfg in configs:
                 result = run_experiment(
-                    task_name=task_name,
-                    cfg=cfg,
-                    model_type=model_type,
-                    n_epochs=n_epochs,
-                    batch_size=batch_size,
-                    lr=3e-4,
-                    device=device,
-                    seed=seed,
-                    n_train=n_train,
-                    n_eval=500,
+                    task_name=task_name, cfg=cfg, model_type=model_type,
+                    n_epochs=int(n_epochs), batch_size=int(batch_size), lr=3e-4,
+                    device=device, seed=seed, n_train=int(n_train), n_eval=500,
                 )
                 all_results.append(result)
                 print(f"  {model_type}/{task_name} seed={seed}: "
-                      f"acc={result['final_accuracy']:.4f} em={result['final_exact_match']:.4f} "
+                      f"acc={result['final_accuracy']:.4f} "
+                      f"em={result['final_exact_match']:.4f} "
                       f"time={result['elapsed_s']}s")
-
-    # Summary table
-    summary = "
-" + "="*70 + "
-"
-    summary += f"{'Task':<15} {'Model':<15} {'Params':>8} {'Acc':>8} {'EM':>8} {'Time':>6}
-"
-    summary += "-"*70 + "
-"
+    # Summary
+    lines = ["", "=" * 70]
+    lines.append(f"{'Task':<15} {'Model':<15} {'Params':>8} {'Acc':>8} {'EM':>8} {'Time':>6}")
+    lines.append("-" * 70)
     for r in all_results:
-        summary += (f"{r['task']:<15} {r['model_type']:<15} {r['n_params']:>8,} "
-                    f"{r['final_accuracy']:>8.4f} {r.get('final_exact_match',0):>8.4f} "
-                    f"{r['elapsed_s']:>5.0f}s
-")
-    print(summary)
-
+        em = r.get("final_exact_match", 0) or 0
+        lines.append(
+            f"{r['task']:<15} {r['model_type']:<15} {r['n_params']:>8,} "
+            f"{r['final_accuracy']:>8.4f} {em:>8.4f} {r['elapsed_s']:>5.0f}s"
+        )
+    print("\n".join(lines))
     return json.dumps(all_results, indent=2)
 
 
-def health_check() -> str:
+def health_check():
     import torch
     return json.dumps({
         "status": "ok",
@@ -134,38 +108,31 @@ def health_check() -> str:
 
 with gr.Blocks(title="FOG Stress Ablation") as demo:
     gr.Markdown("# FOG Stress Ablation (ZeroGPU)")
-    gr.Markdown("Runs 4 stress tasks: conditional, intersection, compose_add, multihop")
-
+    gr.Markdown("4 stress tasks: conditional, intersection, compose_add, multihop")
     with gr.Row():
         with gr.Column():
             seeds_input = gr.Textbox(label="Seeds (comma-sep)", value="42, 43, 44")
             epochs_input = gr.Slider(10, 200, value=100, step=10, label="Epochs")
             n_train_input = gr.Slider(1000, 10000, value=5000, step=1000, label="Train samples")
             batch_input = gr.Slider(64, 512, value=256, step=64, label="Batch size")
-
         with gr.Column():
             task_dropdown = gr.Dropdown(
                 ["conditional", "intersection", "compose_add", "multihop"],
-                label="Single task", value="conditional"
+                label="Single task", value="conditional",
             )
             run_one_btn = gr.Button("Run Single Task", variant="secondary")
             run_all_btn = gr.Button("Run ALL 4 Tasks", variant="primary")
-
     output_box = gr.Textbox(label="Results JSON", lines=20)
-
     run_one_btn.click(
         fn=run_fog_task,
         inputs=[task_dropdown, seeds_input, epochs_input, n_train_input, batch_input],
-        outputs=output_box,
-        api_name="run_task",
+        outputs=output_box, api_name="run_task",
     )
     run_all_btn.click(
         fn=run_all_tasks,
         inputs=[seeds_input, epochs_input, n_train_input, batch_input],
-        outputs=output_box,
-        api_name="run_all",
+        outputs=output_box, api_name="run_all",
     )
-
     health_btn = gr.Button("Health Check")
     health_output = gr.Textbox(label="Health", lines=3)
     health_btn.click(fn=health_check, outputs=health_output, api_name="health")