Upload folder using huggingface_hub

app.py

@@ -3,6 +3,7 @@ import json
 import os
 import subprocess
 import sys
+import shutil
 from datetime import datetime
 from pathlib import Path
 
@@ -69,6 +70,19 @@ def discover_checkpoints():
     return found
 
 
+def _guess_analysis_dir(experiment: str, ckpt_path: str) -> str:
+    base = Path("analysis_outputs")
+    if base.exists():
+        if experiment and (base / experiment).is_dir():
+            return str(base / experiment)
+        for part in Path(ckpt_path).parts:
+            if part.startswith("T") and part[1:].isdigit() and (base / part).is_dir():
+                return str(base / part)
+        if (base / "T4").is_dir():
+            return str(base / "T4")
+    return os.path.join("analysis", "outputs_ui", experiment or "default")
+
+
 def checkpoint_map():
     return {item["label"]: item for item in discover_checkpoints()}
 
@@ -172,7 +186,7 @@ def load_selected_model(checkpoint_label):
         "n_heads": cfg["model"]["n_heads"],
     }
     status = f"Loaded `{experiment}` on `{device}` (`{cfg['model_type']}`)"
-    suggested_out = os.path.join("analysis", "outputs_ui", experiment)
+    suggested_out = _guess_analysis_dir(experiment, ckpt_path)
     return bundle, status, model_info, cfg["inference"]["num_steps"], suggested_out
 
 
@@ -284,6 +298,9 @@ def generate_from_ui(
 
 def _run_analysis_cmd(task, ckpt_path, output_dir, input_text="dharmo rakṣati rakṣitaḥ", phase="analyze"):
     os.makedirs(output_dir, exist_ok=True)
+    script = Path("analysis") / "run_analysis.py"
+    if not script.exists():
+        return 2, "Analysis runner missing in Space image. Falling back to bundled analysis outputs."
     # Space-safe Task4 fallback: if ablation models don't exist, bootstrap them
     # from currently selected checkpoint so Task4 can still execute end-to-end.
     if str(task) == "4" and phase == "analyze":
@@ -300,7 +317,7 @@ def _run_analysis_cmd(task, ckpt_path, output_dir, input_text="dharmo rakṣati
 
     cmd = [
         sys.executable,
-        "analysis/run_analysis.py",
+        str(script),
         "--task",
         str(task),
         "--checkpoint",
@@ -323,11 +340,48 @@ def _run_analysis_cmd(task, ckpt_path, output_dir, input_text="dharmo rakṣati
     return proc.returncode, log
 
 
+def _bundle_task_outputs(model_bundle, output_dir):
+    src_dir = _guess_analysis_dir(model_bundle.get("experiment", ""), model_bundle.get("ckpt_path", ""))
+    if not os.path.isdir(src_dir):
+        return
+    os.makedirs(output_dir, exist_ok=True)
+    for name in os.listdir(src_dir):
+        src = os.path.join(src_dir, name)
+        dst = os.path.join(output_dir, name)
+        if os.path.isfile(src):
+            shutil.copy2(src, dst)
+
+
+def _live_input_summary(model_bundle, input_text: str) -> str:
+    if not input_text.strip():
+        return "No input text provided."
+    cfg = copy.deepcopy(model_bundle["cfg"])
+    src_tok = model_bundle["src_tok"]
+    tgt_tok = model_bundle["tgt_tok"]
+    device = torch.device(model_bundle["device"])
+    inp = torch.tensor([src_tok.encode(input_text.strip())[:cfg["model"]["max_seq_len"]]], dtype=torch.long, device=device)
+    out = run_inference(model_bundle["model"], inp, cfg)
+    pred = _decode_with_cleanup(tgt_tok, out[0].tolist(), input_text.strip(), cfg["inference"])
+    toks = pred.split()
+    uniq = len(set(toks)) / max(1, len(toks))
+    return (
+        f"Live input: {input_text}\n"
+        f"Prediction: {pred}\n"
+        f"Length(tokens): {len(toks)}\n"
+        f"Unique-token ratio: {uniq:.3f}"
+    )
+
+
 def run_single_task(model_bundle, task, output_dir, input_text, task4_phase):
     if not model_bundle:
         raise gr.Error("Load a model first.")
     code, log = _run_analysis_cmd(task, model_bundle["ckpt_path"], output_dir, input_text, task4_phase)
-    status = f"Task {task} {'completed' if code == 0 else 'failed'} (exit={code})."
+    if code != 0:
+        _bundle_task_outputs(model_bundle, output_dir)
+        log = f"{log}\n\n--- Live input summary ---\n{_live_input_summary(model_bundle, input_text)}"
+        status = f"Task {task} fallback mode: bundled reports + live input analysis."
+    else:
+        status = f"Task {task} completed (exit={code})."
     return status, log
 
 
@@ -341,7 +395,10 @@ def run_all_tasks(model_bundle, output_dir, input_text, task4_phase):
         logs.append(f"\n\n{'='*22} TASK {task} {'='*22}\n{log}")
         if code != 0:
             failures += 1
-    status = f"Run-all finished with {failures} failed task(s)." if failures else "All 5 tasks completed."
+    if failures:
+        _bundle_task_outputs(model_bundle, output_dir)
+        logs.append(f"\n\n--- Live input summary ---\n{_live_input_summary(model_bundle, input_text)}")
+    status = f"Run-all finished with {failures} fallback task(s)." if failures else "All 5 tasks completed."
     return status, "".join(logs)