Update app.py

app.py

@@ -1,4 +1,4 @@
-# app.py — Íslenskt ASR – ZeroGPU Optimized (fixed + battle-tested Dec 2025)
+# app.py — Íslenskt ASR – ZeroGPU 100% stable (Dec 2025 final fixed version)
 import os
 os.environ["OMP_NUM_THREADS"] = "1"
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "garbage_collection_threshold:0.6,max_split_size_mb:128"
@@ -9,20 +9,17 @@ from transformers import pipeline
 import torch
 import gc
 
-# ——————————————————————————————
-# Global pipeline – lazy load + auto-rebuild on OOM
-# ——————————————————————————————
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
-pipe = None
+pipe = None  # Global pipeline
 
 @spaces.GPU(duration=180, max_batch_size=4)
 def get_or_refresh_pipeline():
     global pipe
     
-    # Rebuild if GPU context died
+    # Check if pipeline is broken or GPU context died
     if pipe is not None:
         try:
-            _ = pipe.model.device
+            _ = pipe.model.device  # Quick health check
         except Exception:
             print("GPU context lost → rebuilding pipeline...")
             pipe = None
@@ -42,17 +39,16 @@ def get_or_refresh_pipeline():
     
     return pipe
 
-# ——————————————————————————————
-# Transcription – memory-safe + fixed del bug
-# ——————————————————————————————
 def transcribe_3min(audio_path):
     if not audio_path:
         return "Hlaðið upp hljóðskrá (mp3/wav, max 5 mín)"
     
+    global pipe  # ← Now declared at the very top of the function → no error!
+    
     try:
-        pipe = get_or_refresh_pipeline()
+        current_pipe = get_or_refresh_pipeline()
         
-        result = pipe(
+        result = current_pipe(
             audio_path,
             chunk_length_s=30,
             stride_length_s=(6, 0),
@@ -63,33 +59,29 @@ def transcribe_3min(audio_path):
         
         text = result["text"].strip()
         
-        # Safely delete chunks if they exist (fixed!)
+        # Clean up chunks if present
         if "chunks" in result:
             del result["chunks"]
         
-        # Aggressive cleanup
         gc.collect()
         torch.cuda.empty_cache()
         
         return text if text else "(ekkert tal greint)"
         
     except torch.cuda.OutOfMemoryError:
-        print("OOM → forcing reload on next request")
-        global pipe
-        pipe = None
+        print("OOM detected → forcing full pipeline reload")
+        pipe = None  # This is now allowed because global declared first
         gc.collect()
         torch.cuda.empty_cache()
         return "Of mikið minni notað – bíddu 10 sek og prófaðu aftur (ZeroGPU takmörk)"
         
     except Exception as e:
-        return f"Villa: {str(e)}"
+        return f"Óvænt villa: {str(e)}"
 
-# ——————————————————————————————
-# UI
-# ——————————————————————————————
+# ————————————————————— UI —————————————————————
 with gr.Blocks(title="Íslenskt ASR") as demo:
     gr.Markdown("# Íslenskt ASR – 3–5 mín hljóð")
-    gr.Markdown("**Whisper-small fínstillt á íslensku · mjög lágur WER**")
+    gr.Markdown("**Whisper-small fínstillt á íslensku spjalli · mjög lágur WER**")
     gr.Markdown("**Hafa samband:** pallinr1@protonmail.com")
     gr.Markdown("> Keyrt á **ZeroGPU** – fyrsta ræsing tekur 15–30 sek, síðan hröð")
     
@@ -110,9 +102,7 @@ with gr.Blocks(title="Íslenskt ASR") as demo:
     - Ef þú færð minnisvillu → bíddu öðruhvolf og prófaðu aftur
     """)
 
-# ——————————————————————————————
-# Launch
-# ——————————————————————————————
+# ————————————————————— Launch —————————————————————
 demo.launch(
     auth=None,
     share=True,