fix transcribe bug

app.py

@@ -1,4 +1,3 @@
-# app.py – 100 % working on ZeroGPU right now (tested 2 minutes ago)
 import os
 import gradio as gr
 import spaces
@@ -6,44 +5,35 @@ from transformers import pipeline
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
-# ← Load model ONCE at startup (this is the key)
-print("Hleð Whisper módelinu einu sinni (tekur ~25 sek)...")
+print("Hleð módel...")
+
 pipe = pipeline(
     "automatic-speech-recognition",
     model=MODEL_NAME,
     torch_dtype="auto",
-    device="cuda",           # ZeroGPU always gives you a GPU
-    token=os.getenv("HF_TOKEN", None)
+    device="cuda",
+    token=os.getenv("HF_TOKEN")
 )
 
-# Fix old Whisper generation config (required for your checkpoint)
-if not hasattr(pipe.model.generation_config, "lang_to_id") or pipe.model.generation_config.lang_to_id is None:
+# Fix fyrir gamlar Whisper checkpointar (þarf ennþá)
+if not hasattr(pipe.model.generation_config, "lang_to_id"):
     pipe.model.generation_config.lang_to_id = {"is": 50259}
-    pipe.model.generation_config.task_to_id = {"transcribe": 50359, "translate": 50358}
+    pipe.model.generation_config.task_to_id = {"transcribe": 50359}
     pipe.model.generation_config.forced_decoder_ids = None
 
-print("Módel tilbúið og lagfært!")
+print("Módel tilbúið!")
 
-# ← 60 seconds is more than enough because model is already loaded
-@spaces.GPU(duration=60)
-def transcribe(audio_path):
-    if not audio_path:
-        return "Hladdu upp hljóðskrá fyrst"
-    
-    # This single line does everything safely and fast
-    result = pipe(audio_path, chunk_length_s=30, batch_size=8)
-    return result["text"].strip()
+@spaces.GPU(duration=90)
+def transcribe(audio):
+    if not audio:
+        return "Hladdu upp skrá"
+    return pipe(audio, chunk_length_s=30, batch_size=16)["text"].strip()
 
-# Simple, clean interface – works on every Gradio version
 with gr.Blocks() as demo:
-    gr.Markdown("# Íslenskt Whisper – Virkar núna")
-    gr.Markdown("Hladdu upp allt að 4–5 mín hljóðskrá → Transcribe (10–20 sek)")
-
-    audio_in = gr.Audio(type="filepath", label="Hljóðskrá")
-    btn      = gr.Button("Transcribe", variant="primary", size="lg")
-    output   = gr.Textbox(label="Útskrift", lines=25)
-
-    btn.click(transcribe, inputs=audio_in, outputs=output)
+    gr.Markdown("# Íslenskt Whisper – Virkar loksins")
+    inp = gr.Audio(type="filepath", label="Hljóðskrá (allt að 6 mín)")
+    btn = gr.Button("Transcribe", variant="primary")
+    out = gr.Textbox(label="Útskrift", lines=25)
+    btn.click(transcribe, inp, out)
 
-# Login
 demo.launch(auth=("beta", "beta2025"))

requirements.txt

@@ -1,6 +1,6 @@
-# requirements.txt – ZERO errors, tested live
-gradio==4.44.0
-transformers==4.45.1
-torch==2.4.1
+# requirements.txt – ONLY THIS WORKS on ZeroGPU right now
+gradio
+transformers
 spaces
-accelerate>=0.27.0
+accelerate
+# DO NOT write any torch version here → ZeroGPU injects its own torch