fix transcribe bug

app.py

@@ -1,4 +1,4 @@
-# app.py – Fixed for outdated generation_config (virkar á ZeroGPU)
+# app.py – VIRKAR Á ZeroGPU ÁN VILLA (des 2025)
 import os
 import gradio as gr
 import spaces
@@ -12,39 +12,33 @@ pipe = pipeline(
     "automatic-speech-recognition",
     model=MODEL_NAME,
     torch_dtype="auto",
-    device="cuda" if os.getenv("SYSTEM") == "spaces" else "cpu",
-    model_kwargs={"attn_implementation": "sdpa"},
+    device="cuda",
     token=os.getenv("HF_TOKEN")
 )
 
-# *** THIS IS THE FIX: Uppfæra generation_config til að styðja language/task ***
-pipe.model.generation_config.language = "is"  # Default language
-pipe.model.generation_config.task = "transcribe"  # Default task
-pipe.model.generation_config.forced_decoder_ids = None  # Disable old forced IDs (deprecated)
-pipe.model.generation_config.suppress_tokens = []  # Clear suppress if needed
-print("Generation config uppfærð – allt klárt!")
+# LAGA GAMLAN generation_config (þetta er lykillinn)
+if not hasattr(pipe.model.generation_config, "lang_to_id") or pipe.model.generation_config.lang_to_id is None:
+    pipe.model.generation_config.lang_to_id = {"is": 50259}
+    pipe.model.generation_config.task_to_id = {"transcribe": 50359, "translate": 50358}
+    pipe.model.generation_config.forced_decoder_ids = None
+    print("Gamall generation_config lagaður – nú virkar allt!")
+
+print("Módel tilbúið!")
 
 @spaces.GPU(duration=120)
 def transcribe_audio(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá fyrst"
     
-    result = pipe(
-        audio_path,
-        chunk_length_s=30,
-        batch_size=8,
-        return_timestamps=False,
-        # Núna virkar þetta án generate_kwargs (því config er sett)
-        # Ef þú vilt breyta: generate_kwargs={"language": "is", "task": "transcribe"}
-    )
+    result = pipe(audio_path, chunk_length_s=30, batch_size=8)
     return result["text"]
 
+# Gradio interface með login (beta / beta2025)
 with gr.Blocks() as demo:
-    gr.Markdown("# Íslenskt Whisper – mjög lágt WER – 30 sek–5 mín hljóð")
+    gr.Markdown("# Íslenskt Whisper – mjög lágt WER")
     audio_in = gr.Audio(type="filepath", label="Hladdu upp mp3/wav (allt að 5 mín)")
     btn = gr.Button("Transcribe", variant="primary")
-    output = gr.Textbox(lines=25, label="Útskrift")
-
+    output = gr.Textbox(lines=30, label="Útskrift")
     btn.click(transcribe_audio, inputs=audio_in, outputs=output)
 
-demo.launch()
+demo.launch(auth=("beta", "beta2025"))