torch and chunk settings

app.py

@@ -5,7 +5,6 @@ from datetime import datetime
 from pathlib import Path
 import tempfile
 import pandas as pd
-import accelerate
 
 import gradio as gr
 import yt_dlp as youtube_dl
@@ -34,7 +33,8 @@ model = AutoModelForSpeechSeq2Seq.from_pretrained(
     MODEL_NAME,
     use_cache=False,
     device_map="auto",
-    low_cpu_mem_usage=False
+    low_cpu_mem_usage=True,
+    attn_implementation="flash_attention_2"
 )
 
 # Flash Attention setup for memory and speed optimization if supported
@@ -56,7 +56,7 @@ pipe = pipeline(
     model=model,
     tokenizer=tokenizer,
     feature_extractor=feature_extractor,
-    chunk_length_s=30,
+    chunk_length_s=900,  # Increased to 15 minutes
 )
 
 def reset_and_update_dataset(new_data):