Update app.py

app.py

@@ -24,6 +24,8 @@ def pipe(file, return_timestamps=False):
         chunk_length_s=30,
         device=device,
         token=auth_token,
+        torch_dtype=torch.float16,
+        model_kwargs={"attn_implementation": "flash_attention_2"} if args.flash else {"attn_implementation": "sdpa"},
     )
     asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids(
         language=lang,
@@ -31,7 +33,7 @@ def pipe(file, return_timestamps=False):
         no_timestamps=not return_timestamps,
     )
     asr.model.config.no_timestamps_token_id = asr.tokenizer.encode("<|notimestamps|>", add_special_tokens=False)[0]
-    return asr(file, return_timestamps=return_timestamps)
+    return asr(file, return_timestamps=return_timestamps, batch_size=24)
 
 def transcribe(file, return_timestamps=False):
     if not return_timestamps: