Cuda no ordinal

app.py

@@ -65,15 +65,15 @@ qwen_model.generation_config = GenerationConfig.from_pretrained(
 )
 
 
-# salmonn_model = SALMONN(
-#     ckpt="./SALMONN_PATHS/salmonn_v1.pth",
-#     whisper_path="./SALMONN_PATHS/whisper-large-v2",
-#     beats_path="./SALMONN_PATHS/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt",
-#     vicuna_path="./SALMONN_PATHS/vicuna-13b-v1.1",
-#     low_resource=False,
-#     device="cuda:0",
-# )
-# salmonn_tokenizer = salmonn_model.llama_tokenizer
+salmonn_model = SALMONN(
+    ckpt="./SALMONN_PATHS/salmonn_v1.pth",
+    whisper_path="./SALMONN_PATHS/whisper-large-v2",
+    beats_path="./SALMONN_PATHS/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt",
+    vicuna_path="./SALMONN_PATHS/vicuna-13b-v1.1",
+    low_resource=False,
+    device="cuda",
+)
+salmonn_tokenizer = salmonn_model.llama_tokenizer
 
 
 diva = AutoModel.from_pretrained("WillHeld/DiVA-llama-3-v0-8b", trust_remote_code=True)
@@ -210,7 +210,7 @@ def transcribe(audio_input, text_prompt, state, model_order):
     initial_responses = [("", "", "")]
     resp_generators = [
         gen_from_via(),
-        #gen_from_salmonn(),
+        gen_from_salmonn(),
         gen_from_qwen(),
     ]
     order = -1
@@ -250,7 +250,7 @@ def transcribe(audio_input, text_prompt, state, model_order):
         resp_2,
         resp_3,
         gr.Button(visible=True),
-        gr.Button(visible=False),
+        gr.Button(visible=True),
         gr.Button(visible=True),
         responses_complete(state),
     )

models/salmonn.py

@@ -44,7 +44,7 @@ class SALMONN(nn.Module):
         speech_qformer_token_num=1,
         speech_qformer_layer=2,
         lora=True,
-        device="cuda:0",
+        device="cuda",
         lora_alpha=32,
         lora_rank=8,
         lora_dropout=0.1,
@@ -138,7 +138,7 @@ class SALMONN(nn.Module):
         wav_path,
         prompt,
         prompt_pattern="USER: <Speech><SpeechHere></Speech> {}\nASSISTANT:",
-        device="cuda:0",
+        device="cuda",
         max_length=200,
         max_new_tokens=128,
         num_beams=1,