Prueba

app.py

@@ -7,16 +7,22 @@ import csv
 from huggingface_hub import login
 
 
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 import torch
 
 model_name = "atorojaen/DeepSeekMisongynyLyrics"
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_compute_dtype=torch.float16,
+)
 
 tokenizer = AutoTokenizer.from_pretrained(model_name)
+
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    device_map="auto",                # o "cuda" si quieres forzarlo en GPU
-    torch_dtype=torch.float16         # si tu modelo lo soporta
+    quantization_config=bnb_config,
+    device_map="auto"
 )
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

requirements.txt

@@ -3,5 +3,5 @@ torch
 transformers
 unsloth
 accelerate
-bitsandbytes
+bitsandbytes==0.42.0
 scipy  # requerido por algunos backends de HF