Update app.py

app.py

@@ -15,7 +15,7 @@ import numpy as np
 import gradio as gr
 
 from collections import Counter
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, Mistral3ForConditionalGeneration
 from peft import PeftModel
 
 # ================================================================
@@ -54,15 +54,15 @@ if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token
 
 print("Loading base model (CPU, float32)...")
-# CPU Basic tidak support bfloat16/4-bit — pakai float32
-# Model akan lebih lambat (~2-5 menit/request) tapi tetap fungsional
-base_model = AutoModelForCausalLM.from_pretrained(
+# Ministral-3-8B menggunakan Mistral3 architecture (VLM)
+# Harus pakai Mistral3ForConditionalGeneration, bukan AutoModelForCausalLM
+base_model = Mistral3ForConditionalGeneration.from_pretrained(
     BASE_MODEL,
-    torch_dtype       = torch.float32,  # ← CPU butuh float32
+    dtype             = torch.float32,  # ← CPU butuh float32
     device_map        = "cpu",
     trust_remote_code = True,
     token             = HF_TOKEN or None,
-    low_cpu_mem_usage = True,           # ← hemat RAM saat loading
+    low_cpu_mem_usage = True,
 )
 
 print("Loading LoRA adapter...")
@@ -360,4 +360,4 @@ pip install gradio_client
                       outputs=outputs)
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()