Fix max_seq_length placement and torch_dtype deprecation

train.py

@@ -63,7 +63,7 @@ print(f"Total conversations: {len(train_dataset)}")
 # Tokenizer
 # ------------------------------------------------------------------
 print("Loading tokenizer...")
-tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token
     tokenizer.pad_token_id = tokenizer.eos_token_id
@@ -76,7 +76,6 @@ model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
     torch_dtype=torch.bfloat16,
     device_map="auto",
-    trust_remote_code=True,
 )
 
 model.gradient_checkpointing_enable()
@@ -92,7 +91,6 @@ args = SFTConfig(
     per_device_train_batch_size=1,
     gradient_accumulation_steps=4,
     learning_rate=2e-5,
-    max_seq_length=MAX_SEQ_LENGTH,
     logging_strategy="steps",
     logging_steps=10,
     logging_first_step=True,
@@ -113,6 +111,7 @@ trainer = SFTTrainer(
     args=args,
     train_dataset=train_dataset,
     processing_class=tokenizer,
+    max_seq_length=MAX_SEQ_LENGTH,
 )
 
 # ------------------------------------------------------------------