Update train_seallm_khm_sum.py

train_seallm_khm_sum.py

@@ -86,11 +86,13 @@ def load_model_and_tokenizer():
         trust_remote_code=True,
     )
 
-    model.gradient_checkpointing_enable()
+    # Disable gradient checkpointing; old transformers breaks autograd here
+    # model.gradient_checkpointing_enable()
 
     return model, tokenizer
 
 
+
 def main():
     train_ds, eval_ds = load_khm_dataset()
     model, tokenizer = load_model_and_tokenizer()
@@ -149,10 +151,11 @@ def main():
         save_total_limit=2,
         lr_scheduler_type="cosine",
         warmup_ratio=0.03,
-        fp16=True,          # safer for old transformers
-        report_to="none",   # remove if this crashes
+        fp16=False,        # turn off mixed precision for CPU
+        report_to="none",
     )
 
+
     trainer = Trainer(
         model=model,
         args=training_args,