.

2025-08-29 09:54:36 +08:00 · 2025-08-29 09:54:36 +08:00 · ba8955b82b
parent fc42feee4f
commit ba8955b82b
1 changed files with 1 additions and 1 deletions
--- a/train_sft_ds.py
+++ b/train_sft_ds.py
@ -487,7 +487,7 @@ def main():
        torch_dtype=dtype,
        low_cpu_mem_usage=True,
        trust_remote_code=True,
-        attn_implementation="eager"
+        attn_implementation="sdpa"
    )

    print(f"GC enabled? {getattr(model, 'is_gradient_checkpointing', False)}", flush=True)