.

2025-08-28 23:19:34 +08:00 · 2025-08-28 23:19:34 +08:00 · fc42feee4f
parent 1e989784d3
commit fc42feee4f
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="sdpa"
+        attn_implementation="eager"
    )

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