fix: force FSDP1 with string value to avoid FSDP2 gather spike

2026-07-02 10:54:17 -04:00
parent f1e016c8af
commit b056ec0306
1 changed files with 1 additions and 1 deletions
--- a/train.py
+++ b/train.py
@@ -169,7 +169,7 @@ def train(config_path):
        eval_steps=config.get("eval_steps", 100),
        bf16=True,
        gradient_checkpointing=False,  # FSDP activation_checkpointing handles it
-        fsdp=True,
+        fsdp="full_shard",  # Force FSDP1 (not FSDP2)
        fsdp_config={
            "sharding_strategy": "SHARD_GRAD_OP",
            "cpu_offload": False,