From b056ec03069b3e3f130aee5a07e503a114d0dc7e Mon Sep 17 00:00:00 2001 From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com> Date: Thu, 2 Jul 2026 10:54:17 -0400 Subject: [PATCH] fix: force FSDP1 with string value to avoid FSDP2 gather spike --- train.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/train.py b/train.py index 2f49c43..894162f 100644 --- 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,