From dfe85c3c82d1d47245b5a922190d9fef3cfe2e41 Mon Sep 17 00:00:00 2001 From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com> Date: Thu, 2 Jul 2026 00:16:33 -0400 Subject: [PATCH] fix: fsdp=True instead of list + add activation_checkpointing --- train.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/train.py b/train.py index a0d0a61..6d348ee 100644 --- a/train.py +++ b/train.py @@ -167,13 +167,14 @@ def train(config_path): eval_steps=config.get("eval_steps", 100), bf16=True, gradient_checkpointing=config.get("gradient_checkpointing", True), - fsdp=["full_shard", "auto_wrap"], + fsdp=True, fsdp_config={ "backward_prefetch": "backward_pre", "forward_prefetch": "true", "cpu_offload": "false", "auto_wrap_policy": "TRANSFORMER_BASED_WRAP", "transformer_layer_cls_to_wrap": "Qwen3_5MoeDecoderLayer", + "activation_checkpointing": True, }, )