fix: disable FSDP, use standard accelerate data parallelism

train.py

@@ -168,20 +168,8 @@ def train(config_path):
         eval_strategy=config.get("eval_strategy", "steps"),
         eval_steps=config.get("eval_steps", 100),
         bf16=True,
-        # gradient_checkpointing removed - FSDP activation_checkpointing handles it
-        fsdp=True,
-        fsdp_config={
-            "sharding_strategy": "SHARD_GRAD_OP",
-            "cpu_offload": False,
-            "activation_checkpointing": True,
-            "limit_all_gathers": True,
-            "sync_module_states": False,  # Don't sync from CPU, FSDP will handle placement
-            "backward_prefetch": "backward_pre",
-            "forward_prefetch": "true",
-            "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
-            "transformer_layer_cls_to_wrap": "Qwen3_5MoeDecoderLayer",
-            "mixed_precision": {"param_dtype": torch.bfloat16, "reduce_dtype": torch.float32, "buffer_dtype": torch.float32},
-        },
+        gradient_checkpointing=True,
+        # No FSDP - use standard accelerate data parallelism
     )
     
     # SFT Trainer (DeepSpeed handles distributed training via config)