feat: restore FSDP with SHARD_GRAD_OP + sync_module_states

train.py

@@ -169,7 +169,17 @@ def train(config_path):
         eval_steps=config.get("eval_steps", 100),
         bf16=True,
         gradient_checkpointing=True,
-        # No FSDP - use standard accelerate data parallelism
+        fsdp=True,
+        fsdp_config={
+            "sharding_strategy": "SHARD_GRAD_OP",
+            "cpu_offload": False,
+            "activation_checkpointing": True,
+            "limit_all_gathers": True,
+            "sync_module_states": True,
+            "mixed_precision": {"param_dtype": torch.bfloat16, "reduce_dtype": torch.float32, "buffer_dtype": torch.float32},
+            "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+            "transformer_layer_cls_to_wrap": "Qwen3_5MoeDecoderLayer",
+        },
     )
     
     # SFT Trainer (DeepSpeed handles distributed training via config)