diff --git a/train.py b/train.py
index a1f8ee1..165c813 100644
--- a/train.py
+++ b/train.py
@@ -38,9 +38,9 @@ def train(config_path):
     errors = []
     
     # ------------------------------------------------------------------
-    # Strategy 1: QLoRA with FSDP (preferred)
+    # Strategy 1: QLoRA 4-bit with FSDP (load to CPU, FSDP shards across GPUs)
     # ------------------------------------------------------------------
-    print("\n[1/4] Trying: 4-bit QLoRA (FSDP)...")
+    print("\n[1/4] Trying: 4-bit QLoRA (FSDP, load to CPU)...")
     try:
         bnb_config = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -56,7 +56,7 @@ def train(config_path):
             trust_remote_code=True,
             low_cpu_mem_usage=True,
         )
-        print("✓ Success: QLoRA 4-bit loaded to CPU")
+        print("✓ Success: QLoRA 4-bit loaded to CPU (FSDP will shard across GPUs)")
     except Exception as e:
         errors.append(("QLoRA 4-bit", e))
         print(f"✗ Failed: {e}")
@@ -156,6 +156,12 @@ def train(config_path):
     from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
     from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
     from functools import partial
+    import torch.distributed as dist
+    
+    # Initialize distributed process group (required for FSDP)
+    if not dist.is_initialized():
+        dist.init_process_group(backend="nccl")
+        print("✓ Distributed process group initialized")
     
     def get_auto_wrap_policy(model):
         from transformers.models.qwen3_5_moe.modeling_qwen3_5_moe import Qwen3_5MoeDecoderLayer
@@ -175,7 +181,7 @@ def train(config_path):
         sync_module_states=True,
         use_orig_params=True,
     )
-    print("✓ Model wrapped with FSDP")
+    print("✓ Model wrapped with FSDP (will be sharded across GPUs during training)")
     
     # Training arguments (no FSDP config - we handle it manually)
     training_args = TrainingArguments(