fix(fsdp2): avoid double-dividing tokens_per_gpu by sp_size (#158)

Author: kcz358

src/lmms_engine/train/fsdp2/fsdp2_trainer.py

@@ -669,7 +669,7 @@ def calculate_training_metrics(
         total_tokens += total_seq_len.item()
 
         tokens_per_second = total_seq_len.item() / delta_time
-        tokens_per_gpu = tokens_per_second / sp_size / world_size
+        tokens_per_gpu = tokens_per_second / world_size
 
         # Log total tokens and total tokens per second
         metrics["train/total_tokens"] = TrainUtilities.format_tokens(total_tokens)