[Perf] AdamWConfig: enable fused AdamW kernel

Pass ``fused=True`` to ``torch.optim.AdamW`` on the non-foreach path. The fused
kernel folds the per-parameter element-wise AdamW update into a single CUDA
launch, cutting launch overhead and optimizer-step time at large parameter
counts.

Author: HAOCHENYE

xtuner/v1/config/optim.py

@@ -63,7 +63,7 @@ def build(self, model):
                 foreach=self.foreach,
             )
         return torch.optim.AdamW(
-            params, lr=self.lr, betas=self.betas, eps=self.eps, weight_decay=self.weight_decay, foreach=self.foreach
+            params, lr=self.lr, betas=self.betas, eps=self.eps, weight_decay=self.weight_decay, foreach=self.foreach, fused=True,
         )