Enhance speed by using nn.layernorm and nn.groupnorm (triton-lang/triton#5712)

config.json

@@ -14,7 +14,7 @@
   "decay_low_rank_dim": 64,
   "eos_token_id": 0,
   "fuse_cross_entropy": true,
-  "fuse_norm": true,
+  "fuse_norm": false,
   "gate_low_rank_dim": 128,
   "head_dim": 64,
   "hidden_act": "sqrelu",