diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 4d860505ca4d..933c36fe8bbd 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1411,15 +1411,10 @@ def _set_default_nsa_kv_cache_dtype(self, major: int, quantization: str) -> str:
             )
 
         if self.kv_cache_dtype == "auto":
-            # TODO: Temporarily set default dtype on B200 as bfloat16 to avoid performance regression.
-            # TODO: Remove this after the performance regression is fixed. (Ref: https://github.com/sgl-project/sglang/issues/21291)
-            if quantization == "modelopt_fp4" and major >= 10 and self.dp_size > 1:
+            if major >= 10:
                 self.kv_cache_dtype = "fp8_e4m3"
             else:
                 self.kv_cache_dtype = "bfloat16"
-            # self.kv_cache_dtype = (
-            #     "fp8_e4m3" if (major >= 10 and self.dp_size > 1) else "bfloat16"
-            # )
             logger.warning(
                 f"Setting KV cache dtype to {self.kv_cache_dtype} for DeepSeek DSA on SM{major} device."
             )
@@ -1450,7 +1445,7 @@ def _set_default_nsa_backends(self, kv_cache_dtype: str, major: int) -> str:
             self.nsa_prefill_backend = "tilelang"
             self.nsa_decode_backend = "tilelang"
         elif kv_cache_dtype == "fp8_e4m3":
-            if self.dp_size == 1 and major >= 10:
+            if major >= 10:
                 self.nsa_prefill_backend = "trtllm"
                 self.nsa_decode_backend = "trtllm"
             else: