diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 244e3d96e..4229f4093 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1424,15 +1424,10 @@ class ServerArgs:
             )
 
         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."
             )
@@ -1463,7 +1458,7 @@ class ServerArgs:
             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: