Update DeepSeek-R1-FP4 default config on blackwell (#11512)

python/sglang/srt/server_args.py

@@ -802,7 +802,32 @@ class ServerArgs:
 
         hf_config = self.get_hf_config()
         model_arch = hf_config.architectures[0]
-        if model_arch in ["GptOssForCausalLM"]:
+        if model_arch in ["DeepseekV3ForCausalLM"]:
+            if is_cuda() and is_sm100_supported():
+                if (
+                    self.attention_backend is None
+                    and self.prefill_attention_backend is None
+                    and self.decode_attention_backend is None
+                ):
+                    self.attention_backend = "trtllm_mla"
+                    logger.info(
+                        "Use trtllm_mla as attention backend on sm100 for DeepseekV3ForCausalLM"
+                    )
+                if not self.enable_dp_attention:
+                    self.enable_flashinfer_allreduce_fusion = True
+                    logger.info(
+                        "Enable FlashInfer AllReduce Fusion on sm100 for DeepseekV3ForCausalLM"
+                    )
+                if (
+                    self.quantization == "modelopt_fp4"
+                    and self.moe_runner_backend == "auto"
+                ):
+                    self.moe_runner_backend = "flashinfer_trtllm"
+                    logger.info(
+                        "Use flashinfer_trtllm as moe runner backend on sm100 for DeepseekV3ForCausalLM"
+                    )
+
+        elif model_arch in ["GptOssForCausalLM"]:
             if (
                 self.attention_backend is None
                 and self.prefill_attention_backend is None