enable DeepSeek V3 shared_experts_fusion in sm90 (#5571)

Author: BBuf

python/sglang/srt/models/deepseek_v2.py

@@ -1427,6 +1427,18 @@ def __init__(
                 assert (
                     self.n_share_experts_fusion == self.tp_size
                 ), f"Shared experts fusion optimization is enabled in DeepSeek V3/R1, set it to {self.tp_size} can get best optimized performace."
+        elif self.n_share_experts_fusion == 0:
+            if (
+                torch.cuda.get_device_capability("cuda") >= (9, 0)
+                and self.config.architectures[0] == "DeepseekV3ForCausalLM"
+                and self.config.n_routed_experts == 256
+                and (not global_server_args_dict["enable_deepep_moe"])
+            ):
+                self.n_share_experts_fusion = self.tp_size
+                global_server_args_dict["n_share_experts_fusion"] = self.tp_size
+                logger.info(
+                    "Deepseek V3/R1 with fp8 can use shared experts fusion optimization when SM version >=90. Shared experts fusion optimization is enabled."
+                )
 
         self.model = DeepseekV2Model(
             config, quant_config, prefix=add_prefix("model", prefix)