Add DeepSeekV4 fused MoE Triton autotune support (#25569)

Co-authored-by: undefined <zhouchen.arrebol@jd.com>
Co-authored-by: xq25478 <xq25478@qq.com>
Co-authored-by: xieminghe.simon <xieminghe.simon@jd.com>

benchmark/kernels/fused_moe_triton/common_utils.py

@@ -85,6 +85,7 @@ def get_model_config(
         "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
+        "DeepseekV4ForCausalLM",
         "Glm4MoeForCausalLM",
         "GlmMoeDsaForCausalLM",
         "MistralLarge3ForCausalLM",

benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py

@@ -35,6 +35,7 @@ from sglang.srt.server_args import (
     set_global_server_args_for_scheduler,
 )
 from sglang.srt.utils import get_device, is_hip, is_xpu
+from sglang.srt.utils.hf_transformers_utils import get_config
 
 _is_hip = is_hip()
 _is_xpu = is_xpu()
@@ -174,8 +175,12 @@ def benchmark_config(
         topk_output.router_logits.copy_(new_topk_output.router_logits)
 
     def run():
+        model_config = get_config(args.model, trust_remote_code=True)
+        architecture = model_config.architectures[0]
+        is_dsv4 = architecture == "DeepseekV4ForCausalLM"
         moe_runner_config = MoeRunnerConfig(
             inplace=True,
+            swiglu_limit=10.0 if is_dsv4 else None,
         )
 
         with override_config(config):