[MoE] Add Comprehensive MoE Integration Tests (#12090)

python/sglang/srt/server_args.py

@@ -1396,7 +1396,7 @@ class ServerArgs:
         if self.moe_runner_backend == "flashinfer_cutlass":
             assert (
                 self.quantization == "modelopt_fp4"
-            ), "modelopt_fp4 quantization is required for Flashinfer MOE"
+            ), "modelopt_fp4 quantization is required for Flashinfer Cutlass MOE"
             assert self.ep_size in [
                 1,
                 self.tp_size,

python/sglang/test/test_utils.py

@@ -58,6 +58,7 @@ DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN = "lmsys/sglang-ci-dsv3-test-NextN"
 
 # NVFP4 models
 DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST = "nvidia/DeepSeek-V3-0324-FP4"
+DEFAULT_MODEL_NAME_FOR_TEST_MOE_NVFP4 = "nvidia/Qwen3-30B-A3B-FP4"
 
 # FP8 models
 DEFAULT_MODEL_NAME_FOR_TEST_FP8 = "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8"
@@ -71,6 +72,10 @@ DEFAULT_MODEL_NAME_FOR_MODELOPT_QUANT_ACCURACY_TEST_FP8 = (
 DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8 = "Qwen/Qwen3-1.7B-FP8"
 DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE = "gaunernst/DeepSeek-V2-Lite-Chat-FP8"
 
+# MXFP4 models
+# Standard MXFP4 MoE test model
+DEFAULT_MODEL_NAME_FOR_TEST_MXFP4_WITH_MOE = "openai/gpt-oss-20b"
+
 # W8A8 models
 DEFAULT_MODEL_NAME_FOR_TEST_W8A8 = "RedHatAI/Llama-3.2-3B-quantized.w8a8"
 DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE = "nytopop/Qwen3-30B-A3B.w8a8"
@@ -950,7 +955,6 @@ def run_score_benchmark(
     )
 
     async def _run_benchmark():
-
         # Load tokenizer for generating test data
         from sglang.srt.utils.hf_transformers_utils import get_tokenizer