Flashinfer TRTLLM-GEN-MoE + Qwen3 (#13489)

python/sglang/srt/models/qwen3_moe.py

@@ -49,6 +49,7 @@ from sglang.srt.layers.moe import (
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
+from sglang.srt.layers.moe.utils import RoutingMethodType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding, get_rope
@@ -111,6 +112,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             intermediate_size=config.moe_intermediate_size,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
+            routing_method_type=RoutingMethodType.Renormalize,
         )
 
         self.gate = ReplicatedLinear(

python/sglang/srt/server_args.py

@@ -1099,6 +1099,29 @@ class ServerArgs:
                 f"Disabling Radix Cache for {model_arch} as it is not yet supported."
             )
             self.disable_radix_cache = True
+        elif model_arch in [
+            "Qwen3MoeForCausalLM",
+            "Qwen3VLMoeForConditionalGeneration",
+        ]:
+            if is_sm100_supported():
+                quantization_config = getattr(hf_config, "quantization_config", None)
+                quant_method = (
+                    quantization_config.get("quant_method")
+                    if quantization_config is not None
+                    else None
+                )
+                if self.quantization is None and quant_method is not None:
+                    self.quantization = quant_method
+                if (
+                    self.quantization == "fp8"
+                    and self.moe_a2a_backend == "none"
+                    and self.moe_runner_backend == "auto"
+                ):
+                    self.moe_runner_backend = "flashinfer_trtllm"
+                    logger.info(
+                        "Use flashinfer_trtllm as MoE runner backend on sm100 for "
+                        f"{model_arch}"
+                    )
         elif model_arch in ["Qwen3NextForCausalLM"]:
             if not self.disable_radix_cache:
                 logger.warning(
@@ -1106,7 +1129,24 @@ class ServerArgs:
                     "overlap schedule currently, try to use --disable-radix-cache if overlap schedule is necessary"
                 )
                 self.disable_overlap_schedule = True
-
+            if is_sm100_supported():
+                quantization_config = getattr(hf_config, "quantization_config", None)
+                quant_method = (
+                    quantization_config.get("quant_method")
+                    if quantization_config is not None
+                    else None
+                )
+                if self.quantization is None and quant_method is not None:
+                    self.quantization = quant_method
+                if (
+                    self.quantization == "fp8"
+                    and self.moe_a2a_backend == "none"
+                    and self.moe_runner_backend == "auto"
+                ):
+                    self.moe_runner_backend = "flashinfer_trtllm"
+                    logger.info(
+                        "Use flashinfer_trtllm as MoE runner backend on sm100 for Qwen3NextForCausalLM"
+                    )
         if is_deepseek_nsa(hf_config):
             if (
                 self.attention_backend is None