fix(modelopt): dispatch NVFP4 MoE on the cached backend, not the live global (#38932)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -3045,6 +3045,7 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
         # tuple). Defer per-attribute access to the branches that actually
         # consume them.
         activation = self.moe_runner_config.activation
+        # Use the cached backend: the global differs under speculative decoding.
         moe_runner_backend = getattr(
             self, "_moe_runner_backend", get_moe_runner_backend()
         )
@@ -3077,8 +3078,11 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
             quant_info = self.get_marlin_quant_info(layer)
             return self.runner.run(dispatch_output, quant_info)
 
-        # FlashInfer TRTLLM FP4 path
-        if self.enable_flashinfer_trtllm_moe and hasattr(layer, "g1_scale_c"):
+        # FlashInfer TRTLLM FP4 path (routed shares the weight prep and the runner)
+        if (
+            moe_runner_backend.is_flashinfer_trtllm()
+            or moe_runner_backend.is_flashinfer_trtllm_routed()
+        ):
             from sglang.srt.layers.moe.moe_runner.flashinfer_trtllm import (
                 FlashInferTrtllmFp4MoeQuantInfo,
             )
@@ -3114,7 +3118,7 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
 
             return self.runner.run(dispatch_output, quant_info)
 
-        if self.enable_flashinfer_cutedsl_moe:
+        if moe_runner_backend.is_flashinfer_cutedsl():
             from sglang.srt.layers.moe.moe_runner.flashinfer_cutedsl import (
                 CuteDslFp4MoeQuantInfo,
                 ensure_cutedsl_wrapper,
@@ -3170,7 +3174,7 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
             )
             return self.runner.run(dispatch_output, quant_info)
 
-        if self.enable_flashinfer_cutlass_moe:
+        if moe_runner_backend.is_flashinfer_cutlass():
             from sglang.srt.layers.moe.moe_runner.flashinfer_cutlass import (
                 FlashInferCutlassMoeQuantInfo,
             )