[NVIDIA] Remove unused get_fused_moe_impl_class function (#9764)

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -1074,16 +1074,3 @@ class FlashInferFP4MoE(FusedMoE):
         )[0]
 
         return result
-
-
-def get_fused_moe_impl_class():
-    """Factory function to get the appropriate FusedMoE implementation class."""
-    if should_use_flashinfer_trtllm_moe() and _is_fp4_quantization_enabled():
-        # Use FP4 variant when FP4 quantization is enabled
-        return FlashInferFP4MoE
-    elif should_use_flashinfer_trtllm_moe():
-        # Use regular FlashInfer variant for non-FP4 FlashInfer cases
-        return FlashInferFusedMoE
-    else:
-        # Default case
-        return FusedMoE

python/sglang/srt/layers/quantization/fp8.py

@@ -635,11 +635,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
             layer.register_parameter("w13_weight_scale_inv", w13_weight_scale)
             layer.register_parameter("w2_weight_scale_inv", w2_weight_scale)
             assert self.quant_config.activation_scheme == "dynamic"
-            if (
+            if self.use_cutlass_fused_experts_fp8:
-                get_bool_env_var("SGLANG_CUTLASS_MOE")
-                and self.cutlass_fp8_supported
-                and (is_sm100_supported() or is_sm90_supported())
-            ):
                 self.ab_strides1 = torch.full(
                     (num_experts,),
                     hidden_size,