[Kernels] Clean up FusedMoeMethodBase and modular kernel setup. Remove extra arguments from modular kernel methods. (#22035)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>

vllm/model_executor/layers/quantization/gptq_marlin.py

@@ -10,7 +10,7 @@ import vllm.model_executor.layers.fused_moe  # noqa
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.layer import (
-    FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported,
+    FusedMoE, FusedMoEConfig, FusedMoEMethodBase, FusedMoeWeightScaleSupported,
     UnquantizedFusedMoEMethod)
 from vllm.model_executor.layers.linear import (LinearMethodBase,
                                                set_weight_attrs)
@@ -375,7 +375,12 @@ class GPTQMarlinLinearMethod(LinearMethodBase):
 class GPTQMarlinMoEMethod(FusedMoEMethodBase):
     """MoE Marlin method with quantization."""
 
-    def __init__(self, quant_config: GPTQMarlinConfig) -> None:
+    def __init__(
+        self,
+        quant_config: GPTQMarlinConfig,
+        moe: FusedMoEConfig,
+    ) -> None:
+        super().__init__(moe)
         self.quant_config = quant_config
         if self.quant_config.quant_type.size_bits == 4:
             self.quant_type = scalar_types.uint4b8
@@ -646,6 +651,8 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
         logical_to_physical_map: Optional[torch.Tensor] = None,
         logical_replica_count: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
+        assert self.fused_experts is None
+
         if enable_eplb:
             raise NotImplementedError(
                 "EPLB not supported for `GPTQMarlinMoEMethod` yet.")
@@ -662,7 +669,8 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
             num_expert_group=num_expert_group,
             custom_routing_function=custom_routing_function,
             scoring_func=scoring_func,
-            e_score_correction_bias=e_score_correction_bias)
+            e_score_correction_bias=e_score_correction_bias,
+            indices_type=self.topk_indices_dtype)
 
         return torch.ops.vllm.fused_marlin_moe(
             x,