[Quant][Perf] Use moe_wna16 kernel by default for MoEs with many experts (#13236)

Signed-off-by: mgoin <mgoin64@gmail.com>

vllm/model_executor/layers/quantization/awq_marlin.py

@@ -17,6 +17,7 @@ from vllm.model_executor.layers.quantization.awq import (AWQConfig,
                                                          is_layer_skipped_awq)
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig, QuantizeMethodBase)
+from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config
 from vllm.model_executor.layers.quantization.utils import replace_parameter
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     apply_awq_marlin_linear, awq_to_marlin_zero_points, check_marlin_supported,
@@ -134,7 +135,12 @@ class AWQMarlinConfig(QuantizationConfig):
                     self.full_config).get_quant_method(layer, prefix)
             return AWQMarlinLinearMethod(self)
         elif isinstance(layer, FusedMoE):
-            return AWQMoEMethod(self)
+            if layer.num_experts > 32:
+                # For MoEs with many experts the moe_wna16 kernel is faster
+                return MoeWNA16Config.from_config(
+                    self.full_config).get_quant_method(layer, prefix)
+            else:
+                return AWQMoEMethod(self)
         return None
 
     @classmethod