[feat]w8a8高吞吐模式先做量化在做dispatch

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe_marlin.py

@@ -527,6 +527,7 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
                 False)
             
             return TritonOrGroupGemmExperts(
-                use_int8_w8a8=envs.VLLM_ENABLE_DEEPEP_HT_DEEPGEMM,
+                use_int8_w8a8=True,
+                per_act_token_quant=True,
                 fused_experts=self.w8a8_groupgemm_contiguous_forward if envs.VLLM_ENABLE_DEEPEP_HT_DEEPGEMM else self.fused_moe_forward
             )