Merge branch 'v0.9.2-dev-ds-wm-1217' into 'v0.9.2-dev-ds'

w8a8 高吞吐模式先量化再dispatch

See merge request dcutoolkit/deeplearing/vllm!303

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

@@ -520,6 +520,8 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
                 False)
             
             return TritonOrGroupGemmExperts(
-                use_int8_w8a8=envs.VLLM_ENABLE_DEEPEP_HT_DEEPGEMM,
+                #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
             )
\ No newline at end of file