Merge branch 'v0.9.2-dev-fth' into 'v0.9.2-dev'

解决w8a8 pp16开启marlin的oom问题

See merge request dcutoolkit/deeplearing/vllm!248

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

@@ -127,7 +127,7 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
             w1_marlin_in = get_w8a8_int8_marlin_weights(layer.w13_weight[ii])
             w1_marlin_list.append(w1_marlin_in)
         w1_marlin = torch.stack(w1_marlin_list, dim=0)
-
+        del w1_marlin_list
         w2_marlin_list = []
         for ii in range(layer.w2_weight.shape[0]):
             w2_marlin_in = get_w8a8_int8_marlin_weights(layer.w2_weight[ii])