From 1974cec23a3839d032d4befe172c88bd1d557151 Mon Sep 17 00:00:00 2001 From: WangQing <2917021186@qq.com> Date: Sun, 4 Jan 2026 11:09:13 +0000 Subject: [PATCH] [ascend] fix awq --- dlinfer/framework/lmdeploy_ext/quants/ascend_awq.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/dlinfer/framework/lmdeploy_ext/quants/ascend_awq.py b/dlinfer/framework/lmdeploy_ext/quants/ascend_awq.py index e9bbcaff..7fe0c290 100644 --- a/dlinfer/framework/lmdeploy_ext/quants/ascend_awq.py +++ b/dlinfer/framework/lmdeploy_ext/quants/ascend_awq.py @@ -132,7 +132,10 @@ def AscendMergedAwqLinear__init__( device: Optional[torch.device] = None, is_tp: bool = True, out_names: Optional[List[int]] = None, + layer_type: str = "attn", + dtype: Optional[torch.dtype] = torch.float16, ): + self.init_tp_args(is_tp, all_reduce=False, colwise=True, layer_type=layer_type) if replicate is None: replicate = tuple(False for _ in all_out_features) @@ -160,6 +163,8 @@ def AscendMergedAwqLinear__init__( device, colwise=True, is_tp=is_tp, + layer_type=layer_type, + dtype=dtype, ) self.qweight.weight_loader = self.weight_loader self.qweight.weight_spliter = self.weight_spliter_wz