Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

fix autogptq exporting issue at group_size=-1

wenhuach21eb0bb2d1e7fd
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index efed7988b..1167d7512 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -1074,8 +1074,12 @@ class AutoRound(object):
if hasattr(m, "scale"):
self.weight_config[n]["scale"] = m.scale
self.weight_config[n]["zp"] = m.zp
- self.weight_config[n]["g_idx"] = torch.tensor(
- [i // self.group_size for i in range(m.weight.shape[1])], dtype=torch.int32, device="cpu")
+ if self.group_size <= 0:
+ self.weight_config[n]["g_idx"] = torch.tensor(
+ [0 for i in range(m.weight.shape[1])], dtype=torch.int32, device="cpu")
+ else:
+ self.weight_config[n]["g_idx"] = torch.tensor(
+ [i // self.group_size for i in range(m.weight.shape[1])], dtype=torch.int32, device="cpu")
delattr(m, "scale")
delattr(m, "zp")
else: