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

remove tricky setting in mxfp4 (#445)

Wenhua Cheng1ae5031eb4f2
diff --git a/auto_round/data_type/mxfp.py b/auto_round/data_type/mxfp.py
index 0f629a027..fb27c3faa 100644
--- a/auto_round/data_type/mxfp.py
+++ b/auto_round/data_type/mxfp.py
@@ -84,9 +84,7 @@ def quant_mx(tensor, bits=4, group_size=-1, v=0, max_scale=1.0,
tensor = tensor.to(torch.float32)
shared_exp = shared_exp.to(torch.float32)
tensor = tensor / (2 ** shared_exp)
- is_mx_fp4 = data_type == "mx_fp4" or ("mx_fp" in data_type and bits == 4)
- multiply = 2 if is_mx_fp4 else 1 ## 2 is a tricky setting
- tensor = tensor + v * multiply
+ tensor = tensor + v
if ebits != 0:
private_exp = floor_ste(torch.log2(torch.abs(tensor) + (tensor == 0).type(tensor.dtype)))