PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
limit the scale minimum value not to 0 (#211)
WeiweiZhang11dd2708d84f2
diff --git a/auto_round/data_type/int.py b/auto_round/data_type/int.pyindex 726a2b9c8..19a2a6865 100644--- a/auto_round/data_type/int.py+++ b/auto_round/data_type/int.py@@ -19,7 +19,7 @@ from auto_round.data_type.register import register_dtype, QUANT_FUNC_WITH_DTYPE@register_dtype("int_asym")def quant_tensor_asym(weight, num_bits=4, v=0, min_scale=1.0, max_scale=1.0, scale_dtype=torch.float16,- weight_min=None, weight_max=None, q_scale_thresh=0.0,**kwargs):+ weight_min=None, weight_max=None, q_scale_thresh=0.0, **kwargs):"""Quantizes and dequantizes weight asymmetrically.Args:@@ -59,6 +59,8 @@ def quant_tensor_asym(weight, num_bits=4, v=0, min_scale=1.0, max_scale=1.0, scawmax[tmp] = +1scale = ((wmax - wmin) / maxq).to(scale_dtype)scale = torch.clamp(scale, min=q_scale_thresh)+ if (scale == 0.).any():+ scale = torch.clamp(scale, min=1e-5)zp = round_ste(-wmin / scale) # pylint: disable=E1130scale = scale.unsqueeze(dim=-1)zp = zp.unsqueeze(dim=-1)@@ -68,7 +70,7 @@ def quant_tensor_asym(weight, num_bits=4, v=0, min_scale=1.0, max_scale=1.0, sca@register_dtype("int_sym")def quant_tensor_sym(weight, num_bits=4, v=0, min_scale=1.0, max_scale=1.0, scale_dtype=torch.float16, weight_min=None,- weight_max=None, q_scale_thresh=0.0,**kargs):+ weight_max=None, q_scale_thresh=0.0, **kargs):"""Quantizes and dequantizes weight symmetrically.Args:@@ -114,9 +116,12 @@ def quant_tensor_sym(weight, num_bits=4, v=0, min_scale=1.0, max_scale=1.0, scalwmax_new[tmp] = +1scale = ((wmax_new - wmin_new) / maxq).to(scale_dtype)scale = torch.clamp(scale, min=q_scale_thresh)+ if (scale == 0.).any():+ scale = torch.clamp(scale, min=1e-5)scale = scale.unsqueeze(dim=-1)zp = torch.full_like(scale, (maxq + 1) / 2)int_w = round_ste(weight / scale + v)q = torch.clamp(int_w + zp, 0, maxq)return scale * (q - zp), scale, zp+