PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix: qwen3-next NVFP4 quantization returns act_max incorrectly (#1470)
Xin He229d1fefca6e
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex 3fbd07ab3..de58ec36a 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -2473,6 +2473,9 @@ class BaseCompressor(object):act_max = torch.max(torch.abs(input), dim=-1).valuesif not hasattr(module, "act_max") or module.act_max.numel() == 0:module.act_max = act_max+ if is_nv_fp(self.act_data_type): ## for nvfp per-tensor input_global_scale calculation usage+ max_val = act_max.max()+ module.act_max = max_val.unsqueeze(0) if max_val.dim() == 0 else max_valelse:act_max = act_max.to(module.act_max.device)if is_nv_fp(self.act_data_type): ## for nvfp per-tensor input_global_scale calculation usage