From e67864d405c3408209779330e066b95ecc4b1ef6 Mon Sep 17 00:00:00 2001 From: Kaihui-intel Date: Fri, 6 Jun 2025 17:18:32 +0800 Subject: [PATCH 1/2] fix g_idx init Signed-off-by: Kaihui-intel --- neural_compressor/transformers/models/modeling_auto.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/neural_compressor/transformers/models/modeling_auto.py b/neural_compressor/transformers/models/modeling_auto.py index b38fa35d8bf..751b756b718 100644 --- a/neural_compressor/transformers/models/modeling_auto.py +++ b/neural_compressor/transformers/models/modeling_auto.py @@ -60,6 +60,7 @@ def build_woq_model(model, quantization_config): bits = quantization_config.bits + g_idx = hasattr(quantization_config, "desc_act") and quantization_config.desc_act is True for n, m in model.named_modules(): if n in quantization_config.modules_to_not_convert: continue @@ -70,7 +71,6 @@ def build_woq_model(model, quantization_config): not getattr(quantization_config, "sym", False), ) use_optimum_format = True - g_idx = hasattr(m, "g_idx") and m.g_idx is not None with init_empty_weights(): new_module = INCWeightOnlyLinear( From ab3c5a9ff1c45c17cb31b30e71151f9c88ab35cf Mon Sep 17 00:00:00 2001 From: "Wang, Chang" Date: Fri, 6 Jun 2025 17:37:10 +0800 Subject: [PATCH 2/2] Update neural_compressor/transformers/models/modeling_auto.py --- neural_compressor/transformers/models/modeling_auto.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/neural_compressor/transformers/models/modeling_auto.py b/neural_compressor/transformers/models/modeling_auto.py index 751b756b718..364b991caa8 100644 --- a/neural_compressor/transformers/models/modeling_auto.py +++ b/neural_compressor/transformers/models/modeling_auto.py @@ -60,7 +60,7 @@ def build_woq_model(model, quantization_config): bits = quantization_config.bits - g_idx = hasattr(quantization_config, "desc_act") and quantization_config.desc_act is True + g_idx = getattr(quantization_config, "desc_act", False) for n, m in model.named_modules(): if n in quantization_config.modules_to_not_convert: continue