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

add warnings for lm_head activation scale fallback (#1728)

Heng Guo99d49574e945
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index 69fbb2b57..985ccfc46 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -2079,6 +2079,12 @@ class BaseCompressor(object):
for layer_name in copy.deepcopy(layer_names):
if layer_name not in layer_inputs:
if self.act_bits < 16 and not self.act_dynamic:
+ if "lm_head" in layer_name:
+ logger.warning_once(
+ "Static activation quantization for lm_head is not fully supported yet. "
+ "If lm_head calibration inputs are missing, activation scale may fall back to unit scale "
+ "or quantization may be skipped."
+ )
# Activation quantization requires collected inputs
msg_prefix = (
f"Activation max hook for layer '{layer_name}' is unavailable due to "
diff --git a/auto_round/wrapper.py b/auto_round/wrapper.py
index a48a654be..9b329624e 100644
--- a/auto_round/wrapper.py
+++ b/auto_round/wrapper.py
@@ -538,6 +538,7 @@ class WrapperLinear(torch.nn.Module):
class WrapperWALayer(torch.nn.Module):
+
def __init__(self, orig_layer, enable_torch_compile=False, device="cpu"):
super(WrapperWALayer, self).__init__()
self.orig_layer = orig_layer
@@ -806,6 +807,19 @@ def unwrapper_layer(model, layer, layer_name, best_params):
if hasattr(layer, "orig_layer"):
orig_layer = layer.unwrapper(best_params)
+ act_max = getattr(orig_layer, "act_max", None)
+ act_scale = getattr(orig_layer, "act_scale", None)
+ if (
+ "lm_head" in layer_name
+ and getattr(layer, "enable_act_quant", False)
+ and not getattr(orig_layer, "act_dynamic", True)
+ and act_scale is not None
+ and act_max is None
+ ):
+ logger.warning_once(
+ "Static activation quantization for lm_head is not fully supported yet. "
+ "lm_head activation statistics are missing, so activation scale falls back to unit scale."
+ )
orig_layer = orig_layer.to("cpu")
set_module(model, layer_name, orig_layer)