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

refine error msg and dump more log in the tuning (#386)

Wenhua Cheng34fc38ef9ef2
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index 1bfc82883..bd9ef47bf 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -965,7 +965,7 @@ class AutoRound(object):
unwrapper_layer(self.model, wrapper_linear, layer_name, best_params)
mv_module_from_gpu(layer, self.low_cpu_mem_usage)
dump_info = f"quantized {layer_name}, loss iter 0: {init_loss:.6f} -> iter {best_iter}: {last_loss:.6f}"
- logger.debug(dump_info)
+ logger.info(dump_info)
def register_act_max_hook(self, model):
def get_act_max_hook(module, input, output):
@@ -1136,7 +1136,7 @@ class AutoRound(object):
f"quantized {len(quantized_layer_names)}/{(len(quantized_layer_names) + len(unquantized_layer_names))} "
f"layers in the block, loss iter 0: {init_loss:.6f} -> iter {best_iter}: {last_loss:.6f}"
)
- logger.debug(dump_info)
+ logger.info(dump_info)
if len(unquantized_layer_names) != 0:
logger.info(f"{unquantized_layer_names} have not been quantized")
with torch.no_grad():
diff --git a/auto_round/backend.py b/auto_round/backend.py
index d64786123..fe05cd797 100644
--- a/auto_round/backend.py
+++ b/auto_round/backend.py
@@ -530,7 +530,7 @@ def get_layer_backend(device, backend, orig_backend, bits, group_size, sym, in_f
try:
require_version(requirement)
except ImportError:
- logger.error(f"pip install {requirement}")
+ logger.error(f"pip install '{requirement}' ")
else:
str_info = requirement()[1]
logger.error(str_info)