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

fix nblocks issues (#432)

Wenhua Chengedc2b034fe12
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index b0955f15b..5c0f0c66e 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -176,7 +176,6 @@ class AutoRound(object):
self.enable_quanted_input = enable_quanted_input
self.enable_minmax_tuning = enable_minmax_tuning
self.nsamples = nsamples
- self.nblocks = nblocks
self.bits = bits
self.enable_norm_bias_tuning = enable_norm_bias_tuning
self.group_size = group_size
@@ -1324,15 +1323,15 @@ class AutoRound(object):
if pbar is None:
pbar = tqdm(range(0, len(block_names), nblocks))
- # for i in pbar:
- for i in range(len(block_names)):
+
+ for i in range(0, len(block_names), nblocks):
if nblocks == 1:
n = block_names[i]
pbar.set_description(f"Quantizing {n}")
m = get_module(model, n)
else:
- names = block_names[i: i + nblocks]
- pbar.set_description(f"Quantizing [{i + 1}-{i + nblocks}]/{len(block_names)}")
+ names = block_names[i: min(i + nblocks, len(block_names))]
+ pbar.set_description(f"Quantizing [{i + 1}-{min(i + nblocks, len(block_names))}]/{len(block_names)}")
modules = [get_module(model, n) for n in names]
m = WrapperMultiblock(modules)