NNCF use group size instead of number of groups and set default group size for int4 to 64

This commit is contained in:
Disty0
2025-05-11 20:38:01 +03:00
parent 03d05b6243
commit 4eedeab9f8
2 changed files with 17 additions and 13 deletions
+16 -11
View File
@@ -62,21 +62,26 @@ def nncf_compress_layer(layer, num_bits, is_asym_mode, torch_dtype=None, quant_c
reduction_axes = [i for i in range(layer.weight.ndim) if i != 1]
else:
reduction_axes = -1
if shared.opts.nncf_compress_weights_num_of_groups > 1 or num_bits == 4:
num_of_groups = shared.opts.nncf_compress_weights_num_of_groups
if shared.opts.nncf_compress_weights_group_size > 0 or (num_bits == 4 and shared.opts.nncf_compress_weights_group_size != -1):
group_size = shared.opts.nncf_compress_weights_group_size
channel_size = layer.weight.shape[-1]
if num_of_groups > channel_size:
group_size = 1
num_of_groups = channel_size
elif num_of_groups == 0:
group_size = 128
if group_size == 0:
group_size = 64
num_of_groups = channel_size // group_size
if group_size >= channel_size:
group_size = channel_size
num_of_groups = 1
else:
group_size = channel_size / num_of_groups
while channel_size % group_size != 0: # find something divisible
num_of_groups -= 1
group_size = channel_size / num_of_groups
num_of_groups = channel_size // group_size
while channel_size % group_size != 0: # find something divisible
num_of_groups -= 1
if num_of_groups <= 1:
group_size = channel_size
num_of_groups = 1
break
group_size = channel_size / num_of_groups
if num_of_groups > 1:
result_shape = layer.weight.shape