simcop2387/AutoGPTQ
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages 1 Wiki Activity

import exllama QuantLinear instead of exllamav2's

Browse source
This commit is contained in:
student686 2023-09-27 11:05:13 +08:00
parent 3b81fb5ea0
commit c1a3013c45
1 changed files with 2 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
auto_gptq/modeling/_utils.py
View file

@ -188,7 +188,7 @@ def pack_model(
warmup_triton: bool = False, warmup_triton: bool = False,
force_layer_back_to_cpu: bool = False force_layer_back_to_cpu: bool = False
): ):
QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits) QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=False, disable_exllamav2=True)
if force_layer_back_to_cpu: if force_layer_back_to_cpu:
model.to(CPU) model.to(CPU)
@ -196,7 +196,7 @@ def pack_model(
logger.info('Packing model...') logger.info('Packing model...')
layers = find_layers(model) layers = find_layers(model)
layers = {n: layers[n] for n in quantizers} layers = {n: layers[n] for n in quantizers}
make_quant(model, quantizers, bits, group_size, use_triton=use_triton, use_cuda_fp16=use_cuda_fp16, desc_act=desc_act) make_quant(model, quantizers, bits, group_size, use_triton=use_triton, use_cuda_fp16=use_cuda_fp16, desc_act=desc_act, disable_exllama=False, disable_exllamav2=True)
qlayers = find_layers(model, [QuantLinear]) qlayers = find_layers(model, [QuantLinear])
for name in qlayers: for name in qlayers:
logger.info(name) logger.info(name)