MCPcopy Create free account
hub / github.com/GreenBitAI/low_bit_llama / make_quant

Function make_quant

model.py:99–110  ·  view source on GitHub ↗
(module, names, name='', groupsize=-1, double_groupsize=-1, bits=4, v1=True, asym=True)

Source from the content-addressed store, hash-verified

97
98
99def make_quant(module, names, name='', groupsize=-1, double_groupsize=-1, bits=4, v1=True, asym=True):
100 if isinstance(module, QuantLinear):
101 return
102 for attr in dir(module):
103 tmp = getattr(module, attr)
104 name1 = name + '.' + attr if name != '' else attr
105 if name1 in names:
106 setattr(
107 module, attr, QuantLinear(tmp.in_features, tmp.out_features, groupsize=groupsize, double_groupsize=double_groupsize, bits=bits, v1=v1, asym=asym)
108 )
109 for name1, child in module.named_children():
110 make_quant(child, names, name + '.' + name1 if name != '' else name1, groupsize=groupsize, double_groupsize=double_groupsize, bits=bits, v1=v1, asym=asym)
111
112
113def model_to_half(model):

Callers 2

load_llama_modelFunction · 0.85
load_llama_model_loraFunction · 0.85

Calls 1

QuantLinearClass · 0.85

Tested by

no test coverage detected