MCPcopy Create free account
hub / github.com/deepspeedai/DeepSpeed / _quantize_param

Function _quantize_param

deepspeed/inference/quantization/utils.py:219–236  ·  view source on GitHub ↗
(param: nn.Parameter, quant_config: Dict)

Source from the content-addressed store, hash-verified

217
218
219def _quantize_param(param: nn.Parameter, quant_config: Dict):
220 assert not hasattr(param, 'weight_quantized'), 'Parameter has already been quantized.'
221 quantizer = Quantizer(quant_config)
222 dequantizer = DeQuantizer(quant_config, param.dtype)
223
224 quantized_weight, quant_scale, quant_min = quantizer.quantize(param.data)
225
226 quantized_weight = quantized_weight.view(param.dtype)
227 quant_scale = quant_scale.view(param.dtype)
228 quant_min = quant_min.view(param.dtype)
229
230 quantized_compat_tensor = concat_to_compat_param(quantized_weight, quant_scale, quant_min)
231 param.data = quantized_compat_tensor
232 param.deconcat = quantized_compat_tensor.deconcat
233
234 param.quantizer = quantizer
235 param.dequantizer = dequantizer
236 setattr(param, 'weight_quantized', True)
237
238
239def wrap_quantized_functional(f):

Callers 1

_post_init_methodMethod · 0.90

Calls 4

quantizeMethod · 0.95
DeQuantizerClass · 0.85
concat_to_compat_paramFunction · 0.85
QuantizerClass · 0.70

Tested by

no test coverage detected