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

Function ds_quantizer

deepspeed/ops/quantizer/quantizer.py:14–29  ·  view source on GitHub ↗
(input, groups=1, bit_num=8, sr=False, asym=False)

Source from the content-addressed store, hash-verified

12
13
14def ds_quantizer(input, groups=1, bit_num=8, sr=False, asym=False):
15 # Load cuda modules if needed
16 global quantizer_cuda_module
17 if quantizer_cuda_module is None:
18 quantizer_cuda_module = QuantizerBuilder().load()
19 if sr:
20 if asym:
21 quantize_func = quantizer_cuda_module.ds_sr_quantize_asym_fp16 if input.dtype == torch.half else quantizer_cuda_module.ds_sr_quantize_asym_fp32
22 else:
23 quantize_func = quantizer_cuda_module.ds_sr_quantize_fp16 if input.dtype == torch.half else quantizer_cuda_module.ds_sr_quantize_fp32
24 else:
25 if asym:
26 quantize_func = quantizer_cuda_module.ds_quantize_asym_fp16 if input.dtype == torch.half else quantizer_cuda_module.ds_quantize_asym_fp32
27 else:
28 quantize_func = quantizer_cuda_module.ds_quantize_fp16 if input.dtype == torch.half else quantizer_cuda_module.ds_quantize_fp32
29 return quantize_func(input, groups, bit_num)

Callers 1

compute_quantizationMethod · 0.90

Calls 2

QuantizerBuilderClass · 0.90
loadMethod · 0.45

Tested by

no test coverage detected