`neural_compressor.compression.pruner.model_slim.auto_slim`

Auto slim.

Module Contents

`model_slim`(model[, dataloader, round_multiplier])	Slim the sparse model automatically.
`model_slim_ffn2`(model[, dataloader, round_multiplier])	Remove some sparse part in the model permanently and obtain acceleration directly.
`model_slim_mha`(model[, dataloader])	Remove some sparse part in the model permanently and obtain acceleration directly.
`parse_auto_slim_config`(model[, dataloader, ...])	Get model slim pruning configs.
`generate_ffn2_pruning_config`(model, dataloader, ...)	Get consecutive linear layers pruning configs.
`generate_mha_pruning_config`(model, dataloader, ...)	Get multi-head attention layers pruning configs.

neural_compressor.compression.pruner.model_slim.auto_slim.model_slim(model, dataloader=None, round_multiplier=32)[source]: Slim the sparse model automatically.

neural_compressor.compression.pruner.model_slim.auto_slim.model_slim_ffn2(model, dataloader=None, round_multiplier=32)[source]

Remove some sparse part in the model permanently and obtain acceleration directly.

Parameters:

model – a sprase model.
round_multiplier (int) – the channel number after slimming should be multiple of this number.

neural_compressor.compression.pruner.model_slim.auto_slim.model_slim_mha(model, dataloader=None)[source]

Remove some sparse part in the model permanently and obtain acceleration directly.

neural_compressor.compression.pruner.model_slim.auto_slim.parse_auto_slim_config(model, dataloader=None, ffn2_sparsity=0.0, mha_sparsity=0.0, **kwargs)[source]: Get model slim pruning configs.

neural_compressor.compression.pruner.model_slim.auto_slim.generate_ffn2_pruning_config(model, dataloader, ffn2_sparsity, **kwargs)[source]: Get consecutive linear layers pruning configs.

neural_compressor.compression.pruner.model_slim.auto_slim.generate_mha_pruning_config(model, dataloader, mha_sparsity, **kwargs)[source]: Get multi-head attention layers pruning configs.