def test_can_broadcast_initialized_precisions_in_distributed_mode(
config_builder, tmp_path):
config = config_builder.build()
ngpus_per_node = torch.cuda.device_count()
config.world_size = ngpus_per_node
torch.multiprocessing.spawn(hawq_dumping_worker,
nprocs=ngpus_per_node,
args=(ngpus_per_node, config, tmp_path),
join=True)
assert not compare_multi_gpu_dump(config, tmp_path,
get_path_to_bitwidth_dump)
def test_hawq_broadcast_avg_traces_in_distributed_mode(tmp_path):
num_data_points = 10
batch_size = 2
config = HAWQConfigBuilder(batch_size=batch_size, num_data_points=num_data_points, image_size=224).build()
ngpus_per_node = torch.cuda.device_count()
config.world_size = ngpus_per_node
torch.multiprocessing.spawn(hawq_dumping_worker,
nprocs=ngpus_per_node,
args=(ngpus_per_node, config, tmp_path),
join=True)
assert not compare_multi_gpu_dump(config, tmp_path, get_path_to_bitwidth_dump)
def test_multiprocessing_distributed_shares_init_scales_signedness_across_gpus(
tmp_path):
num_init_steps = 10
config = get_squeezenet_quantization_config()
config['compression']['initializer'] = {
'range': {
'num_init_steps': num_init_steps
}
}
ngpus_per_node = torch.cuda.device_count()
config.world_size = ngpus_per_node
torch.multiprocessing.spawn(scale_signed_dumping_worker,
nprocs=ngpus_per_node,
args=(ngpus_per_node, config, tmp_path),
join=True)
assert not compare_multi_gpu_dump(config, tmp_path,
get_path_after_broadcast)
assert not compare_multi_gpu_dump(config, tmp_path,
get_path_path_after_train_iters)