print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_profile(
nn.parallel.DataParallel(AsymmetricQuantizer(
QuantizerConfig(QuantizationParams(bits=NBITS))).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
# CUDA DataParallel high batch
# wall time
print()
print("CUDA DataParallel high batch")
print("------------------------------------------------")
print("Pytorch Symmetric(cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_wall(
nn.parallel.DataParallel(ReferenceQuantize(NBITS).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
print()
print("Custom Symmetric (cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_wall(
nn.parallel.DataParallel(SymmetricQuantizer(
QuantizerConfig(QuantizationParams(bits=NBITS))).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
print()
print("Custom Assymetric (cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_wall(
run_profile(
nn.parallel.DataParallel(AsymmetricQuantizer(
DefaultedPTQuantizerSpec(
num_bits=NBITS, scale_shape=per_tensor_scale_shape)).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
# CUDA DataParallel high batch
# wall time
print()
print("CUDA DataParallel high batch")
print("------------------------------------------------")
print("Pytorch Symmetric(cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_wall(
nn.parallel.DataParallel(ReferenceQuantize(NBITS).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
print()
print("Custom Symmetric (cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))
run_wall(
nn.parallel.DataParallel(SymmetricQuantizer(
DefaultedPTQuantizerSpec(
num_bits=NBITS, scale_shape=per_tensor_scale_shape)).cuda(),
device_ids=device_ids), HIGH_BATCH_INPUT_SIZE,
'cuda', GPU_RUNS_HIGH_BATCH)
print()
print("Custom Asymmetric (cuda {0}) DataParallel impl:".format(device_ids))
print("input size: {0}".format(HIGH_BATCH_INPUT_SIZE))