MLPerf Common - a collection of common MLPerf tools

MLPerf Logging

MLPerf common can be installed via pip install by adding the following line to the requirements.txt file:

git+https://github.com/NVIDIA/mlperf-common.git

Integration using torch.distributed (pytorch)

In mlperf_logger.py module define:

from mlperf_common.logging import MLLoggerWrapper
from mlperf_common.frameworks.pyt import PyTCommunicationHandler

mllogger = MLLoggerWrapper(PyTCommunicationHandler(), value=None)

Then use mllogger by importing from mlperf_logger import mllogger in other modules.

Integration using MPI (horovod/hugectr/mxnet/tensorflow)

In mlperf_logger.py global module define:

from mlperf_common.logging import MLLoggerWrapper
from mlperf_common.frameworks.mxnet import MPICommunicationHandler

mllogger = MLLoggerWrapper(MPICommunicationHandler(), value=None)

Then use mllogger by importing from mlperf_logger import mllogger in other modules.

Optionally, you can pass an MPI communicator during the initialization of MPICommunicationHandler().

comm = MPI.COMM_WORLD
mllogger = MLLoggerWrapper(MPICommunicationHandler(comm), value=None)

by default, MPICommunicationHandler() creates a global communicator.

Logging additional metrics

MLPerf logger can be used to track additional non-required metric, for example throughput. The recommended way is to add a line such as:

mllogger.event(key='tracked_stats', metadata={'step': epoch}, value={"throughput": throughput, "metric_a": metric_a, 'metric_b': metric_b})

where throughput is recommended to be samples per second, logged every epoch or as often as it is reasonable for a given benchmark. Additional metrics, metric_a and metric_b, can represent any numerical value that requires logging. The key tracked_stats and an increasing value for step are required.

Scaleout Bridge

init_bridge

Instead of previous sbridge = init_bridge(rank), initialize sbridge as follows:

from mlperf_common.frameworks.pyt import PyTNVTXHandler, PyTCommunicationHandler

sbridge = init_bridge(PyTNVTXHandler(), PyTCommunicationHandler(), mllogger)

or, for horovod/tf/mxnet:

from mlperf_common.frameworks.mxnet import MXNetNVTXHandler, MPICommunicationHandler

sbridge = init_bridge(MXNetNVTXHandler(), MPICommunicationHandler(), mllogger)

and start your profiling as usual

sbridge.start_prof()
sbridge.stop_prof()

EmptyObject

Current ScaleoutBridgeBase class replaces previous EmptyObject class, so just replace EmptyObject() with ScaleoutBridgeBase().

Mount check

Get mount info

get-mount-info.sh prints a description of given directory and takes one argument: paths_to_verify that contains paths separated by commas.

Example of use:

get-mount-info.sh "/data,/checkpoints"

Example output:

declare -a directory_sizes
declare -a number_of_paths_in_dir
# ----------
directory_sizes+=(
",3743665220,2"
"coco2014,1445240,3"
"laion-400m,3742219976,2"
"coco2014/val2014_512x512_30k,1412012,30000"
"laion-400m/webdataset-moments-filtered,876410532,833"
"laion-400m/webdataset-moments-filtered-encoded,2865809440,832"
)
number_of_paths_in_dir+=(6)
# ----------
directory_sizes+=(
",52403072,5"
"clip,7704556,3"
"inception,93392,1"
"sd,20888768,2"
)
number_of_paths_in_dir+=(4)

number_of_paths_in_dir for each path specified contains the number of subdirectories in it.
The fields in directory_sizes contain 3 values separated by commas:
relative path path to the directory, its size in KB, the number of directories and files inside it.

Verify mounts

verify-mounts.sh checks if a given directory is consistent with a description generated with get-mount-info.sh and takes one argument: paths_to_verify that contains paths separated by commas.