Running SDK on a Wafer-Scale Cluster¶

Note

The Cerebras Wafer-Scale Cluster appliance running Cerebras ML Software 2.2 supports SDK 1.1. See here for SDK 1.1 documentation.

The Cerebras Wafer-Scale Cluster appliance running Cerebras ML Software 2.3 supports SDK 1.2, the current version of SDK software.

In addition to the containerized Singularity build of the Cerebras SDK (see Installation and Setup) for installation information), the SDK is also supported on Cerebras Wafer-Scale Clusters running in appliance mode.

This page documents some modifications needed to your code to run on a Wafer-Scale Cluster. For more information about setting up and using a Wafer-Scale Cluster, see the documention here.

Summary¶

The Cerebras Wafer-Scale Cluster is our solution to training massive neural networks with near-linear scaling. The Wafer-Scale Cluster consists of one or more CS systems, together with special CPU nodes, memory servers, and interconnects, presented to the end user as a single system, or appliance. The appliance is responsible for job scheduling and allocation of the systems.

There are two types of SDK jobs that can run on the appliance: compile jobs, which are used to compile code on a worker node, and run jobs, which either run the compiled code on a worker node using the simulator, or run the code on a real CS system within the appliance.

We will walk through some changes necessary to compile and run your code on a Wafer-Scale Cluster. Modified code examples for supporting a Wafer-Scale Cluster can be requested from developer@cerebras.net.

Note that there are currently some limitations for running SDK jobs on a Wafer-Scale Cluster. Unlike ML jobs, SDK jobs can only use a single worker node and CS system.

Setup and Installation¶

First, learn about the components of the Cerebras Wafer-Scale Cluster here.

The user interacts with the Wafer-Scale Cluster via a user node. Start by setting up a Python virtual environment on the user node:

$ python3.8 -m venv sdk_venv
$ source sdk_venv/bin/activate

Next, install the cerebras_appliance and cerebras_sdk Python wheels in the virtual enviroment, specifying the proper Cerebras Software release:

(sdk_venv) $ pip install --upgrade pip
(sdk_venv) $ pip install cerebras_appliance==2.3.0
(sdk_venv) $ pip install cerebras_sdk==2.3.0

Compiling¶

As an example, we’ll walk through porting GEMV 1: A Complete Program. In the containerized SDK setup, this code is compiled with the following command:

cslc ./layout.csl --fabric-dims=8,3 --fabric-offsets=4,1 --memcpy --channels=1 -o out

To compile for the Wafer-Scale cluster, we use a Python script which launches a compile job:

import json
from cerebras.sdk.client import SdkCompiler

# Instantiate copmiler
compiler = SdkCompiler()

# Launch compile job
artifact_path = compiler.compile(
    ".",
    "layout.csl",
    "--fabric-dims=8,3 --fabric-offsets=4,1 --memcpy --channels=1 -o out",
    "."
)

# Write the artifact_path to a JSON file
with open("artifact_path.json", "w", encoding="utf8") as f:
    json.dump({"artifact_path": artifact_path,}, f)

The SdkCompiler::compile function takes three arguments:

the directory containing the CSL code files,

the name of the top level CSL code file that contains the layout block,

the compiler arguments,

and the output directory or output file for the compile artifacts.

The last argument can either be a directory, specifying the location to which compile artifacts will be copied with default file name; or a file name, explicitly specifying the name and location for the compile artifacts. The function returns the compile artifact path. We write this artifact path to a JSON file which will be read by the runner object in the Python host code.

Just as before, simply pass the full dimensions of the target system to the --fabric-dims argument to compile for a real hardware run.

The SdkCompiler() constructor takes two optional kwargs:

resource_cpu: amount of CPU cores on the WSC’s management node used by the compile job in units of 1/1000 CPU (default: 24000, or 24 cores)

resource_mem: number of bytes of memory requested from the management node for the compile job (default: 64 << 30, or 64 GiB)

If SDK compilation jobs on the WSC are often waiting in the queue behind other jobs, such as ML execute or run jobs, this is typically because not enough resources are available on the management node. These kwargs can be used to request less resources from the management node and increase the number of simultaneously running jobs.

Running¶

In the containerized SDK setup, our Python host code for running is as follows:

import argparse
import numpy as np

from cerebras.sdk.runtime.sdkruntimepybind import SdkRuntime, MemcpyDataType, MemcpyOrder

# Read arguments
parser = argparse.ArgumentParser()
parser.add_argument('--name', help="the test compile output dir")
parser.add_argument('--cmaddr', help="IP:port for CS system")
args = parser.parse_args()

# Matrix dimensions
M = 4
N = 6

# Construct A, x, b
A = np.arange(M*N, dtype=np.float32).reshape(M, N)
x = np.full(shape=N, fill_value=1.0, dtype=np.float32)
b = np.full(shape=M, fill_value=2.0, dtype=np.float32)

# Calculate expected y
y_expected = A@x + b

# Construct a runner using SdkRuntime
runner = SdkRuntime(args.name, cmaddr=args.cmaddr)

# Load and run the program
runner.load()
runner.run()

# Launch the init_and_compute function on device
runner.launch('init_and_compute', nonblock=False)

# Copy y back from device
y_symbol = runner.get_id('y')
y_result = np.zeros([1*1*M], dtype=np.float32)
runner.memcpy_d2h(y_result, y_symbol, 0, 0, 1, 1, M, streaming=False,
order=MemcpyOrder.ROW_MAJOR, data_type=MemcpyDataType.MEMCPY_32BIT, nonblock=False)

# Stop the program
runner.stop()

# Ensure that the result matches our expectation
np.testing.assert_allclose(y_result, y_expected, atol=0.01, rtol=0)
print("SUCCESS!")

If this file is named run.py and the compilation output is in the directory out, we run it with the command:

cs_python run.py --name out

To run on hardware, we specify an IP address with the --cmaddr flag.

For appliance mode, we need a few modifications to the host code:

import json
import os

import numpy as np

from cerebras.appliance.pb.sdk.sdk_common_pb2 import MemcpyDataType, MemcpyOrder
from cerebras.sdk.client import SdkRuntime

# Matrix dimensions
M = 4
N = 6

# Construct A, x, b
A = np.arange(M*N, dtype=np.float32).reshape(M, N)
x = np.full(shape=N, fill_value=1.0, dtype=np.float32)
b = np.full(shape=M, fill_value=2.0, dtype=np.float32)

# Calculate expected y
y_expected = A@x + b

# Read the artifact_path from the JSON file
with open("artifact_path.json", "r", encoding="utf8") as f:
    data = json.load(f)
    artifact_path = data["artifact_path"]

# Instantiate a runner object using a context manager.
# Set simulator=False if running on CS system within appliance.
with SdkRuntime(artifact_path, simulator=True) as runner:
    # Launch the init_and_compute function on device
    runner.launch('init_and_compute', nonblock=False)

    # Copy y back from device
    y_symbol = runner.get_id('y')
    y_result = np.zeros([1*1*M], dtype=np.float32)
    runner.memcpy_d2h(y_result, y_symbol, 0, 0, 1, 1, M, streaming=False,
    order=MemcpyOrder.ROW_MAJOR, data_type=MemcpyDataType.MEMCPY_32BIT, nonblock=False)

# Ensure that the result matches our expectation
np.testing.assert_allclose(y_result, y_expected, atol=0.01, rtol=0)
print("SUCCESS!")

In particular, note that:

The imports have changed to reflect appliance modules.

We read the path of our compile artifacts from the JSON file generated when compiling.

We no longer need to specify a CM address when running on real hardware. Instead, we simply pass a flag to the SdkRuntime constructor specifying whether to run in the simulator or on hardware.

load() and run() are replaced by start().

We can use a context manager for the runner object. If we do so, the start() and stop() functions are implicit, and we do not need to explicitly call them.

Without a context manager, you must call start() and stop() explicitly.

Running with SdkLauncher¶

We also provide an SdkLauncher class for running host code directly from a worker node within the appliance. This class allows files to be staged on the appliance before running the same host code that you would with the Singularity container.

We demonstrate below using SdkLauncher to run the host code for the example above. We also include a demonstration of stage for transferring a file to the appliance.

import json
import os

from cerebras.sdk.client import SdkLauncher

# read the compile artifact_path from the json file
with open("artifact_path.json", "r", encoding="utf8") as f:
    data = json.load(f)
    artifact_path = data["artifact_path"]

# artifact_path contains the path to the compiled artifact.
# It will be transferred and extracted in the appliance.
# The extracted directory will be the working directory.
# Set simulator=False if running on CS system within appliance.
with SdkLauncher(artifact_path, simulator=True) as launcher:

    # Transfer an additional file to the appliance,
    # then write contents to stdout on appliance
    launcher.stage("additional_artifact.txt")
    response = launcher.run(
        "echo \"ABOUT TO RUN IN THE APPLIANCE\"",
        "cat additional_artifact.txt",
    )
    print("Test response: ", response)

    # Run the original host code as-is on the appliance,
    # using the same cmd as when using the Singularity container
    response = launcher.run("cs_python run.py --name out")
    print("Host code execution response: ", response)

    # Fetch files from the appliance
    launcher.download_artifact("sim.log", "./output_dir/sim.log")

SDK Documentation (1.2.0)

Running SDK on a Wafer-Scale Cluster

Contents