aiu-fms-testing-utils 0.2.0

Spyre FMS Testing Utils

Installation Guide for AIU FMS Testing Utilities

This guide provides instructions for installing the aiu-fms-testing-utils package.

Installation with CPU-Only PyTorch

To install the package with the CPU-only version of PyTorch, run the following command:

pip3 install aiu-fms-testing-utils --extra-index-url=https://download.pytorch.org/whl/cpu

Installation with Default PyTorch

To install the package with the platform's default PyTorch version, execute:

pip3 install aiu-fms-testing-utils

Verify the PyTorch Version

To ensure compatibility with aiu-fms-testing-utils, verify that the correct PyTorch version is installed.

1. Check the PyTorch Version

   pip show torch
   

   Expected Output for CPU-Only PyTorch:

   Name: torch
   Version: 2.7.1+cpu
   

   For the CPU-only version, the version string should include the +cpu postfix (e.g., 2.7.1+cpu).

2. Corrective Action for Default PyTorch

   If the installed PyTorch version does not include the +cpu postfix (e.g., 2.7.1 without +cpu), it indicates that the default PyTorch version (which may include CUDA support) was installed. To resolve this, uninstall torch and reinstall aiu-fms-testing-utils with the CPU-only version:

   pip3 uninstall torch -y
   pip3 install aiu-fms-testing-utils --extra-index-url=https://download.pytorch.org/whl/cpu
   

Setting Up the Development Environment from Source

To set up the development environment for aiu-fms-testing-utils from source, follow these steps.

In this directory, checkout the Foundation Model Stack (FMS) and the FMS Model Optimizer:

git clone https://github.com/foundation-model-stack/foundation-model-stack.git
git clone https://github.com/foundation-model-stack/fms-model-optimizer.git

Install both FMS, FMS-Model-Optimizer and aiu-fms-testing-utils:

cd foundation-model-stack
pip install -e .
cd ..

cd fms-model-optimizer
pip install -e .
cd ..

pip install -e .

Running in OpenShift

Use the pod.yaml file to get started with your OpenShift allocation

• Modify the ibm.com/aiu_pf_tier0 values to indicate the number of AIUs that you want to use
• Modify the namespace to match your namespace/project (i.e., oc project)

Start the pod

oc apply -f pod.yaml

Copy this repository into the pod (includes scripts, FMS stack)

oc cp ${PWD} my-workspace:/tmp/

Exec into the pod

 oc rsh my-workspace bash -l

When you are finished, make sure to delete your pod:

oc delete -f pod.yaml

Setup the environment in the container

Verify the AIU discovery has happened by looking for output like the following when you exec into the pod:

---- IBM AIU Device Discovery...
---- IBM AIU Environment Setup... (Generate config and environment)
---- IBM AIU Devices Found: 2
------------------------
[1000760000@my-workspace ~]$  echo $AIU_WORLD_SIZE
2

Inside the container, setup envars to use the FMS:

export HOME=/tmp
cd ${HOME}/aiu-fms-testing-utils/foundation-model-stack/
# Install the FMS stack
pip install -e .

Run with AIU instead of, default, senulator.

export FLEX_COMPUTE=SENTIENT
export FLEX_DEVICE=PF

Optional envars to supress debugging output:

export DTLOG_LEVEL=error
export TORCH_SENDNN_LOG=CRITICAL
export DT_DEEPRT_VERBOSE=-1

How to use Foundation Model Stack (FMS) on AIU hardware

The scripts directory provides various scripts to use FMS on AIU hardware for many use cases. These scripts provide robust support for passing desired command line options for running encoder and decoder models along with other use cases. Refer to the documentation on using different scripts for more details.

The examples directory provides small examples aimed at helping understand the general workflow of running a model using FMS on AIU hardware.

Common Errors

Pod connection error

Errors like the following often indicate that the pod has not started or is still in the process of starting.

error: unable to upgrade connection: container not found ("my-pod")

Use oc get pods to check on the status. ContainerCreating indicates that the pod is being created. Running indicates that it is ready to use.

If there is an error the use oc describe pod/my-workspace to see a full diagnostic view. The Events list at the bottom will often let you know what the problem is.

torchrun generic error

Below is the generic torchrun failed program trace. It is not helpful when trying to find the problem in the program. Instead look for the actual error message a little higher in the output trace.

[2024-09-16 16:10:15,705] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: 1) local_rank: 0 (pid: 1479484) of binary: /usr/bin/python3
Traceback (most recent call last):
  File "/usr/local/bin/torchrun", line 8, in <module>
    sys.exit(main())
  File "/usr/local/lib64/python3.9/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
    return f(*args, **kwargs)
  File "/usr/local/lib64/python3.9/site-packages/torch/distributed/run.py", line 812, in main
    run(args)
  File "/usr/local/lib64/python3.9/site-packages/torch/distributed/run.py", line 803, in run
    elastic_launch(
  File "/usr/local/lib64/python3.9/site-packages/torch/distributed/launcher/api.py", line 135, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/usr/local/lib64/python3.9/site-packages/torch/distributed/launcher/api.py", line 268, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
./roberta.py FAILED
------------------------------------------------------------
Failures:
  <NO_OTHER_FAILURES>
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
  time      : 2024-09-16_16:10:15
  host      : ibm-aiu-rdma-jjhursey
  rank      : 0 (local_rank: 0)
  exitcode  : 1 (pid: 1479484)
  error_file: <N/A>
  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
============================================================

Additional warnings

You may see the following additional warnings/notices printed to the console. They are normal and expected at this point in time. The team will work on cleaning these up.

CUDA extension not installed.
using tensor parallel
ignoring module=Embedding when distributing module
[WARNING] Keys from checkpoint (adapted to FMS) not copied into model: {'roberta.embeddings.token_type_embeddings.weight', 'lm_head.bias'}