This is the code for the Discord bot we'll be using to queue jobs to a cluster of GPUs that our generous sponsors have provided. Our goal is to be able to queue kernels that can run end to end in seconds that way things feel interactive and social.
The key idea is that we're using Github Actions as a job scheduling engine and primarily making the Discord bot interact with the cluster via issuing Github Actions and and monitoring their status and while we're focused on having a nice user experience on discord.gg/gpumode, we're happy to accept PRs that make it easier for other Discord communities to hook GPUs.
- Supported Schedulers
- Local Development
- Available Commands
- Using the Leaderboard
- Testing the Discord Bot
- How to Add a New GPU to the Cluster
- Acknowledgements
- GitHub Actions
- Modal
- Slurm (not implemented yet)
Important
Do not fork this repository. Instead, directly clone this repository to your local machine.
Important
Python 3.11 or higher is required.
After, install the dependencies with pip install -r requirements.txt.
To run and develop the bot locally, you need to add it to your own "staging" server. Follow the steps here and here to create a bot application and then add it to your staging server.
Below is a visual walk-through of the steps linked above:
-
The bot needs the
Message Content IntentandServer Members Intentpermissions turned on.Click here for visual.
-
The bot needs
applications.commandsandbotscopes.Click here for visual.
-
Finally, generate an invite link for the bot and enter it into any browser.
Click here for visual.
Note
Bot permissions involving threads/mentions/messages should suffice, but you can naively give it Administrator since it's just a test bot in your own testing Discord server.
The leaderboard persists information in a Postgres database. To develop locally, set Postgres up on your machine. Then start a Postgres shell with psql, and create a database:
$ psql -U postgres
Password for user postgres: ********
psql (16.6 (Ubuntu 16.6-1.pgdg22.04+1))
Type "help" for help.
postgres=# CREATE DATABASE clusterdev;
We are using Yoyo Migrations to manage tables, indexes, etc. in our database. To create tables in your local database, apply the migrations in src/discord-cluster-manager/migrations with the following command line:
yoyo apply src/discord-cluster-manager/migrations \
-d postgresql://user:password@localhost/clusterdev
Click here for a transcript of a yoyo apply session
$ yoyo apply . -d postgresql://user:password@localhost/clusterdev
[20241208_01_p3yuR-initial-leaderboard-schema]
Shall I apply this migration? [Ynvdaqjk?]: y
Selected 1 migration:
[20241208_01_p3yuR-initial-leaderboard-schema]
Apply this migration to postgresql://user:password@localhost/clusterdev [Yn]: y
Save migration configuration to yoyo.ini?
This is saved in plain text and contains your database password.
Answering 'y' means you do not have to specify the migration source or database connection for future runs [yn]: n
Applying migrations to our staging and prod environments also happens using yoyo apply, just with a different database URL.
To make changes to the structure of the database, create a new migration:
yoyo new src/discord-cluster-manager/migrations -m "short_description"
...and then edit the generated file. Please do not edit existing migration files: the existing migration files form a sort of changelog that is supposed to be immutable, and so yoyo will refuse to reapply the changes.
We are following an expand/migrate/contract pattern to allow database migrations without downtime. When you want to make a change to the structure of the database, first determine if it is expansive or contractive.
- Expansive changes are those that have no possibility of breaking a running application. Examples include: adding a new nullable column, adding a non-null column with a default value, adding an index, adding a table, etc.
- Contractive changes are those that could break a running application. Examples include: dropping a table, dropping a column, adding a not null constraint to a column, adding a unique index, etc.
After an expansive phase, data gets migrated to the newly added elements. Code also begins using the newly added elements. This is the migration step. Finally, when all code is no longer using elements that are obsolete, these can be removed. (Or, if adding a unique or not null constraint, after checking that the data satisfies the constraint, then the constraint can be safely added.)
Expand, migrate, and contract steps may all be written using yoyo.
Create a .env file with the following environment variables:
DISCORD_DEBUG_TOKEN: The token of the bot you want to run locallyDISCORD_TOKEN: The token of the bot you want to run in productionDISCORD_DEBUG_CLUSTER_STAGING_ID: The ID of the "staging" server you want to connect toDISCORD_CLUSTER_STAGING_ID: The ID of the "production" server you want to connect toGITHUB_TOKEN: A Github token with permissions to trigger workflows, for now only new branches from discord-cluster-manager are tested, since the bot triggers workflows on your behalfDATABASE_URL: The URL you use to connect to Postgres.
Below is where to find these environment variables:
Note
For now, you can naively set DISCORD_DEBUG_TOKEN and DISCORD_DEBUG_CLUSTER_STAGING_ID to the same values as DISCORD_TOKEN and DISCORD_CLUSTER_STAGING_ID respectively.
-
DISCORD_DEBUG_TOKENorDISCORD_TOKEN: Found in your bot's page within the Discord Developer Portal:Click here for visual.
-
DISCORD_DEBUG_CLUSTER_STAGING_IDorDISCORD_CLUSTER_STAGING_ID: Right-click your staging Discord server and selectCopy Server ID:Click here for visual.
-
GITHUB_TOKEN: Found in Settings -> Developer Settings (or here). Create a new (preferably classic) personal access token with an expiration date to any day less than a year from the current date, and the scopesrepoandworkflow.Click here for visual.
-
DATABASE_URL: This contains the connection details for your local database, and has the formpostgresql://user:password@localhost/clusterdev.
Run the following command to run the bot:
python src/discord-cluster-manager/bot.py --debug
Then in your staging server, use the /verifyruns command to test basic functionalities of the bot and the /verifydb command to check database connectivity.
Note
To test functionality of the Modal runner, you also need to be authenticated with Modal. Modal provides free credits to get started. To test functionality of the GitHub runner, you may need direct access to this repo which you can ping us for.
TODO. This is currently a work in progress.
/run modal <gpu_type> which you can use to pick a specific gpu, right now defaults to T4
/run github <NVIDIA/AMD> which picks one of two workflow files
/resync to clear all the commands and resync them
/ping to check if the bot is online
The main purpose of the Discord bot is to allow servers to host coding competitions through Discord. The leaderboard was designed for evaluating GPU kernels, but can be adapted easily for other competitions. The rest of this section will mostly refer to leaderboard submissions in the context of our GPU Kernel competition.
Note
All leaderboard commands have the prefix /leaderboard, and center around creating, submitting to,
and viewing leaderboard statistics and information.
/leaderboard create {name: str} {deadline: str} {reference_code: .cu or .py file}
The above command creates a leaderboard named name that ends at deadline. The reference_code
has strict function signature requirements, and is required to contain an input generator and a
reference implementation for the desired GPU kernel. We import these functions in our evaluation
scripts for verifying leaderboard submissions and measuring runtime. In the next mini-section, we
discuss the exact requirements for the reference_code script.
Each leaderboard name can also specify the types of hardware that users can run their kernels on.
For example, a softmax kernel on an RTX 4090 can have different performance characteristics on an
H100. After running the leaderboard creation command, a prompt will pop up where the creator can
specify the available GPUs that the leaderboard evaluates on.
The Discord bot internally contains an eval.py script that handles the correctness and timing
analysis for the leaderboard. The reference_code that the leaderboard creator submits must have
the following function signatures with their implementations filled out. InputType and
OutputType are generics that could be a torch.Tensor, List[torch.Tensor], etc.
depending on the reference code specifications. We leave this flexibility to the leaderboard creator.
# Reference kernel implementation.
def ref_kernel(input: InputType) -> OutputType:
# Implement me...
# Generate a list of tensors as input to the kernel
def generate_input() -> InputType:
# Implement me...
# Verify correctness of reference and output
def check_implementation(custom_out: OutputType, reference_out: OutputType) -> bool:
# Implement me...The Discord bot internally contains an eval.cu script that handles the correctness and timing
analysis for the leaderboard. The difficult of CUDA evaluation scripts is we need to explicitly
handle the typing system for tensors. The reference.cu that the leaderboard creator submits must have
the following function signatures with their implementations filled out:
The main difference is we now need to define an alias for the type that the input / outputs are. A
simple and common example is a list of FP32 tensors, which can be defined using a pre-defined array of
const ints called N_SIZES, then define an array of containers, e.g.
std::array<std::vector<float>, N_SIZES>.
// User-defined type for inputs, e.g. using input_t = std::array<std::vector<float>, IN_SIZES>;
using input_t = ...;
// User-defined type for outputs, e.g. using output_t = std::array<std::vector<float>, OUT_SIZES>;
using output_t = ...;
// Generate random data of type input_t
input_t generate_input() {
// Implement me...
}
// Reference kernel host code.
output_t reference(input_t data) {
// Implement me...
}
// Verify correctness of reference and output
bool check_implementation(output_t out, output_t ref) {
// Implement me...
}/leaderboard submit {github / modal} {leaderboard_name: str} {script: .cu or .py file}
The leaderboard submission for Python code requires the following function signatures, where
InputType and OutputType are generics that could be a torch.Tensor, List[torch.Tensor],
etc. depending on the reference code specifications.
# User kernel implementation.
def custom_kernel(input: InputType) -> OutputType:
# Implement me...Deleting a leaderboard:
/leaderboard delete {name: str}
List all active leaderboards and which GPUs they can run on:
/leaderboard list
List all leaderboard scores (runtime) for a particular leaderboard. (currently deprecated. Doesn't support multiple GPU types yet)
/leaderboard show {name: str}
Display all personal scores (runtime) from a specific leaderboard.
/leaderboard show-personal {name: str}
We plan to add support for the PyTorch profiler and CUDA NSight Compute CLI to allow users to profile their kernels. These commands are not specific to the leaderboard, but may be helpful for leaderboard submissions.
If you'd like to donate a GPU to our efforts, we can make you a CI admin in Github and have you add an org level runner https://github.com/organizations/gpu-mode/settings/actions/runners
- Thank you to AMD for sponsoring an MI250 node
- Thank you to NVIDIA for sponsoring an H100 node
- Thank you to Nebius for sponsoring credits and an H100 node
- Thank you Modal for credits and speedy spartup times
- Luca Antiga did something very similar for the NeurIPS LLM efficiency competition, it was great!
- Midjourney was a similar inspiration in terms of UX