Forge S3 bucket

[allenwang28]

Context

Forge builds on top of vLLM, Monarch, and TorchTitan. Our current build pipeline involves building wheels of stable/pinned versions of each package against the latest PyTorch nightly. For our initial release, we plan to include a zipped version or a venv of these packages, along with a helper script to help users get started. We plan to not include this with the PyTorch index for now, until our APIs are more stable / when we know which stable versions of each package we want to pin to etc.

This issue tracks the creation of a custom S3 bucket for hosting Forge specific artifacts.

Packages / Specifics

Forge will run a nightly CI that builds the following packages using PyTorch nightly:

Package	Version	Size
Monarch (meta-pytorch/monarch)	v0.10.0	~40MB
vLLM (vllm-project/vllm)	nightly	~0.72GB
torchtitan (pytorch/torchtitan)	nightly	~1MB
torchstore (meta-pytorch/torchstore)	nightly	~30KB

Install journey

For our initial release, we want the experience to be as simple as:

conda create -n forge python=3.10
conda activate forge
./scripts/install.sh

Ideally, the install script can be replaced with just

uv pip install -e .

if we can link all of the dependency wheels correctly in our pyproject.toml.

Otherwise, we could package it as a zip file and have the install script download, extract and install the wheels too.

Ideal S3 bucket structure

Ideally can look something like either:

$ aws s3 ls s3://forge-wheelhouse
s3://forge-wheelhouse/cuda-12-8/9-26-25/forge-0.1.0-py3-none-any.whl
s3://forge-wheelhouse/cuda-12-8/9-26-25/torchtitan-0.1.0-py3-none-any.whl
s3://forge-wheelhouse/cuda-12-8/9-26-25/monarch-0.1.0-py3-none-any.whl
s3://forge-wheelhouse/cuda-12-8/9-26-25/vllm-0.1.0-py3-none-any.whl
s3://forge-wheelhouse/cuda-12-8/9-26-25/torchstore-0.1.0-py3-none-any.whl
s3://forge-wheelhouse/cuda-12-6/9-26-25/... # only build for CUDA < 13 for now

or

$ aws s3 ls s3://forge-wheelhouse
s3://forge-wheelhouse/cuda-12-8/9-26-25/forge.zip

depending on if the uv path would work

[huydhn]

I have a qq as you mention building wheels of stable/pinned versions of each package against the latest PyTorch nightly, we now have vLLM nightly wheel built with PyTorch nightly and they can be installed with uv pip install vllm xformers flashinfer-python --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu128. Does this work for your use case? I also see that there is a nightly wheel for torchtitan on https://download.pytorch.org/whl/nightly/cu128 already. So, if this works, this will resolve 2 out of 4 packages

[allenwang28]

@huydhn yeah I saw that! I think we could use it longer term - my honest concern is that we've been relying on v0.10.0. I'm afraid of updating it for what may or may not break, esp as we're so close to launch.

[atalman]

@allenwang28 why do we need cuda-12-8/9-26-25/ ? Can we label build the date and version in the filename like pytorch nightly ?

We can perhaps host the binaries you looking for in separate folder like this:
download.pytorch.org/whl/nightly/forge

[atalman]

How many cuda/ python version we are talking about here ?

[huydhn]

@huydhn yeah I saw that! I think we could use it longer term - my honest concern is that we've been relying on v0.10.0. I'm afraid of updating it for what may or may not break, esp as we're so close to launch.

I see. Agree that sticking with v0.10.0 is best for stability. On the other hand, you could also pin night version if one is working, i.e. dev20250926. I don't fully endorse that route though, just table an option to keep in mind

[allenwang28]

Can we label build the date and version in the filename like pytorch nightly ?

oh that makes sense, that is definitely fine by me too!

How many cuda/ python version we are talking about here ?

It's probably fine to just have 1 CUDA and 1 Python for now. We have been just running CUDA 12.9 w/ Python 3.10 so far. If it's more helpful to only start with 1 tuple, we'll probably go Python 3.12 + CUDA 12.8.

We can perhaps host the binaries you looking for in separate folder like this: download.pytorch.org/whl/nightly/forge

S3 and/or a separate folder is fine by me, how would I set that up?

[atalman]

Here is an example of separate folder: https://download.pytorch.org/whl/test/variant Do you already have the workflow that builds the binaries and dependnecies ?

[ebsmothers]

@atalman bumping this thread as I am helping Allen take a look at this. Do you have an example of a script to upload wheels to a custom folder like download.pytorch.org/whl/test/variant as you mentioned?