pyfg wheels are x86_64-only — please publish aarch64 (ARM64) builds, or open-source the FG library

[wondermen]

Summary

pyfg is a hard dependency of TorchEasyRec, but every published wheel is
linux_x86_64-only. There is no aarch64/arm64 build, no pure-Python
fallback and no sdist, so TorchEasyRec cannot be installed on any ARM64
Linux host. This blocks ARM-based training and offline feature-generation
workloads.

I would like to ask for either:

1. aarch64 wheels for the currently supported CPython versions, or
2. open-sourcing the FG library (source + build instructions) so the
   community can build it for architectures you do not target.

Option 1 is the smaller ask and would unblock us immediately; option 2 is
the durable fix, since it also covers architectures beyond ARM.

[tiankongdeguiji]

Thanks for reporting this.

Could you share a bit more about the ARM64 environment and workload you are targeting?

In particular, it would be helpful to know whether you need ARM64 mainly for:

• CPU-only offline feature generation / preprocessing,
• CPU-only training or inference,
• or ARM64 + NVIDIA GPU training/inference, for example on Grace/GH200 systems.

TorchEasyRec also depends on several native packages besides pyfg, such as graphlearn, fbgemm-gpu, and some GPU-specific components, so the required ARM64 support may differ quite a bit depending on the target scenario.

Understanding your actual use case would help us determine the minimum set of dependencies that need ARM64 builds.

[wondermen]

currently is cpu only,include all stage，trainning/inference/feature generating.

[wondermen]

I noticed https://github.com/alibaba/FeatureGenerator has the feature func used by torcheasyrec,but it's latest update is 5 years ago, so is there any plan to update it to latest?

[wondermen]

@tiankongdeguiji I'd appreciate your time and consideration

[tiankongdeguiji]

@wondermen
Thanks for the clarification.

For the CPU-only ARM64 scenario you described, including training, inference, and feature generation, we currently do not have plans to add ARM64 support for these TorchEasyRec workloads.

Regarding [alibaba/FeatureGenerator](https://github.com/alibaba/FeatureGenerator), it is not the same project as the pyfg library currently used by TorchEasyRec, even though some of the feature functions may look similar. Therefore, updating that repository would not provide the current pyfg implementation.

At the moment, we also do not have plans to open-source pyfg.