S3 Bucket to host llama prebuilts for RPi

[psiddh]

Context

S3 Bucket for hosting prebuilts for NEW Raspberry Pi ExecutorTorch Tutorials

We're developing comprehensive Raspberry Pi tutorials for deploying Llama models using ExecutorTorch. Our goal is to provide friction-free deployment by offering prebuilt ARM64 binaries that work out-of-the-box on Raspberry Pi 4/5.

This issue tracks the creation of a dedicated S3 bucket for hosting Raspberry Pi specific ExecutorTorch artifacts.

Install Journey (Quick Deploy)

On RPi devices:

# One-line install
curl -sSL https://download.pytorch.org/executorch/rpi/v1/install_rpi.sh | bash -s pi5

# Run model immediately and infer
./llama_main --model_path model.pte --tokenizer_path tokenizer.model --prompt "Hello"

Ideal S3 Bucket Structure

s3://executorch-artifacts/rpi/v1/install_rpi.sh
s3://executorch-artifacts/rpi/pi4/v1/llama_main
s3://executorch-artifacts/rpi/pi4/v1/libllama_runner.so
s3://executorch-artifacts/rpi/pi4/v1/libextension_module.so
s3://executorch-artifacts/rpi/pi5/v1/llama_main
s3://executorch-artifacts/rpi/pi5/v1/libllama_runner.so
s3://executorch-artifacts/rpi/pi5/v1/libextension_module.so

Public Access Requirements

The S3 bucket should be configured for public read access to enable direct wget/curl downloads, following the same pattern as existing PyTorch artifacts on download.pytorch.org.

[malfet]

I don't think we can put llama weights anywhere that can bypass authentication
cc: @cbilgin , @mergennachin

[cbilgin]

Llama weights and models will be part of HF, and be provided through the hummingbird initiative. What Sid is asking should be strictly the runners and libs, not the model itself. This should be no different than providing prebuilt libs for iOS and Android apps. @psiddh can you confirm that you are not suggesting to put any pte file here?

[psiddh]

Yes just to clarify , the ask is not for llama weights , but just the files that were requested in as part of the section "Ideal S3 Bucket Structure" above