Roadmap
Language Pipes is actively developed. This page collects what’s planned so you can see where the project is headed — and where a contribution would land well. Nothing here is a dated commitment; it’s a direction.
Want to help with any of these? Open a GitHub Discussion or an issue, and see CONTRIBUTING.
Model support
- More architectures. New model families are added regularly. Adding one is self-contained and a great first contribution — see adding a new model architecture and the current supported models.
Quantization and formats
- INT4 quantization. An 8-bit path exists today via
LP_8_BIT_MODE(LLM.int8 throughbitsandbytes); the goal is broader, better-supported quantization including INT4. Default inference currently runs infp16. - GGUF format support. Weights currently must be
safetensors. GGUF support would open up the large ecosystem of pre-quantized community models.
Robustness
- Failover and rerouting. Today a layer node dropping mid-generation fails the request (see Failure Modes and Limitations). Automatic rerouting to another node hosting the same layers is a natural next step.
- Integrity of layer computation. The current threat model names — but does not yet solve — the problem of a malicious layer node returning subtly steered tensors (Attack Vector #5 in Privacy). Redundant recomputation or verifiable-computation approaches are open design work.
Open research
- Quantify recovery difficulty vs. capture depth. The privacy threat model argues that inverting hidden states back to a prompt becomes harder as capture happens deeper in the network, but this has not been measured. Capturing hidden states at layers 1, 5, 10, and 20, attempting recovery, and reporting success rates would turn an expectation into a result — and is a contribution Petals never made.
Good first issues
Several of the items above are well-shaped for a first contribution; especially
adding a model architecture, which has a natural, ready supply of candidates.
Look for the good first issue
label on the issue tracker.