github ↗

sponsor

Support Raven.

Raven is built in public. Sponsorship funds the road to 1.0.

github sponsors ↗email about sponsoring

Sponsors

corporate sponsor

Ahrefs

Building tools to help you grow your search traffic.

past sponsor

Tarides

Secure-by-design infrastructure and tooling for a better digital world. Raven wouldn't be here without them.

individual sponsors

Thank you.

To all our individual sponsors on GitHub, for their support.

01 — why

One language from research to production.

Python's monopoly on scientific computing forces a choice: ship everything in Python, or prototype in Python and rewrite for production. OCaml lets you prototype as quickly and scale the same code to production; strong typing catches bugs before they crash your ML pipeline, while eager kernels and JIT compilation aim at NumPy and PyTorch performance.

Raven brings that stack to OCaml: train models with automatic differentiation and JIT compilation today, and, at 1.0, deploy them as one static binary that needs nothing on the host beyond the GPU's kernel driver. Read the introduction.

02 — what it funds

The road to 1.0.

full roadmap →
  1. 01Training: fine-tune Llama 3.1 70B at 90% or better of PyTorch's throughput, with LoRA on one 8×H100 node and in full across two
  2. 02Deployment: serve open-weight models behind an OpenAI-compatible API from a dependency-free static binary at 90% or better of vLLM's throughput
  3. 03Mixed precision, memory-efficient attention, and GEMMs at cuBLAS throughput
  4. 04Distributed training: data parallelism, FSDP-style sharding, and multi-node collectives
  5. 05AOT compilation to standalone binaries (CPU and GPU)
  6. 06Inference engine with KV cache, continuous batching, and PagedAttention
  7. 07Post-training quantization (INT8/INT4) and mixture-of-experts serving

03 — other ways to help

Not only money.

developers

  • Try it out: test Raven with your workflows and report issues
  • Contribute code: see the contributing guide
  • Share feedback: what would make you switch from Python? Tell us
  • Spread the word: star the repo, share it with your team, write about it

companies

individuals

from doc/support-raven.mdedit ↗