Plans for the new Burn AI engine (after burn-lm)? Looking to contribute: serving, training, systems work

[ailakshya]

Hi Burn team 👋

I saw burn-lm was archived to make way for a new Burn-powered AI engine that can be embedded in apps or run as a standalone server. That is very close to what I want to build next, so I'd rather work with you than duplicate effort.

About me

• M.Tech (CSE, AI/ML) student at Manipal University Jaipur. My main interest is systems-level programming: inference runtimes, GPU kernels, memory management, schedulers and batching.
• Built infergo (https://github.com/ailakshya/infergo), a Go inference runtime that serves LLMs (GGUF), embeddings (ONNX) and detection (TorchScript) from one binary behind an OpenAI-compatible API. Continuous batching gives 12x the throughput of llama-cpp-python at 8 concurrent users, and the Go layer adds only 0.7 ms (0.3%) over raw llama.cpp. C++/CUDA core, 252 tests.
• Built and published TensorFlow 2.20 with CUDA for Blackwell GPUs (compute capability 12.0), which official releases did not support.
• First author on three ML manuscripts on model compression (distillation + pruning) and multimodal optimization.
• Side project: a cold storage business system where I run a YOLOv11 + TensorRT CCTV counting pipeline on a self-hosted RTX 5070 Ti.

What I'd like to do
Rebuild the infergo ideas natively in Rust on Burn: serving (continuous batching, paged KV cache, OpenAI-compatible API) plus a built-in training layer, so one binary can fine-tune and serve. I'm especially keen on the low-level parts: CubeCL kernels, memory and scheduling.

Questions

1. What is the scope of the new engine: LLM only, or multi-model (embeddings, detection, etc.)? Will it cover serving features like continuous batching and paged KV cache?
2. Is there a timeline, or a public repo/branch to follow?
3. Are you open to outside contributors on it? Which areas would help most (CubeCL kernels, serving, models, training pipelines)?
4. If I build separately for now, is burn + tracel-ai/models (minilm-burn, yolox-burn) + burn-onnx the recommended base?

Happy to share infergo benchmarks (RTX 5070 Ti, CUDA 12.8) if useful.

Thanks for Burn!

[laggui]

Hey! Your interests definitely overlap closely with what we're building.

1. The engine covers inference and training across multiple model types. We already have text generation, speech recognition and image generation. Continuous batching, paged KV caching and an OpenAI-compatible API are already implemented.

2. The repository is still private, so there isn't a public branch to follow yet. We don't have a release date to share at this point, we want to make sure the performance and experience is solid before we publish it this time.

3. Contributions are welcome across burn/cubek/cubecl, and the new engine once it's public. Our kernels use cubecl across hardware backends, so performance work that preserves portability is especially valuable. We've upstreamed many improvements that directly target the engine in burn/cubek/cubecl already.

4. Burn + tracel-ai/models is a reasonable starting point. burn-onnx is useful for bringing additional models into burn, though getting competitive performance may require profiling and optimization after import. Serving and scheduling remain separate work.