zengxiao-he / tessera
From teacher to tiles — a from-scratch LLM distillation & serving engine: custom Triton/CUDA kernels, FSDP distillation, paged-KV continuous batching, speculative decoding, a Rust gateway, a JAX oracle, and interpretability tooling.
Why this job: topics: inference-engine, quantization
What it does
A small, from-scratch LLM stack built around one goal: distill a large teacher into a small student, then serve that student efficiently. Keeping that goal end-to-end means touching most of the pieces that matter in practice — custom GPU kernels, sharded training, an inference engine, quantization, and a serving front end — without any of it being a toy.
Get started
pip install -e ".[dev,gpu]"
# adds Triton
pytest -m gpu
# Triton kernels vs the torch reference- Maintenance
- Active
- Last commit 5 months ago
- Stars
- 556
- +2 this week
- Licence
- Custom
- Read before business use
- Downloads for
- None listed
- No release files
- Latest version
- None yet
- No releases
- Built with
- Python
- 88% of the code
Is it looked after?
- Contributors
- Not stated
- Last 90 days
- Issues
- 0% closed
- 2 still open
- Pull requests
- 0 merged
- 1 waiting
- Releases
- None
- Never published
- No contributing guide
- No security policy
- No code of conduct
Similar repos
| Repo | Stars | Last commit | Licence | Downloads for |
|---|---|---|---|---|
| pegainfer-project / pegainfer | 738 | yesterday | Apache-2.0 | Linux |
| LMCache / LMCache | 12,321 | today | Apache-2.0 | Linux |
| syv-ai / HyperQwen | 2,639 | today | Apache-2.0 | No downloads |
More: languages, topics and project record
Languages
- Python 88%
- Rust 8%
- Cuda 4%
Topics
- cuda
- flash-attention
- fsdp
- inference-engine
- jax
- knowledge-distillation
- kv-cache
- llm
- mechanistic-interpretability
- ml-systems
- paged-attention
- pytorch
Project record
- JobLocal LLMs
- StatusListed
- Commits29
- Forks8
- Watching5
- Stars in 28 days+3
- CreatedJun 5, 2026
- First seen by usNov 7, 2026
- CheckedNov 8, 2026