lightseekorg / TorchSpec
A PyTorch native library for training speculative decoding models
Why this job: works in the areas we cover, but does not say which job it does
What it does
TorchSpec is a torch-native speculative decoding training framework. We introduce a disaggregated way of training speculative decoding draft models where inference and training are fully decoupled and stream hidden states directly from inference engine groups to distributed training workers via Mooncake store, allowing each side to scale independently.
Get started
./examples/qwen3-8b-single-node/run.sh - Maintenance
- Active
- Last commit today
- Stars
- 268
- +1 this week
- Licence
- MIT
- Allows business use
- Downloads for
- None listed
- No release files
- Latest version
- None yet
- No releases
- Built with
- Python
- 97% of the code
Is it looked after?
- Contributors
- 6
- Last 90 days
- Issues
- 88% closed
- 6 still open
- Pull requests
- 150 merged
- 2 waiting
- Releases
- None
- Never published
- Contributing guide (opens in new tab)
- No security policy
- No code of conduct
Similar repos
| Repo | Stars | Last commit | Licence | Downloads for |
|---|---|---|---|---|
| aws-neuron / torch-neuronx | 137 | 13 days ago | Apache-2.0 | Linux |
More: languages, topics and project record
Languages
- Python 97%
- Shell 2%
- Dockerfile <1%
- Rust <1%
Topics
- eagle3
- fsdp
- lightseek
- llm
- pytorch
- mooncake
- vllm
- tokenspeed
- dspark
Project record
- JobOther AI tools
- StatusListed
- Commits150
- Forks70
- Watching1
- Stars in 28 days+11
- CreatedFeb 21, 2026
- First seen by usNov 7, 2026
- CheckedNov 12, 2026