From the repository, docs/labels.md
Labels
Every issue and pull request gets one type label and one or more area labels. Open issues also get a priority. A label from the other groups is added when it applies. tools/gh_labels.sh creates and updates these labels on GitHub, and this table is its --table output: change the script, then run bash tools/gh_labels.sh --table and paste the result here.
GitHub's own duplicate, invalid, wontfix, good first issue and help wanted stay as they are.
| Group | Label | Meaning |
|---|---|---|
| type | bug |
Something is broken: a wrong answer, a crash, a hang, a wrong number |
| type | feature |
A new capability, or a change to how something works |
| type | perf |
Speed or memory: tok/s, time to first token, prefill, GPU memory |
| type | docs |
README, docs/ or the manual |
| type | refactor |
Code that changes shape and keeps its behaviour and output |
| type | test |
Tests, test tools or the test suite |
| type | security |
A security fix or hardening. Report a vulnerability in private: SECURITY.md |
| type | question |
A question about using TandemLLM |
| area | engine/core |
Engine: model code, loaders, the decode and verify loop |
| area | engine/kernels |
Engine: GPU kernels, matrix products, attention |
| area | engine/drafting |
Engine: drafters and speculation (lookup, StairCut, MTP, DFlash2) |
| area | engine/cache |
Engine: KV cache, prefix cache, disk cache |
| area | server/api |
Server: the OpenAI API, tool calls, streaming, structured output |
| area | server/conductor |
Server: the conductor (session monitor, deciders, policies) |
| area | server/proxy |
Server: the proxy in front of the engine and other backends |
| area | chat |
Chat service (chatd): agent loop, sandbox, files, tools |
| area | dash/setup |
Dashboard: Setup Guide |
| area | dash/chat |
Dashboard: Chat |
| area | dash/usage |
Dashboard: Usage |
| area | dash/performance |
Dashboard: Performance |
| area | dash/dev |
Dashboard: Dev (logs and request inspector) |
| area | dash/playground |
Dashboard: Playground |
| area | dash/conductor |
Dashboard: Conductor |
| area | dash/memory |
Dashboard: memory screens (conductor memory, disk cache) |
| area | dash/backends |
Dashboard: Backends |
| area | dash/system |
Dashboard: System |
| area | models/new |
Support for a model that TandemLLM does not run yet |
| area | models/flint |
Model: Recursive Flint |
| area | models/qwen3.8 |
Model: Qwen3.8 27B and Qwen3.8 Flash-Next |
| area | models/lfm2.5 |
Model: LFM2.5-24B-A2B |
| area | models/nemotron |
Model: Nemotron 3.5 Lightning, Bolt and Super VL |
| area | models/ornith |
Model: Ornith 1.5 122B |
| area | quant/exl3 |
Weights: EXL3 packs |
| area | quant/nvfp4 |
Weights: NVFP4 packs |
| area | quant/bf16 |
Weights: BF16 |
| area | quant/community |
Weights: community packs (models/ |
| area | install |
install.sh and the first start on a new machine |
| area | ops |
Service scripts, profiles, watchdog, release tooling |
| area | docs/manual |
The manual (manual/, LaTeX) |
| area | bench |
Benchmarks and measurement tools |
| area | ci |
GitHub workflows, issue forms, repository settings |
| priority | p0 |
Broken for everyone, or data loss: first |
| priority | p1 |
Next release |
| priority | p2 |
Normal |
| priority | p3 |
When there is time |
| status | needs-triage |
New: no type, area, priority or milestone yet |
| status | needs-repro |
Needs steps that reproduce it |
| status | needs-info |
Waits for an answer from the reporter |
| status | blocked |
Waits for another issue or an outside change |
| hardware | hw/dgx-spark |
Seen on an NVIDIA DGX Spark (GB10) |
| hardware | hw/rtx-pro-6000 |
Seen on an NVIDIA RTX PRO 6000 |
| hardware | hw/other |
Seen on another GPU or machine |
| process | run-gpu |
A maintainer runs the GPU tests on a DGX Spark |
| process | skip-changelog |
Left out of the generated release notes |