llmopt: effect-driven tile compiler research
Bundle OKF 0.2 · 144 conceitos · tungd/llmopt
Open source Repository Open in the app JSON
About
# llmopt: effect-driven tile compiler research
The product boundary is a model-neutral PyTorch/Dynamo capture and a declared
`model.llmopt` execution contract. The current Metal backend uses W4A16 weights
plus grouped-Q8 KV/recurrent state; those formats are backend capabilities, not
a built-in model profile or the target weight policy. GGUF with Unsloth Dynamic
mixed quantization is the target weight-distribution path; FX capture, not a
GGUF architecture identifier, defines executable topology. Q8-weight and FP16-KV material below is historical
evidence, not an available compiler, package, runtime, CLI, or benchmark
configuration.
## Orientation
* [Architecture](architecture.md) - current Dynamo/FX to OCaml planning pipeline.
* [Complete OCaml serving goal](goal-serving-runtime.md) - requirement-by-requirement end state and evidence map.
* [Probe models](probe-models.md) - model coverage without product defaults.
* [LFM2.5 probe](target-lfm25.md) - hybrid probe facts and probe-only
Details
- Kind
- OKF bundles
- Topic
- Developer tools
- Publisher
- tungd
- Origin
- okf_github
- Category
- dados
- Version
- 0.2
- Added
- 2026-09-09 19:04:11
- Updated
- 2026-09-09 19:04:11
- Origin id
tungd/llmopt:.okf/index.md