Training (Fine-tuning Service) — foundations RFC

Status

*ccepted*— ratificada 20260509 (mesmo dia da abertura) como parte da onda piloto de bootstrap servicesai. Implementação iniciada em `servicesaitraining; tickets em servicesaitrainingbacklogpending/{001..005}`.

Summary

Foundation pra finetuning — LoRA, QLoRA, SFT, DPO. Substitui setup adhoc atual (libras tem ML worker spike one-off).

Motivation

libras/ tem ML worker spike isolado (#001). Sem service, cada finetune vira projeto nextof-kind. Centralização ganha cache de checkpoints, GPU scheduling, MLflow integration.

Scope

In

• LoRAQLoRASFT pipeline
• GPU job scheduling
• Checkpoint registry
• MLflow tracking

Out (yet)

• Pretraining (escopo proibitivo)
• Distributed multi-node (v2)

Initial design

Surfaces

• backend/ — Go API + GPU job orchestrator
• app/ — não aplicável v1

Key APIs

• POST /v1/train/jobs — submit job
• GET /v1/train/jobs/{id} — status/logs
• POST /v1/train/checkpoints — registrar checkpoint

Dependencies

• services/ai/dataset — input datasets versionados
• services/ai/modelreg — modelos base
• services/ai/runtime — serving pós-train
• infra/data/kdb-blob — checkpoint storage
• infra/observe — GPU metrics

Relation to existing sectors

• Consome dataset, produz checkpoints pra modelreg
• Bloqueador implícito de #001 libras

Selfhostedfirst analysis (5 gates)

• *1 Feature parity* zero
• *2 Performance* N/A
• *3 Stability* N/A
• *4 Capability* axolotlunslothtransformers viáveis
• *5 Criticalpath readiness* bloqueia escala de domainspecific models (libras, koda, vertical pieces)

Open questions

• Q1: GPU on-prem (s.k.lin?) vs cloud burst?
• Q2: Adopt MLflow direct ou wrap?

Next steps

1. Ratificar esta RFC (1 round de comments).
2. Criar sector dir services/ai/training/ com koder.toml, README.md, skeleton.
3. Abrir tickets de implementação em services/ai/training/backlog/pending/.
4. Registrar em meta/docs/stack/registries/self-hosted-pairs.md se substituir externo.