Training (Fine-tuning Service): foundations

Status

accepted — ratificada 2026-05-09 (mesmo dia da abertura) como parte da onda piloto de bootstrap servicesai. Implementação iniciada em `servicesaitraining; tickets em servicesaitrainingbacklogpending/{001..005}`.

Summary

Foundation pra fine-tuning — LoRA, QLoRA, SFT, DPO. Substitui setup ad-hoc atual (libras tem ML worker spike one-off).

Motivation

libras/ tem ML worker spike isolado (#001). Sem service, cada fine-tune vira projeto next-of-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

Self-hosted-first analysis (5 gates)

• G1 Feature parity: zero
• G2 Performance: N/A
• G3 Stability: N/A
• G4 Capability: axolotlunslothtransformers viáveis
• G5 Critical-path readiness: bloqueia escala de domain-specific 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.