Add llm.starintel.actor fine-tune control plane and cheapest-GPU training jobs #4

New issue

Open

opened 2026-09-17 01:13:08 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-17 01:13:08 +00:00

Owner

Goal

Make llm.starintel.actor own StarIntel fine-tuning end-to-end. star-synthdata will submit POST /v1/training/jobs requests rather than talking directly to GPU vendors.

Contract

Accept kind: starintel.fine_tune.v1 with:

immutable dataset URI + SHA-256 + format
base model
trainer (axolotl) and method (qlora)
bounded recipe
compute policy: strategy=cheapest_compatible, provider allow-list, minimum VRAM, max GPU count, spot preference, hard USD budget
artifact publication policy
idempotency key

Replaying the same idempotency key MUST NOT provision another GPU.

State machine

queued -> pricing -> provisioning -> staging -> training -> validating -> publishing -> succeeded

Terminal: rejected | failed | cancelled | budget_exhausted.

Every terminal path MUST destroy provisioned compute and persist teardown evidence.

Provider layer

Implement provider adapters behind the actor, beginning with Runpod and a Vast-compatible adapter. Selection should rank effective estimated job cost after checking GPU VRAM/CUDA compatibility, storage, expected runtime, egress, availability and spot/preemption constraints. Provider credentials remain actor-side.

Worker

Initial worker should:

download corpus bundle from S3/HTTPS
verify SHA-256 before extraction
materialize OpenAI-messages JSONL
run the checked-in StarIntel Axolotl 8B QLoRA recipe
evaluate held-out data
upload LoRA adapter + metrics/logs
register a model alias only after validation
report lifecycle events back to the actor

Default should publish the adapter, not a merged full checkpoint.

Safety / cost invariants

hard budget checked before provisioning and during training
no unbounded retry loops
teardown on success, error, timeout, cancellation and budget exhaustion
provider API keys never enter dataset artifacts or GitHub Actions logs
training job and artifact metadata are durable/auditable

Client

Companion dataset-factory branch: lost-rob0t/star-synthdata:feature/dataset-factory-llm-actor-20260916.

## Goal Make `llm.starintel.actor` own StarIntel fine-tuning end-to-end. `star-synthdata` will submit `POST /v1/training/jobs` requests rather than talking directly to GPU vendors. ## Contract Accept `kind: starintel.fine_tune.v1` with: - immutable dataset URI + SHA-256 + format - base model - trainer (`axolotl`) and method (`qlora`) - bounded recipe - compute policy: `strategy=cheapest_compatible`, provider allow-list, minimum VRAM, max GPU count, spot preference, hard USD budget - artifact publication policy - idempotency key Replaying the same idempotency key MUST NOT provision another GPU. ## State machine `queued -> pricing -> provisioning -> staging -> training -> validating -> publishing -> succeeded` Terminal: `rejected | failed | cancelled | budget_exhausted`. Every terminal path MUST destroy provisioned compute and persist teardown evidence. ## Provider layer Implement provider adapters behind the actor, beginning with Runpod and a Vast-compatible adapter. Selection should rank effective estimated job cost after checking GPU VRAM/CUDA compatibility, storage, expected runtime, egress, availability and spot/preemption constraints. Provider credentials remain actor-side. ## Worker Initial worker should: 1. download corpus bundle from S3/HTTPS 2. verify SHA-256 before extraction 3. materialize OpenAI-messages JSONL 4. run the checked-in StarIntel Axolotl 8B QLoRA recipe 5. evaluate held-out data 6. upload LoRA adapter + metrics/logs 7. register a model alias only after validation 8. report lifecycle events back to the actor Default should publish the adapter, not a merged full checkpoint. ## Safety / cost invariants - hard budget checked before provisioning and during training - no unbounded retry loops - teardown on success, error, timeout, cancellation and budget exhaustion - provider API keys never enter dataset artifacts or GitHub Actions logs - training job and artifact metadata are durable/auditable ## Client Companion dataset-factory branch: `lost-rob0t/star-synthdata:feature/dataset-factory-llm-actor-20260916`.