hanzo-ml/trainer

Fork 0

mirror of https://github.com/hanzo-ml/trainer.git synced 2026-09-24 00:31:34 +00:00

Files

T

History

0b97a47048 feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

Adds one transformers + PEFT + TRL ClusterTrainingRuntime per method that
fine-tunes ANY HuggingFace causal LM. The base model + dataset come from the
TrainJob's model/dataset initializers (hf:// or s3://, secretRef for private/gated
repos); every hyperparameter is read from the trainer container env that the Hanzo
Cloud broker sets. Adapters are merged into the base on completion, so one serving
path (KServe HuggingFace runtime) works for lora/qlora/full alike; output is
written to OUTPUT_DIR (the org's S3 Space when s3://).

- manifests/base/runtimes/hanzo_finetune.yaml  3 ClusterTrainingRuntimes (+kustomization)
- cmd/trainers/hanzo/finetune.py               env-driven entrypoint (LoRA/QLoRA/full)
- cmd/trainers/hanzo/finetune_test.py          pytest for the pure config/format helpers
- cmd/trainers/hanzo/{Dockerfile,requirements.txt,README.md}
- examples/yaml/hanzo-finetune-qlora.yaml      example TrainJob (the broker's output)

kustomize build clean · ruff clean · 14 pytest pass. Image built in CI
(ghcr.io/hanzoai/finetune-runtime), not locally.

Co-authored-by: Hanzo Dev <[email protected]>

2026-06-29 23:53:55 -07:00

Dockerfile

feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

2026-06-29 23:53:55 -07:00

finetune_test.py

feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

2026-06-29 23:53:55 -07:00

finetune.py

feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

2026-06-29 23:53:55 -07:00

README.md

feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

2026-06-29 23:53:55 -07:00

requirements.txt

feat(runtimes): hanzo-ft-{lora,qlora,full} — env-driven HF fine-tuning runtimes (#2 )

2026-06-29 23:53:55 -07:00

README.md

Hanzo fine-tuning trainer

The node entrypoint for the hanzo-ft-{lora,qlora,full} ClusterTrainingRuntimes (manifests/base/runtimes/hanzo_finetune.yaml). One env-driven image fine-tunes any HuggingFace causal LM with LoRA, QLoRA (4-bit) or full fine-tuning using transformers + peft + trl.

How it runs

A TrainJob references a runtime (e.g. hanzo-ft-qlora) and supplies:

initializer.model.storageUri / initializer.dataset.storageUri — the base model and dataset (hf://… or s3://…). The runtime's model/dataset initializers download them to /workspace/model and /workspace/dataset before this entrypoint runs. A secretRef (the hf-token Secret) authenticates private or gated repos.
trainer.resourcesPerNode — the GPU count.
trainer.env — every hyperparameter (see the env contract in finetune.py).

The entrypoint trains, then writes a servable HuggingFace-format model to OUTPUT_DIR (LoRA/QLoRA adapters are merged into the base, so KServe's HuggingFace runtime can serve any method). When OUTPUT_DIR is an s3:// URI the result is uploaded to the org's S3 Space.

Build

The image is built in CI (not locally) and pinned by the operator:

docker build . -f cmd/trainers/hanzo/Dockerfile -t ghcr.io/hanzoai/finetune-runtime:0.1.0

Test

The pure config/format helpers are unit-tested (no GPU):

cd cmd/trainers/hanzo && pytest finetune_test.py