mirror of
https://github.com/hanzo-ml/trainer.git
synced 2026-10-30 14:46:16 +00:00
0b97a47048
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]>
60 lines
1.8 KiB
YAML
60 lines
1.8 KiB
YAML
# Example TrainJob produced by the Hanzo Cloud fine-tuning broker
|
|
# (hanzoai/ai `/v1/finetune/jobs`). It QLoRA-fine-tunes a HuggingFace base model on
|
|
# a HuggingFace dataset using the `hanzo-ft-qlora` ClusterTrainingRuntime. The
|
|
# broker fills the same fields from the console's new-job panel:
|
|
# - initializer.model / initializer.dataset .storageUri → the picked HF repos
|
|
# (a secretRef is added for private/gated repos, from the in-namespace hf-token
|
|
# Secret the broker provisions)
|
|
# - trainer.resourcesPerNode → the chosen GPU count
|
|
# - trainer.env → the recommended (or custom) hyperparameters + OUTPUT_DIR
|
|
# (the org's S3 Space, so checkpoints land in the tenant's bucket)
|
|
apiVersion: trainer.kubeflow.org/v1alpha1
|
|
kind: TrainJob
|
|
metadata:
|
|
name: llama-3-1-8b-instruct-demo
|
|
namespace: ft-acme
|
|
spec:
|
|
runtimeRef:
|
|
name: hanzo-ft-qlora
|
|
apiGroup: trainer.kubeflow.org
|
|
kind: ClusterTrainingRuntime
|
|
initializer:
|
|
model:
|
|
storageUri: hf://meta-llama/Llama-3.1-8B-Instruct
|
|
secretRef:
|
|
name: hf-token
|
|
dataset:
|
|
storageUri: hf://tatsu-lab/alpaca
|
|
secretRef:
|
|
name: hf-token
|
|
trainer:
|
|
numNodes: 1
|
|
resourcesPerNode:
|
|
limits:
|
|
nvidia.com/gpu: "1"
|
|
env:
|
|
- name: METHOD
|
|
value: qlora
|
|
- name: TASK
|
|
value: instruct
|
|
- name: EPOCHS
|
|
value: "3"
|
|
- name: LEARNING_RATE
|
|
value: "0.0002"
|
|
- name: BATCH_SIZE
|
|
value: "2"
|
|
- name: GRAD_ACCUM
|
|
value: "8"
|
|
- name: MAX_SEQ_LEN
|
|
value: "2048"
|
|
- name: LORA_RANK
|
|
value: "16"
|
|
- name: LORA_ALPHA
|
|
value: "32"
|
|
- name: QUANT_4BIT
|
|
value: "true"
|
|
- name: GRADIENT_CHECKPOINTING
|
|
value: "true"
|
|
- name: OUTPUT_DIR
|
|
value: s3://hanzo-finetunes/acme/llama-3-1-8b-instruct-demo/
|