Hanzo Gym

LoRA fine-tuning and GRPO for language models, in Rust on hanzo-ml. One binary for CUDA, Metal and CPU. Configs use Axolotl's keys, so a config from zooai/gym runs here unchanged where the feature exists and fails at load where it does not.

cargo install hanzo-gym --features metal   # or --features cuda
gym train examples/qwen3-lora.yml   # Qwen3-0.6B, LoRA on q/v, 20 steps
gym merge examples/qwen3-lora.yml

base_model: Qwen/Qwen3-0.6B
chat_template: qwen3
datasets:
  - path: mlabonne/FineTome-100k
    type: chat_template
    field_messages: conversations
    message_property_mappings: { role: from, content: value }
adapter: lora
lora_r: 16
lora_alpha: 32
lora_target_linear: true
sequence_len: 2048
micro_batch_size: 1
gradient_accumulation_steps: 8
num_epochs: 1
learning_rate: 2e-4
lr_scheduler: cosine
warmup_ratio: 0.03
output_dir: ./outputs/qwen3-lora

The adapter is written in PEFT layout, so it loads in Hanzo Engine, vLLM and peft without conversion. rl: grpo switches the run to GRPO with the reward functions named under trl:.

Scope

Implemented: LoRA on Qwen3 and Llama, SFT on chat_template, alpaca and completion datasets, cosine/linear/constant schedules with warmup, gradient accumulation and clipping, gradient checkpointing, GRPO with built-in and HTTP verifiers, PEFT adapter save and merge.

Not implemented: QLoRA, full fine-tuning, multi-GPU, DPO/KTO/ORPO, sample packing. A config that asks for one of these is rejected at load.

Hosted / TrainJob

ghcr.io/hanzoai/gym (built from Dockerfile.cuda by .hanzo/workflows/image.yml, tagged by git tag) is the trainer container of the Kubeflow TrainJob that POST /v1/ai/finetune/jobs submits. Its command is gym trainjob, which reads the run from the environment the broker (hanzoai/ai cluster/finetune.go) sets, with the broker's defaults when a variable is unset:

Variable	Config key	Default
BASE_MODEL	base_model, when MODEL_DIR is not a directory	—
MODEL_DIR	base_model, the initializer's copy of the model	/workspace/model
DATASET_DIR	one dataset per .jsonl / .json / .parquet file under it	/workspace/dataset
OUTPUT_DIR	output_dir	/workspace/output
TASK	chat → chat_template, instruct → alpaca, completion → completion	—
METHOD	lora → adapter: lora, lora_target_linear: true	qlora
EPOCHS	num_epochs (whole number)	3
LEARNING_RATE	learning_rate	2e-4
BATCH_SIZE	micro_batch_size	2
GRAD_ACCUM	gradient_accumulation_steps	8
MAX_SEQ_LEN	sequence_len	2048
LORA_RANK, LORA_ALPHA, LORA_DROPOUT	lora_r, lora_alpha, lora_dropout	16, 32, 0.05
WARMUP_RATIO, WEIGHT_DECAY	warmup_ratio, weight_decay	0.03, 0.01
QUANT_4BIT, GRADIENT_CHECKPOINTING	rejected when true	false
HF_TOKEN	read by hf-hub for gated repos; never logged	—

METHOD=lora is what this image serves today (the hanzo-ft-lora runtime). qlora and full, QUANT_4BIT=true and GRADIENT_CHECKPOINTING=true fail at startup with a message naming the runtime or limit, so the TrainJob goes Failed instead of training something other than what was asked. Precision is bf16: auto. The resolved config is logged as YAML before training.

Apache-2.0.