Frameworks are isolated into reproducible optional environments because their
CUDA, PyTorch, and Transformer Engine constraints differ. Install exactly one
backend extra per image, then invoke it through oshun-train so licensing,
dataset validation, scaling-law declarations, memory estimation, evaluation,
safety scanning, and registry publication remain mandatory.
bash
uv sync --extra axolotl --extra efficiency
oshun-train --framework axolotl --method qat --model Qwen/Qwen3-8B \
--dataset registry://edge/v1 --output /runs/qwen3-edge
Production images must lock this manifest, record the lock digest in the run card, disable network access after artifact staging, and mount datasets and base models read-only. NeMo-RL multi-node jobs additionally require NCCL topology validation and an immutable Megatron Bridge conversion report.