RunPod Plugin

mere-runpod runs mere.run workflows on ephemeral RunPod pods owned by the user.

The first supported workflow is FLUX.2 Klein LoRA training.

Why This Is A Plugin

The core mere.run repo is local-first. RunPod creates paid remote compute, so it belongs in an explicit companion plugin. The plugin still runs the normal mere.run image train-lora command remotely, which keeps training behavior aligned with local runs.

Required Inputs

• RUNPOD_API_KEY
• SSH private/public key registered with RunPod
• paired dataset directory
• recipe id or recipe JSON file
• Linux CUDA mere.run build pack

For Klein recipes, set HF_TOKEN or HUGGING_FACE_HUB_TOKEN in the environment or in the --env-file before running. The plugin fails preflight if the token is missing.

Commands

mere-runpod doctor
mere-runpod volume ensure \
  --name mere-klein-cache \
  --data-center-id US-KS-2 \
  --size-gb 512 \
  --dry-run
mere-runpod plan --recipe klein-style-lora --data ./dataset --output ./runs/foo
mere-runpod run \
  --recipe klein-style-lora \
  --data ./dataset \
  --output ./runs/foo \
  --build-pack ./build-packs/mere-run-linux-cuda.tar.gz \
  --network-volume-id <volume-id> \
  --data-center-id US-KS-2
mere-runpod cleanup ./runs/foo/run.json

Cleanup

The plugin terminates the pod by default after fetching the final LoRA, archive, log, and small metadata files. Intermediate checkpoint safetensors are skipped unless --fetch-checkpoints is passed. Pass --keep-pod only for active debugging.

Build Pack Strategy

The plugin requires a build pack for real remote runs. This avoids spending H100 minutes compiling Swift and CUDA dependencies on every run. The build pack must be a prebuilt Linux CUDA package tarball, such as one produced on Linux by scripts/package-linux.sh.

Bootstrap/source packs that contain source.tar.gz and a wrapper bin/mere.run compile on the paid pod. The plugin rejects those packs by default; pass --allow-bootstrap-build-pack only for an intentional slow debug run.

Network Volume Strategy

For fast repeated training, use a RunPod network volume as the persistent model cache:

mere-runpod volume list
mere-runpod volume ensure --name mere-klein-cache --data-center-id US-KS-2 --size-gb 512 --dry-run
mere-runpod volume ensure --name mere-klein-cache --data-center-id US-KS-2 --size-gb 512

volume ensure --dry-run prints the billable storage request without calling RunPod. Running it without --dry-run returns JSON with volume.id and volume.dataCenterId. Pass those values to run as --network-volume-id and --data-center-id. The volume replaces the pod's normal /workspace disk, so the plugin's existing /workspace/mere-runpod/models and /workspace/mere-runpod/hub paths survive after the pod is terminated. The plugin also caches Linux CUDA build packs under /workspace/mere-runpod/build-packs by SHA, so repeated runs avoid re-uploading the same package.

Remote CUDA training defaults to MLX_CUDA_USE_CUDNN_SDPA=0 and MLX_CUDA_GRAPH_CACHE_SIZE=4096. Override either environment variable before mere-runpod run only when deliberately testing alternate MLX CUDA graph behavior.

RunPod network volumes for pods are Secure Cloud resources and must be attached when the pod is created. They are tied to a data center, so choose one with the H100/A100 availability you plan to use.