ADAPT-Chase's picture
Add files using upload-large-folder tool
93be2a2 verified
# Elizabeth Containers for Vast.ai / Cast.ai
## Images
- Serving image: `docker/serve/Dockerfile`
- Training image: `docker/train/Dockerfile`
## Model Mount
- Mount your model directory into the container:
- Host: `/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft`
- Container: `/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft`
- Or mount at `/models` and set `MODEL_PATH=/models/qwen3-8b-elizabeth-sft`.
## Serving Container
- Build:
- `docker build -f docker/serve/Dockerfile -t <repo>/elizabeth-serve:latest .`
- Run (local test):
- `docker run --gpus all -p 8000:8000 -v /models/qwen3-8b-elizabeth-sft:/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft:ro --env ELIZABETH_API_KEY=elizabeth-secret-key-2025 <repo>/elizabeth-serve:latest`
- Fallback seeding:
- From HF (private): set `PREFER_HF=1 HF_TOKEN=<token> HF_ORG=LevelUp2x MODEL_NAME=qwen3-8b-elizabeth-checkpoints`
- From seed node: set `SEED_HOST=user@host` (rsync)
## Training Container
- Build:
- `docker build -f docker/train/Dockerfile -t <repo>/elizabeth-train:latest .`
- Run (mount ckpts):
- `docker run --gpus all -v /models/qwen3-8b-elizabeth-sft:/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft:ro -v /data/ckpts/elizabeth:/data/ckpts/elizabeth <repo>/elizabeth-train:latest`
- Exec your training:
- `accelerate launch your_train.py --model /data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft --output_dir /data/ckpts/elizabeth`
## Vast.ai
- Specify the image (pushed to Docker Hub or GHCR).
- Mount a persistent NVMe volume at `/models` with the model files.
- Env:
- `ELIZABETH_API_KEY=...` (serving)
- `MODEL_PATH=/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft`
- Optional: `HF_TOKEN`, `PREFER_HF=1`, `SEED_HOST` for bootstrap.
## Promote Checkpoint
- After training, validate then promote:
- `bash scripts/promote_checkpoint.sh /data/ckpts/elizabeth/checkpoint-XXXX /models/qwen3-8b-elizabeth-sft`
- Restart serving container or let health-monitor reload.