swaleio/transformers-gpu repository overview

⁠transformers-gpu

Swale-built image backing the transformers, trl-sft, and gpu-smoke-test task definitions. It layers the Hugging Face training stack on top of the official PyTorch CUDA runtime, so a workflow can run inference, supervised fine-tuning, or arbitrary GPU Python without a build step.

• Base: pytorch/pytorch:2.13.0-cuda13.0-cudnn9-runtime (torch 2.13.0, CUDA 13.0, cuDNN 9)
• Installed (pinned): transformers==4.57.1, accelerate==1.11.0, datasets==4.4.1, trl==0.24.0, peft==0.18.0, sentencepiece==0.2.1, huggingface_hub[hf_xet]==0.36.0
• User: non-root swale (uid 1000)
• Hub cache — defaults to the Hugging Face location under the home directory, which the platform mounts writable and per-task, so it is discarded with the task. A script that wants the weights reused by later tasks in the run sets HF_HOME (or from_pretrained(cache_dir=...)) to a workspace path itself — this image is driven by a free-form script, so nothing here forces the default.
• WORKDIR /mnt/workspace, no ENTRYPOINT

⁠Invocation

The image declares no entrypoint (the PyTorch base leaves it unset), so a task definition's exec.args are the full command line — one image serves a fixed CLI task (["trl", "sft", "--config", "${{inputs.config}}"]) and a free-form script task (["bash", "-lc", "… exec python /tmp/run.py"]) alike. bash, trl, and python are all on PATH.

Hugging Face authentication is not the image's job: a definition that needs it maps its token input to HF_TOKEN via its own exec.env (an omitted token resolves to an empty HF_TOKEN, which the Hugging Face stack treats as no token).

⁠GPU

This is a CUDA image. Tasks that reference it require a GPU compute type — the workflow selects the compute type, and scheduling on a CPU type fails at runtime with CUDA errors (gpu-smoke-test exists to surface exactly that, fast).

⁠Building

From the repository root:

docker build -t docker.io/swaleio/transformers-gpu:1.0.0 images/transformers-gpu

CI pins the real base digest and publishes the image; task definitions reference it as docker.io/swaleio/transformers-gpu@sha256:<digest>.

Built from images/transformers-gpu⁠ in the swaleio/task-definitions⁠ repository. For how a task definition references an image, see the task definition reference⁠ in the Swale documentation⁠.

The Dockerfile and entrypoint are covered by the MIT License⁠. The software this image packages keeps its own licenses, which ship inside the image.