calvinmoody1/codontransformerstuff
ASU CodonTransformer
This Gradio Space exposes the ASU wiki's CodonTransformer as two API endpoints: /organisms and /predict. Select ZeroGPU as the Space hardware in the Space settings.
The public CodonTransformer model is used when no fine-tuned checkpoint is present. The ASU yarn build:hf-space export automatically includes the local CodonTransformer source and ADP1 checkpoint. The source package is copied from the sibling CodonTransformer checkout, including its ADP1 and checkpoint compatibility changes. Those source files are public in this Space. The source is bundled directly because the upstream package metadata fails to build on Linux due to its case-sensitive package-directory name. If publishing code and model separately, upload the checkpoint to a public Hugging Face model repository and set these Space variables:
CODONTRANSFORMER_MODEL_REPO_ID:owner/model-repositoryCODONTRANSFORMER_MODEL_FILENAME:a_baylyi_adp1/a_baylyi_adp1.ckpt
The checkpoint is downloaded at Space startup. Do not put a Hugging Face token in the wiki or in this Space's public source.
Free personal accounts can host up to two ZeroGPU Spaces if the account is in good standing, has a verified email, and is at least 30 days old. Free ZeroGPU has per-user daily GPU quotas, a shared queue, and cold starts. Unauthenticated visitors currently receive 2 GPU minutes per day; signed-in free Hugging Face account holders receive 5 minutes per day. Calls through the public wiki do not send a Hugging Face credential and should be treated as unauthenticated. A prediction reserves up to 90 seconds of GPU runtime; hardware-specific quota factors can increase the charged amount. Long inputs may exceed this runtime ceiling. This is a zero-cost proof of concept, not dependable public production hosting.
The Space source and checkpoint repository must both be public for unauthenticated wiki visitors to use them.
The Space image supplies and pins its own Gradio, spaces, Uvicorn, and PyTorch versions. The requirements file intentionally does not request separate versions of those packages; an additional, newer Gradio or spaces release can conflict with the versions selected by the ZeroGPU image during dependency resolution. Keep sdk_version at 6.16.0: newer Gradio releases require huggingface-hub>=1.2, while CodonTransformer's transformers<5 requires huggingface-hub<1.0. Gradio 6.16 accepts Hub versions below 2, so the resolver can select a version satisfying both packages. Runtime dependencies from the CodonTransformer project metadata are listed in requirements.txt.