Hugging Face models from a Python function

Two examples of a RunR Python function using models deployed from Models (one click from the Hugging Face browser) and bound to the project in its Models tab.

File	Runtime	What it does
`main.py`	Python (XaaS)	calls the model services over HTTP with the standard library: embeddings through `EMBED`, a chat answer through `CHAT`
`weights/main.py`	PyTorch (FaaS, with `requirements.txt`)	loads the weights of the model bound as `EMBED` with Mount the model files from `RUNR_MODEL_EMBED_PATH` and computes the embedding in the function itself

Set up#

Models → Browse Hub: deploy sentence-transformers/all-MiniLM-L6-v2 (task sentence-similarity, engine TEI or transformers, CPU) and a chat model (for example Qwen/Qwen2.5-0.5B-Instruct; vLLM or TGI on a GPU, or transformers on CPU).
Project → Models: bind the embedding service as EMBED and the chat service as CHAT. For weights/main.py tick Mount the model files (the model needs an offline snapshot in the project's region: Models → Library → Mirror to region).
Paste main.py into the editor and deploy. Bindings are applied on deploy.

Call it#

curl -s -X POST https://<function>.<your function domain>/ -d '{"text": "hello"}'
# {"dims": 384, "head": [...]}

curl -s -X POST https://<function>.<your function domain>/ -d '{"action": "chat", "text": "What is RunR?"}'
# {"answer": "..."}

Environment of a bound model#

Variable	Example
`RUNR_MODEL_EMBED_URL`	`http://m-minilm.runr-projects.svc.cluster.local`
`RUNR_MODEL_EMBED_MODEL`	`all-MiniLM-L6-v2` (the `model` field of OpenAI-style requests)
`RUNR_MODEL_EMBED_API`	`tei`, `openai`, `runr` or `custom`
`RUNR_MODEL_EMBED_TASK`	`sentence-similarity` (the pipeline tag)
`RUNR_MODEL_EMBED_OPENAI_BASE`	`http://m-minilm.runr-projects.svc.cluster.local/v1`
`RUNR_MODEL_EMBED_PATH`	`/models/embed` (with Mount the model files only)
`RUNR_MODELS`	`EMBED,CHAT` (every alias of the project)

The calls stay inside the cluster; an offline region runs these examples unchanged. The full reference is in docs/huggingface.md.