SSE chunks echo the requested model, not always the base

per-model tokenizer: the shared one scrambled the v5-lineage finetune

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

api: gradio_client Client(token=) not hf_token= — fix token-path crash

api: token-attributed ZeroGPU quota + structured errors on /v1 and /blah

ft8 weights + rp1.1 ft serving default

two tracks, one Space: base + ANDREA-120M-ft with per-model serving defaults

serving defaults: repetition_penalty=1.2 + no_repeat_ngram_size=3 (Track A, grid-validated)

api: real SSE streaming + real usage token counts

api: ssr_mode=False so the python server owns 7860

api: disable SSR so added routes are reachable on 7860

api: attach routes to gradio app (7860 is owned by the runner)

api: OpenAI-compatible + blah.dev routes alongside the Gradio UI

fix: redirect HF_MODULES_CACHE to a writable path for trust_remote_code

ANDREA-120M chat demo on ZeroGPU

initial commit