Production-ready: VRAM-sized batches, frontend API URL fix, repo hygiene

Add temporary /api/debug/gpu to read real VRAM for batch sizing

Restore demo.launch() (required for ZeroGPU) + attach router to live demo.app

Bump Gradio to 5.50.0 and mount UI at /ui so the API works

Mount Gradio under our FastAPI app so /api routes actually register

Symmetric encoder batching: ENC_BATCH cap + chunking + GPU probe

Disable Gradio SSR so /api routes aren't shadowed by the SSR catch-all

Fix API routes shadowed by Gradio SSR catch-all

Probe: report actual prompt_tokens to confirm worst-case length

Serialize inference behind a process-wide lock

Probe: binary-search exact max batch (not just powers of two)

Encoder: use batch word_ids (drop per-sentence re-tokenize); per-model DEC_BATCH + GPU probe

Fix encoder LoRA scaling (read alpha from ckpt); chunk decoder batch

Fix ZeroGPU startup: load PEFT adapter on CPU; torch_dtype->dtype

Stop tracking tsconfig.tsbuildinfo; ignore one-off scripts

Replace models: new encoder + gemma-3-1b/qwen3 decoders, drop llama3

Remove flash_attention_2 (not installed on HF Space)

Add GitHub Actions workflow to sync to HF Space

Debugging

restrict decoder inference to only emojis

Encoder support for non-word-based languages

Replace dart with tick, make light mode a bit darker

Trying faster inference with flash attention

Font Improvements

Dark mode added

UI improvements, background added

Fix: use return value of demo.launch() to get FastAPI app

Switch to demo.launch() for ZeroGPU compatibility

Remove double spaces.GPU wrapping — inference.py already applies it

Fix model IDs and update decoder dropdown labels

Fix HF Space startup: run uvicorn unconditionally, guard spaces import

Run uvicorn explicitly to keep HF Space process alive

Pin huggingface-hub<1.0.0 to fix HfFolder compatibility with gradio

Upgrade to Gradio 5 to fix huggingface_hub compatibility

Fix HF Space README color

Add HF Spaces README metadata

UI improvements and HF Spaces preparation

Next.js and React implementation. UI improvements. Cloze test removed from the website.

Implemented and deployed encoder + decoder approac

Current website

Phase 3 completed

Phase 1 and 2 are complete

Phase 0 completed