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

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

Symmetric encoder batching: ENC_BATCH cap + chunking + GPU probe

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

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

Remove flash_attention_2 (not installed on HF Space)

Debugging

restrict decoder inference to only emojis

Encoder support for non-word-based languages

Trying faster inference with flash attention

Fix model IDs and update decoder dropdown labels

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