Run on ZeroGPU: spaces @GPU + CUDA llama.cpp (n_gpu_layers=-1)

Fix CPU perf: pin n_threads=2 (cpu-basic), tighter token budget

Rebuild as native Gradio app: server-side LFM2 (GGUF) rollout

Bake React routes into Gradio app before launch via _app=

Disable Spaces hot-reload so React UI serves (fixes 7861 + hijack)

Serve React UI via Gradio's own launched server (cpu-basic)

Restore real app: FastAPI(React+API) + mounted Gradio on cpu-basic

diag: vanilla gradio app to discover HF launch mechanism

Fix gradio Space startup: override demo.launch() to serve the full app

Fix gradio Space startup: serve at import time on GRADIO_SERVER_PORT

Redeploy as gradio SDK Space (no custom Docker)

Revert "exp: drop "Predict next frame:" suffix — append-only prompt format"

exp: drop "Predict next frame:" suffix — append-only prompt format

perf: multi-frame v2 — no_repeat_ngram_size kills repetition loops

docs: refresh dtype comment — FA now lit, not bypassed

Revert "perf: emit N frames per generate() — amortize prefill across rollout"

fix: re-export ONNX without gqa_attention_bias subgraph (LFM2.5-VL parity)

perf: emit N frames per generate() — amortize prefill across rollout

perf: cascade graphOpt basic→disabled, recover safe fusions

perf: switch WebGPU to q4f16 (fp16 compute, ~2x speedup) with q4 fallback

fix: disable ORT graph optimizer to preserve float32 bias types on WebGPU

fix: force device=wasm, ORT WebGPU breaks GQA float32 type constraint

deploy: fix WebGPU model loading — use q4 dtype, add debug logging