fallback script on zerogpu expiry

extend zerogpu script duration

fix(zerogpu): normalize tiny-model script speakers

fix(zerogpu): persist raw rank diagnostics and script indexes

fix(zerogpu): keep fallback rank items above threshold

refactor: unify storage layer and modularize Gradio web app

feat: ZeroGPU backend for Hugging Face Spaces deployment (#4)

feat: add scheduled job backend and resumable pipeline sync

feat: implement automatic CPU fallback for ZeroGPU queue lease execution failures

Optimize ZeroGPUBackend model loading by caching weights in system CPU RAM at startup/request start, moving to GPU instantly only during leases to bypass Expired ZeroGPU proxy token timeouts

feat: implement budget-aware resilient queue for ZeroGPU backend

Fix ZeroGPU timeouts by increasing chunk size and reducing top items

Increase ZeroGPU duration to 120s

Apply user's zerogpu.patch

Fix ZeroGPU: Use snapshot_download and lazy loading pattern

Fix ZeroGPU: module-level model loading with bfloat16, drop bitsandbytes

Fix: prevent low-level CUDA init outside of spaces.GPU decorator on ZeroGPU

fix: load quantized model with device_map='auto' and explicit max_memory for ZeroGPU emulation

fix: change device_map from auto to cuda for ZeroGPU compatibility

fix: resolve NVML assertion failure in ZeroGPU backend by loading globally

fix: pin torch==2.4.0 to fix NVML crash on ZeroGPU

fix: force eager attention to avoid NVML crash in SDPA on ZeroGPU

fix: lazy-load model inside @spaces.GPU for ZeroGPU

fix: move model to CUDA inside @spaces.GPU function

Restrict model loading to CUDA-enabled environments on ZeroGPUBackend

Support dynamic devices (MPS/CPU) on ZeroGPUBackend for local macOS testing

Improve model load error logging in ZeroGPUBackend by printing the full traceback

Refactor ZeroGPUBackend to use BAML's native render and parse API

Implement switchable LLM backends (BAML & ZeroGPU) for HF Spaces deployment