- Added --vae-tiling flag to enable VAE tiling for lower VRAM usage
- Added --clip-on-cpu flag to run CLIP on CPU to save VRAM
- Both options work with stable-diffusion-cpp-python

- Added --image-seed argument to set default seed for image generation
- Updated diffusers and sd.cpp code to use request seed or CLI default seed
- Priority: request seed > CLI default seed > random

- Added model_key initialization before sd.cpp loading in on-demand section
- Added model_key assignment before adding model to manager

- Fixed model_key variable scope issue in GGUF->sd.cpp fallback
- Fixed model_path undefined in diffusers preloading section
- These fixes prevent startup crashes when using --loadall

- Reordered the image generation backend priority to try torch/diffusers first
- If torch/diffusers fails (ImportError or other error), fallback to stable-diffusion-cpp-python
- If both backends fail, return a helpful error message with installation instructions
- Added dynamic loading of sd.cpp model if not pre-loaded

- Add debug output to trace audio model registration at startup
- Add debug output when audio endpoint checks for audio_model
- Fix global load_mode to be updated at startup based on --loadall/--loadswap flags

- Fix missing indentation in async with semaphore block
- Fix invalid elif syntax in load_mode determination
- Fix request.steps reference (field doesn't exist in request model)

- Without --loadall: serialize all requests (one at a time)
- With --loadall: allow one concurrent request per model

These parameters are not available in the installed version of
stable-diffusion-cpp-python. The engine will auto-detect the
architecture from file headers.

- Added is_huggingface_model_id() helper to detect HF model IDs
- Added download_huggingface_model() to download from HF Hub
- Updated download logic to handle model IDs, URLs, and local paths

- Renamed CLI argument from --clip-l-path to --llm-path
- Updated all references from args.clip_l_path to args.llm_path
- Changed config dictionary key from 'clip_l_path' to 'llm_path'
- Added model_type='z-image' and backend='vulkan' parameters

The bug was that the code checked 'if clip_l_path:' but clip_l_path
was just set to None above, so the download block was always skipped.

Changed to check 'if args.clip_l_path:' to properly detect when the
--clip-l-path CLI argument is provided and trigger the download/caching.

- Changed incorrect 'elif clip_l_path:' to proper 'if clip_l_path:' logic
- Fixed vae_path assignment to use args.vae_path instead of undefined vae_path variable
- This ensures CLIP LLM and VAE paths are properly passed when using
  --clip-l-path and --vae-path CLI arguments with Vulkan backend

These are txt2img() parameters, not constructor parameters