- Add 'all' as a valid backend option
- Change default from 'nvidia' to 'all'
- Add comprehensive 'all' backend section that installs:
  - Base requirements
  - PyTorch with CUDA (nvidia backend)
  - llama-cpp-python with CUDA and Vulkan support
  - stable-diffusion-cpp-python with OpenCL
  - Additional requirements
- Detect available hardware (CUDA, Vulkan, OpenCL) and enable accordingly
- Show summary of available backends after installation

- Store original backend before switching to vulkan for GGUF files
- Pass original_backend to VulkanBackend constructor
- Add force_cuda flag that triggers CUDA environment setup
- Set CUDA_VISIBLE_DEVICES when force_cuda is True
- Update success/error messages to reflect actual backend used
- Add debug output for CUDA detection

The local import of os inside the HTTPS block caused Python to treat os as a local variable throughout the main() function.

- Add python-multipart to requirements.txt, requirements-nvidia.txt, requirements-vulkan.txt
- Add llama-cpp-python to requirements-nvidia.txt for GGUF support
- When using CUDA/nvidia backend with GGUF file, automatically use llama-cpp-python

- Add --vision-model for image/video-to-text models
- When --file-path is set, return URL by default, base64 only if explicitly requested
- Add --https flag with auto-certificate generation
- Add --privkey and --pubkey for custom certificates

- Add --image-cfg-scale CLI option (default 1.0)
- Add get_cfg_scale() helper that auto-detects VRAM
- If Vulkan and VRAM < 16GB, use cfg_scale=1.0 automatically

- Add --image-backend, --audio-backend, --tts-backend CLI args
- Add opencl to backend choices
- Add OpenCL build target in build.sh

- When --debug is enabled, show full command line coderai was called with
- Add --nopreload flag to disable model preloading at startup
- When --nopreload is not specified, skip checking for preloaded sd.cpp models (forces load in worker thread to avoid Vulkan context issues)
- Fix image model preloading to respect --nopreload flag

This helps prevent the GPU from staying at 100% utilization after
image generation by allowing the Vulkan driver to transition from
compute state to idle state.

- Use keep_clip_on_cpu instead of clip_on_cpu
- VAE tiling is handled internally in newer builds
- Add n_threads using psutil.cpu_count() for optimal performance

- 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