- Removed redundant 'import os' statements inside functions (lines 4522, 4926, 5005)
- Added back missing 'from llama_cpp import Llama' that was accidentally deleted
- Global 'import os' at line 12 is now the only one

This fixes the UnboundLocalError when running --list-cached-models or other CLI options.

- --list-cached-models: List all cached models with sizes
- --remove-all-models: Remove all cached models
- --remove-model <modelid>: Remove specific model by name/hash (partial match)

- Check if downloaded file is valid GGUF (magic bytes = 'GGUF')
- If not valid, show clear error that URL is wrong (returns HTML instead)
- Explain that URL must be direct download link ending in .gguf

- Enable verbose=True in llama.cpp to see actual error
- Print GGUF model file size for debugging
- Add try/except with traceback to see detailed errors

- Check if model is URL before any processing
- Use original model name with query params for URL download
- Strip query params only for HuggingFace repo ID parsing
- Added more debug output to trace issues

- Strip query parameters from model name before processing
- Handle URLs with ?download=true or other query params

- Detect if image model is GGUF (ends with .gguf or contains 'gguf')
- If GGUF, load using llama.cpp (same as text Vulkan models)
- If diffusers model, load using Stable Diffusion pipeline
- Fixed both locations where image model preloading happens
- Now supports both GGUF and diffusers image generation models

- Fixed bug where image model wasn't actually being loaded when --loadall was specified
- The code only printed messages but never loaded the diffusers pipeline
- Now actually loads the Stable Diffusion pipeline using diffusers library
- Tries StableDiffusionXLPipeline first, falls back to generic DiffusionPipeline
- Moves to GPU if CUDA available, enables attention slicing for memory efficiency
- Also fixes second location where image model is the only configured model

- Debug command line output was already implemented

- Fixed undefined variable bug where model_name wasn't defined in scope
- Fixed duplicate model loading when using --loadall/--loadswap with multiple models
- First model is now only loaded once (skipped in loop if already loaded)
- Loadall mode now properly preloads all models in VRAM respecting offload strategy
- Loadswap mode properly loads additional models to RAM
- Ondemand mode doesn't reload first model

Feature 1: --debug now shows full command line as first output
Feature 2: --loadall with multiple models now preloads all in VRAM

- Move whisper-server check before audio_model check
- Now whisper-server will be used if available, regardless of audio_model setting
- Also update multi_model_manager.audio_models with cached path

- WhisperServerManager.start() now returns actual model path (useful for URL -> cached path)
- Update audio_models[0] with cached path after downloading
- Store actual_model_path in current_model instead of original URL

- Changed default port from 8081 to 8744 (less common)
- Check if port is available before using, auto-find available port if needed
- Download URL models before starting whisper-server (use model cache)

- Add WhisperServerManager class to manage whisper-server subprocess
- Add --whisper-server argument to specify whisper-server binary path
- Add --whisper-server-port argument for port configuration (default 8081)
- Modify audio transcription endpoint to proxy to whisper-server
- Add cleanup on shutdown to stop whisper-server
- Model can stay loaded in VRAM as long as the server runs

Fix remaining occurrences of model_name (singular) being used instead
of model_names (list) in main function.

Fixed bug where model_name (singular) was used instead of model_names (list)
in several places, causing UnboundLocalError when running without --model.

- CLI (coder): Add --alias argument to create model aliases
- Config: Add model_aliases dict and resolve_model() method
- Server (coderai): Add server-side alias support with --model-alias
- Aliases are resolved in both client and server when making API calls
- Aliases appear in /v1/models endpoint
- Aliases are persisted in config file

- Add QueueManager class to track waiting requests
- Send 'waiting for model...' frames with time counter at regular intervals
- Send 'Model starting' frame when model begins processing
- Add x_queue_info field to streaming response frames for queue status
- Track queue position and wait time for each client

- Add support for multiple --audio-model arguments (action='append')
- Add support for multiple --image-model arguments (action='append')
- Add 'audio' alias pointing to first audio model
- Add 'vision'/'image' aliases pointing to first image model
- Update MultiModelManager to store audio_models and image_models as lists
- Add audio_model and image_model properties for accessing first model
- Update get_model_for_request to handle aliases
- Update list_models to show all models and aliases
- Fix remaining references in main function to use list-based variables

- Changed --model to -m
- Changed --output to -otxt (output as text)
- Changed --device to -dev
- Changed --file to -f for input audio

The args variable was not accessible in the create_transcription function,
causing a NameError when using --whisper-cpp CLI option. This fix adds
global_args to store the parsed arguments for access in endpoint functions.

- Modified build.sh to build whispercpp with Vulkan support
- Added --audio-vulkan-device argument to specify GPU device for Whisper
- Added Vulkan detection and logging for Whisper transcription
- Set GGML_VULKAN_DEVICE environment variable for GPU selection

faster-whisper doesn't support GGUF format (it's llama.cpp format).
Now detects GGUF files by extension and goes directly to whispercpp.

- Add faster_whisper_failed flag to properly track failures
- When faster-whisper throws non-ImportError (e.g., GGUF not supported),
  now falls back to whispercpp instead of failing
- Applies to both pre-loading and transcription endpoint

- Add specific detection for 'invalid ELF' / 'Mach-O' architecture mismatch errors
- Improve error messages to mention both options:
  - Install PyTorch + faster-whisper
  - Use built-in whispercpp model (tiny/base/small/medium/large)
- Fix critical bug: now raises HTTPException instead of returning None

- Recognize built-in model names: tiny, base, small, medium, large-v1, large
- Allow pre-loading these models directly without file path

- Add better error detection for 'not a valid preconverted model' errors
- Provide clear guidance to users about whispercpp limitations
- Suggest installing faster-whisper with PyTorch or using built-in model names
- Update both transcription endpoint and pre-loading code

- Update transcription endpoint to try faster-whisper first, then whispercpp
- Update pre-loading code to support both backends
- Add whispercpp to all requirements files (vulkan, nvidia, default)
- Remove broken llama.cpp fallback (llama.cpp cannot transcribe Whisper)