- Add offload_strategy to kwargs in _load_default_model and _load_model_by_name
- Fix parameter name: ram -> manual_ram_gb to match backend expectation
- Also pass load_in_4bit, load_in_8bit, and max_gpu_percent

- Add 'none' to --offload-strategy choices in cli.py
- In cuda.py backend:
  - _get_vram_percentages_for_strategy() returns None for 'none' strategy
  - _get_vram_percentages_for_gpu() skips VRAM detection for 'none'
  - load_model() loads directly on GPU without max_memory constraints
- Add startup status message in main.py for --offload-strategy none

- Add --no-ram CLI option to force model loading without CPU RAM spilling
- Implement --no-ram behavior for:
  - llama-cpp-python: n_gpu_layers=-1, use_mmap=False, ignore --n-ctx
  - HuggingFace transformers: device_map='cuda:0', low_cpu_mem_usage=True
  - Diffusers: force full GPU loading
  - sd.cpp: maximize GPU usage
- Propagate flag through model manager
- Add startup banner message

- Add get_all_allowed_identifiers() to MultiModelManager returning all valid
  model identifiers (default model + short name + aliases, audio, tts, image,
  vision models, and custom aliases)
- Rewrite is_allowed_model() to check against the full allowed set with
  support for prefixed forms and short-name matching
- Add validation in request_model() that rejects unknown models with an error
  message listing all available models
- Fix get_model_for_request() to reject loading arbitrary models not in the
  allowed set
- Update all API endpoints (text, images, tts, transcriptions) to check for
  the error key and return HTTP 404 when a disallowed model is requested

- Try GGUF pattern first for HuggingFace model IDs
- Fall back to snapshot_download for entire repo (transformers/diffusers models)
- Works for both GGUF models and full HuggingFace repos

- Remove auto-detection logic, just use download_model from cache
- User can specify --download-file-pattern for non-GGUF models

- Scan HuggingFace repo to detect available file patterns
- Try multiple patterns (.gguf, .safetensors, .bin, .pt, .pth)
- Default to .gguf if nothing found

- Add --download-model argument to download a model (URL or HuggingFace ID) to cache
- Add --download-file-pattern argument to specify file pattern for HF downloads
- Use download_model from codai.models.cache module
- Model downloads to appropriate cache and exits without starting server

text.py's set_global_args() was only setting its local global_args but
not calling state.set_global_args(). This meant _load_default_model()
and _load_model_by_name() got None from get_global_args(), so CLI flags
like --flash-attn, --n-gpu-layers, --ram were not passed to backends.

- Add flash_attn extraction from global_args in _load_default_model()
- Add flash_attn extraction from global_args in _load_model_by_name()
- Now --flash-attn flag will properly enable Flash Attention 2 when loading models

- ModelParserDispatcher: Only log parser selection when actually used for parsing
- ModelParserAdapter: Defer dispatcher creation until first use
- Fixes noisy 'model_name=None, selected parser: ApexBig50Parser' during initialization

- loadall: pre-load image models into VRAM at startup (with OOM fallback)
- loadswap: pre-load image models into CPU RAM at startup (first model stays in VRAM)
- Audio and TTS models are cached at startup, loaded into memory on first request
  (they use specialized loading mechanisms via faster-whisper and kokoro)

- Default mode changed to ondemand (pre-load first model, unload/load on switch)
- loadswap: load first model in VRAM, others in CPU RAM, swap on switch
- loadall: try to load all models in VRAM, offload to CPU RAM if OOM
- --nopreload: skip pre-loading in any mode, load on first request
- request_model() now properly handles all three modes
- Added _move_model_to_cpu() and _move_model_to_vram() for loadswap
- Fixed NameError: model_manager reference in request_model() (was using global singleton instead of self)
- Updated CLI help text for --loadall, --loadswap, --nopreload

- Added request_model() method to MultiModelManager that handles:
  1. Alias resolution (image, audio, tts, vision, default, custom aliases)
  2. VRAM management (unloading previous models in ondemand mode)
  3. Checking if model is already loaded

- Simplified codai/api/images.py:
  - Uses request_model() for model resolution and VRAM management
  - Extracted helper functions: _is_gguf_model(), _load_diffusers_pipeline(),
    _generate_with_diffusers(), _generate_with_sdcpp(), _load_sdcpp_model()
  - Removed duplicated sd.cpp generation code
  - Fixed semaphore scope (all generation now inside semaphore block)

- Simplified codai/api/tts.py:
  - Uses request_model() instead of duplicated VRAM management code
  - Removed duplicate get_cached_model_path() and get_model_cache_dir() wrappers

- Simplified codai/api/transcriptions.py:
  - Uses request_model() instead of duplicated VRAM management code

- Simplified codai/api/text.py:
  - Both /v1/chat/completions and /v1/completions use request_model()
  - Removed duplicated VRAM management blocks

- **Model Manager**: Central coordinator for model lifecycle, alias resolution, loading/unloading
- **Cache Module**: Handles downloading, caching, and storage of models
- **API Modules**: Request models from Model Manager (not directly from cache)

Key changes:
- Removed resolve_and_load_model() from cache - moved logic to Model Manager
- Model Manager now downloads/caches models at startup when registered
- API modules use multi_model_manager.load_model() instead of cache functions
- Proper separation: Cache=storage, Manager=lifecycle coordination, APIs=requests

This fixes the incorrect direct API-to-cache coupling and establishes proper architectural boundaries.

- Added resolve_and_load_model() function to codai.models.cache
- Simplified codai/api/images.py by removing 100+ lines of complex model resolution logic
- API modules now use single centralized function for all model loading
- Eliminates code duplication across API endpoints
- All model resolution logic now managed in one place

- Added check in sd.cpp fallback to skip HF model IDs that are likely diffusers models
- Prevents sd.cpp from trying to download non-GGUF files like .gitattributes for diffusers models
- Tongyi-MAI/Z-Image-Turbo and similar diffusers models now handled correctly by diffusers library
- GGUF models still work with sd.cpp as before

- Updated codai/api/images.py to use cache module functions directly
- Updated codai/api/tts.py to use centralized load_model() function
- Removed proxy method calls that were causing AttributeError
- All model loading/downloading now goes through codai.models.cache

- Updated load_model() to handle three input types:
  1. Local files: Use directly without caching
  2. URLs: Download to cache if not cached, then use
  3. HF model IDs: Download via HF API if not cached, then use
- Updated get_cached_model_path() to validate local files
- Enhanced module documentation to reflect new capabilities
- All model types (text, image, audio, etc.) can now use any input type

- Updated remove_cached_model() to remove entire repo directories when matching by repo_id
- Previously only removed individual files, now removes complete repository cache
- Handles both files and directories in removal process
- More thorough cleanup of HuggingFace cached models

- Added unified load_model() function as main entry point for model loading
- Updated WhisperServerManager to use centralized load_model() instead of inline logic
- Removed proxy methods from MultiModelManager - use cache module directly
- All cache functions now work seamlessly with both GGUF and HF model caches
- Improved separation of concerns: cache module handles all caching/downloading

- Updated get_cached_model_path() to check both coderai and HF caches
- Updated download_model() to handle both URLs and HF model IDs automatically
- Made download_huggingface_model() consistent with unified API
- Updated module docstring to reflect unified cache functionality
- All cache functions now work seamlessly with both cache types

- Updated remove_cached_model() to search by repo_id for HuggingFace models
- Moved cache management options (--list-cached-models, --remove-model, --remove-all-models) to run before heavy imports
- Improved cache operations to use centralized functions in codai.models.cache module
- Fixed model removal to work with full repo IDs like 'TheBloke/Llama-2-7B-GGUF'

- Add list_cached_models_info() function to codai.models.cache module
- Move cache listing logic from main.py to the cache module
- Update main.py to use the centralized function early (before heavy imports)
- Improves code organization and avoids unnecessary imports for --list-cached-models

- CoderAI cache: Shows individual GGUF files with sizes
- HuggingFace cache: Uses HF API (scan_cache_dir) to show model-level info, not individual files
- Shows model names, sizes, revision counts - not thousands of individual files
- Much more useful and readable output

- Added code to print individual cached model files with sizes
- Previously only showed cache directory headers and summary
- Now shows each file with format: [cache_name] filename (size MB)
- Matches the format used by --remove-model command

- Updated get_all_cache_dirs() to properly find HuggingFace hub directory
- Now checks for ~/.cache/huggingface/hub/ instead of just ~/.cache/huggingface/
- This fixes --list-cached-models not showing HuggingFace cached models

- Removed the GGUF-only restriction on sd.cpp fallback
- Some HF models may be GGUF even without 'gguf' in the name
- Let sd.cpp attempt loading and fail gracefully if incompatible
- This allows sd.cpp to work as a proper fallback for any model type

- Added check to only attempt sd.cpp fallback for GGUF models
- Tongyi-MAI/Z-Image-Turbo is a diffusers model, not GGUF, so sd.cpp should be skipped
- sd.cpp only supports GGUF models, diffusers models use the diffusers pipeline
- This prevents unnecessary sd.cpp resolution attempts for incompatible model types

- Added proxy methods to MultiModelManager class for cache module functions
- These methods are called by images.py sd.cpp fallback path
- Fixes AttributeError: 'MultiModelManager' object has no attribute 'get_cached_model_path'

- Enhanced the HF model resolution logic in images.py sd.cpp fallback path
- Now checks for ANY cached file from the repo first (not just GGUF files)
- Falls back to checking for cached GGUF files specifically
- Last resort: downloads the first file in the repo as fallback
- Better error handling and logging throughout the resolution process
- This should resolve models that are already cached even if the exact GGUF filename isn't known

- Enhanced model resolution for sd.cpp fallback path
- Added multiple fallback strategies:
  1. Try HuggingFace GGUF resolution (existing)
  2. Fallback to direct file path check
  3. Fallback to cached model lookup
  4. Last resort: attempt download as URL
- Better error logging and handling
- Ensures model loading attempts all possible resolution paths before failing

- Added model resolution and unload logic to /v1/audio/transcriptions
- Added model resolution and unload logic to /v1/audio/speech (TTS)
- Now ALL endpoints (text, image, audio, TTS) properly handle model switching
- In ondemand mode, ANY model type switch triggers unload first (e.g., text->audio, TTS->image, etc.)

- Added resolve_model_name() to MultiModelManager to properly resolve model aliases
- Added get_currently_loaded_model_name() to track what's actually in VRAM
- Updated /v1/chat/completions, /v1/completions, and /v1/images/generations
- Now correctly compares resolved canonical names before deciding to unload
- Handles all aliases (default, image, audio, tts) and custom aliases
- Works across ALL model types: text->text2, image->image2, text->image, etc.

- Added unload_all_models() to MultiModelManager that handles ALL model types:
  ModelManager, diffusers pipelines, sd.cpp StableDiffusion, and any other objects
- Text endpoints now properly unload image models before loading text models
- Image endpoints now properly unload text models before loading image models
- The rule: in ondemand mode, if the model in VRAM differs from the requested
  model (regardless of type), fully unload before loading the new one
- Includes gc.collect(), torch.cuda.empty_cache(), and 1s settle delay

- In ondemand mode (no --load-all or --loadswap specified), when a new model
  is requested, the current model in VRAM is now fully unloaded before loading
  the new one. This ensures clean model switching.
- Added cleanup logic to both /v1/chat/completions and /v1/completions endpoints
- Added same logic to image generation endpoints (diffusers and sd.cpp paths)
- Cleanup includes: model cleanup, gc.collect(), torch.cuda.empty_cache()

Root cause: The refactored code was hardcoding torch.float16 for CUDA,
ignoring the --image-precision bf16 CLI argument. The Z-Image-Turbo model
requires bfloat16 precision - using float16 causes NaN values in the
image processor, resulting in all-black images.

Also restored the original model loading logic with:
- GGUF model detection (skip diffusers for GGUF)
- OOM retry with progressive memory optimization
- use_safetensors=True
- Sequential CPU offload support

- Changed default image size from 512x512 back to 1024x1024 to match original coderai
- Changed NaN handling from 0.5 to 0.0 to match original coderai