- Add Qwen-specific tool call parsing in ToolCallParser
- Support for Instruct-style: <tool_call>{JSON}</tool_call>
- Support for Coder-style: <tool_call><function=name><parameter=k>v</parameter></function></tool_call>
- Add model_name attribute to ToolCallParser for model-specific parsing
- Update ModelManager.load_model to set model name on tool parser
- Fix duplicate method definitions in ToolCallParser class

- Pass response_format to llama.cpp create_chat_completion
- Supports {'type': 'json_object'} for JSON output mode
- Applied to both streaming and non-streaming responses

When chat_template is 'default' (embedded GGUF template), allow llama.cpp
to handle tool messages via create_chat_completion instead of forcing
manual formatting. This allows the GGUF's native template to be used
even with tools.

When chat_template is 'default', it means llama.cpp detected an embedded
template in the GGUF model. Don't fall back to manual formatting - instead
let llama.cpp's create_chat_completion use its internal template handling.

- Add provider object with provider_name and provider_id
- Add system_fingerprint (null)
- Add logprobs in choices (null)
- Add native_finish_reason in choices
- Add usage.prompt_tokens_details with cached_tokens and audio_tokens
- Add usage.completion_tokens_details with reasoning_tokens and audio_tokens
- Apply to both streaming and non-streaming responses

- Directly set chat_template to known template names (qwen3, qwen, llama3, etc.)
  instead of trying to load non-existent HuggingFace tokenizers
- Add use_manual condition to use manual formatting when chat_template is set
  but hf_tokenizer is None (applies to both generate_chat and generate_chat_stream)
- This ensures GGUF models loaded from URLs with known templates use proper
  <|im_start|> formatting instead of failing on create_chat_completion

When HF tokenizer loading fails, try known template names based on model name:
- Qwen models: try qwen3, qwen templates
- Llama models: try llama3, llama templates
- Phi models: try phi template
- Mistral models: try mistral template

This helps when the tokenizer can't be loaded but we know the model family.

The model_backend_types attribute was not being initialized properly due to
incorrect indentation, causing 'MultiModelManager' object has no attribute
'model_backend_types' error when trying to load models on-demand.

- Add uppercase quantization suffixes (_Q4_K_M, etc.) to handle cached GGUF filenames
- Add progressive fallback to try shorter model names when tokenizer loading fails
- Example: Qwen3.5-27B-Uncensored-HauhauCS-Aggressive -> try Qwen3.5-27B-Uncensored -> Qwen3.5-27B -> Qwen3.5 -> Qwen
- Add warning when all tokenizer loading attempts fail (will use manual formatting instead)

- Added _aggressive_vram_cleanup method to properly clear VRAM
- Moves model to CPU before deletion
- Deletes pipeline, vae, text_encoder, tokenizer explicitly
- Multiple rounds of gc.collect()
- Uses torch.cuda.synchronize() before clearing cache
- Increased delay to 5 seconds after cleanup

- Now can specify template directly: --hf-chat-template "model:template"
- Updated check_hf_chat_template to return tuple (should_use, template_name)
- Updated _load_huggingface_tokenizer to accept template_name parameter
- Updated README with new syntax and template examples

- Added 'auto' as a valid value for --hf-chat-template
- When --hf-chat-template auto is used, it auto-detects and applies HF template to all models
- Updated README with new syntax

- Changed --hf-chat-template from boolean to action=append
- Added check_hf_chat_template() function for model-specific checking
- Updated _finalize_chat_template_detection to use new function
- Updated README with new syntax

- Added --hf-chat-template CLI flag to use transformers apply_chat_template
- Added _load_huggingface_tokenizer() to load HF tokenizer for GGUF models
- Added _format_messages_hf() method for HF chat template formatting
- Updated generate_chat and generate_chat_stream to use HF tokenizer when available
- Updated format_messages to check for HF tokenizer first
- Added documentation in README.md

- Added --reply-filters CLI flag to make content filtering optional
- Supports comma-separated values: --reply-filters malformed,tool_calls
- Supports model-specific filters: --reply-filters text:malformed --reply-filters image:tool_calls
- Supports specific model names: --reply-filters text:llama-3.1:malformed
- Added check_reply_filter() and check_single_filter() helper functions
- Updated stream_chat_response and generate_chat_response to use new filtering
- Updated ToolCallParser._filter_malformed_content for conditional filtering
- Added documentation in README.md

- Give more time for Vulkan memory to be freed after unloading image models

- Add garbage collection and torch.cuda.empty_cache() after unloading image models
- Add a small delay to allow VRAM to be freed before loading new model
- This should help prevent OOM errors when switching between image and text models

- Remove stripping in strip_tool_calls_from_content function
- Whitespace (spaces, newlines) are valid content and should be preserved

- When reloading a default model that was loaded from a URL,
  check for cached file path and use it instead of the URL

- Replace NaN and Inf values with valid values before saving
- Clip image values to valid range [0, 1] to prevent black images

- When 'default' model is requested but not loaded (was unloaded for image model),
  the code now tries to reload the default model
- Cleanup image models first to free VRAM, then reload the text model

- Add --image-cpu-offload CLI flag for explicit sequential CPU offload
- Enable sequential CPU offload only on 3rd OOM retry or when --image-cpu-offload is set

- Skip URLs when listing the default model in list_models()
- This prevents download URLs from appearing in available models list

- Enable sequential CPU offload if --offload-strategy or --offload-dir is specified
- Add retry logic: on OOM, retry with attention_slicing, then with sequential_offload
- Clear CUDA cache between retry attempts

- Add --image-precision with choices: bf16, f32, f16, f8
- bf16 recommended for modern GPUs (RTX 30/40 series) to avoid NaN issues
- Enable VAE tiling for diffusers when --vae-tiling is specified

- Changed torch_dtype from float16 (when CUDA available) to float32
- This prevents NaN/Infinity values in image output that cause black/corrupted images
- FP16 can cause numerical overflow on some models like SDXL

- Add 'steps' parameter to ImageGenerationRequest (overrides quality-based default)
- Add 'guidance_scale' parameter to ImageGenerationRequest (overrides CLI --image-cfg-scale)
- Use request values in diffusers pipeline call

- Import time module inside try block with alias to avoid UnboundLocalError
- This prevents Python's exception handling from affecting variable scope

- Remove duplicate 'image' entry in list_models()
- Remove vision: alias (user doesn't want it)
- Skip URLs in loaded models listing (they're download sources)
- Add full traceback to diffusers error for debugging

- Recursively scan huggingface cache directory (hub/, xet/, etc.)
- Also fix remove-model to search recursively in huggingface cache

- Add get_all_cache_dirs() to find GGUF, HuggingFace, and Diffusers caches
- Update --list-cached-models to show all cache locations
- Update --remove-all-models to clean all cache directories
- Update --remove-model to search across all caches
- Add better error handling for diffusers image extraction

- When --file-path is set and --url is 'auto', use the Host header
  from the request (what the client used to connect) instead of
  the client's IP address
- This ensures the returned URL points to the correct server