refactor: enhance model unloading and memory management for improved GPU efficiency

refactor: enable CPU offload and adjust model loading for improved performance

refactor: disable KV cache to prevent quality degradation after multiple requests

refactor: enable 8-bit quantization and adjust device map for improved model loading diagnostics

refactor: disable 8-bit quantization and set device map to CPU when GPU is unavailable

refactor: enable 8-bit quantization for improved memory efficiency in Transformers model loading

refactor: disable 8-bit quantization and CPU offload for optimized model loading on T4 GPUs

fix: improve error handling during model loading and fallback for quantization failures

refactor: remove unused model configurations and streamline model creation logic

refactor: remove runtime installation of llama-cpp-python, now pre-installed via requirements.txt

feat: Add main backup and simplified service implementations with API endpoints

fix: streamline CPU offload handling in model loading for better memory management

feat: add CPU offload support for Transformers model to optimize memory usage

feat: add Transformers model support with GPU optimization and 8-bit quantization

feat: add model size and polish support to model info

fix: use prebuilt CUDA wheel for llama-cpp-python

refactor: defer llama-cpp-python install to runtime

feat: add GPU-enabled Dockerfile.gpu for HF Spaces CUDA support

fix: graceful fallback for llama-cpp-python installation on HF Spaces

fix: enable CUDA compilation for llama-cpp-python

perf: defer llama-cpp-python build to runtime startup

feat: enable GPU acceleration for Bielik GGUF models

update Dockerfile and README.md to replace Qwen2.5-3B and Gemma-2-2B with Bielik-1.5B-GGUF; adjust model loading instructions in the API documentation

update HuggingFaceInferenceAPI comment for clarity; change huggingface_hub version to minimum required

refine GBNF grammar for car advertisement; ensure compact JSON output and improve gap-item structure

add model management methods to ModelRegistry; include model listing, loading, and unloading functionalities

add HuggingFace Inference API model; implement async initialization and text generation with caching

add GBNF grammar for car advertisement gap filling; update LlamaCppModel to support loading grammar from file

add GBNF grammar utilities for structured LLM output; integrate grammar in model generation

enhance infill processing to handle custom messages; return cleaned output directly when provided

update llama-cpp-python installation to version 0.3.16 for improved compatibility

install llama-cpp-python at runtime to avoid build issues in HuggingFace Spaces; update requirements.txt to reflect this change

update LlamaCppModel initialization parameters and enable verbose logging for model loading; update llama-cpp-python requirement

enhance error handling in LlamaCppModel initialization; include full traceback on failure

add get_info method to return model details for /models endpoint

add debug logging for batch infill and model generation processes; update bielik model configuration

increase context size and improve message handling in LlamaCppModel

adding-bielik-gguf

adding custom prompt messages

fix: Add optional attributes parameter to create_infill_prompt and update InfillItem schema

fix: Refine infill prompt instructions for clarity and improved user guidance

fix: Simplify infill prompt creation and enhance JSON parsing logic

fix: Improve infill prompt instructions for clarity and structured JSON output

fix: Enhance infill prompt instructions for clarity and structured JSON output

fix: Update infill prompt instructions for clarity and structured JSON output

update prompts again

fix: Refine infill prompt instructions for clarity and strict JSON output

feat: Add flexible prompt with 2 and 3 gap examples plus explicit instruction to only fill specified gaps

fix: Adjust prompt example to use 2 gaps instead of 3 to match common use case

fix: Improve infill prompt with real example to prevent model from copying generic placeholder