- Updated .gitignore to include new model directories.
- Removed version specification from existing docker-compose.yml for llama-cpp-server.
- Created new docker-compose.yml for ai_stack/llamacpp.server with GPU support and environment configurations.
- Added Gemma4_(E2B)_Vision.ipynb notebook for data preparation, training, and inference of the Gemma 4 model.
- Implemented WhisperModel from faster_whisper for transcribing audio files.
- Configured model to run on GPU with FP16 for improved performance.
- Added functionality to detect language and print transcription segments with timestamps.
- Added SQLite database to store chat messages.
- Created HTML interface for the chat application using Bootstrap.
- Developed backend logic in FastAPI to handle chat requests and responses.
- Implemented TypeScript for dynamic message rendering and interaction.
- Integrated logging with Logfire for monitoring.
- Configured Ollama model for OpenAI-compatible API usage.
- Established a question graph for evaluating user responses.
- Created a Jupyter notebook to address vector dimension mismatch errors in the RAG database.
- Implemented a solution to reset the database schema to match the current embedding model dimensions.
- Updated rag.py to use the Ollama model with OpenAI-compatible endpoints.
- Added functionality for building the search database and inserting document sections with correct embeddings.
- Included Pydantic models for validation of embedding dimensions.
- Enhanced error handling for asyncpg DataError exceptions during database operations.
- Reset execution counts for cells in `ollama.ipynb` to reflect the correct order.
- Adjusted output tokens in the usage statistics for accuracy.
- Added a new cell to query the top five highest-grossing animated films of 2025 using DuckDuckGo search.
- Updated `pyproject.toml` to include `duckduckgo` as a dependency for the new search functionality.
- Enhanced `uv.lock` to reflect the addition of `duckduckgo-search` and its dependencies.
- Updated the Modelfile path in `pull-deepseek-model.sh` to use an absolute path for model files.
- Simplified the model loading process by removing unnecessary directory changes and directly referencing the mounted Modelfile.
- Added a new Jupyter notebook `ollama.ipynb` that demonstrates the usage of the Ollama API with a sample query and error handling for unsupported tools.
- Created README.md for Ollama Docker setup with instructions for running and using models.
- Added docker-compose.yml for Ollama service configuration.
- Introduced scripts for model setup, including pulling models from Hugging Face and loading them into Ollama.
- Implemented AI launcher script to choose between Ollama and LlamaCPP backends.
- Added GPU support scripts for LlamaCPP with detailed logging and error handling.
- Included simple model pull script for easier model management.
- Established a structured approach for model management and integration with Ollama API.
This script allows users to easily enable or disable NVIDIA persistence mode and adjust the performance state of the GPU. It includes options to set the clock speeds and reload the NVIDIA Unified Memory kernel module. Usage instructions are provided for clarity.
- Introduced a new Jupyter notebook `unsloth_nb.ipynb` for utilizing the Unsloth library.
- Implemented code to load and patch FastLanguageModel and FastQwen2Model for improved training speed.
- Configured models to support 4-bit quantization for reduced memory usage.
- Included detailed logging for model loading and system specifications.
furyhawk