Models & Labs
Finetuning Multi-Vector Models with Sentence Transformers
Hugging Face Bloghigh confidence
Why it matters
- →Multi-vector models preserve fine-grained token-level signals, enhancing retrieval accuracy.
- →The approach allows for domain-specific finetuning, improving performance in specialized fields.
- →It provides a practical guide for building custom retrieval models on consumer hardware.
Hugging Face has released a detailed guide on training and finetuning multi-vector embedding models using Sentence Transformers. These models, which utilize a ColBERT-style late interaction retrieval method, are designed to enhance retrieval performance by maintaining token-level detail. The blog post outlines the necessary components for training, such as model architecture, datasets, and loss functions, enabling users to customize models for specific domains. This method is particularly beneficial for specialized fields, offering improved retrieval accuracy over general-purpose models.
Read originalMore in Models & Labs
Models & Labsmodels
vLLM v0.28.0 Release: Major Performance Enhancements
The v0.28.0 release of vLLM introduces substantial improvements in performance and functionality, particularly for the Kimi-K3 model. With the addition of Decode Context Parallel support and fused FlashKDA decode kernels, the update significantly enhances processing speed and efficiency. DeepSeek V4 now includes sparse MLA support and advances in speculative decoding, offering better execution on both NVIDIA and AMD hardware. These updates make vLLM more robust and adaptable, providing developers with enhanced tools for deploying and executing models on a broader range of hardware configurations.
vLLM Releases
Models & Labsmodels
llama.cpp b10657 Release Expands Platform Support
The b10657 release of llama.cpp brings new OpenCL binary kernels, enhancing performance and compatibility across a wide range of systems. This update includes specific improvements for Apple Silicon, with KleidiAI support, and Vulkan on Ubuntu, making it more accessible for developers using these platforms. While no new model architectures are introduced, the release focuses on strengthening llama.cpp's capabilities as an inference runtime, particularly for those not using NVIDIA hardware. With ROCm 7.14 support on Ubuntu and CUDA 12 and 13 DLLs for Windows, llama.cpp continues to evolve as a versatile tool for AI model deployment. This release underscores the commitment to broadening hardware compatibility and optimizing performance across different environments.