github llama.cpp · 6h ago · inference

llama.cpp b11284 fixes OpenVINO GET_ROWS on quantized weight views

from English

The llama.cpp project released build b11284, which includes a fix for the OpenVINO backend to correctly handle GET_ROWS operations on quantized weight views.

The update resolves view_src when collecting weight Constants, preventing views over quantized weights from becoming dynamic typed Parameters.
The row offset of the view is now folded into the gather indices instead of slicing the dequantization subgraph.
This change lifts the rejection of quantized src0 views with nonzero offsets, allowing them to work with OpenVINO.

The release provides binaries for macOS, Linux, Windows, Android, and openEuler across CPU, GPU (CUDA, Vulkan, ROCm, SYCL), and NPU backends.

Importance 1/3 Trust 1/3 llama.cpp Inference efficiency Local inference

Read original

github llama.cpp · 9h ago · 2 views

llama.cpp adds GLM-5.3-Flash support with k-pool and MTP fixes

The llama.cpp project has added initial support for the GLM-5.3-Flash (GLM5-Next) model architecture, including migration to hybrid-index memory and early Multi-Token Prediction (MTP) capabilities.

github llama.cpp · 12h ago · 2 views

llama.cpp b11272 optimizes Hexagon concat op

The llama.cpp project released build b11272, which includes optimizations for the Hexagon backend. The primary change focuses on optimizing the concatenation operation to improve performance on Qualcomm Snapdragon hardware.

github llama.cpp · 1d ago · 7 views

llama.cpp b11260 adds FP32 GELU_ERF and GEGLU_ERF support

The llama.cpp project has released version b11260, introducing support for FP32 GELU_ERF and GEGLU_ERF activation functions in the Hexagon backend. This update also includes optimizations to reduce register pressure within the associated kernels.

github llama.cpp · 1d ago · 1 view

llama.cpp b11258 removes built-in UI service worker when UI is not served

The llama.cpp project released build b11258, which includes a fix for the built-in web interface's service worker behavior. Previously, using flags like `--path` or `--no-ui` caused browsers to retain cached versions of the UI because the server returned a 404 for `/sw.js`, which does not trigger service worker removal.

media Hugging Face Forums · 2d ago · 4 views

Krea 2 LoRA Studio enables training on free Colab T4

Jordan Epstein has released Krea 2 LoRA Studio, a project designed to make the full Krea 2 LoRA workflow usable on a free Google Colab T4 GPU. The tool addresses memory constraints by implementing a staged workflow that manages cache preparation, training, and preview generation separately.