github llama.cpp · 2d ago · inference

llama.cpp b10947 fixes Nemotron-H MoE zero-divisor crash

from English

The llama.cpp project released version b10947, which addresses a critical load-time crash in the Nemotron-H model. The update guards against a zero divisor error that occurred when the NextN/MTP tail loop derived expert FFN size from per-layer arrays holding zero values for non-MoE layers.

Fixes SIGFPE crash by reporting malformed metadata instead of dividing by zero when expert_feed_forward_length is missing.
Provides binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, OpenCL, Vulkan, OpenVINO, SYCL, ROCm), and openEuler.

This change prevents the application from dying with a signal error during checkpoint loading, ensuring stable operation for models with mixed MoE and dense layers.

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

Read original

github llama.cpp · 1h ago · 1 view Live

llama.cpp b10984 adds CUDA row-contiguous SUM_ROWS support

The llama.cpp project released version b10984, introducing CUDA support for row-contiguous SUM_ROWS operations. This update also organizes the code to include GGML_OP_MEAN for handling row-contiguous tensors via a shared kernel.

github llama.cpp · 2h ago · 1 view

llama.cpp b10985 restricts RPC hash-cache to weights

The llama.cpp project released build b10985, which modifies the RPC backend to limit hash-based caching exclusively to weight tensors. This change addresses an issue where activation data was unnecessarily hashed and stored, leading to excessive disk usage.

github llama.cpp · 3h ago · 4 views

llama.cpp b10982 adds Vulkan sparse Flash Attention support for DSV4/GLM

The llama.cpp project has released build b10982, which introduces support for sparse Flash Attention via the Vulkan backend. This update specifically targets DSV4 and GLM models, allowing them to leverage sparse attention mechanisms on compatible GPU hardware.

github llama.cpp · 4h ago · 2 views

llama.cpp b10981 fixes stateful decode for Gemma-4 and optimizes GPU MoE inference

The llama.cpp b10981 release addresses critical correctness issues in the OpenVINO backend's stateful decoding path, specifically fixing support for models like Gemma-4 that use per-layer head sizes. It also introduces optimizations for GPU Mixture of Experts (MoE) inference and improves performance through KV state relayouts.

github llama.cpp · 5h ago · 1 view

llama.cpp adds Metal FA kernels for MiniCPM3's HSK=96, HSV=64

The llama.cpp project has added Flash Attention (FA) kernel support for the specific configuration used by MiniCPM3 on Apple Silicon. This change addresses a crash where the `-fa auto` flag would abort due to a missing kernel instantiation for attention key length 96 and value length 64.