github llama.cpp · 1h ago · inference

llama.cpp b10427 fuses SYCL FFN ops for up to 12.4% throughput gain

from English

The llama.cpp project released version b10427, which includes a performance optimization for Intel Arc GPUs via SYCL. The update implements a fusion of the gate, up, and GLU operations within the q4_K dense feed-forward network (FFN) layer.

Measured on an Arc Pro B70 using llama-bench, throughput increased by 2.8% for qwen2.5-3B-Instruct and 2.0% for gemma-2-2b-it.
Batched inference benchmarks on qwen2.5-3B showed significant gains: +3.4% at batch size 1, rising to +10.1% at B=2, +10.7% at B=4, and +12.4% at B=8.

This release provides precompiled binaries for macOS, Linux, Windows, Android, and openEuler across CPU, CUDA, Vulkan, ROCm, and OpenVINO backends.

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

Read original

github llama.cpp · 13h ago · 1 view

llama.cpp b10413 adds auto-detection of draft spec type from GGUF metadata

The llama.cpp project released version b10413, which introduces automatic detection of the speculative decoding specification type directly from draft GGUF model metadata.

github llama.cpp · 18h ago · 1 view

llama.cpp b10410 removes separate fp32 type promotion in gemm non-oneDNN path

The llama.cpp project has released version b10410, which includes a code change to the SYCL backend. This update removes the separate fp32 type promotion in the GEMM non-oneDNN path and instead uses automatic fp16 promotion.

github llama.cpp · 19h ago · 4 views

llama.cpp b10408 adds Q3_K ESIMD kernel and moves control to runtime

The llama.cpp project released version b10408, which introduces a DMMV Q3_K ESIMD kernel for Intel GPUs. This release also includes new Q4_K and Q6_K ESIMD kernels and refactors the codebase to allow ESIMD control at runtime rather than compile time.

github llama.cpp · 3d ago · 6 views

llama.cpp b10353 fixes silent wrong results in CUDA and Metal ROLL kernels

The llama.cpp project released build b10353, which addresses a critical bug in the ggml_roll operation on CUDA and Metal backends. Previously, permuted (non-contiguous) source tensors produced silently incorrect results because these backends ignored stride information.

github llama.cpp · 7d ago · 8 views

llama.cpp b10321 fixes NORM/RMS_NORM for partial simdgroup row lengths

The llama.cpp project released version b10321, which includes a critical fix for GGML_OP_NORM and GGML_OP_RMS_NORM operations on Metal. The update resolves an issue where threadgroups sized to leave a partial SIMD group resulted in incorrect mean and variance calculations due to dropped partial sums.