Qeravio
Canonical AI event

MKL-FA softmax loads coalesced in SYCL update

The SYCL optimization for MKL-FA softmax improves load efficiency. The update is available for multiple platforms, including Ubuntu with CUDA 12 and Windows with Vulkan.

21 Sept 20261 verified claims1 sources1 observations
What happened

The official source reports this update: b11073: sycl : coalesce MKL-FA softmax loads instead of one work-item per row.

Why it matters

This official update documents a development concerning b11073: sycl : coalesce MKL-FA softmax loads instead of one work-item per row. Its practical significance depends on the scope and evidence stated by the source.

What to watch next

Read the official source update and verify its stated scope, evidence and timing before acting on it.

Connected knowledge

Entities affected by this event

Llama Cppsubject
Continue this topic
infrastructureggml-metal drops _all suffix from fusion op pattern vectors21 Sept 2026infrastructureHexagon DMA handling updated to support 64-bit mappings.21 Sept 2026infrastructureApple's Metal framework updated to resolve macOS 27 SDK deprecation warnings.21 Sept 2026infrastructureNVIDIA Volta GPU MMVQ to MMQ crossover tuning applied.21 Sept 2026infrastructureWebGPU adds fused GDN + cpy to macOS, Linux, Android, Windows, and openEuler.21 Sept 2026infrastructureGemma 4 FA tuning added for Ampere GPUs and newer.21 Sept 2026
Evidence trail

Sources behind the event

llama.cpp official releasesb11073: sycl : coalesce MKL-FA softmax loads instead of one work-item per row
Editorial presentation

Read the public story

Open story