AI Briefingggml v1.0.0 Released with FP8 Support and Backend Optimizations
·2026.11.14 11:28
Key point
The release removes deprecated APIs and adds FP8 support, YaRN truncation controls, and performance improvements across CUDA, Metal, Vulkan, and other backends.
Details
- ggml v1.0.0 marks a milestone release that removes unofficial public APIs and legacy DIAG_MASK operations while introducing FP8 support and YaRN truncation settings.
- Performance optimizations are applied across all major backends, including CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU, and CPU.
- The RPC protocol is updated to version 9, improving loading speeds and handling handshake failures gracefully.
- CUDA backend gains support for OCP FP8, NVFP4 gate/up matmuls, and FlashAttention extensions for specific head sizes.
- Metal backend implements MoE routing fusion and fixes out-of-bounds access errors in GROUP_NORM and FlashAttention kernels.
- Vulkan backend adds Intel FlashAttention prefill kernels and enables large-tile split-K optimizations for RDNA3/4.
- CPU backend transitions MUL_MAT to F32 src1 and adds specific kernels for SpacemiT and ARM NEON architectures.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Unsloth Releases Dynamic V3.1 Quantizations for Qwen3.8-27B and New 1-bit Data Types for Qwen3.8-2.4TNVIDIA Tutorial: Serving GLM-5.3 NVFP4 on Dual DGX Stations with vLLM