The ggml 1.0.0 release page shows a signed commit and a long list of backend enhancements.
*ggml 1.0.0 prunes dead API, adds configurable YaRN truncation and FP8 support, and rewrites core scheduling. The update targets CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon, promising faster inference across heterogeneous hardware.*
On 13 Nov at 08:34, the ggml‑org team pushed version 1.0.0 to GitHub, signing the commit with GPG key B5690EEEBB952194. The release strips the deprecated public API, eliminates legacy DIAG_MASK ops, and introduces configurable YaRN truncation alongside FP8 support. It also delivers a sweeping performance overhaul across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends. These changes target quantized matmuls, MoE fusion and FlashAttention pathways, promising faster inference for large language models on a wide hardware spectrum. Developers must adopt the new RPC protocol v9 and adjust to the removed API elements, or risk incompatibility. The timing aligns with industry moves toward lower‑precision formats and hardware‑specific kernels, positioning ggml as a pivotal open‑source layer for next‑gen AI workloads.
The release removes the legacy DIAG_MASK_INF and DIAG_MASK_ZERO ops and prunes the deprecated public API. It bumps the RPC protocol to version 9, rejecting incompatible peers and handling unreachable endpoints gracefully. New functions ggml_rope_set_truncate and ggml_rope_get_truncate let developers control YaRN correction-bound rounding. The backward pass for TRI is implemented, and precision tests for MUL_MAT src1 are added with gating via ggml_prec_set_acc. These changes tighten the interface and force downstream projects to update their integration points.
Across CUDA and Metal, ggml adds few‑row quantized matmuls for batch sizes 2‑24 and BF16 few‑row matmuls, plus MMQ fused FP8 support and NVFP4 gate/up matmuls with GLU. FlashAttention receives an sm70 head‑size‑256 config and MMA head sizes 40/72. Vulkan gains an Intel FlashAttention prefill kernel and batched tensor copies. SYCL introduces a device‑side MoE scheduler. OpenCL adds decomposed Adreno FlashAttention and new GEMV kernels for q6_k, q8_0 and multi‑column formats. WebGPU sees batched tensor API support and a maxBufferSize query. Hexagon expands op coverage with Q5_0/Q5_1 support and optimized conv_2d/conv_3d. Collectively, these backends cut latency and improve throughput for large language model inference.
The expanded kernel set lets developers run the same ggml model on GPUs, CPUs, and specialized accelerators without rewriting code. Metal’s fused MoE routing stays inside a single command buffer, while Vulkan’s large‑tile split‑K and RDNA3/4 support enable high‑throughput AMD pipelines. SYCL’s scheduler offloads cache misses to the CPU and copies results asynchronously, smoothing memory pressure. OpenCL’s Adreno FlashAttention and Hexagon’s conv_3d broaden mobile and embedded deployment. By skipping unused inputs during scheduling, the runtime reduces memory traffic on all platforms, making ggml a more efficient bridge between model graphs and heterogeneous hardware.
The commit creating v1.0.0 is signed with GitHub’s verified signature (GPG key ID B5690EEEBB952194), confirming author authenticity. The RPC protocol upgrade to v9 enforces faster loading and graceful failure on failed handshakes, reducing attack surface from malformed peers. Deprecated API removal forces developers to migrate, preventing legacy code from lingering in production. Scheduler improvements that avoid uploading unused inputs also limit exposure of unnecessary data. Overall, the release tightens both functional and security posture while demanding immediate compliance from existing ggml users.
The release adds configurable YaRN truncation and FP8 support, plus exact rounding for Q4_1, Q5_1 and Q8_0 formats.
CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon all received new kernels, including few‑row quantized matmuls, MoE fusion, FlashAttention extensions and Adreno FlashAttention.
The RPC protocol is bumped to version 9, rejecting incompatible peers and handling unreachable endpoints or failed handshakes gracefully.
Version 1.0.0 marks a decisive step for ggml, turning it from a flexible tensor library into a hardened, cross‑platform inference engine. By tightening the API, expanding backend kernels and enforcing signed releases, the project signals readiness for production‑grade AI deployments. The onus now lies on downstream projects to integrate the new RPC protocol and leverage the performance gains, or risk falling behind as the ecosystem coalesces around lower‑precision, hardware‑aware computation.
Sources: Hacker News, GitHub release v1.0.0 (https://github.com/ggml-org/ggml/releases/tag/v1.0.0)