llama.cpp b10090 adds depthwise conv2d kernel for faster WebGPU AI inference
New WebGPU depthwise conv2d kernel speeds up LLM inference in-browser…
Deep Dive
ggml-org's llama.cpp release b10090 adds a CONV_2D_DW (depthwise convolution) kernel for the WebGPU backend, ported from the Vulkan backend, with assistance from Claude Opus-4.8. The release also updates supported ops tables. Pre-built binaries are available for macOS, Linux, Windows, Android, and openEuler across various architectures and backends.
Key Points
- Adds CONV_2D_DW (depthwise conv2d) kernel to WebGPU backend, ported from Vulkan with help from Claude Opus-4.8
- Enables more efficient neural network inference directly in browsers via WebGPU, reducing CPU fallbacks
- Ships with ready-to-use binaries for macOS, Windows, Linux, Android, iOS, and openEuler across CPU, Vulkan, CUDA, ROCm, OpenVINO, SYCL, and more
Why It Matters
Boosts browser-based LLM performance, making AI more accessible without local installs or expensive hardware.