b10090

ggml-org/llama.cppb10090Jul 22, 2026by github-actions[bot]same day as b10089

What changed since b10089

low risk

b10090 adds a WebGPU depthwise conv2d (CONV_2D_DW) kernel ported from the Vulkan backend, with no other changes noted.

New features

  • WebGPU backend now implements GGML_OP_CONV_2D_DW, enabling depthwise 2D convolution to run on GPU instead of falling back to CPU.

Improvements

  • WebGPU supported-ops tables were updated to reflect the newly available CONV_2D_DW operation.
  • The kernel is a direct port of Vulkan's conv2d_dw.comp, keeping behavior consistent across GPU backends.

Upgrade notes

This is a backend capability addition only; no config, model, or API changes are required. Users running depthwise-conv models (e.g. vision/multimodal encoders) on WebGPU should see those ops stay on-device rather than falling back to CPU. CUDA GET_ROWS quantization work from b10089 is unchanged.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Adds a depthwise convolution 2D (CONV_2D_DW) kernel to the WebGPU backend, ported from the Vulkan backend implementation.

Key Highlights

  • Added CONV_2D_DW (depthwise conv2d) kernel for WebGPU
  • Ported implementation from Vulkan backend's conv2d_dw.comp
  • Updated supported ops tables for WebGPU

Release Notes (excerpt)

<details open> webgpu : add CONV_2D_DW (depthwise conv2d) kernel (#25847) * webgpu : add CONV_2D_DW (depthwise conv2d) kernel Implement GGML_OP_CONV_2D_DW for the WebGPU backend, ported from the Vulkan backend's conv2d_dw.comp. Assisted-by: Claude Opus-4.8 * Remove unnecessary comments in webgpu support * update supported ops tables, triggered by adding webgpu CONV_2D_DW </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories