ggml_vulkan_stage_handoff: Hand a pipeline stage's activation tensor to the next stage's...

View source: R/tensor_parallel.R

ggml_vulkan_stage_handoffR Documentation

Hand a pipeline stage's activation tensor to the next stage's input

Description

Copies the contents of src (a Vulkan-backed activation tensor produced by one pipeline stage, living on its device) into dst (the next stage's input tensor, allocated on another device) via host staging. This is the single cross-device transfer per forward pass that pipeline parallelism needs.

Usage

ggml_vulkan_stage_handoff(src, dst)

Arguments

src

Source tensor (stage N output), Vulkan-backed.

dst

Destination tensor (stage N+1 input), Vulkan-backed, same size.

Value

Invisibly 0 on success; errors on a shape/buffer mismatch.


ggmlR documentation built on July 14, 2026, 1:08 a.m.