]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
vulkan: optimize UMA buffer operations and fix driver hangs (llama/16059)
authorGiuseppe Scrivano <redacted>
Sun, 21 Sep 2025 06:31:55 +0000 (08:31 +0200)
committerGeorgi Gerganov <redacted>
Thu, 25 Sep 2025 08:56:34 +0000 (11:56 +0300)
commita4e2c167cf783aaa63c266f5511b528ac246d610
tree504eec597c983ed671e0b74df39a3f61afa20bc8
parent06e5cfd9e95c840363aaa71c4ca281048f4fffa3
vulkan: optimize UMA buffer operations and fix driver hangs (llama/16059)

* vulkan: optimize UMA buffer operations and fix driver hangs

The previous implementation was blocking the GPU for extended periods,
causing the i915 driver to reset the context due to the hangcheck
protection.

[32628.443070] i915 0000:00:02.0: [drm] GPU HANG: ecode 12:1:85dffffb, in llama-server [194114]
[32628.443091] i915 0000:00:02.0: [drm] llama-server[194114] context reset due to GPU hang

* vulkan: implement deferred_memset on UMA

---------

Signed-off-by: Giuseppe Scrivano <redacted>
src/ggml-vulkan/ggml-vulkan.cpp