Commit 8bf88ba93 for llama.cpp
commit 8bf88ba9368e70fe3ba1b0001b34e70ffc713716
Author: Hongqiang Wang <wangh@qti.qualcomm.com>
Date: Thu Nov 19 20:58:46 2026 -0800
vulkan: disable subgroup arithmetic on Qualcomm's driver (#30826)
On Adreno, subgroupAdd() after divergent code returns wrong sums. The
fused ADD+RMS_NORM partial-sum path hits this, so the decode-step norm
is garbage and models emit stray tokens after long prompts.
Clear subgroup_arithmetic for the Qualcomm proprietary driver, which
also turns off add_rms_fusion and the other subgroupAdd paths.
test-backend-ops on Adreno 850 and X2-90: ADD_RMS_NORM 20/25 -> 25/25,
RMS_NORM_MUL_ADD 33/45 -> 45/45. llama 3B Q4_0 on Adreno 850: pp512
and tg128 unchanged.
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ecd46d828..cd8ef9ba3 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -4735,6 +4735,10 @@ vk_device ggml_vk_get_device(size_t idx) {
device->subgroup_arithmetic = false;
}
#endif
+ // Workaround for subgroupAdd() returning wrong sums on Adreno after divergent code
+ if (device->driver_id == vk::DriverId::eQualcommProprietary) {
+ device->subgroup_arithmetic = false;
+ }
device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
#ifdef __APPLE__