[Bug] Cannot apply lora immediately anymore (vulkan) #1010
Description
Git commit
commit 2034588
Operating System & Version
Windows 10
GGML backends
Vulkan
Command-line arguments used
unchanged
Steps to reproduce
As mentioned in #969 (comment)
Non-runtime applications of LoRA in Vulkan now fail with an assert in ggml_vk_mul_mat_q_f16
This even applies when wtype is set to q4_0 which worked before
ggml/src/ggml-vulkan/ggml-vulkan.cpp:6552: GGML_ASSERT(y_non_contig || !qy_needs_dequant) failed
both my lora and model are .safetensors, runtime converted to q4_0
model is a SDXL model.
previously, this combination worked fine. running in immediate mode works, though.
What you expected to happen
LoRA applied successfully
What actually happened
GGML Assert occurred. Reproducible by @wbruna
Logs / error messages / stack trace
No response
Additional context / environment details
No response