Replies: 1 comment 2 replies
-
Layers are not duplicated, but the available memory to an iGPU is up to the driver. You can look at what your driver is doing by using |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Vulkan is a great blessing for those of us running on consumer hardware with iGPUs. I have an Intel Iris igpu that actually turns out to be pretty decent with Vulkan acceleration. Many thanks to @0cc4m and others who have made this happen. You are amazing!
I want to understand something that I'm seeing on my laptop. The setup has 16GB of main memory, but it looks like Vulkan can't access all of it. For example, there are models that fit into the combined memory when I lower the -ngl value, but fail to run at higher -ngl values due to failed Vulkan memory allocations (ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory).
This is surprising because regardless of whether the model layer is on Vulkan or on the CPU, it sits on the same system memory. This leads me to think that somehow the part of system memory that Vulkan can access is lower than the total amount of memory.
Does this make sense? Or is there something else going on? Are layers replicated twice (once for Vulkan and once again for CPU) even for UMA?
Beta Was this translation helpful? Give feedback.
All reactions