Hardware
- GPUs: 2x Intel Arc B70 Pro (Battlemage G31), 32 GB each (64 GB total)
- CPU: Intel 12th Gen (Alder Lake)
- OS: Ubuntu 26.04 LTS, kernel 7.0.0-27
Issue
Using the Level Zero API (via llama.cpp SYCL backend), the following errors occur:
-
UR_RESULT_ERROR_DEVICE_LOST (error 20): When attempting to allocate and initialize buffers on dual GPUs, the device connection is lost. This happens during memset operations after successful device enumeration. Even a single GPU can trigger this when larger allocations are made.
-
UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY (error 39): Occurs when loading models that should fit within available VRAM (64 GB total across 2 GPUs). The system reports the model at ~33 GB, but fails on allocation.
Stack Trace
The crash originates in ggml_backend_sycl_buffer_clear during memset via a SYCL stream.
Workaround
The Vulkan backend (Mesa 26.0.3) works correctly with dual GPUs on the same hardware. The issue appears specific to the Level Zero / SYCL path.
Willing to help
Happy to provide additional logs, run diagnostic tools, or test patches.
Hardware
Issue
Using the Level Zero API (via llama.cpp SYCL backend), the following errors occur:
UR_RESULT_ERROR_DEVICE_LOST (error 20): When attempting to allocate and initialize buffers on dual GPUs, the device connection is lost. This happens during memset operations after successful device enumeration. Even a single GPU can trigger this when larger allocations are made.
UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY (error 39): Occurs when loading models that should fit within available VRAM (64 GB total across 2 GPUs). The system reports the model at ~33 GB, but fails on allocation.
Stack Trace
The crash originates in ggml_backend_sycl_buffer_clear during memset via a SYCL stream.
Workaround
The Vulkan backend (Mesa 26.0.3) works correctly with dual GPUs on the same hardware. The issue appears specific to the Level Zero / SYCL path.
Willing to help
Happy to provide additional logs, run diagnostic tools, or test patches.