KoboldCpp on AMD Radeon: Vulkan, ROCm and which to use
KoboldCpp AMD setup in one line: download koboldcpp-nocuda.exe (Windows) or koboldcpp-linux-x64-nocuda (Linux) and choose Use Vulkan: this is the official path and works on Windows and Linux without extra software. ROCm has no official Windows build: Windows users can try a third-party fork, which lags the current release, and Linux users the official experimental rocm-rolling build. We have no AMD card in our lab yet, so we have no measurements of Vulkan against ROCm; compare both on your own card before switching.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
KoboldCpp AMD: the short answer by operating system
AMD paths by operating system
Windows
Vulkan
koboldcpp-nocuda.exe (85.7 MB)
Official. Recommended.
Windows
ROCm (hipBLAS)
koboldcpp_rocm.exe from the third-party fork
Not from the KoboldCpp project; lags behind
Linux
Vulkan
koboldcpp-linux-x64-nocuda (102.4 MB)
Official. Recommended.
Linux
ROCm (hipBLAS)
koboldcpp-linux-x64-rocm (618 MB, rocm-rolling)
Official, experimental
We have not built a per-card table (which Radeon works with what), because that needs the cards in hand and we have none. Card-specific facts below come from user reports and are labelled.
Vulkan: the official AMD path
Download the nocuda build
Take koboldcpp-nocuda.exe (Windows, 85.7 MB) or koboldcpp-linux-x64-nocuda (Linux, 102.4 MB) from the GitHub releases. Do not take the CUDA build: on a PC without NVIDIA it works but starts slower F02.
Check that a Vulkan driver is present
The normal Radeon driver from AMD includes Vulkan. Nothing else to install.
Terminal
$ vulkaninfo --summary
Your Radeon should be listed. If the only device is a software renderer such as llvmpipe, install the Mesa Vulkan driver for your distribution first.
Choose Use Vulkan
In the launcher, set Backend to Use Vulkan. On the command line use Selects the Vulkan backend, which works on NVIDIA, AMD and Intel GPUs. Optionally list device indices; leave empty to auto-detect.DefaultnoneLauncher Quick Launch tab: Backend (dropdown: Use CPU / Use CUDA / Use hipBLAS (ROCm) / Use Vulkan / Use CPU (Old CPU) / Use Vulkan (Old CPU) / Use Vulkan (Older CPU) / Failsafe Mode (Older CPU); only options whose library ships in the binary are listed)Sincev1.56All details.
Leave GPU layers on auto and launch
Keep How many model layers to offload to the GPU. -1 (default) turns on autofit; 0 keeps everything on the CPU.Default-1Launcher Quick Launch tab: GPU Layers (layer-count hint shown next to it)Sincev1.23betaAll details at -1. See GPU layers.
Check it worked
The console should show a device and Vulkan0 buffer lines, as in this log from a Vulkan device we could test (a software one, because we have no Radeon):
ggml_vulkan: Found 1 Vulkan devices:1load_tensors: Vulkan0 model buffer size = 235.76 MiB2llama_kv_cache: Vulkan0 KV buffer size = 51.00 MiB3
ggml_vulkan: Found 1 Vulkan devices:: A device was found. Your Radeon's name appears here.
load_tensors: Vulkan0 model buffer size = 235.76 MiB: Weights on the device.
llama_kv_cache: Vulkan0 KV buffer size = 51.00 MiB: Context memory on the device.
Without a usable device, the log says ggml_vulkan: No devices found. and KoboldCpp runs on the CPU, with the line offloaded 25/25 layers to GPU still printed F35F17. Trust the buffer lines: GPU layers, step 3.
We have not measured speeds on a Radeon. When we can, this section gets a speed table for an 8B, a 12B and a 24B model.
ROCm on Windows: a third-party fork
When we last looked, the newest fork release was v1.104.yr0-ROCm, dated about 30 December 2025 (the year is inferred; we could not re-check the page on the test date). KoboldCpp itself has had 20 releases since, up to 1.124.2. That gap matters: for example autofit started choosing layers by itself in 1.108, and the fork predates it F65. The fork’s assets are koboldcpp_rocm.exe, koboldcpp_rocm_b2.exe and _files.zip variants; we do not know which one suits which card and say so rather than guess. Keep the official Vulkan build next to the fork so you can compare speed on your own card with Runs a benchmark instead of starting the server; optionally appends results to a CSV file.DefaultnoneLauncher Hardware tab: Run Benchmark (button)Sincev1.57All details.
ROCm on Linux: the official rocm-rolling build
The project publishes koboldcpp-linux-x64-rocm (about 618 MB) under the rocm-rolling pre-release on its GitHub page, with per-version copies named koboldcpp-linux-x64-rocm-<tag>. It is experimental. It is built in an AMD ROCm container and moved to ROCm 7.1 in 1.101.1. ROCm support for WMMA (matrix instructions) was removed in 1.118. The same --usecuda flag selects the ROCm path in these builds (the old --usehipblas is an alias). Arch Linux has a third-party package, koboldcpp-hipblas, in the AUR. We ran it on a Linux machine with no AMD GPU F74. With no flag it started on its CPU path (Unable to detect VRAM, Auto Selected Default Backend, offloaded 0/25 layers) and served normally, with nothing extra installed. With --usecuda it loaded the ROCm library, printed ggml_cuda_init: failed to initialize ROCm: no ROCm-capable device is detected and rocBLAS error: Could not initialize Tensile host: No devices found, and then aborted with exit code 134. So if this build closes straight after those two lines, ROCm cannot see a GPU: check your driver, or use Vulkan. We could not test it on a real Radeon, so we do not know which packages or groups it needs there, or whether it beats Vulkan.
Unsupported gfx targets and HSA_OVERRIDE_GFX_VERSION
Every AMD GPU chip has a name such as gfx1031. ROCm only supports some of them officially. Users report the “not supported” message for the RX 6700 XT and 6800M (gfx1031, issue #441) and for gfx1032 cards (issue #655). Vulkan does not have this problem, so the simplest fix is to use it.
Integrated GPUs and APUs
On some integrated graphics, such as the 780M in recent Ryzen chips, autofit cannot read the memory and prints Unable to detect VRAM (issue #1748, and 25 issues for the message in general). Set How many model layers to offload to the GPU. -1 (default) turns on autofit; 0 keeps everything on the CPU.Default-1Launcher Quick Launch tab: GPU Layers (layer-count hint shown next to it)Sincev1.23betaAll details by hand: an integrated GPU shares system RAM, so how much you can offload depends on how much RAM the BIOS sets aside for it. Use the calculator with that amount as the memory, and see GPU layers, laptops and integrated GPUs.
Shared memory spill on Windows
On Windows a model that is slightly too big for VRAM can spill into “Shared GPU memory” and become very slow instead of failing. Users report it on an RX 9070 (issue #1413). Watch Task Manager’s shared memory while it generates. The fix is fewer layers or less context: the shared memory section. We have not measured it on AMD.
Known AMD problems
Problems users report on AMD, with the fix
Garbled output on Vulkan in 1.101.1
Issues #1826, #1108 (users report)
Try the next release, or --noflashattention. Keep the previous file and compare with the benchmark.
Restart KoboldCpp; check Task Manager for other GPU programs.
Runs on the CPU although you chose Vulkan
Tested, F17 and F35
Look for Found 1 Vulkan devices and Vulkan0 lines: not using the GPU.
Old advice to ignore
“Use CLBlast for AMD.” CLBlast was removed in 1.107. Use Vulkan.
Fork-only guides from 2023 and 2024. They describe a fork that is many releases behind.
koboldcpp_rocm.exe presented as official. It is a third-party build.
Frequently asked questions
Does KoboldCpp work with AMD GPUs?
Yes. The official download for AMD is the nocuda build with the Vulkan backend, on Windows and on Linux. Choose Use Vulkan in the launcher or start with --usevulkan. KoboldCpp's own documentation recommends Vulkan for AMD users.
Vulkan or ROCm: which is faster on AMD?
We do not know yet: we have no AMD card in our lab, so we have not measured either. Start with Vulkan because it works everywhere without extra software, and try ROCm only if you want to compare on your own card. Run --benchmark with each build and the same settings.
Is there an official ROCm build for Windows?
No. KoboldCpp publishes a ROCm build only for Linux, as an experimental download under the rocm-rolling release. For Windows there is a third-party fork by YellowRoseCx that is not made by the KoboldCpp project and lags behind the current release.
How do I fix "gfx1031 not supported"?
That error comes from ROCm, not Vulkan, and it means ROCm does not support your card's chip by default. The simplest fix is to use Vulkan. On Linux, users report setting HSA_OVERRIDE_GFX_VERSION=10.3.0 for gfx1031 and gfx1032 cards; we have not tested it, and wrong values can crash.
Does KoboldCpp work on an RX 580?
Vulkan works on older cards in principle, and the Polaris generation is one of them, but we have not tested one. ROCm does not support it officially. Use the nocuda build with Vulkan, keep the model small and expect to put some layers on the CPU.
Why doesn't KoboldCpp detect my Radeon iGPU's VRAM?
Autofit needs to read the card's memory, and it cannot on some integrated graphics, so it prints Unable to detect VRAM. Set --gpulayers yourself, using the calculator, and make sure KoboldCpp picked the right device if both an integrated and a separate GPU are listed.