"hipErrorNoBinaryForGpu: Unable to find code object for all current devices!"

The program (usually PyTorch) has no compiled code for your AMD GPU's gfx target. Either install a build that includes it, or, for cards no build covers, use the override workaround.

What you see

On ROCm 6.x and earlier, the HIP runtime logs:

hipErrorNoBinaryForGpu: Unable to find code object for all current devices!

The error PyTorch raises uses the same wording as the CUDA one, or reports an invalid function:

RuntimeError: HIP error: no kernel image is available for execution on the device

RuntimeError: HIP error: invalid device function

Why it happens

AMD code is compiled per gfx target, with no compatibility between targets: code for gfx1030 does not run ongfx1031, even though both are RDNA 2. So the build must include your exact target. Find yours with rocminfo | grep gfx.

Which PyTorch build covers your card

A build existing isn't the same as AMD supporting the card: PyTorch keeps compiling some older targets into builds for ROCm releases that no longer support them. The install checker picks the right pairing for your card and explains which kind it is.

Cards no build covers

For targets no PyTorch build includes, like the RX 6700 XT's gfx1031, the usual workaround is HSA_OVERRIDE_GFX_VERSION, which makes ROCm treat the card as a supported target from the same generation. It's unofficial and can break in subtle ways; seeROCm on consumer GPUs for when it works. For llama.cpp, the Vulkan backend avoids the question entirely.

Sources: each PyTorch ROCm build's gfx list from PyTorch's .ci/docker/manywheel/build.sh at each release; AMD's per-card support lists from the ROCm documentation (see ROCm versions); error text from ROCm's HIP runtime (hip_error.cpp, and hip_code_object.cpp in ROCm 6.3).

← All error fixes