← News
Runtime

Windows AMD Vulkan regression reports: an RX 7800 XT workaround does not transfer to RX 9070 XT

Two community reports describe Flash Attention crashes with the proprietary AMD Vulkan driver on Windows. Workarounds differ between the GPUs; this is not evidence of a universal AMD failure.

Source: llama.cpp GitHub issue及窗口内评论

On 15 November UTC, llama.cpp issue #32353 reported a Windows Vulkan crash, followed by a second user's RDNA4 report. These are community observations, not AraoAI reproductions or proof of one root cause across all AMD GPUs. A successful short prompt does not establish stability at realistic input lengths.

The first setup was an RX 7800 XT 16GB, Windows 11 Pro 10.0.26300, and the proprietary AMD driver amdvlk64.dll 32.0.32015.16001. It used self-built b12421, commit 75fd7f9ec, and the Gemma4-26B-A4B Uncensored-HauhauCS-Balanced IQ4_XS derivative—not the original Gemma4 weights. The author reported that longer prompts crashed during prefill while b12370 worked in that test.

The important distinction appears in the comments. Forcing GGML_VK_FORCE_FA_FLIP=kvsep worked for the RX 7800 XT reporter. A second reporter, using an RX 9070 XT, official b12425 binaries and gpt-oss-20b MXFP4, saw both direct and kvsep paths crash; disabling FA worked there. Different models and builds prevent treating either workaround as a universal fix or combining their speeds into a GPU comparison.

Keep a known-working build before testing an update. Record the full weight filename, driver, launch command and triggering input length; change one setting at a time on the same machine. When testing FA disabled, verify that the KV types are supported and document any additional configuration changes. A workaround is not a confirmed upstream fix, and a fault in the driver module is not a completed source-code bisection.

This is a compatibility notice, not a clean benchmark or a claim about untested cards. Original commands, logs and both reports: https://github.com/ggml-org/llama.cpp/issues/32353