A CUDA 13.x build of llama.cpp needs a driver that runs CUDA 13.x
CI / test (push) Failing after 9m48s

The release carries some GPUs' code as PTX, which the driver compiles,
and a driver older than the toolkit refuses it at the first kernel,
after the build has already listed the GPU. On cuda01 (Turing, driver
615 = CUDA 13.4) b12606's CUDA 13.5 build installed and then aborted
loading Inkling 2 Mini with "the provided PTX was compiled with an
unsupported toolchain". The driver's CUDA, which NVML already reports
per GPU, is now carried on the host and the CUDA 13.x candidate is
passed over for CUDA 12.8 when the driver is older.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-11-21 14:01:16 +01:00
1 parent 09ebe77513
commit 6765d430d8
5 files changed
+65 -13

No files matched your search

+15
View File
@@ -289,6 +289,21 @@ func TestTheBuildFollowsTheHardware(t *testing.T) {
}
}
// Found on cuda01 (Turing, driver 615, which runs CUDA 13.4): b12606's
// CUDA 13.5 build listed the GPU, then failed its first kernel with "the
// provided PTX was compiled with an unsupported toolchain".
turing := discovery.Host{Arch: "amd64", NVIDIADriver: "615.71.09", MinComputeCapability: 7.5}
for _, c := range []struct {
driverCUDA string
want string
}{{"13.4", "cuda-12.8"}, {"13.5", "cuda-13.5"}, {"14.0", "cuda-13.5"}, {"", "cuda-13.5"}} {
turing.DriverCUDA = c.driverCUDA
got, err := llamaCppCandidates("b12606", "", turing)
if err != nil || got[0].Variant != c.want {
t.Errorf("driver for CUDA %q: %+v, %v; want %s first", c.driverCUDA, got, err, c.want)
}
}
slow := ubuntu
slow.CPUFlags = nil
st := (&Installer{Host: slow}).ready(Status{}, Build{CPUFlags: []string{"avx2"}}, "/x/llama-server", "")