[sreeram-11]

Problem

ccache could reuse old GPU code on ROCm 10.1.

• When ccache's quick check misses, it falls back to a second check that compares the preprocessed source.
• On ROCm 10.1 that preprocessed output only contains the CPU code, so a change that only touches GPU code (for example inside #if defined(RDNA3)) looks unchanged, and an object with the old GPU code is reused.
• Saw this on PR Speedup Windows ROCm builds #34's re-run: upstream changed GPU-only code in mma.cuh, and 10.1 reused 69 cached objects in 5 min while 7.14 and 10.0 rebuilt them.
  • The first PR run run #1 was slow and that was expected. Both the hash setting and the ccache version changed, so nothing old can match.
  • However, in the second PR run (run #2), upstream moved from 1a24df16 (b11846) to edc35384 (b11851), and changed GPU code. ROCm 7.14 and 10.0 recompiled correctly but 10.1 did not (it re-used stale GPU code).

---

FIX
Turn on ccache depend_mode in ubuntu-24-rocm and windows-rocm for all ROCm versions. ccache then only reuses a result when the source file and every file the compiler says it read are unchanged, and never uses the second check.

---

Notes

• 7.14 and 10.0 aren't affected today. This change keeps them safe once their compiler picks up the same behaviour.
• Small speed cost: in recent runs the second check usually saved 0-7 files, occasionally ~80-90.
• To confirm it's active, every ROCm job's ccache stats should show Preprocessed: 0 (Ubuntu: Post ccache step log; Windows: "CCache Statistics" on this PR's run Summary page).

[sreeram-11]

This PR and #34 are meant to go together.

• Speedup Windows ROCm builds #34 makes Windows ROCm builds fast again. Reinstalling ROCm no longer throws away the whole ccache. In the re-run of this PR, Windows ROCm reinstalled from scratch and still built in 9m 1s / 8m 51s / 11m 9s (7.14/10.0/10.1), down from 2-4h.
• This PR makes sure the speed doesn't come from skipping work. With depend_mode, ccache only reuses a result when the source and every file it includes are unchanged. All six ROCm jobs show Preprocessed: 0, so the check that earlier let 10.1 reuse old GPU code is no longer used.

[superm1]

Make sure to send upstream too.

[sreeram-11]

Make sure to send upstream too.

Upstream PR submitted and merged: ggml-org/llama.cpp#31231

Upstream is still on ROCm 10.0. PR that bumps ROCm to 10.1 is still open ggml-org/llama.cpp#30652.