[test] chore(turboquant): bump fork pin to rebase/upstream-sync-april-2026

[mudler]

Move the TurboQuant llama.cpp fork pin from feature/turboquant-kv-cache (627ebbc6) to rebase/upstream-sync-april-2026 (7f320bb8), picking up the upstream-sync work on the fork.

Testing TheTom/llama-cpp-turboquant#101

cc @TheTom

[apollosenvy]

Pulled and sanity-checked on AMD / ROCm 7.2.1 (7900 XTX, gfx1100, wave32). HIP path looks good on the new pin.

Build

cmake -B build-hip -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1100 -DCMAKE_BUILD_TYPE=Release
cmake --build build-hip -j 12

Clean build, 0 errors, only a benign -Wmissing-prototypes in ggml-turbo-quant.c. The three fattn-vec-instance-f16-turbo{2,3,4}_0.cu template instances added in b8b1d49b3 link in correctly, so no more undefined references on ROCm.

Smoke tests (Qwen3-8B-Q4_K_M, -fa on -ngl 99)

K type	V type	pp t/s	tg t/s	VRAM used	Output
q8_0	turbo4	223.3	89.9	~12.2 GiB	coherent
f16	turbo3	228.8	90.9	~11.9 GiB	coherent

Both ran to the -n limit with sensible completions. No OOM, no crash, no hipError, no NaNs.

What this covers from PR #101's community-testing checklist

• AMD HIP build with variadic shfl macros
• OOM fix (d7b533446 / 4d754604e) validated: turbo3 + turbo4 V-cache both fit and decode on 24 GiB
• F16-K + TURBO-V dispatch case (58bbe5518) exercised cleanly
• turbo V unpad gate on V type (156592051) behaves correctly with turbo V + f16 K

Not covered here: gfx1200 (RX 9060 XT), head_dim > 256 (Gemma 4 full-attention), multi-GPU, prefill-heavy workloads against long contexts.

CI note

The only pull_request failure on this PR is backend-jobs (cublas, 13, ...) which timed out at 6 h inside the webui cp -rT rename into tools/grpc-server/ (GitHub Actions wall-clock ceiling, not a build error). Unrelated to the pin itself. The fork's own llama.cpp CI matrix (android-arm64 / ggml-ci-* / etc.) is failing upstream-wide, not specific to this rebase.

LGTM for the HIP slice. Ship it.

cc @TheTom @mudler

[TheTom]

Thank you!

[mudler]

Closing ad this seems now fixed and merged already in #9497 . Thanks for your support!