feat(llama-swap-strix): llama-swap with the Strix Halo q8_0-KV FA patch (#25494)

[fastlorenzo]

What

New app llama-swap-strix: llama-swap carrying ggml-org/llama.cpp#25494 (unmerged) — dequant q8_0 KV once in the Vulkan coopmat1 flash-attention path instead of once per workgroup. That is the exact configuration every generative route on ai-box (Strix Halo / gfx1151) runs: kv: q8_0 + -fa 1.

Author-reported on a Qwen3-MoE-30B-A3B with q8_0 KV (≈ ai-box's coder route):

	pp512 @32k	pp512 @65k	tg32 @32k	tg32 @65k
stock	200.2 t/s	99.1 t/s	36.71	26.27
patched	282.0	166.2	36.73	27.50

Greedy output byte-identical; cost is a scratch buffer that scales with KV size (~268 MB @128k).

How

Two build stages reproduce upstream llama.cpp's own .devops/vulkan.Dockerfile (same base, packages, cmake flags, plus the embedded web UI), applying patches/*.patch to the checked-out release tag first, then graft the patched llama-server and the ggml backend .sos onto the stock llama-swap image ai-box already runs.

• The Vulkan FA code is in libggml-vulkan.so, not the binary — a binary-only copy would build, run, and benchmark as a null result. container_test.go asserts the libs are present and patches-applied.txt is non-empty.
• Grafting onto the same BASE_IMAGE keeps the eventual A/B (ai-bench sweep coder-fa-patch on ai-box) to one variable.
• The PR's tests/test-backend-ops.cpp hunk is dropped (doesn't apply to b10331; built with LLAMA_BUILD_TESTS=OFF).

Notes for review

• VERSION (llama.cpp tag) and BASE_IMAGE (the llama-swap build from that same tag) must move together. Renovate is disabled for this app in .renovaterc.json5 so a bot can't bump one alone and silently invalidate every benchmark.
• amd64 only — this exists for one gfx1151 APU.
• Temporary: delete this app when #25494 merges upstream.

Local checks: hadolint clean, go vet ./apps/llama-swap-strix/..., docker buildx bake --print. The image has not been built end-to-end yet — first CI build is the real test (npm ci for the web UI stage, and whether a full Vulkan compile fits the runner).

🤖 Generated with Claude Code