Eval bug: RX 7800 XT Vulkan decode and MTP slower on 84e76d8a2 than c77ae695c; #27952 prefill gain not seen #29410

Name and Version

Regressed build:

version: 0.5.0-dev (build 1099, commit 84e76d8a2)
built with MSVC 19.44.35227.0 for x64

Previous build, same machine and flags:

version: 0.4.1-dev (build 953, commit c77ae695c)

84e76d8a2 is 146 commits after c77ae695c and includes #27952. This is not a bisect. Prefill did not show the Strix Halo speedup, and decode / MTP got slower.

Operating systems

Windows 11

Which llama.cpp modules do you know to be affected?

llama-bench, llama-server (Vulkan)

Hardware

• GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver
• Log line: matrix cores: KHR_coopmat, uma: 0, shared memory: 32768
• CPU: Ryzen 5 7600X, 32 GB RAM
• Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf)

Command line

Empty-context bench (no speculative decoding):

llama-bench -m Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf -dev Vulkan0 -ngl 99 -ncmoe 18 -lm none -fa on -p 512 -n 32 -r 2

MTP decode at depth (llama-server, then tg128 on a filled KV):

llama-server -dev Vulkan0 -ngl 99 -lm none --n-cpu-moe 18 -m Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf -fa on -b 2048 --ubatch-size 2048 --cache-type-k q4_0 --cache-type-v q4_0 --reasoning off --spec-type draft-mtp --spec-draft-n-max 3 -c 163840 -t 12

Depths measured: 48000 and 130000 tokens. Acceptance stayed 1.00 on both builds.

Problem description

#27952 reports about +32% prompt processing for Qwen3.6-35B-A3B Q4_K_M on Strix Halo (RDNA3.5), and the new int8 coopmat1 MMQ path is enabled for q4_K on RDNA3. On this discrete RX 7800 XT the same quant did not get faster, and token generation got slower.

MTP acceptance was 1.00 at both depths on both builds, so the draft tokens are accepted. Each step just takes longer. The 48k MTP point fell by about half (55.8 → 25.8 t/s). The no-MTP tg32 point fell about 15% (42.4 → 36.1). pp512 is flat (1174 → 1151), not the +32% from the Strix Halo table.

Vulkan commits in c77ae695c..84e76d8a2 also include #28415 (IQ4_XS), #28822 (IQ3_S), #29139 (symbol visibility), #24406 (Intel Xe flash attention), #29328 (Adreno), #29365 (conv alignment) and #28501 (mul_mat_id hoist 256→512). The daily model is Q4_K_M with 256 experts, so the IQ kernels and the 512-expert hoist should not be the hot path. #27952 is the change that retargets q4_K MMQ on RDNA3. Decode uses the vec path, and #25666 (do not enable MMVQ for speculative decode on AMD) is still open, so the MTP drop may be a different commit in the same range.

Related but not the same report: #28752 is a Linux prefill regression after #28190. Here prompt processing at pp512 did not regress; generation and MTP did.

Happy to retest a single commit or a build with #27952 reverted if that helps narrow it.