#1793 bench: community report - a single Radeon Pro VII 16 GB (gfx906) on the distribution's ROCm 6.2

说明

One Radeon Pro VII 16 GB (Vega 20, gfx906 - the MI50's ISA) on a Xeon E5-2666 v3, built on the host with the distribution's own ROCm 6.2 instead of the community rocm-gfx906 Docker image, and measured at the 131,072-token context with KV streaming.

**What is new.** The two gfx906 reports in the tree are 2x MI50 16 GB and 1x MI50 32 GB. This is a single 16 GB card - the cheap end of this family - and it also carries the power/temperature readings and the per-request cache behaviour the earlier two do not. (#1596's "16 GB-limit" case limits a 32 GB card in software, which is a different measurement.)

**Results.** On 0.1.40.4: decode 24.2 / 24.0 / 23.7 tok/s, prompt 350 / 339 / 284 tok/s at 4,096 / 32,768 / 128,000 prompt tokens; 9 of 9 speed requests read their whole prompt; needles 6/6 at 32k and 128k. Against the 32 GB MI50 that is about a third less decode at the same prompt speed, and the engine log says why: this card's expert cache holds 7,245 of the model's 12,288 experts (19,304 on the 32 GB card), with a steady-state decode hit rate of 87.7-89.6 % against 97-99.5 %.

**Issue #1691 did not reproduce.** Two 0.1.41 sweeps (9 requests each) and its needle run showed no `verify: timed out at layer N`, no decode collapse and no engine exit; `grep -ci 'timed out'` over the whole 0.1.41 engine log is 0. The conditions differ from that report (ROCm 6.2 against 10.x, IQ2_XS against UD-IQ4_XS, int8 KV against q4_0, 131,072 against 262,144), so this narrows the bug rather than closing it. The README lists what it does not rule out.

**ROCm 6.2 needs four host patches**: the gfx906 branch's CUDA-compat include path, the 24 CUDA->HIP name mappings that 0.1.40.x's split headers lost, a `cudaInitDevice` shim, and `HIPBLAS_V2` plus a `cublasSetWorkspace` no-op. They are in `patches/` with the reasoning in `BUILD.json`. Happy to split them into their own PR if that is easier to review - the report can land on its own.

Two notes. The engine and build logs are committed past `.gitignore` (`*.log`); `TRIMMED.md` describes the folder's trimming, and I will drop the two build logs if you would rather keep a bench folder under a megabyte. And the board is a third-party (HUAQIANGBEI-refurbished) 16 GB Vega 20 sold under the Radeon Pro VII name - the VBIOS is where `rocminfo` gets that name, the ISA is gfx906 either way - so I have kept the raw `lspci`/`rocminfo` output in the folder rather than paraphrase it.