#2932 [gfx906] 0.1.45: Flash-Next answers differ from 0.1.43.1 and 32k decode is ~13 % slower; the wide top-k (#1749) and STRATA_TOPK_OLD=1 give different outputs

Descrição

**Setup:** 2x Radeon Pro VII 16 GB (gfx906, wave64), Ryzen 9 3950X, 64 GB RAM, Ubuntu 24.04, kernel 6.8, ROCm 7.14 in a docker build image (clang 23), ggml at the pinned 3cf0325.
Build as in docs/AMD_HIP.md: `cmake -DSTRATA_HIP_GFX906=ON -DCMAKE_HIP_ARCHITECTURES=gfx906 -DSTRATA_BUILD_TESTS=OFF -DSTRATA_GGML_DIR=...`, targets `strata strata-device`.
Model Qwen3.8-Flash-Next GSQ-RCO IQ3_S native pack, layer split 27 over both cards, `STRATA_ARENA_MMAP=1`, `--pcie-frac 0`, MTP spec 4, KV int8 + `--kv-resident 32768`.

Built from v0.1.45 (477ac13) with the three gfx906 build fixes from #2459 (we hit the same breaks: the cublas maps, `GGML_USE_HIP` on `strata_mmq_w4a4`, and `moe_f16_kq.cu` in the gfx906 `strata_mmq`). With them it builds and serves (`strata 0.1.45`).

## Observed, not localised: answers and 32k decode change on gfx906 (0.1.43.1 -> 0.1.45)

Interleaved A/B (old, new, old, new; same config, fixed-nonce bench + 6 greedy prompts x 768 tokens x 2; `STRATA_ROUTE_TAIL_SKIP=0` on both):

| | prompt 4k / 32k tok/s | decode 4k / 32k (256 tok) | decode bench | hit |
|---|---|---|---|---|
| 0.1.43.1 (two arms) | 395 / 557, 393 / 551 | 43.1 / 51.3, 42.9 / 51.3 | 44.3, 44.7 | 89.5 % |
| 0.1.45 (two arms) | 394 / 564, 393 / 557 | 43.9 / 44.8, 43.7 / 44.8 | 43.7, 44.0 | 89.9 % |

- Each version reproduces itself exactly, arm to arm, but **12 of 12 greedy decode-bench texts differ between the versions**, from character 9-259. The release notes say Flash-Next answers are byte-identical; the identity gates listed are RTX 3060 / P100 / R9700 / A750, not gfx906.
- 32k: the 256-token reply to the same prompt takes 6258 vs 5227 ms with 229 vs 208 verify windows (drafts accepted 139 vs 162), so roughly +9 % per window on top of the different text.
- `STRATA_TOPK_OLD=1` on 0.1.45 gives a **third** set of texts (matches neither version), although #1749's docs say the wide and the original top-k give the same ids. On gfx906 they do not, possibly a wave64 issue in the wide kernel. It does not explain the whole difference.
- Separate good news: `--pipeline-windows 2` with `STRATA_ARENA_MMAP=1` no longer stalls on 0.1.45 (it stalled on 0.1.41 / 0.1.43.1 with "verify window (pipelined): the CPU experts of layer 22"), but it gives no decode gain on the mmap arena here (43.0 vs 44.0).

Happy to run any parity test or switch you want on the cards.