engine cherry-pick plan — nathan/strix-halo-vulkan → ggml-org master

date: 2026-08-31. base for counting: merge-base 9f0d017ef (#27235 era). nathan branch tip: ad914eb65. #27742 landed in master as squash 6c84c7d5d. nathan's branch = the #27742 development history + his strix-halo patch stack + three master merges he already did + the qwen4exp runtime continued past the squash point.

counts

175 commits on 6c84c7d5d..nathan/strix-halo-vulkan (reverse order in docs/nathan-175-commits.txt)
~40 of them are the #27742 development history — skip, master's squash 6c84c7d5d already carries that content
~8 are master commits that reached his branch via his three master merges (muse glimmer #26841/#26879, motif-3, dspark #27508, kv-cell #27762) — skip, master has them
4 are CI/toolbox release plumbing — skip (fork-specific)
~10 are merge commits — skip (resolved by the one big merge below)
~115 genuine candidates, grouped below

recommendation: one merge, not 115 cherry-picks

the qwen4exp runtime commits and the vulkan shader stack interleave (the sparse-FA shaders are prerequisites for the qwen4exp QSA gather path; the FACP refactor renames classes the later commits use). piecemeal cherry-picking breaks the build between commits. instead:

git checkout -b haloq38flash-engine ggml-org/master   # or origin/master
git merge nathan/strix-halo-vulkan
# resolve conflicts once: ggml-vulkan mostly takes THEIRS (the perf stack),
# src/llama*.cpp mixed, everything else master
cmake -B build -DGGML_VULKAN=ON && cmake --build build -j 24

nathan already merged master into his branch three times (aaf4fba83, b7b85da9c, f94fad0e8/add19980d) — the reverse merge is the same operation he proved works, and conflicts concentrate in the files he owns.

group A — vulkan fa/mmq perf stack (~45, oldest first)

the coopmat1 FA rework, dequant-once scratch, contiguized KV, mul_mat_id tile probes, f16-B path, q5_K/q4_K scale caches, wave32, LDS pad tuning, the six env-gated perf flags now default-on. cherry-pick as a block, oldest first; acd14737e FACP and 892924042 single source of truth are the load-bearing refactors the later ones sit on. skip 681675530 (marked NEGATIVE result).

group B — dsv4 lightning indexer + sparse fa gather (~25)

890550c0a indexer kernels + indexed sparse FA, 5dfc01ff6 gather-to-compact decode, the sparse prefill split/tile/cache cluster, quantised K/V inside the gathers (8b66f91c6, 7b63cbd6b, 6b2cade31), small-batch union (8115df4c7..31202f9df). written for deepseek v4, powers qwen4exp's QSA the same way. NOTE: b65c360c7 fixes multi-sequence — keep.

group c — fused hyper-connection ops + command buffers (~5)

2041049a4 fused HC pre/comb/post (the 3550→2800 dispatch win), e709b949e command buffers bounded by memory traffic, 18239a695 perf-logger flush, 0f80b884d/8a8fee776 UMA copy path.

group d — hip/"ggml-cuda" rdna3.5 tuning (~12)

64e5c14f1 kernel tuning, f074165ae quantized-KV FA, MMQ tile tuning (71ac6c1d9, f70839f9a, 86e3f34fc), Q8_1 activation cache (a649f1634), WMMA indexer (8209c8954), tiled FA (e88b92eff), GDN tune (910f0f25d), NaN fix (4ea44eef2), tests (b1282d2af, 6e7b355cb). named ggml-cuda because the hip backend rides the cuda code paths.

group e — qwen4exp runtime past the squash point (~25)

what master's squash does NOT have:

be71d63c9 quantized KV cache in the QSA attention path (the q8_0 kv fix our rocmfpx build lacks)
631b9ffb1 decode-graph reuse + host-side PLE gather (graphs reused 68 vs 0)
354390810 + 39817c476 NextN/MTP draft: sidecar AND in-file loading
f32aca1c1/79c2d2cad/3849d54b8/d763facad/fdf96fcea indexer cache in llama_memory_hybrid_idx, slots, names
87f31259a/c04b3ff4b/8f58c2f0a PLE history per context + iterator fix
05f6575ab/25a796300/cdd2e47ae indexer cache save/restore + slots
c1d5b2d0e/bd92a90c4/671203688 random-access mmap advice for the gather table (the ple-ssd-streaming primitive)
024b7ad93 QSA bias per block, 7073ae357 hparams shrink, 1486f6b88 non-unified KV in QSA, a80d678ad image placeholder hash, 562cb00bc/42d976771 tensor-split segments, d6f65ff28 graph budget
quantizer: 7a4d5960d PLE streaming (independently written — same fix as our banding), 9e2d2eb84 --tensor-type names the PLE (the flag the 91g quant used), 5beb9965b f16 fallback for odd ncols, 5096585d6 exact output buffer
tests: 171ddb8df/086457e7b/77953f1e1

group f — speculative decoding fixes (~10)

the 7-bug stack behind "spec decode works end to end": 53fd8b48c GDN state graph order, 9c5d899ff/f25eefeaf/a17e8432b MTP rollback full checkpoints (apply→revert→reapply), 08a325524 checkpoints on device, 0eb528051 draft trimming for mtmd, 64e2b680a dflash cache alignment, 397ef7c72 no_vocab special tokens.

group g — optional, other archs (~10, default skip)

dspark bailingmoe3 (2586f6edd), dflash2 (015f09c8a/0b0f35d0e), motif-3 (4c7f96093/be54e2891/a359e55c9) — only if wanted; they ride along in the merge anyway.

verification after the merge

build vulkan, zero errors
our depth bench on the 91g quant: 0/8k/32k/128k — expect >= the nathanw1014 numbers (29.9/24.1/20.1 plain, 53.1/56.4/18.6 mtp) plus master's 771 commits
greedy oracle: with 39817c476 spec decode + the rollback fixes, the 6-line divergence on our current engine should close (his fork is the one the 7-bug fix stack was written for)
lazy ple: --tensor-read-lazy auto must log "lazy read enabled" for per_layer_token_embd — the load_mode=none hardcode does not exist on master's path
then 256k mtp: no thrash expected (66g resident with lazy ple)

candidates for upstreaming after validation

radix/sparse top-k fa, fused hc epilogs, gdn concat fix, the q8_0-kv-in-qsa fix, the lazy ple plumbing, the converter trap note (hc norms).

5.9 KiB Raw Blame History