[k0valik]

Fork replay — 🟡 REVIEW

upstream	https://github.com/Niko1221/Strata.git (main)
foreign fork	https://github.com/eddoursul/Strata (custom)
foreign commits	539 (a0116b2782a5..c4bcb6e81ff3)
their fork point	282 commit(s) behind upstream; -282 (their fork point is older than our accepted baseline)
replayed onto	93caa81eb4b9 (HEAD)
raw change	143 files changed, 19869 insertions(+), 45622 deletions(-)
replayed change	** 119 files changed, 18700 insertions(+), 44856 deletions(-)**
filtered by policy	15 dropped, 3 ours, 0 protected, 5 unclassified
source filter	1 file(s), +0/-0 lines skipped as non-source — the reading surface stays code (--include-data to include)
replay branch	replay/githubproxy.fjygbaifeng.eu.org-eddoursul-Strata-custom — 10 squashed commit(s), 118 file(s)
verification	0 structural leak(s); pristine files: 74 exact / 0 mismatched; both-sides-changed: 45; conflict-marked: 38
published	replay/githubproxy.fjygbaifeng.eu.org-eddoursul-Strata-custom pushed to origin — a pull request opens from this report

Attention

🟡 The foreign fork touched fork-sensitive areas (build integration,
dependency integration, or runtime data) — worth a look before adopting anything:

• CMakeLists.txt (M, mirror)

Fork divergences touched by this range

These mirrored paths already differ from their fork point — upstream moved
since they forked, or this fork patched them — and the foreign fork changed
them too. The three-way apply kept this fork's version and layered their
change on top, and it did so without a conflict: were the edits on the same
lines, the file would have landed in the manual list instead. A clean apply
is a statement about the text, not about the meaning: if the foreign fork
fixed the same thing its own way, the replayed file can hold a duplicated or
contradictory fix. Read the replayed file, not just its status.

• CMakeLists.txt (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/expert_source.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/mtp.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/native_dense.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/native_head.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/pinned.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/core/verify.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/cpu/native_expert.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/fused_gr.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/iq_kernels.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/native_qsa_indexer.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/native_router.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/kernels/verify_kernels.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/prefill/gemm.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/prefill/kernels.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/prefill/moe_mmq.hpp (foreign fork M; this fork's copy differs from their fork point)
• include/strata/prefill/prefill.hpp (foreign fork M; this fork's copy differs from their fork point)
• src/core/expert_cache.cpp (foreign fork M; this fork's copy differs from their fork point)
• src/core/expert_source.cpp (foreign fork M; this fork's copy differs from their fork point)
• src/core/layer.cpp (foreign fork M; this fork's copy differs from their fork point)
• … and 25 more

Why each divergence exists is recorded under "Active source divergences" in
FORK_NOTES.md; this list comes from git and does not guess intent.

What changed

• src — 69 integrated
• include — 44 integrated
• serve — 7 suppressed by fork policy
• (repository root) — 1 integrated, 6 suppressed by fork policy
• examples — 5 suppressed by fork policy
• tools — 3 integrated
• docs — 3 suppressed by fork policy
• tests — 2 integrated
• bench — 2 suppressed by fork policy

Foreign fork commits (539, showing the oldest 10 and newest 25)

• f2a08d4b2da9 Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install
• 28c47b103ea9 Swift 1.5 support (UkisAI's fine-tune): split shards, per-model chat template, engine v0.1.1
• 1ee8b666c13e README: a short, beginner-friendly front page (details in docs/DETAILS.md); pictures in WebP/TIFF/AVIF
• 0aaa75465679 Prompt cache for --serve
• 8029fa9df743 Unsloth Dynamic 3.0 model files
• 240c8be7e692 Short prompt parts through verify windows
• 83d83e77ce00 Build without compiler warnings
• 8d4210ff1abc MTP draft head from Q6_K output heads
• 7317dadbd888 Speculation benchmarks on fixed text (--spec-follow)
• 7efc36cf3bc4 One adaptive VRAM tier for both decode loops; routing concentration in --stats

504 commits omitted; the whole range is a0116b2782a5..c4bcb6e81ff3.

• b75794e86d0e Elastic K/V (--kv-grow): the K/V takes VRAM as the context grows, the expert cache holds the rest
• d45fb025cadc kv-grow: with --second-gpu too (the VMM arena is the main GPU's cache alone)
• bb8fbc6bf9d8 kv-grow: off with --mmap-experts (every expert in RAM, as upstream's)
• 4fd80b24a19b Linux startup: the expert arena filled with preadv and page-locked layer by layer
• c354a0c1f570 experts only in VRAM (--vram-pin-gib, --second-gpu-pin-gib): a model whose experts outgrow the RAM
• 9a9463ad5e7f review fixes: the PCIe share asks whether a layer is mapped, the pinned slots in every loan probe, the dequant test registered
• ae64b1e4ebff Merge pull request Fork: uncensored model choices (OrcaRouter, mradermacher, RVN) in the… Niko1221/Strata#12 from txiki739/tools-build-fix: strata-load and gr_parity compile again
• 6845bd18007a gr_parity: without upstream's STRATA_QFUSE check, which the 0.1.43 merge brought for a read this branch does not have
• 5660c8364362 Merge pull request CPU pool: fix dangling else in Linux physical_cores() (fixes #13, root cause of #11) Niko1221/Strata#14 from txiki739/ple-ram-ready-wait: --ple-io ram, READY once the PLE table is in RAM
• 2d21182be9ae Merge pull request Add optional multi-GPU expert caches and compact transfers Niko1221/Strata#16 from txiki739/pool-fused: a native layer's gate/up and down rows as one phase (STRATA_POOL_FUSED=1, opt-in)
• be1009b290a5 pool_test --stress: the native formats through the fused phase too (bitwise the serial runs)
• 2771a7757e3a Merge pull request kv: add Q4_0 KV cache mode with FWHT-256 Hadamard rotation Niko1221/Strata#21 from txiki739/adapt-decay: --adapt-decay F, refused outside (0, 1)
• ad6e0eccc9e2 CPU experts: upstream's AVX-VNNI probe in CMake (the AVX-2 rows' vpdpwssd / vpdpbusd copies, IQ4_NL down rows of 2+ tokens on this CPU), iq_avx2_parity and q8k_quant_parity built, native_expert_parity's AVX-512 rows only where the CPU runs them
• c8a4d02f48dc Merge pull request Linux: CPU pool spawns ~16k workers and hangs at start (dangling else in pool.cpp, root cause of #11) Niko1221/Strata#13 from txiki739/linux-fast-arena-load: Linux startup, the expert arena filled with preadv and page-locked layer by layer
• 288499abe35f expert_source: the Windows read loop's offset renamed (it hid Linux: CPU pool spawns ~16k workers and hangs at start (dangling else in pool.cpp, root cause of #11) Niko1221/Strata#13's per-layer callback: C4458)
• 4ddc1e60afca Merge pull request serve: per-request temperature/top_p/top_k/min_p/penalties sampling Niko1221/Strata#19 from txiki739/kv-grow: elastic K/V (--kv-grow), the K/V takes VRAM as the context grows, the expert cache holds the rest
• e74be388b71e Merge pull request Unable to prepare the model, python version and ram requirements clarifications Niko1221/Strata#20 from txiki739/vram-pinned-experts: experts only in VRAM (--vram-pin-gib, --second-gpu-pin-gib)
• 7c82fe6418bb Merge pull request generate: residual control-vector projection from GGUFs Niko1221/Strata#15 from txiki739/resplit-gguf-native-dense: a re-split GGUF may name the architecture on every shard; a later shard must name the first one's
• b9e5fabaab46 Merge pull request a bit broken in latest update Niko1221/Strata#17 from txiki739/cpu-iquant-kernels, its content in this branch with upstream 0.1.43 and ad6e0ec
• 0848772103a9 docs: --kv-grow, --vram-pin-gib, STRATA_POOL_FUSED and --adapt-decay measured on this machine; examples/README no longer lists the elastic K/V as left out
• 27eb21b5f081 examples: --kv-grow in both configs (the 3090's cache 7,835 -> 8,787 IQ3_S and 4,602 -> 5,330 UD-Q4_K_XL slots while a conversation is short)
• 2b419a5fabed Merge the rebased heads of pull requests Fork: uncensored model choices (OrcaRouter, mradermacher, RVN) in the… Niko1221/Strata#12, Linux: CPU pool spawns ~16k workers and hangs at start (dangling else in pool.cpp, root cause of #11) Niko1221/Strata#13, CPU pool: fix dangling else in Linux physical_cores() (fixes #13, root cause of #11) Niko1221/Strata#14, Add optional multi-GPU expert caches and compact transfers Niko1221/Strata#16, serve: per-request temperature/top_p/top_k/min_p/penalties sampling Niko1221/Strata#19, Unable to prepare the model, python version and ram requirements clarifications Niko1221/Strata#20 and kv: add Q4_0 KV cache mode with FWHT-256 Hadamard rotation Niko1221/Strata#21: their content is in the merges above
• 212f34533c39 Merge upstream main (a0116b2, engine 0.1.43.1): server and bench fixes
• cc69de9ca2f4 PLE: the mapped prompt gather asks for its next window's rows ahead (upstream 62497bf)
• c4bcb6e81ff3 STRATA_TRACE=2: each startup mark waits for the device and names the first step after which it reports an error (upstream 8b271f2)

ours — 3 path(s) the foreign fork changed that this fork does not take

  .gitignore (M)
  AGENTS.md (D)
  README.md (M)

dropped — 15 path(s) the foreign fork changed that this fork does not take

  UPDATE.bat (D)
  bench/results/2026-10-04-expert-math/README.md (A)
  bench/results/2026-10-04-expert-math/results.json (A)
  docs/COMPARISON.md (A)
  docs/DETAILS.md (M)
  docs/SECOND_GPU.md (M)
  serve/frontend.py (M)
  serve/server.py (M)
  serve/test_choose.py (A)
  serve/test_mcp.py (M)
  serve/test_responses.py (M)
  serve/test_tiers.py (A)
  serve/test_winjob.py (M)
  setup.py (M)
  update.sh (D)

mirror — 120 path(s) taken from the foreign fork

  CMakeLists.txt (M)
  data/draft_vocab.bin (M)
  include/strata/core/adaptive_policy.hpp (A)
  include/strata/core/adaptive_tier.hpp (A)
  include/strata/core/expert_cache.hpp (M)
  include/strata/core/expert_source.hpp (M)
  include/strata/core/layer.hpp (M)
  include/strata/core/mtp.hpp (M)
  include/strata/core/native_dense.hpp (M)
  include/strata/core/native_head.hpp (M)
  include/strata/core/pinned.hpp (M)
  include/strata/core/prompt_cache.hpp (A)
  include/strata/core/second_gpu.hpp (A)
  include/strata/core/session.hpp (M)
  include/strata/core/split_head.hpp (A)
  include/strata/core/verify.hpp (M)
  include/strata/kernels/bf16_gemv.hpp (M)
  include/strata/kernels/cpu/native_expert.hpp (M)
  include/strata/kernels/cpu/pool.hpp (M)
  include/strata/kernels/elementwise.hpp (M)
  include/strata/kernels/expf_contract.hpp (A)
  include/strata/kernels/fused_gr.hpp (M)
  include/strata/kernels/gr.hpp (M)
  include/strata/kernels/iq_kernels.hpp (M)
  include/strata/kernels/kv_q4.hpp (M)
  include/strata/kernels/kv_q8.hpp (M)
  include/strata/kernels/native_gr_norm.hpp (M)
  include/strata/kernels/native_mmvq.hpp (M)
  include/strata/kernels/native_moe.hpp (M)
  include/strata/kernels/native_ple_postops.hpp (M)
  include/strata/kernels/native_qsa.hpp (M)
  include/strata/kernels/native_qsa_indexer.hpp (M)
  include/strata/kernels/native_rope.hpp (M)
  include/strata/kernels/native_router.hpp (M)
  include/strata/kernels/ngram.hpp (M)
  include/strata/kernels/ple.hpp (M)
  include/strata/kernels/qsa_decode_attn.hpp (M)
  include/strata/kernels/qsa_select.hpp (M)
  include/strata/kernels/shared_expert.hpp (M)
  include/strata/kernels/verify_kernels.hpp (M)
  include/strata/prefill/experts.hpp (A)
  include/strata/prefill/gemm.hpp (M)
  include/strata/prefill/kernels.hpp (M)
  include/strata/prefill/moe_mmq.hpp (M)
  include/strata/prefill/prefill.hpp (M)
  include/strata/spec/suffix_drafter.hpp (M)
  src/artifact/dequant.cpp (M)
  src/core/adaptive_tier.cpp (A)
  src/core/expert_cache.cpp (M)
  src/core/expert_source.cpp (M)
  src/core/layer.cpp (M)
  src/core/load_main.cpp (M)
  src/core/mtp.cpp (M)
  src/core/native_dense.cpp (M)
  src/core/native_head.cpp (M)
  src/core/pinned.cu (M)
  src/core/prompt_cache.cpp (A)
  src/core/second_gpu.cpp (A)
  src/core/session.cpp (M)
  src/core/split_head.cpp (A)
  src/core/verify.cpp (M)
  src/kernels/cpu/kq_avx2.cpp (M)
  src/kernels/cpu/native_expert.cpp (M)
  src/kernels/cpu/pool.cpp (M)
  src/kernels/cpu/pool_test.cpp (M)
  src/kernels/cuda/elementwise.cu (M)
  src/kernels/cuda/fused_gr.cu (M)
  src/kernels/cuda/fwht.cuh (A)
  src/kernels/cuda/gdn.cu (M)
  src/kernels/cuda/gr.cu (M)
  src/kernels/cuda/hit_plan.cuh (A)
  src/kernels/cuda/iq_dot.cuh (A)
  src/kernels/cuda/iq_kernels.cu (M)
  src/kernels/cuda/kv_q4.cu (M)
  src/kernels/cuda/kv_q8.cu (M)
  src/kernels/cuda/mmvf_multi.cuh (A)
  src/kernels/cuda/native_bf16.cu (M)
  src/kernels/cuda/native_gr_norm.cu (M)
  src/kernels/cuda/native_mmvq.cu (M)
  src/kernels/cuda/native_moe.cu (M)
  src/kernels/cuda/native_ple_postops.cu (M)
  src/kernels/cuda/native_qsa.cu (M)
  src/kernels/cuda/native_qsa_indexer.cu (M)
  src/kernels/cuda/native_rope.cu (M)
  src/kernels/cuda/native_router.cu (M)
  src/kernels/cuda/ple.cu (M)
  src/kernels/cuda/qsa_decode_attn.cu (M)
  src/kernels/cuda/qsa_select.cu (M)
  src/kernels/cuda/quantize_act.cu (M)
  src/kernels/cuda/s2_gemv_quads.cu (M)
  src/kernels/cuda/shared_expert.cu (M)
  src/kernels/cuda/verify_kernels.cu (M)
  src/kernels/cuda/xor_scatter.cuh (A)
  src/kernels/dequant_q8_0_identity.cu (A)
  src/kernels/expert_grouped_parity.cu (A)
  src/kernels/gr_parity.cpp (M)
  src/kernels/mmvq_il_parity.cu (A)
  src/kernels/native_expert_parity.cpp (M)
  src/kernels/ngram.cpp (M)
  src/kernels/qsa_parity.cpp (M)
  src/kernels/qsa_select_parity.cu (A)
  src/kernels/quantize_act_parity.cpp (M)
  src/kernels/rope_parity.cpp (M)
  src/kernels/router_top10_parity.cpp (M)
  src/kernels/shared_expert_parity.cpp (M)
  src/kernels/swiglu_expf_parity.cu (A)
  src/kernels/verify_parity.cpp (M)
  src/prefill/experts.cpp (A)
  src/prefill/gemm.cu (M)
  src/prefill/kernels.cu (M)
  src/prefill/moe_mmq.cu (M)
  src/prefill/prefill.cpp (M)
  src/program/generate.cpp (M)
  src/spec/suffix_drafter.cpp (M)
  src/spec/suffix_drafter_test.cpp (M)
  tests/core/adaptive_policy_test.cpp (A)
  tests/core/file_expert_source_layout_test.cpp (A)
  tools/iq_pack.py (M)
  tools/mtp_pack.py (M)
  tools/mtp_rt.py (M)

Groups (one squashed commit each)

• include/strata — 44 applied (30 clean, 14 conflict-marked) file(s)
• root — 1 applied (0 clean, 1 conflict-marked) file(s)
• src/artifact — 1 applied (1 clean, 0 conflict-marked) file(s)
• src/core — 13 applied (7 clean, 6 conflict-marked), 1 not representable file(s)
• src/kernels — 46 applied (35 clean, 11 conflict-marked) file(s)
• src/prefill — 5 applied (1 clean, 4 conflict-marked) file(s)
• src/program — 1 applied (0 clean, 1 conflict-marked) file(s)
• src/spec — 2 applied (2 clean, 0 conflict-marked) file(s)
• tests — 2 applied (2 clean, 0 conflict-marked) file(s)
• tools — 3 applied (2 clean, 1 conflict-marked) file(s)

Where the mass is (top 10 files by churn on the replay)

Start reading here or skip accordingly: one rewritten file explains more
than ten thousand added lines of data.

    7983  +2977/-5006  src/program/generate.cpp
    2666  +942/-1724  src/core/verify.cpp
    2530  +683/-1847  src/prefill/prefill.cpp
    2188  +247/-1941  src/kernels/cuda/fused_gr.cu
    1931  +270/-1661  src/prefill/kernels.cu
    1615  +980/-635  src/kernels/cuda/qsa_decode_attn.cu
    1608  +431/-1177  src/kernels/cuda/qsa_select.cu
    1578  +1563/-15  src/kernels/verify_parity.cpp
    1424  +601/-823  src/kernels/cuda/iq_kernels.cu
    1170  +218/-952  src/kernels/cuda/native_mmvq.cu

Applied with conflict markers — resolve, do not discard

These files carry this fork's version and the foreign fork's version in one file,
separated by conflict markers — exactly like a pull request with conflicts. The
work in them is not rejected: read each file, decide what to keep, and resolve
the markers consciously. Only mirrored paths ever reach this state; dropped,
ours and protected paths never enter the replay at all.

CMakeLists.txt
include/strata/core/expert_source.hpp
include/strata/core/mtp.hpp
include/strata/core/native_dense.hpp
include/strata/core/native_head.hpp
include/strata/core/pinned.hpp
include/strata/core/verify.hpp
include/strata/kernels/fused_gr.hpp
include/strata/kernels/iq_kernels.hpp
include/strata/kernels/native_router.hpp
include/strata/kernels/verify_kernels.hpp
include/strata/prefill/gemm.hpp
include/strata/prefill/kernels.hpp
include/strata/prefill/moe_mmq.hpp
include/strata/prefill/prefill.hpp
src/core/mtp.cpp
src/core/native_dense.cpp
src/core/native_head.cpp
src/core/pinned.cu
src/core/session.cpp
src/core/verify.cpp
src/kernels/cpu/native_expert.cpp
src/kernels/cpu/pool.cpp
src/kernels/cuda/fused_gr.cu
src/kernels/cuda/iq_kernels.cu
src/kernels/cuda/native_mmvq.cu
src/kernels/cuda/native_qsa_indexer.cu
src/kernels/cuda/native_router.cu
src/kernels/cuda/qsa_decode_attn.cu
src/kernels/cuda/qsa_select.cu
src/kernels/cuda/verify_kernels.cu
src/kernels/qsa_parity.cpp
src/prefill/gemm.cu
src/prefill/kernels.cu
src/prefill/moe_mmq.cu
src/prefill/prefill.cpp
src/program/generate.cpp
tools/iq_pack.py

Not representable as a merge (a human decides)

src/core/expert_source.cpp

Not source (skipped; --include-data to include)

Measurement dumps, binaries and other non-source payloads under mirrored
paths. Kept out of the replay so the reading surface stays code; nothing
is lost -- adopting one is a single checkout away.

  - `data/draft_vocab.bin` (M)

Unclassified paths (excluded; they need a policy decision first)

examples/README.md
examples/iq3_s.json
examples/run-iq3_s.bat
examples/run-ud-q4_k_xl.bat
examples/ud-q4_k_xl.json

Inspect locally

git log --oneline HEAD..replay/githubproxy.fjygbaifeng.eu.org-eddoursul-Strata-custom
git diff HEAD...replay/githubproxy.fjygbaifeng.eu.org-eddoursul-Strata-custom --stat
git log --oneline a0116b2782a5..c4bcb6e81ff3   # the original commits

Review artifact — produced by fork-replay.sh. The replayed commits exist
for reading: building and testing them is a human step, exactly as for a
synchronization candidate. Nothing here is evidence that any of it works.

---

No upstream code was executed while preparing this synchronization: it reads git
objects and applies .fork/path-policy.

Building and running the engine against a candidate is a manual step (see
FORK_NOTES.md); this workflow runs no engine CI.