Skip to content

replay (review only): githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 — filtered squash of 920 commit(s) from https://github.com/eelgaev/Strata-AC922 - #4

Closed
k0valik wants to merge 10 commits into
mainfrom
replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922
Closed

k0valik wants to merge 10 commits into
mainfrom
replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922

Conversation

@k0valik

@k0valik k0valik commented Oct 15, 2026 •

Copy link
Copy Markdown
Owner

Fork replay — 🟡 REVIEW

upstream https://github.com/Niko1221/Strata.git (main)
foreign fork https://github.com/eelgaev/Strata-AC922 (ac922)
foreign commits 920 (fb58e0dbc839..1f029aadc044)
their fork point 622 commit(s) behind upstream; -622 (their fork point is older than our accepted baseline)
replayed onto 93caa81eb4b9 (HEAD)
raw change 72 files changed, 4700 insertions(+), 573 deletions(-)
replayed change ** 65 files changed, 4033 insertions(+), 545 deletions(-)**
filtered by policy 6 dropped, 1 ours, 0 protected, 0 unclassified
replay branch replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 — 10 squashed commit(s), 65 file(s)
verification 0 structural leak(s); pristine files: 27 exact / 0 mismatched; both-sides-changed: 38; conflict-marked: 18
published replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 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/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/core/weights.hpp (foreign fork M; this fork's copy differs from their fork point)
  • include/strata/kernels/cpu/expert_layout.hpp (foreign fork M; this fork's copy differs from their fork point)
  • include/strata/kernels/fused_gdn.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/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/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)
  • src/core/mtp.cpp (foreign fork M; this fork's copy differs from their fork point)
  • src/core/pinned.cu (foreign fork M; this fork's copy differs from their fork point)
  • src/core/session.cpp (foreign fork M; this fork's copy differs from their fork point)
  • src/core/verify.cpp (foreign fork M; this fork's copy differs from their fork point)
  • … and 18 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 — 36 integrated
  • include — 26 integrated
  • tools — 2 integrated, 2 suppressed by fork policy
  • (repository root) — 1 integrated, 3 suppressed by fork policy
  • docs — 2 suppressed by fork policy
Foreign fork commits (920, 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
  • 6da1f667e865 Engine v0.1.2: streaming tool calls, live status, real cancel; fix requests stalling on a full GPU
  • dd03276deaa9 serve: fix zero-token replies from engine-queue race between requests
  • 23f6b15b6b80 serve: conversation cache - keep the chat between requests, read only what is new
  • c47e290d35ce cpu pool: sleep between requests instead of spinning (fixes replay (review only): githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 — filtered squash of 920 commit(s) from https://github.com/eelgaev/Strata-AC922 #4)
  • 9d9351b9abdd serve: read short prompt parts through the decode windows
  • 80a1f1327b52 Merge branch 'pr-7'
  • 30acb0c7a89e Merge branch 'pr-8'

885 commits omitted; the whole range is fb58e0dbc839..1f029aadc044.

  • 0514771fa4c9 serve: commit only the tokens a verify window hands out (upstream PR serve: commit only the tokens a verify window hands out Niko1221/Strata#652)
  • ed6c4c835e5d layer split: the last stage books the draft head; the VRAM check reads every stage
  • f7c3b9d0f39d docs: the fork's text on top of upstream's README; IBM_AC922.md refreshed
  • 4f0ccbb25794 README: peak prompt and decode numbers at 256K context (4x V100)
  • 6ead04805874 Record upstream's rewritten history: a1641e9 is 6f32ec0 (merged in ef9f8b4)
  • 6de0c9fae6a8 Merge upstream 0.1.40 part 1: the Aurora gfx1151 merge (bc6b6f8)
  • 0c4af3dcb592 Merge upstream 0.1.40 part 2: through f96392d (Aurora decode kernels build for CUDA)
  • 0d19fd3a0d8c Merge upstream 0.1.40 part 3: through ce37456 (f-140 merges, HIP dense MMQ, gfx906 compat)
  • 8e9539669d9a Merge upstream 0.1.40 part 4: through 458bcbf (Feat/k8v4 kv streaming Niko1221/Strata#705 K8V4 KV streaming, HIP FP16 activation images)
  • 29d292e7ba08 Merge upstream 0.1.40 part 5: through 1048d23 (KV prefetch, session files, elastic K/V)
  • e37287733c8d Merge upstream 0.1.40 part 6: through b55c48b (4-way split scoring, UD-Q6_K_XL, AVX-VNNI)
  • a615e1efa9a2 Merge upstream 0.1.40 part 7: through 4552a7f (MIN_ENGINE 0.1.40)
  • a4648144cc81 Merge upstream 0.1.40 part 8: through 1664236 (deferred arena registration opt-in)
  • 4c8923e89327 Merge upstream 0.1.40 part 9: through 023dd01 (perf(cuda): fused decode/verify/MTP kernels and graph launch reductions on 0.1.39 Niko1221/Strata#783 decode performance series)
  • a86e4b1ed10a Merge upstream 0.1.40 part 10: through eb4e0c4 (docs, community benchmarks)
  • 5f849f9f6157 Merge upstream 0.1.40 part 11: through 536db01 (our PR qsa_prompt_attn: Volta (sm_70) m8n8k4 kernel - keep the hi and lo halves in separate accumulator chains Niko1221/Strata#800 upstream, SIMT block scores)
  • 34883d498674 Merge upstream 0.1.40 part 12: through e2de691 (0.1.40.1 follow-ups)
  • 4d1a00ac09a4 Merge upstream 0.1.40.2 (part 13): through c4c3221
  • 53af82c84715 Merge upstream 0.1.40.3 (part 14): through d5ea713
  • 62bbcfb54893 tools/mtp_experts.py: the BF16 reference directory is an argument, no private path
  • 3776e7a971ce Merge upstream 0.1.41 part 1: through 073e2fe (f-141-quick)
  • c4889cd3e366 Merge upstream 0.1.41 part 2: through c27998f (f-141-cuda)
  • 9688f05b9be8 Merge upstream 0.1.41 part 3: through 9b5e9ab (fixes, ports, multi-GPU, batches, research hooks)
  • 8dc93d5b4ce0 Merge upstream 0.1.41 part 4: through fb58e0d (Version 0.1.41, v0.1.40.4, stage pin opt-in)
  • 1f029aadc044 docs: merged up to upstream v0.1.41; the launch switches the merges added
ours — 1 path(s) the foreign fork changed that this fork does not take
  README.md (M)
dropped — 6 path(s) the foreign fork changed that this fork does not take
  docs/DGX_SPARK.md (A)
  docs/IBM_AC922.md (A)
  setup.py (M)
  setup.sh (M)
  tools/test_setup_golden.py (M)
  tools/test_setup_unsloth.py (M)
mirror — 65 path(s) taken from the foreign fork
  CMakeLists.txt (M)
  include/strata/core/expert_cache.hpp (M)
  include/strata/core/expert_source.hpp (M)
  include/strata/core/fp16_mode.hpp (A)
  include/strata/core/load_converted.hpp (A)
  include/strata/core/mtp.hpp (M)
  include/strata/core/pinned.hpp (M)
  include/strata/core/verify.hpp (M)
  include/strata/core/vmm.hpp (M)
  include/strata/core/weight_form.hpp (A)
  include/strata/core/weights.hpp (M)
  include/strata/hip_compat/cuda_runtime.h (M)
  include/strata/kernels/bf16_gemv.hpp (M)
  include/strata/kernels/cpu/expert_layout.hpp (M)
  include/strata/kernels/cpu/kq_vsx.hpp (A)
  include/strata/kernels/fused_expert.hpp (A)
  include/strata/kernels/fused_gdn.hpp (M)
  include/strata/kernels/fused_gr.hpp (M)
  include/strata/kernels/gr.hpp (M)
  include/strata/kernels/ple.hpp (M)
  include/strata/kernels/shared_expert.hpp (M)
  include/strata/kernels/verify_kernels.hpp (M)
  include/strata/platform/cpu_relax.hpp (A)
  include/strata/platform/thread_affinity.hpp (A)
  include/strata/prefill/gemm.hpp (M)
  include/strata/prefill/kernels.hpp (M)
  include/strata/prefill/prefill.hpp (M)
  src/core/expert_cache.cpp (M)
  src/core/expert_source.cpp (M)
  src/core/graph.cpp (M)
  src/core/layer.cpp (M)
  src/core/load_converted.cpp (A)
  src/core/mtp.cpp (M)
  src/core/pinned.cu (M)
  src/core/session.cpp (M)
  src/core/verify.cpp (M)
  src/kernels/cpu/expert_layout.cpp (M)
  src/kernels/cpu/kq_vsx.cpp (A)
  src/kernels/cpu/native_expert.cpp (M)
  src/kernels/cpu/pool.cpp (M)
  src/kernels/cpu/portable.cpp (A)
  src/kernels/cpu/router_dot_parity.cpp (M)
  src/kernels/cuda/dequant_bf16.cu (M)
  src/kernels/cuda/elementwise.cu (M)
  src/kernels/cuda/fused_expert.cu (A)
  src/kernels/cuda/fused_gdn.cu (M)
  src/kernels/cuda/fused_gr.cu (M)
  src/kernels/cuda/gr.cu (M)
  src/kernels/cuda/iq_kernels.cu (M)
  src/kernels/cuda/native_bf16.cu (M)
  src/kernels/cuda/ple.cu (M)
  src/kernels/cuda/qsa_prompt_attn.cu (M)
  src/kernels/cuda/qsa_select.cu (M)
  src/kernels/cuda/shared_expert.cu (M)
  src/kernels/cuda/verify_kernels.cu (M)
  src/kernels/ngram.cpp (M)
  src/kernels/qsa_prompt_attn_parity.cpp (M)
  src/ngram/ple_reader.cpp (M)
  src/platform/direct_file.cpp (M)
  src/prefill/gemm.cu (M)
  src/prefill/kernels.cu (M)
  src/prefill/prefill.cpp (M)
  src/program/generate.cpp (M)
  tools/mtp_experts.py (A)
  tools/vision/CMakeLists.txt (M)

Groups (one squashed commit each)

  • include/strata — 26 applied (25 clean, 1 conflict-marked) file(s)
  • root — 1 applied (0 clean, 1 conflict-marked) file(s)
  • src/core — 9 applied (4 clean, 5 conflict-marked) file(s)
  • src/kernels — 21 applied (14 clean, 7 conflict-marked) file(s)
  • src/ngram — 1 applied (0 clean, 1 conflict-marked) file(s)
  • src/platform — 1 applied (0 clean, 1 conflict-marked) file(s)
  • src/prefill — 3 applied (2 clean, 1 conflict-marked) file(s)
  • src/program — 1 applied (0 clean, 1 conflict-marked) file(s)
  • tools — 1 applied (1 clean, 0 conflict-marked) file(s)
  • tools/vision — 1 applied (1 clean, 0 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.

     372  +372/-0  src/kernels/cuda/fused_expert.cu
     360  +256/-104  src/kernels/cuda/fused_gr.cu
     323  +314/-9  src/program/generate.cpp
     301  +248/-53  src/prefill/prefill.cpp
     277  +277/-0  src/kernels/cuda/qsa_prompt_attn.cu
     267  +236/-31  src/kernels/cuda/verify_kernels.cu
     227  +181/-46  src/core/verify.cpp
     198  +198/-0  src/kernels/cpu/portable.cpp
     194  +194/-0  src/kernels/cpu/kq_vsx.cpp
     188  +178/-10  src/core/expert_source.cpp

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/kernels/verify_kernels.hpp
src/core/expert_source.cpp
src/core/layer.cpp
src/core/mtp.cpp
src/core/pinned.cu
src/core/verify.cpp
src/kernels/cpu/native_expert.cpp
src/kernels/cpu/pool.cpp
src/kernels/cuda/dequant_bf16.cu
src/kernels/cuda/fused_gr.cu
src/kernels/cuda/iq_kernels.cu
src/kernels/cuda/verify_kernels.cu
src/kernels/ngram.cpp
src/ngram/ple_reader.cpp
src/platform/direct_file.cpp
src/prefill/prefill.cpp
src/program/generate.cpp

Inspect locally

git log --oneline HEAD..replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922
git diff HEAD...replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 --stat
git log --oneline fb58e0dbc839..1f029aadc044   # 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.

Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 26 file(s) under include/strata -- 25 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  include/strata/kernels/verify_kernels.hpp

Original authors: 4EverBuilder,Andy,Anon,BlueKingMuch

Original commits touching these paths (152 subject(s)):
  - Niko1221#253: the 8 GiB arena pin cap only under WDDM; STRATA_ARENA_PIN_GIB; the split's prompt chunk steps down
  - Niko1221#284 follow-up: wait for the async commit on an event, at every point that touches the session afterwards
  - Niko1221#315 follow-up: its variants and switch renamed (plain / split / staged, STRATA_HC_SPLIT) - no clash with STRATA_GR_V3
  - Niko1221#340 prototype: a split's streamed ring is 96 slots when most experts are resident (STRATA_SPLIT_RING); STRATA_SPLIT_SMALL_OWN=S keeps an S-token chunk's buffers per lending stage for the session; --no-prefill-borrow on a split books the ring (it no longer starves the draft head)
  - Niko1221#465 engine: batch slots for one GPU - warn instead of refuse, windows over the active slots, slot conversation cache, prompts read beside decoding slots, BYIELD
  - Niko1221#583 default on (owner): STRATA_RING_BYTES=0 opts out; DETAILS.md with the measured quality
  - Niko1221#583: the byte-budget ring only where it gains; short prompts keep 0.1.39's ring
  - AMD HIP backend for gfx1100 (PR Niko1221#121, squashed onto its 0.1.21 base)

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: include/strata
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under root -- 0 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  CMakeLists.txt

Original authors: Andy,BlueKingMuch,Chanyeong Lim,Daehun

Original commits touching these paths (152 subject(s)):
  - Niko1221#136: fused int8 tensor-core experts for the Q2_0 pack's prompt path, grouped on the GPU (opt-in STRATA_PF_FUSED=1)
  - Niko1221#136: the fused int8 prompt experts for the native IQ packs (gate/up IQ2_XXS IQ2_XS IQ2_S IQ3_XXS IQ3_S IQ4_XS, down Q2_0 IQ4_NL; opt-in STRATA_PF_FUSED=1)
  - Niko1221#290 follow-up: a BF16-row parity test (native_expert_parity --bf16-embd)
  - Niko1221#296 follow-up: only the Q5_0 MMQ instance and case stay (Q4_K / Q5_1 come with S5's STRATA_MMQ_KQUANTS)
  - Niko1221#296 follow-up: ple_q5_parity / q5_projection_parity build only with STRATA_BUILD_TESTS; the doc says what runs where
  - A fatal CUDA fault in the prompt path exits at once instead of hanging until the watchdog (Niko1221#224); CMake warns when sm_120 is built with CUDA < 13.0 (Niko1221#220, Niko1221#224); iq_pack writes native_experts.txt under a temporary name, so a stop part-way leaves no partial pack (Niko1221#172)
  - AMD HIP backend for gfx1100 (PR Niko1221#121, squashed onto its 0.1.21 base)
  - AMD HIP: RDNA2 gfx1030 (tested), gfx(1031, 1032) untested.

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: root
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 9 file(s) under src/core -- 4 clean,
5 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/core/expert_source.cpp
  src/core/layer.cpp
  src/core/mtp.cpp
  src/core/pinned.cu
  src/core/verify.cpp

Original authors: 4EverBuilder,Abhinand B,Andy,Anon

Original commits touching these paths (158 subject(s)):
  - "The draft head does not fit" says what to change (Niko1221#474)
  - Niko1221#253: the 8 GiB arena pin cap only under WDDM; STRATA_ARENA_PIN_GIB; the split's prompt chunk steps down
  - Niko1221#280 follow-up: the float64 RoPE table is opt-in (STRATA_ROPE_TABLE=1); a freed session's table is forgotten
  - Niko1221#284 follow-up: wait for the async commit on an event, at every point that touches the session afterwards
  - Niko1221#293 follow-up: the drafter stores rotated INT8 K/V as INT8; the rotation is part of the snapshot format
  - Niko1221#465 engine: batch slots for one GPU - warn instead of refuse, windows over the active slots, slot conversation cache, prompts read beside decoding slots, BYIELD
  - Niko1221#646 on 0.1.39b: the layer split and the CPU path as before, the resident plan for 8-token windows
  - A 6 GB card starts with a small expert cache, or says what is short (Niko1221#496)

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/core
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 21 file(s) under src/kernels -- 14 clean,
7 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/kernels/cpu/native_expert.cpp
  src/kernels/cpu/pool.cpp
  src/kernels/cuda/dequant_bf16.cu
  src/kernels/cuda/fused_gr.cu
  src/kernels/cuda/iq_kernels.cu
  src/kernels/cuda/verify_kernels.cu
  src/kernels/ngram.cpp

Original authors: 4EverBuilder,Andy,BlueKingMuch,Chanyeong Lim

Original commits touching these paths (151 subject(s)):
  - Niko1221#152 as an opt-in (the owner's choice): STRATA_IQ_MT_MIN=1 makes CPU expert rounding independent of the drafting; the default stays 0.1.29's rule (IQ3_S decoded 1-3% slower with it on); docs; the width test runs the opt-in
  - Niko1221#289 follow-up: a placeholder GPU name when the properties query fails; HIP builds compile (no device_code_error probe)
  - Niko1221#315 follow-up: its variants and switch renamed (plain / split / staged, STRATA_HC_SPLIT) - no clash with STRATA_GR_V3
  - Niko1221#646 on 0.1.39b: the layer split and the CPU path as before, the resident plan for 8-token windows
  - Niko1221#646: the single-matrix mmvq does not stage the IQ grids (10-20% slower on the RTX 5070 at 3-8 columns)
  - AMD HIP backend for gfx1100 (PR Niko1221#121, squashed onto its 0.1.21 base)
  - AMD HIP: RDNA2 gfx1030 (tested), gfx(1031, 1032) untested.
  - C-1, C-2: the prompt path's QSA select grid reaches only the active blocks; the indexer appends a chunk at once

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/kernels
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under src/ngram -- 0 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/ngram/ple_reader.cpp

Original authors: BlueKingMuch,Eli Elgaev,Konstantinos Korres,Niko1221

Original commits touching these paths (5 subject(s)):
  - AMD HIP backend for gfx1100 (PR Niko1221#121, squashed onto its 0.1.21 base)
  - Helper threads leave the host thread's core (Linux)
  - PLE: the n-gram table in FP8 E4M3, as Qwen ships it
  - Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install
  - ple: keep the SSD awake while rows are read (STRATA_SSD_KEEPALIVE)

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/ngram
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under src/platform -- 0 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/platform/direct_file.cpp

Original authors: Eli Elgaev,Niko1221

Original commits touching these paths (3 subject(s)):
  - C-4: the prompt's embedding in one gather, and the n-gram table read by several threads
  - Helper threads leave the host thread's core (Linux)
  - Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/platform
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 3 file(s) under src/prefill -- 2 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/prefill/prefill.cpp

Original authors: Andy,Anon,BlueKingMuch,Chen Bao

Original commits touching these paths (123 subject(s)):
  - Niko1221#136 P3: with the fused experts the prompt path's MoE buffers hold only what the fused kernels need
  - Niko1221#136: fused int8 tensor-core experts for the Q2_0 pack's prompt path, grouped on the GPU (opt-in STRATA_PF_FUSED=1)
  - Niko1221#136: the fused experts take a 1024-slot streamed ring by default (two layers' experts)
  - Niko1221#136: the fused experts take a 512-slot streamed ring (a whole layer's experts) by default
  - Niko1221#136: the fused int8 prompt experts for the native IQ packs (gate/up IQ2_XXS IQ2_XS IQ2_S IQ3_XXS IQ3_S IQ4_XS, down Q2_0 IQ4_NL; opt-in STRATA_PF_FUSED=1)
  - Niko1221#253: the 8 GiB arena pin cap only under WDDM; STRATA_ARENA_PIN_GIB; the split's prompt chunk steps down
  - Niko1221#280 follow-up: the rope kernels without the table are 0.1.31's code exactly (template <bool TAB>)
  - Niko1221#281 follow-up: the saturated FP16 SwiGLU product keeps a NaN a NaN

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/prefill
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under src/program -- 0 clean,
1 with conflict markers.

Conflict-marked files -- a pull request with conflicts: resolve them,
do not discard the work:
  src/program/generate.cpp

Original authors: Andy,Anon,BlueKingMuch,Constantin

Original commits touching these paths (241 subject(s)):
  - Niko1221#253: the 8 GiB arena pin cap only under WDDM; STRATA_ARENA_PIN_GIB; the split's prompt chunk steps down
  - Niko1221#272 follow-up: --pool-affinity all (the pool's layout so far) is the default; auto and p-cores are opt-in
  - Niko1221#282 follow-up: chunks above 8192 are opt-in (--prefill auto:16384 / auto:32768, or STRATA_PREFILL_AUTO_MAX)
  - Niko1221#284 follow-up: wait for the async commit on an event, at every point that touches the session afterwards
  - Niko1221#289 follow-up: a placeholder GPU name when the properties query fails; HIP builds compile (no device_code_error probe)
  - Niko1221#293 follow-up: the drafter stores rotated INT8 K/V as INT8; the rotation is part of the snapshot format
  - Niko1221#340 prototype: STRATA_SPLIT_SMALL_MAX=M - a split request of at most M prompt tokens reads in the stages' own S-token chunks (nothing lent)
  - Niko1221#340 prototype: a layer split refills every stage's prompt loan concurrently (queue all, then wait); STRATA_REFILL_SERIAL=1: stage after stage; STRATA_TRACE times the loan and the refill

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: src/program
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under tools -- 1 clean,
0 with conflict markers.

Original authors: Eli Elgaev

Original commits touching these paths (3 subject(s)):
  - STRATA_MTP_EXPERTS=<file>: the drafter experts from an MTP GGUF, native kernels
  - tidy: no host name or private path in the sources
  - tools/mtp_experts.py: the BF16 reference directory is an argument, no private path

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: tools
Fork-Replay-Automated: true
Squashed from 920 commit(s) fb58e0d..1f029aa on https://github.com/eelgaev/Strata-AC922, restricted to the
paths this fork still mirrors: 1 file(s) under tools/vision -- 1 clean,
0 with conflict markers.

Original authors: Eli Elgaev,Niko1221

Original commits touching these paths (3 subject(s)):
  - Merge upstream 0.1.40 part 12: through e2de691 (0.1.40.1 follow-ups)
  - Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install
  - Support aarch64 and NVIDIA DGX Spark (GB10, unified memory)

Review artifact produced by fork-replay.sh (a spike adapted from
sync-upstream.sh): a filtered, squashed replay for reading -- not built,
not tested, and not evidence that any of it works.

Fork-Replay-Source: https://github.com/eelgaev/Strata-AC922
Fork-Replay-Branch: ac922
Fork-Replay-Base: fb58e0d
Fork-Replay-Tip: 1f029aa
Fork-Replay-Group: tools/vision
Fork-Replay-Automated: true
@k0valik
k0valik force-pushed the replay/githubproxy.fjygbaifeng.eu.org-eelgaev-Strata-AC922-ac922 branch from 1ecf008 to 04a2991 Compare October 15, 2026 02:54
@k0valik k0valik closed this Oct 31, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant