[Triage] 30 open PRs: merge / rebase / author buckets, merge order and acceptance criteria (2026-10-21) #153

Activity

1. hsliuustc0106 commented on Oct 21, 2026

   

   hsliuustc0106

   on Oct 21, 2026

   ContributorAuthor

   More actions

   Merge manifest v1 - 4 wrong-base PRs + #118 verdict

   Snapshot: main = 0410e84 (2026-10-21; CI run 39614284188 success). Scope: the 4 PRs whose base is not main (#122, #123, #127, #143) and the fate of #118. Bots do not merge, retarget or push; this is a list for a human/Codex operator.

   Verdict 1 - #118 is not duplicated; it must be re-landed

   • git cherry -v origin/main origin/codex/qwen-vision-27b -> 9 non-merge commits, all +: no patch-equivalent exists in main. 10 maintenance merge commits ride on top (content covered by those 9). Nothing to discard as a duplicate.
   • Re-land is conflict-free at 0410e84: git merge-tree --write-tree origin/main origin/codex/qwen-vision-27b exits 0 with no CONFLICT; a scratch git merge --no-ff gives 19 commits / 60 files / +10129 -76, no conflicts.
   • Do not cherry-pick the 9 commits: the first one (ef1fd63 "jemm: add native decision worker...") already conflicts in Cargo.toml (verified at 0410e84). Merge the branch instead.
   • jemm: add native decision worker and pinned weight export #118 itself is spent (it was merged into the feature branch at 14:09Z): open a new PR codex/qwen-vision-27b -> main.

   Verdict 2 - #127 must be merged before the #118 re-land

   • cua_s1: retain bounded vision cache across screenshot grids #127 retargeted onto main: clean (git merge-tree --write-tree origin/main pr127 exits 0; 11 files, +785 -68).
   • cua_s1: retain bounded vision cache across screenshot grids #127 onto the re-landed base: 1 conflict in src/models/qwen3_5/native/src/vision/model.rs (re-land commit db400fb "Preserve all-image projection shapes in native 27B vision" touches the same file).
   • So cua_s1: retain bounded vision cache across screenshot grids #127 goes first; if jemm: add native decision worker and pinned weight export #118 lands first, cua_s1: retain bounded vision cache across screenshot grids #127 needs that one conflict resolved.

   Verdict 3 - #122/#123/#143 depend on the re-land

   Files they modify do not exist in main (checked per PR file list): #122 7/7 modified files missing in main, #143 3/28, #123 1/1. After the re-land, retargeting is conflict-free:

   • jemm: load and validate checksum-covered calibration #122 after re-land: merge-tree clean; own diff 7 files, +124 -26.
   • docs(jemm): document CUDA Graph controls and validation limits #123 after jemm: load and validate checksum-covered calibration #122: merge-tree clean; 1 file.
   • fix(cua-s1): match text reference and preserve merged compatibility #143 after re-land: merge-tree clean; own diff 28 files, +1811 -55.

   Order and per-step state

   step	operator action	PR	head SHA	CI on that head	state	depends on
   1	retarget base -> main, update branch to trigger CI, merge	#127	03b111ce	no run on head SHA	APPROVED / CONFLICTING	-
   2	new PR codex/qwen-vision-27b -> main, review, merge	#118 re-land	branch tip 8945af81	needs new PR CI	merged into feature branch only	-
   3	retarget base -> main, update branch, merge	#122	5d630bc2	run 38667515611 success	APPROVED / MERGEABLE	step 2
   4	retarget base -> main, update branch, merge	#123	7013822a	run 38667525027 success	APPROVED / MERGEABLE	step 3
   5	retarget base -> main, update branch, merge	#143	3d150b01	run 38675516723 success	APPROVED / MERGEABLE	step 2

   Acceptance for every step

   1. The new head SHA (after retarget/update, not the SHA above) has CI rust green: cargo fmt --all --check, cargo clippy --workspace --locked --all-targets -- -D warnings, cargo test --workspace --locked, cargo build --workspace --release --locked; and CI benchmarks green.
   2. Review is APPROVED with no unresolved threads; record the CI run id + SHA in the PR.
   3. One merge at a time; after each merge main CI must be green; next PR rebased onto the new main.
   4. No Bot merges, retargets or pushes branches (team policy); operator is human/Codex.

   Caveats

   • The CI run ids above were produced on the pre-retarget head SHAs; they do not transfer to a new head. CI must be re-run after the base change. Note that editing a PR's base does not trigger pull_request CI; "Update branch" or a rebase is needed.
   • cua_s1: retain bounded vision cache across screenshot grids #127 has no CI run at all on 03b111ce; do not quote an older run for it.
   • CUDA / accuracy / latency: not run (no NVIDIA GPU, no model weights locally). CI success is not accuracy evidence.
   • Local baseline at 0410e84 was reproduced and persisted: logs/cargo-test-0410e84-2026-10-21.log (workspace) - cargo test --workspace --locked = 198 passed / 0 failed / 48 ignored, exit 0, 63 ok blocks, 0 FAILED; cargo fmt --all --check exit 0 (macOS Apple M4, no GPU, cold target dir 2.5 GiB; 11 GiB free afterwards). 48 ignored tests are the GPU/weights-gated ones. This is a local reference, not a merge gate; the gate is CI on the new head SHA.
   • Note: the earlier cargo-test-0410e84.log at workspace root is the ENOSPC linker-failure run from 22:33 and is not evidence of a green baseline; the green run is the logs/ file above.
   • Re-verify main SHA and mergeable before executing; this snapshot is 2026-10-21 ~23:00 +08.

   Repro commands

   git fetch origin codex/qwen-vision-27b
   for n in 122 123 127 143; do git fetch origin pull/$n/head:pr$n; done
   git cherry -v origin/main origin/codex/qwen-vision-27b
   git merge-tree --write-tree origin/main origin/codex/qwen-vision-27b
   git merge-tree --write-tree origin/main pr127
   git merge-tree --write-tree pr118base pr122     # pr118base = scratch merge of the branch into main
   git diff --stat origin/main...pr127
2. hsliuustc0106 commented on Oct 21, 2026

   

   hsliuustc0106

   on Oct 21, 2026

   ContributorAuthor

   More actions

   Addendum (2026-10-21 ~23:15 +08) - approvals, dependencies, re-land plan

   Approvals check. Currently approved in this stack: #127, #122, #123, #143 (4, not 3) - plus #147, which is approved but base=main/CONFLICTING and outside the stack. Bot does not approve; the count is only the read-only state.

   Cross-PR dependencies found during the read-only reviews (new):

   1. #93 and #131 both change benchmarks/bench.py, benchmarks/README.md, tests/benchmarks/test_score_wire.py, and neither is based on the other (git merge-base --is-ancestor = false). Merge bench: validate rounded Score and isolated zero-fit results #93 first, then rebase bench: preserve complete serving evidence and offline summaries #131; merging both as-is gives an add/add conflict on tests/benchmarks/test_score_wire.py.
   2. #148 and #151 both add entries to README.md, docs/supported-models.md, recipe/README.md. Either order, but the second must be rebased to re-apply its registry lines.
   3. The re-land of jemm: add native decision worker and pinned weight export #118 is now planned as a flat cherry-pick of the 9 deduped commits; commit list, 6-way conflict map, 14-file fidelity residual and acceptance criteria are in [Re-land] #118 jemm native worker: 9 deduped commits onto main, conflict map and fidelity check #154.

   Updated order for the current batch:
   #127 -> #118 re-land (#154) -> #122 -> #123 -> #143, then #93 -> #131, then #148 / #151 (second rebased).

   Read-only diff reviews posted (not approvals): #93, #131, #148, #151. Each comment states scope and what was not run (CUDA / accuracy / latency: 未跑 - no NVIDIA GPU, no model weights). Bots do not approve, retarget, push or merge; a human/Codex executes.
3. hsliuustc0106 commented on Oct 21, 2026

   

   hsliuustc0106

   on Oct 21, 2026

   ContributorAuthor

   More actions

   Bucket A clarification (2026-10-21 ~23:20 +08)

   To keep the board from reading "7 mergeable", split bucket A by approval:

   • Merge batch (approved): cua_s1: retain bounded vision cache across screenshot grids #127, jemm: load and validate checksum-covered calibration #122, docs(jemm): document CUDA Graph controls and validation limits #123, fix(cua-s1): match text reference and preserve merged compatibility #143 - order #127 -> #118 re-land (see #154) -> #122 -> #123 -> #143.
   • Review queue (no approval, NOT in the merge batch): bench: validate rounded Score and isolated zero-fit results #93, bench: preserve complete serving evidence and offline summaries #131, D3: serve images and videos through the Python reference worker #148, Vela: serve Vela 2.0 0.3B through a Python reference worker #151. Read-only review notes are already posted on each PR; approval must come from Codex or a human. Do not count these as "可合".
   • valen: export the trained stack to an unmerged safetensors artifact set #147 is approved but base=main / CONFLICTING, so it is outside this stack.
   • jemm: add native decision worker and pinned weight export #118 re-land decision changed to whole-branch merge PR (see [Re-land] #118 jemm native worker: 9 deduped commits onto main, conflict map and fidelity check #154 decision update); CI must be re-run on the new PR head, and cua_s1: retain bounded vision cache across screenshot grids #127's head 03b111ce needs a fresh CI run before it can be quoted.
4. twu3202 commented on Oct 21, 2026

   

   twu3202

   on Oct 21, 2026

   Contributor

   More actions

   #98, #99, #129 and #134: I'm refreshing them onto main 32bdd2a now that #155 is in. #98 goes first (ABI 9, including #155's fake test library). #99 then adds the early refusal of the LoRA/reference profiles in forward_fixed/forward_shared agreed there, with a CPU test, and #129 and #134 follow. The merge order stays #98 → #99 → #129 → #134; I'll post the new heads here once CI is green.
5. twu3202 commented on Oct 21, 2026

   

   twu3202

   on Oct 21, 2026

   Contributor

   More actions

   New heads on main 32bdd2a: #98 c2b1ab7, #99 866bed5, #129 d7418ce, #134 8b8882e. #99 adds the profile check in forward_fixed/forward_shared, and #129 adds it to forward_multimodal_readout. CI and the sm_89 reruns pass; the descriptions have the numbers. Merge order unchanged: #98 → #99 → #129 → #134.