ahrefs/ocannl

OCANNL: OCaml Compiles Algorithms for Neural Networks Learning

Stars: 116
Forks: 8
Open beginner issues: 0
Indexed issues: 124

Dominant language: OCaml
License: BSD-2-Clause
Last GitHub push: Sep 18, 2026
Latest indexed: Sep 20, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed
PR merge metrics: No merged PRs in 30d

Beginner friendly Everything

Open All

124 open issues indexed

Timing-cache provenance omits GPU drivers, selected CPU runtimes, and HIP compiler headers Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#1026 · Opened Sep 20, 2026
tools/sweep.sh: a staging-side check that the lab lock contract still matches wake-lab.sh, right after the routine's fetch Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ahrefs/ocannl#1025 · Opened Sep 19, 2026
Ll_test.cycle's modulus sets the value set and the blind-axis property at once, so fixing a blind fixture cannot preserve its values Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#1024 · Opened Sep 18, 2026
The errexit-exempt family has a scan for `! cmd` but not for a bare `[ A ] && [ B ]`, and one such assertion was silent in exactly the case it existed for Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ahrefs/ocannl#1023 · Opened Sep 18, 2026
Autotune.report carries no source digest, so the placement store's outcome guard is recomputed and pinned only indirectly Open

Difficulty 4/5 3-5 days Newbie friendliness 62/100

ahrefs/ocannl#1022 · Opened Sep 18, 2026
placement_store.ml replays only in-process: the decision problem's identity needs a cross-process leg Open

Difficulty 3/5 1-2 days Newbie friendliness 76/100

ahrefs/ocannl#1021 · Opened Sep 18, 2026
A knob to bypass the placement-decision store while replaying the schedule cache: a warm-cache run can no longer measure the placement A/B Open

Difficulty 4/5 3-5 days Newbie friendliness 72/100

ahrefs/ocannl#1020 · Opened Sep 18, 2026
Operand fixtures that hand-roll the flat index bypass Ll_test.cycle's blind-axis guard, and one is blind today Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

ahrefs/ocannl#1018 · 2 comments · Opened Sep 18, 2026
flip_candidates lists a node's readings as independent records; consumers should receive one group per node Open

Difficulty 5/5 Over a week Newbie friendliness 42/100

ahrefs/ocannl#1017 · Opened Sep 18, 2026
Affine.access carries neither scalar gatedness nor local effects, so every between-two-points hazard question re-walks the raw code Open

Difficulty 5/5 Over a week Newbie friendliness 48/100

ahrefs/ocannl#1016 · 1 comment · Opened Sep 18, 2026
Non_virtual exit codes are described in four places with no owner: one stale reachability claim had six copies, and the phase table accepts entries no row exercises Open

Difficulty 5/5 Over a week Newbie friendliness 38/100

ahrefs/ocannl#1015 · Opened Sep 17, 2026
Windows CI: the _opam cache restore fails on cygwin symlinks, so every job re-extracts 893 MB and then re-saves it Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ahrefs/ocannl#1014 · 1 comment · Opened Sep 17, 2026
README's retained-compatibility API list is an unchecked restatement Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

ahrefs/ocannl#1013 · Opened Sep 17, 2026
Conventions note: the one-owner rule is stated for code and tests but not for prose, and a doc entry restated the provenance representation stale twice in a row Open

Difficulty 2/5 1-3 hours Newbie friendliness 86/100

ahrefs/ocannl#1012 · 2 comments · Opened Sep 17, 2026
Flip-candidate pricing re-implements the inliner's binding rules; price from a synthetic instantiation through inline_computation and the shared post-virtualization pipeline Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#1011 · 1 comment · Opened Sep 17, 2026
lib/ has no interfaces, so every removal from the user-facing half is arguable in both directions Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#1010 · Opened Sep 17, 2026
dead_export_scan sees neither type declarations nor lib/ Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

ahrefs/ocannl#1009 · 1 comment · Opened Sep 17, 2026
The pytorch and tinygrad runners duplicate the measurement loop, so one protocol defect lands in two files Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ahrefs/ocannl#1008 · Opened Sep 17, 2026
Benchmark report columns: peak device memory, %-of-peak for the dominant kernel, kernel-sum vs step (split from #720 leg 5) Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

ahrefs/ocannl#1006 · 3 comments · Opened Sep 17, 2026
A dune action whose run is designed to fail must inherit neither stream, and nothing enforces it Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ahrefs/ocannl#1005 · Opened Sep 16, 2026
tools/test-run.sh takes the fleet's run-time correctness slot itself Open

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

ahrefs/ocannl#1004 · Opened Sep 16, 2026
Block-tiled online attention: open the key axis of the scan to blocking so the two contractions run on tensor cores Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

ahrefs/ocannl#1003 · Opened Sep 16, 2026
Fused attention backward: recompute the probabilities from the saved (max, normalizer) and drop the [seq, seq] gradient buffers Open

Difficulty 5/5 Over a week Newbie friendliness 28/100

ahrefs/ocannl#1002 · Opened Sep 16, 2026
Design question: the compilation pipeline cannot loop, and the placement lattice is what makes the phase order rigid Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

ahrefs/ocannl#1000 · Opened Sep 16, 2026
simplify_llc mixes exact canonicalization with ungated rounding-changing algebra (FMA contraction, constant reassociation); the profiles cannot state what they deliver Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#998 · 3 comments · Opened Sep 16, 2026
A gated claim must report through p, not Verdict.pass_fail, or its skip breaks the golden — and nothing enforces it Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

ahrefs/ocannl#997 · Opened Sep 16, 2026
A claim bypassed by its contention hatch leaves no record, so a unit that habitually bypasses it reads as covered Open

Difficulty 4/5 3-5 days Newbie friendliness 47/100

ahrefs/ocannl#996 · Opened Sep 16, 2026
tools/test-run.sh: record the checkout's HEAD and dirty state in the run directory at launch Open

Difficulty 3/5 1-2 days Newbie friendliness 74/100

ahrefs/ocannl#992 · 1 comment · Opened Sep 15, 2026
Conventions note: the controller's clock is not the box's, for anything that correlates a local timestamp with a remote artifact Open

Difficulty 1/5 1-3 hours Newbie friendliness 85/100

ahrefs/ocannl#991 · Opened Sep 15, 2026
tools/sweep.sh is ~1900 lines; extract the unit-state cursor and the fingerprint logic the way box-jobs.sh and dxg-window.sh were Open

Difficulty 4/5 3-5 days Newbie friendliness 64/100

ahrefs/ocannl#990 · Opened Sep 15, 2026
The sweep harness cannot reach a remote unit end to end: everything remote is pinned at seams or with synthetic sidecars Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ahrefs/ocannl#989 · 1 comment · Opened Sep 15, 2026
CI's shell-harness tier has no single local command, so a PR ships a red the author could not have run Open

Difficulty 3/5 1-2 days Newbie friendliness 72/100

ahrefs/ocannl#988 · 1 comment · Opened Sep 15, 2026
Measure what dune width rog-nv's discrete GPU tolerates behind the WSL2 dxg bridge Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ahrefs/ocannl#987 · Opened Sep 15, 2026
tools/remote-verify.sh is the next uncapped entry point into a GPU suite on a dxg host, and its header hides that --test requires --backend Open

Difficulty 3/5 1-2 days Newbie friendliness 78/100

ahrefs/ocannl#986 · Opened Sep 15, 2026
float_to_half_emulated has no exhaustive check, unlike the fp8 codecs Open

Difficulty 3/5 1-2 days Newbie friendliness 68/100

ahrefs/ocannl#985 · Opened Sep 15, 2026
sweep dxg window: the dmesg fallback's wall-clock bounds, and whole-second window edges (residuals from gh-ocannl-979) Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

ahrefs/ocannl#984 · Opened Sep 15, 2026
env_var_deps misses a dependency-expanded runner under chdir and accepts a missing ambient dependency Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ahrefs/ocannl#982 · Opened Sep 14, 2026
Delete test-run.sh's transitional pre-606 legacy-lock handling once no such run can be in flight Open

Difficulty 3/5 1-2 days Newbie friendliness 76/100

ahrefs/ocannl#966 · Opened Sep 10, 2026
Cross-statement thread conflicts on a pinned device cell are unjudged, and the tensorized zero→lane-0 store-back relies on a threadgroup-only barrier ordering device memory Open

Difficulty 5/5 Over a week Newbie friendliness 32/100

ahrefs/ocannl#963 · 3 comments · Opened Sep 9, 2026
Loop-carried recurrences have no high-level surface: cumulative ops and top-k as primitives, or a first-class scan in Assignments Open

Difficulty 5/5 Over a week Newbie friendliness 38/100

ahrefs/ocannl#952 · Opened Sep 8, 2026
Census each seeded register-tile geometry's accumulator loop: gh-614's spill should be reported by the -march matrix, not found by an asm read Open

Difficulty 4/5 3-5 days Newbie friendliness 63/100

ahrefs/ocannl#948 · 1 comment · Opened Sep 7, 2026
Register_tile.default ranks by a fitted constant while the tuner can now time the geometry: derive the peel weight from Ir.Cost_model and seed rm Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ahrefs/ocannl#947 · 2 comments · Opened Sep 7, 2026
Editorial changelog passes have no inventory of the window's public interface changes, so reviewers re-derive it Open

Difficulty 4/5 3-5 days Newbie friendliness 70/100

ahrefs/ocannl#946 · 1 comment · Opened Sep 7, 2026
A launch parameter never assigned runs at index 0, and nothing distinguishes that from a caller who meant zero Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ahrefs/ocannl#940 · Opened Sep 5, 2026
The conv over-cap control pins that a seed is refused, not which dimension carries it or at what extent Open

Difficulty 3/5 1-2 days Newbie friendliness 72/100

ahrefs/ocannl#939 · Opened Sep 5, 2026
ocannl_config.reference should carry the macOS dispatch_apply trap and its cc_parallel_grid=none escape hatch Open

Difficulty 1/5 Under an hour Newbie friendliness 88/100

ahrefs/ocannl#934 · Opened Sep 4, 2026
Agent-note citation policy leaves hashless PR/issue labels unenforced Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#932 · Opened Sep 4, 2026
The symbolic-extent test surface is forward-only: no gh-490 backward path is covered Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ahrefs/ocannl#929 · Opened Sep 4, 2026
lhs cells past a bound symbolic extent have no stated contract Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ahrefs/ocannl#928 · 1 comment · Opened Sep 4, 2026
Outside_sweep verdicts have no machine-checked coverage owner Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ahrefs/ocannl#926 · Opened Sep 4, 2026

A short digest of beginner-friendly GitHub issues.