Repository navigation
v0.0.5 — unit-level static partition (auto on matching GPUs), per-unit ledger, receiver checks - #6
Closed
alejandrogzi wants to merge 4 commits into
Closed
alejandrogzi wants to merge 4 commits into
alejandrogzi wants to merge 4 commits into
Conversation
…rds (experimental rounds 88/90/90b) Sync of experimental HEAD (alejandrogzi/hspZ @ 04d5985) into production. Sources are experimental HEAD with the production packaging re-applied (file headers, clap attribute, hspZ naming). No dependency changes. - round 90: HSPZ_UNIT_PARTITION splits the frozen plan's work units by count quotas (whole bins first, then contiguous query slices of as few bins as possible, one extra reference build per split bin); --time gains the per-unit ledger, and every run prints schedule lines; the receiver rejects duplicate/out-of-plan units and verifies completeness after joining the workers, reporting a worker's own error first. - round 90b: unset = auto — partition only with W>=2, one worker per device, and every used device the same class (equal SM count and L2, nominal clocks within 10%); the reason is printed in the schedule line. 1/0 force. - round 88 also lands here: v0.0.4 was synced from ce1de86 and cherry-picked b44cc04's window budget only, so -B 0 automatic layout (W>=2; W=1 is exactly the 500 Mbp plan), the ref-buckets inconclusive-block bound, and the --time gap/phase attribution were never in production. Round 90 builds on that code. - the round-89 find_hits overlap prototype (043bd30, reverted upstream) is absent. Verified: cargo oxide build --release; 138 tests; chr20 digest 70152625c1a75cb3 (3,638 HSPs); schedule: policy=whole-bin W=1 ... (auto: W=1).
…ver checks - version 0.0.5 in Cargo.toml and the hspZ entry of Cargo.lock only; features unchanged (default = simd-prelude, find-hits-warp, ref-loc-buckets) and no dependency changes (experimental's [dependencies] unchanged since b44cc04). - changelog 0.0.5: unit-level static partition with the measured seed-and-filter numbers (-11.9% on 2x RTX 4090, 1,400 -> 1,233 s; -5.1% on 4x RTX 4090, 718 -> 681 s at MAX_HITS=16,711,680; identical output), the auto rule (class match, 10% clocks, one worker per device), the per-unit ledger, and the receiver checks; the entry also records the -B 0 automatic layout, the inconclusive-block bound and the --time attribution that land with round 88.
- gpus.mdx: the runtime tuning knobs section is retitled (0.0.5) and gains the HSPZ_UNIT_PARTITION=auto|0|1 row (device-class guard, whole-bin at W=1, time-sliced workers and mixed classes; -11.9% on 2x RTX 4090 at MAX_HITS 16,711,680, -5.1% on 4x RTX 4090). "Both knobs" -> "Every knob". - docs/src/generated/cli-reference.txt regenerated from the 0.0.5 binary and byte-identical: no CLI text changed. - docs validate (9 journeys, 43 slides) and astro build (16 pages) pass.
Member
Author
|
Superseded: rounds 83-96 are collapsed into a single v0.0.4 release (one branch, one changelog entry, one PR). Nothing here is dropped — the replacement PR carries this branch's tree plus the on-disk reference index and multi-GPU batching. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #5 (v0.0.4). Merge and tag v0.0.4 first; this PR then retargets to master.
Syncs the tested experimental tree (alejandrogzi/hspZ
experimental@ 04d5985) into production with the production packaging kept (clap attribute, file headers). Sources are experimental HEAD's byte-for-byte except those headers.What changes
HSPZ_UNIT_PARTITION=0|1forces either policy. Output bytes are unchanged.--timeandschedule:lines on every run (policy, reason, per-worker visits).-B 0automatic layout (opt-in measurement tool; changes the frozen plan) and autotune/ledger diagnostics from round 88.Measurements (canonical hg38 × mm39,
MAX_HITS=16,711,680, ABBA, same host, frozen output digest68296b830fdef311in every arm)Exactness: 138 unit tests (incl. a brute-force partition oracle), chr20 pair and a 5×3 two-worker ZLUDA plan byte-identical under auto / forced 0 / forced 1, Kaggle 2× T4 identity (four arms, frozen digest), and the two 4090 campaigns above. Reviews: grok-4.6 and codex gpt-6-astra, both PORT WITH FIXES, all fixes included. Caveat stated in the changelog: the static-attribute guard cannot detect same-model cards with different sustained clocks (on a heterogeneous T4 pair the balance depended on device order).
Release check:
Cargo.tomlversion 0.0.5 (release.yml asserts tag == version).🤖 Generated with Claude Code