omnia-model bench example panics (assertion failed: index < dim) #28
Breadway
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
How it was found: while estimating tiny-model training time on hermes, I ran the documented command
cargo run --release -p omnia-model --example bench -- nano 32(and-- tiny 32) on branch docs/model-architecture at 21a130a.What happens: both sizes panic immediately in burn-ndarray 0.21.0 (
src/ops/base.rs:927:28,assertion failed: index < dim). No throughput number is printed.Confidence: the failure is certain (reproduced for nano and tiny). The cause is not diagnosed. My guess, about 60% confidence, is that the example builds its samples or batch options in a way that went stale after the recent observation changes (6-player support, OMNI-26, or the hygiene pass, OMNI-24), so a gather index exceeds a dimension. I did not read the example or capture a backtrace.
Why it matters: the bench example is the only isolated throughput check. docs/RESULTS.md section 4 throughput figures cannot be reproduced with it, and I fell back to timing
omnia diagnoseinstead.Suggested fix: run with RUST_BACKTRACE=1, find the gather or index op, and make the example build its batch the same way
omnia-trainingdoes. Consider a CI smoke run of the example so it cannot rot again.Fix ready: #30 (branch fix/omni-28-bench-index,
645e14a). Root cause was a hardcoded rel vocab of 40 in the example's diagnostic embedding lookups; rel ids now reach 41, so any batch with registers panicked. Both nano 32 (227 samples/s) and tiny 32 (82 samples/s) print throughput; full release test suite passes.Independent review PASS on PR #30 head
645e14a(base1913272), merged to main as0710de2. Root cause confirmed by measurement: real nano batch carries max rel_ss=41, rel_as=40 (rel::VOCAB=42) while the example built dim-40 tables, hence the index < dim panic. Fix uses rel::VOCAB in all three diagnostic lookups plus trainer-canonical BatchOpts. Verified: bench nano 32 (87/94/78/84 samples/s) and tiny 32 (31/34/34/34) print throughput with no panic; bench_path_smoke passes; cargo test --workspace --release all green (training 10/10, OMNI-29 seeded test passed first run); clippy clean for omnia-model with only 4 pre-existing omnia-dsl lints workspace-wide; fmt drift pre-existing repo-wide, new test file clean. Review detail on PR: #30 (comment). Card left for coordinator verification, not moved to done.Verified on integrated main
0710de2(merge of PR30, parents 1913272+645e14a, base unchanged so tree equals reviewed head): reviewer ran nano/tiny bench runs plus full release workspace suite at the head; no CI workflows in repo.