[gspark-etri]

Background

ATOM B1/B4 serving tests exposed lane-dependent Qwen3-4B outputs, an invalid Gemma causal export, and low decode occupancy on RNGD→ATOM. This evidence-only draft records the failures, isolated corrections, performance tradeoffs and reproducible audits. It does not change production serving defaults.

Changes

• Re-audited the four requested models across GPU B4 and ATOM B1/B4, with native, whole-KV and layer-streamed routes:33 within-route HTTP passes,1 original Qwen3-4B B4 failure and2 uncollected dependent PD routes out of36 combinations. Raw file hashes and cross-route reference differences are retained. Gemma B4 stream evidence uses the instrumented causal-corrected run and reaches at most3 active lanes.
• Gemma causal-export reproduction and corrected-artifact evidence. Short-input whole-transfer ABBA finds1.89–2.01× throughput with producer width1024 versus8192; four-lane short-output occupancy remains unproven.
• Qwen3-0.6B whole/stream long-input ABBA:528 exact responses,21–28% throughput improvement at2K,20.8% regression at8K/32, and near-equal throughput at8K/128. No global streaming default is justified.
• Four-model cancellation/reuse and producer-death evidence. A stricter dense-model follow-up triggers after a real victim KV import and preserves9216 survivor-token and48 replacement-token comparisons. Gemma hybrid B4 intentionally defers partial imports.
• Qwen3-4B same-card SDK reproduction and attention-output diagnostics. The all-core diagnostic graph passes264 native HTTP responses against its own references and sequential/concurrent frozen quality gates. It remains an isolated diagnostic export, not a production fix.

Verification

The published auditors recompute raw token/text comparisons and timing summaries, check device binding and compiled geometry, and preserve failed or missing cases. The four-model inventory rescored576 frozen quality responses without postprocessing. Strict answers out of36, in native-GPU/RNGD-GPU/native-ATOM/RNGD-ATOM order: Qwen3-0.6B32/34/33/33, Qwen3-4B36/36/36/36, KONI31/31/31/31, Gemma0/0/0/0. These scores must not be replaced with post-hoc content extraction.

The all-core Qwen3-4B sequential and concurrent quality runs each pass36/36 strict answers and formats, match36/36 historical B1 token arrays, and stop at EOS. The concurrent run has three12-request cohorts and64 actual four-lane decode steps. Its real localhost concurrency self-check passes. Detailed source hashes, exact commands, failed attempts, artifact hashes and independent audits are in the experiment document and evidence270 directory.

Changed-file sensitive scans and whitespace checks pass. The full-tree sensitive scan retains7 findings in6 unchanged baseline files, verified against HEAD. No full repository suite is claimed for this evidence-only update. Previously recorded targeted CPU/fault validation is documented with its separate hardware limits.

Compatibility and limits

The serving runtime depends on the PR266 experimental stack; this branch is not a standalone implementation on main. Production integration belongs to the engine owner. The Gemma export fix uses isolated artifacts and requires issue271/upstream review. Maintainer review and merge are required.

Within-route parity is not equivalence to native decoding. Qwen3-4B all-core references match historical B1 in4/8 performance conditions and original B4 in2/8. The earlier single-run8K/128 throughput16.438tok/s versus historical18.242tok/s does not show a gain; cards differed. Same-card ABBA is now complete: all-core mean throughput62.894 versus75.976 tok/s at512/128 (-17.22%) and16.489 versus18.253 tok/s at8K/128 (-9.66%). Original arms each fail12token+12text checks; both candidate arms pass66responses. Four identical bindings, raw audits and cleanup pass. Two timed batches per variant/condition, not a confidence interval. Different cross-variant outputs preclude equivalent-output performance claims. Candidate RNGD→ATOM whole/stream compatibility now passes528responses and8/8 pinned-reference comparisons with actual four-lane calls251/504 and cleanup0. Historical native references remain4/8. Exploratory single-pair stream throughput is2.55–29.24% higher than whole; this is not a statistical or resource-equivalent native speedup. The separate72answer PD quality gate now passes:36/36 strict answers and formats per route, all EOS, full tokens equal historical PD and native candidate. PD quality occupancy reaches only2whole/3stream lanes; full-four-lane PD semantic quality is not established. The original B4 numerical failure remains recorded.

The frozen structured-answer set is not general model quality. Gemma strict JSON failures, native/PD numerical differences, short-output under-occupancy and workload-specific streaming regressions remain explicit. Host timelines do not prove physical DMA overlap. Large raw logs and site configuration remain outside Git; local research-hub registration is recorded, remote archive synchronization is unverified. No issue is closed or default promoted.

Related Issue

Refs #270, #271, #225, #265, #201; RBLN-SW/optimum-rbln#827.

AI assistance: generated and verified with Codex; maintainer review and merge remain required.

[gspark-etri]

Closing to keep the remote at main and optimus-rm. This is example/experiment work with no library change to land; the branch is archived locally and can be restored if any part is needed on optimus-rm. The tracking Issue stays open.