Investigate standalone native RNGD batch-dependent KONI output #274

Problem

Standalone native RNGD serving of KONI-8B returns batch-dependent greedy token sequences. This occurs without an external PD transport, so it must be separated from the research RNGD→GPU drift tracked in #261.

Evidence

Installed furiosa-llm/compiler 2026.4.0 (b1baffa75 / ae27d44cd8), KONI snapshot 0172bf7e73977ae213613529f3d2d5400a33f678, one RNGD, TP8, max context9216, prefix cache off. Native FXB: 105/105 successful entries, SHA256 255e0929e98d85cc244846a6e6a1f76dc5d9d78167bc13c4e5e59eb4c6555e41.

Full matrix: input128/512/2048/8192 × output32/128 × client concurrency1/4/16. All24 conditions completed,344 responses including8warmups,168 token checks and168 timed SSE responses. Against same-server single-request references: c1 zero mismatches in8 token and8 text checks; c4 13/32 token and13/32 text mismatches; c16 50/128 token and50/128 text mismatches. Temperature0, ignore_eos=True for fixed-length performance controls. Raw result SHA256 fe90c32fa91dee3987cac7fcb0253daee43f8f7e181267dd89fc81b9dd6086fb.

Separate sequential frozen synthetic quality:31/36 strict answers,34/36 formats,all36 stopped,zero request errors,explicit EOS128001/128009. This does not establish concurrent quality. Native JSON request-header defect in the measurement client was reproduced and fixed before this matrix; failed initial run is retained.

Hypothesis and scope

Profile-guided decode batching changes batch composition and may change the selected execution shape or arithmetic. This is a hypothesis, not an identified engine defect. Compare default vs --no-profile-guided-decode-batching on the same physical RNGD in ABBA order, with the same model/FXB/input IDs and no other scheduling changes. Preserve each arm's own c1 references, compare cross-arm references separately, and retain every mismatching output. Start with representative short128 and long8192 inputs/output128/c1/4/16, then expand only if an effect is observed. Do not silently serialize requests or redefine exact parity to obtain a pass.

Done when

Reproduce the batch dependence; establish whether this public setting changes it; publish exact commands, raw hashes, same-device evidence, repeated timing, full token comparisons and limitations. If a candidate improves correctness, validate held-out concurrent quality and the broader matrix before recommending it. Any production integration is a separate owner-maintained change under CONTRIBUTING §8; no native binary modification is planned.

Owner

Codex for experiment/reproduction; aicomp maintainer for any integration decision. Related #201, #261, #263. No claim of equivalent-output speedup or a completed four-model validation.