Investigate Qwen3-4B RNGD/GPU pre-EOS drift after matched 8K prefill #256

Problem

Qwen3-4B live RNGD prefill to the paged GPU decoder diverges before EOS for an 8192-token prompt. A matched full-prefill GPU control moves the first divergence later but does not eliminate it for128output tokens. This must not be marked transport-correctness plus full model equivalence or used for an equivalent-output speedup claim.

Evidence

Checkpoint1cfa9a7208912126459214e8b04321603b3df60c; BF16; RNGD Torch/kernel SDK2026.4; GPU Torch2.13 on A10080GB; producer width8192; GPU4slots/context9216/cache-off. Live HTTP matrix completes24timed batches (inputs128/8192,outputs32/128,c1/4,3repeats), but original nativeGPU chunk256 parity fails for every8192-input condition at output index12, before EOS151645/151643. Short input passes.

Controlled single-model tests: nativeGPU chunk256 reproduces original reference, nativeGPU chunk8192 and retained RNGD KV through the same PDBatcher reproduce the live32-token sequence. The retained export has exactly the same input IDs. Independent nativeGPU chunk8192 HTTP control then completes24batches with its own output references, no RNGD outputs supplied. Comparing live results to that control:15/20nonstream token checks and45/60timed text responses match. The remaining8192input/output128cases diverge at index107 (108th token), before EOS, identically at c1/4. Original failures and all control outputs are retained.

Local artifacts: local201-qwen4b-matrix-a.json, local201-qwen4-diagnostic.json, local201-qwen4b-native-full.json, local201-qwen4b-full-prefill-audit.json; experiment2026-10-24-live-rngd-gpu-preflight.md. These are currently local research artifacts; publication pending.

Scope / hypothesis

Separate execution-shape numerical effects from remaining producer/decode precision differences with identical teacher-forced prefixes and exact KV readback. Compare retained versus live KV and logit margins near the107boundary; do not force a fixture token or relabel the failed original comparison as passed. Current evidence does not identify a faulty operator.

Done when

Reproducible diagnosis identifies the source, with any correction validated across all four-model fixtures and supported conditions; alternatively explicitly document a numerical compatibility boundary backed by broader evidence. Both original and matched-control failed results remain visible. No default performance claim without equivalent outputs.

Owner

Codex; maintainer review. Related#201,#202,#217,#248; analogous Gemma numerical investigation#233.