[gspark-etri]

Background

Prefix reuse changes greedy outputs in the mixed128/2048/8192 HTTP fixture. A homogeneous512/2048 cache matrix passed but did not exercise the original partial-prefix history.

Changes

Add four-model page64/page256 evidence and current-runtime cache-off controls, with raw artifact hashes, frozen revisions, reproduction commands and limitations. No engine or default-setting changes.

At chunk256, page64 reproduces concurrent output differences for Qwen3-0.6B, Qwen3-4B and KONI. Page256 has12/12 within-run matches for each selected model and3/3 serial reference matches to each model's current page256/cache-off control. Qwen0.6 reverse-order repetitions retain the difference. Gemma's current page64 within-run failure does not reproduce; its archived-reference difference is after EOS.

Verification

• Experimental runtime a3497e5; evidence branch based on main fd89265.
• One A10080GB PCIe, driver580.178.04; torch2.13.0+cu130, Transformers5.15.0; BF16,4slots, context9216, chunk256; exact model revisions in summary.json.
• Fifteen runs complete:10cache-on and5cache-off. All HTTP lifecycle checks pass. Independent saved-array audit counts180concurrent responses, with17mismatches confined to page64 cache-on arms. Cache counters verify reuse and GPU cleanup is empty.
• KONI current cold controls match across page sizes but differ from old128/2K records; historical drift is kept separate from the cache effect.
• Diagnostic source copies verified to change only import path and page-size parameter. Staged sensitive-data scan and git diff --check pass.

Additional partial-prefix controls

The same four models also run280/536-token seeds, which share one/two complete256-token pages with later2K prompts. Sixteen current cacheoff/cacheon runs pass HTTP lifecycle checks; independent saved-array audit gives192/192 concurrent matches and24/24 paired serial references equal. Cache hits are positive in every enabled arm, zero in every disabled arm; GPU cleanup is empty. See experiments/cache-history-212/seed-controls.md and seed-summary.json. This adds two prefix lengths, not eviction/host-reload coverage or general deterministic output guarantees.

Device eviction and host reload controls

Four frozen models, one execution slot,36device pages and two distinct8K prompts:12off/device/host arms completed on2026-10-25 and were independently audited on2026-10-30. All72response arrays match within-arm references; all48device/host responses match cacheoff and16/16A/B reference arrays agree. Each host arm records61offloaded/87reloaded blocks and22272host-hit tokens. Device-only revisits recompute; host revisits reuse7936tokens. Actual removal events, allocation returns and zero remaining live references are checked. A verifier negative control rejects missing reload evidence.

Includes portable diagnostic fixture, raw hashes and verifier in experiments/cache-history-212/pressure-controls.md. Published verifier passes all12raw runs and rejects4missing-reload controls; existing-output guard tested before model load. No host-tier eviction, multi-slot transfer, NPU transport or stable performance claim. All12children exited0; no contemporaneous global GPU-idle snapshot is claimed.

Full native GPU condition grid

Adds the 2026-10-30 matched page256/chunk256 cache-off/on grid for all four selected models: input128/512/2048/8192, output32/128, concurrency1/4/16, BF16 and four decode slots. Eight fresh server arms complete192conditions2752responses (64warmups,1344token checks,1344SSE checks) with zero mismatches against each model's fresh cache-off references. Cache-hit counters and reference hashes are independently checked. All children and coordinator exit0; a fresh GPU-idle snapshot is captured before the next experiment.

Runtime8f9f2a05178bcbb859c97dce2e5ee1bbc23eca3e, one A10080GB PCIe, driver580.178.04, torch2.13.0+cu130, Transformers5.15.0. At8K/128/c4, warm-prefix cache-off/on throughput is Qwen0.6 62.42/185.98, Qwen4 40.59/134.29, Gemma59.14/149.25 and KONI40.82/158.97tok/s. These are single ordered repeated-prefix samples, not cold-input or stable speedup claims.

Includes portable fixture, example config, full timing/hash summary, auditor and12selfchecks. Published fixture changes only require an explicit config and reject existing outputs before model load; both guards tested. Original unsupported-metadata preflight failure is documented separately. Prior full weight-hash identity evidence is distinguished from current metadata-hash/file-size rechecks; weights were not rehashed during this grid. Staged sensitive and whitespace checks pass. See experiments/cache-history-212/full-grid-controls.md.

Compatibility and limits

Page256/chunk256 is a scoped configuration mitigation candidate, not a general numerical fix. Coarser pages reduce reuse in this fixture. Single-burst timings do not establish a stable speedup. Other prompt histories, host-capacity eviction, concurrent reload/fault controls, numerical controls and semantic quality remain open; this does not certify RNGD/ATOM paths. The documented runtime is experimental, not current main.

Related Issue

Related to #212 and #201; neither is closed.

Prepared and verified with Codex assistance. Maintainer review/merge required.