[gspark-etri]

Background

Offline RNGD KV export checks did not prove that requests could trigger actual prefill and continue through the ATOM HTTP scheduler. The existing worker was Qwen-specific; Gemma was a one-shot sweep and KONI additionally needed a single-block scheduler path.

Changes

• Reuse callable Gemma computation in both its sweep and the live worker; add the validated partitioned KONI adapter and an explicit experimental long-Qwen factory.
• Keep the existing SubprocessPrefillWorker shared-memory/JSON protocol. Factories are trusted local CLI code, never request data. Infer Llama head_dim only when absent; preserve explicit Qwen/Gemma geometry.
• Add actual worker and HTTP validators using input/EOS fixtures and native token baselines. The HTTP validator never loads saved KV and covers serial, four queued clients, cancellation/disconnection and reuse.

Verification

• Implementation3855050/74d0d3b; long evidence04dab7c; core scheduler dependencies feat: support fresh hybrid KV imports in ATOM PD scheduler #232 (1e54786) and feat: support cache-off ATOM PD with one KV block #237 (df02223), SIGPIPE-safe HTTP fix: avoid native SIGPIPE handlers on HTTP disconnect #215 (9e442cd), corrected Gemma artifact experiment: isolate Gemma native global-prefill causality discrepancy #225/upstream#827.
• Four models at short, medium and long lengths on real RNGD->ATOM:14normal/concurrent/reuse responses per model/condition match native tokens through EOS,168total;192remote prefills including cancellation requests,0local fallbacks. Final lane ownership clean. Gemma/Qwen B1/context9216/k3; KONI uses its existing B1/context9216/k1 artifact. Prompt lengths23–143,1526–1643 and6026–6428; compiled widths128/512/2048/8192. Concurrency queues on B1, not hardware batch4.
• Gemma callable refactor also matches4/4 original full K/V/logit exports; actual worker5/5 K/V/token matches plus oversized-request recovery.
• CPU adapter/geometry/layer/single-block-helper suite28passed plus5subtests; scoped Ruff/diff/sensitive checks pass. Separate scheduler suite in#237:43passed. Failed initial geometry/factory/PD-capacity checks are retained in experiment records.
• SDKs: producer furiosa-torch2026.4.0, torch2.10.0+cpu, transformers5.19.0; decoder optimum-rbln0.11.0.post1, rebel-compiler0.11.0, torch2.11.0+cpu, transformers5.8.1. Pinned model revisions, exact commands and JSON in serving_matrix/experiments and evidence234.

Compatibility and limits

Draft, part of#234. Stacked on#228 and includes the existing#200/#199 worker dependency. Shared SDK installations and communication libraries are unchanged; candidate core/HTTP modules are loaded process-locally for hardware tests. Full repository sensitive scan retains seven unchanged baseline documentation findings; changed files are clean.

Short/medium/long fresh PD is demonstrated for all four models. Prefill-time hardware cancellation/worker-failure recovery, cache restoration, hardware batch>1, cross-host live RNGD-to-GPU, actual transfer/compute overlap and controlled latency/throughput remain open. GPU-to-GPU live HTTP and numerical controls are recorded separately under#240/#241. Recorded latencies are observations, not speedup evidence. Native-output parity does not certify answer quality; Gemma numerical mismatch#233 remains unresolved.

Related Issue

Part of#234; related#199/#201/#227/#231/#233/#236. No overall-goal completion claim.