[gspark-etri]

Background

Direct model runs do not establish HTTP behavior under concurrent arrivals, streaming cancellation/disconnect and prefix reuse. This adds real socket-level evidence for the four selected GPU checkpoints.

Changes

Add a harness using the existing serve_batched factory and owner loop, a direct teacher-forced cache-history diagnostic, and an actual HTTP logit trace. Preserve raw successful/failed runs, a compact summary, exact setup and research records. No serving engine behavior is changed.

Verification

Harness16f58f7 on engine prerequisite21f2c87; A10080GB PCIe, torch2.10.0+cu128 / Transformers5.1.0. Qwen3-0.6B, Qwen3-4B, KONI8B, Gemma3-1B;4 paged slots, context9216, inputs128/2048/8192,32 fixed output tokens, concurrency4/8.

All four cache-disabled BF16 configurations pass lifecycle/count checks;48/48 concurrent outputs match serial baselines. Cache-enabled BF16 also passes lifecycle checks when cancellation is compared at the same warm-cache state, but raw concurrent equality is31/48. The initial KONI cold-baseline cancellation assertion failed and is retained. Gemma's observed difference is after EOS.

Actual Qwen HTTP trace captures a BF16 tie18.375/18.375 that becomes18.5/18.25 with a different reused-prefix boundary. Actual HTTP FP32 controls on all four models pass lifecycle checks and48/48 concurrent equality. The direct diagnostic did not itself reproduce the HTTP argmax change; it is not used as substitute evidence.

Changed Python files pass Ruff, diff whitespace and sensitive-data checks. Hardware scripts enforce completion counts, cancellation drain, subsequent-request output, readiness,400 rejection and positive cache hits. All owned GPU processes ended.

Compatibility and limits

Stacked on PR210. This is GPU HTTP validation, not ATOM/RNGD HTTP or full-goal completion. Fixed-output fixtures ignore EOS and use strong synthetic prefix reuse. Single-burst timing is not a statistical capacity study. FP32 is a correctness control, not a proposed performance default. Broader semantic quality, NPU hybrid cache restoration and true compute/transfer overlap remain open.

Related Issue

Related to #211, #212 and #201. Generated with Codex; maintainer review and merge required.

[gspark-etri]

Landed on optimus-rm via #286. Closing; the branch is archived locally (experiment records and evidence included).