[gspark-etri]

Background

The GPU HTTP matrix did not establish ATOM request lifecycle or cache behavior. Extend the same production-factory harness to the four selected models: Qwen3-0.6B, Qwen3-4B, KONI-8B and Gemma3-1B.

Changes

Add an isolated ATOM backend option, artifact KV-block configuration and explicit host-cache checks. Publish measured results and limitations. Stacked on #213; engine prerequisite #210. The two fixed evidence files include the independent HTTP disconnect fix in #215.

Verification

• ATOM cache-off: all four models pass concurrency4/8, explicit cancellation, disconnect/drain/reuse, invalid-input and readiness checks; 48/48 concurrent outputs match each model's serial baseline.
• Both Qwen models: device-cache and host-cache cases each pass12/12 concurrent equality. Device-only prefix hits19584; host runs prefix hits33792 and host hits16128 (0.6B)/8064 (4B).
• Inputs128/2048/8192, output32, one B1 lane, context9216. Clients are queued, not hardware batching. Synthetic fixed-output fixtures, not semantic-quality tests.
• Engine21f2c87, harness16f58f7 extended by3ea0493; ATOM SDK0.11.0, optimum-rbln0.11.0.post1, torch2.11.0+cpu. Full commands/revisions/method in opt_prime/examples/pd/http_matrix/experiments/2026-10-23-atom-http211.md and JSON in evidence/.
• Ruff, Python compile, diff and sensitive-data checks pass. Hardware results exercise the new branches.

Compatibility and limits

GPU remains the default. ATOM precision comes from its artifact. Gemma uses corrected causal-prefill graphs and #210; hybrid prefix/host/PD restoration remains unsupported. KONI uses one KV block because the three-block graph does not compile. No general throughput claim from single bursts, no complete RNGD/heterogeneous validation, and no real compute/transfer overlap claim.

Related Issue

Related to #211, #201, #207 and #208. Full matrix remains ongoing.

AI-assisted implementation and verification with Codex; maintainer review and merge required.