[gspark-etri]

Background

The existing research RNGD prefill implements Qwen2 semantics. Qwen3-0.6B needs Q/K normalization and a query projection width2048 independent of hidden_size1024, so enabling its model name in that graph would be incorrect.

Changes

Add a separate experimental denseQwen3 graph, strict SDK-independent model checks, finite prefill/ATOM runners, and measured validation records. Existing Qwen2/model registration and communication ABI are unchanged. Host-mediated KV import reuses the original and portable importers from PR195.

Verification

• CPU contract tests:2passed; unsupported model/quantization/RoPE/sliding/bias/head configurations rejected.
• Ruff format/check, py_compile, git diff --check.
• RNGD2026.4,fusion4/Mesh2,width128; ATOM rebel0.11.0,optimum-rbln0.11.0.post1,torch2.11.0+cpu,B1/context512/3KV blocks.
• Qwen3-0.6B,three natural prompts,thinking disabled,fixed16output tokens:3/3first tokens matchCPU/ATOM;3/3PD sequences match nativeATOM;3/3full decode-logit sequences exact between importers;336/336KV slot readbacks exact.
• Relative RMS versus CPU FP32:K0.470–0.491%,V1.757–1.911%,last logits1.164–1.604%. First RNGD prefill1.321s,subsequent0.723/0.714s including hostKV/logits: no native-performance parity claim.
• Initial compile rejectedBF16 epsilon/exponent inputs; corrected toFP32 as required by rms_dsl. Failure is retained in the research record.

Compatibility and limits

Experimental finite runner, not production serving registration. Requires optimus-rm/rngd_pd and PR195's portable importer/validated container launcher onPYTHONPATH. Only0.6B,width128 is hardware-validated. NoDMA/source-overlap claim; export follows complete prefill. Longer inputs,concurrency,cancellation,recovery and other sizes remain unvalidated. The shared incorrectIncheon answer demonstrates that token parity is not factual-quality validation. Fixed-length checks continue pastEOS.

README contains reproduction commands; validation.json contains exact token and error measurements. Existing shared installations/artifacts were not modified; dedicatedPod removed.

Related Issue

Related #197, #194, #195.