[gspark-etri]

Background

KONI-8B full-context RNGD prefill failed RMS lowering and fused-graph mapping at 2K tokens even with 128 GiB host memory (#219/#220). This experiment keeps all 32 layers and all input tokens while splitting the compiled stages. The same implementation now executes8K and hands off to GPU and ATOM.

Changes

• Compile reusable RMS row tiles, full-context attention and FFN separately.
• Return token-major KV explicitly at the compiled attention boundary; runtime permutations avoid the reproduced output-layout failure.
• Add independent HF layer/KV/causality tests and an actual-device attention regression.
• Extend the handoff checker to an exclusively owned static B1 ATOM single-block artifact, reusing existing import arithmetic. Record first divergent token logits.
• Include experiment records, research tracking and reviewed machine-readable evidence.

Verification

Baseline 6787628; implementation 720e8dc. RNGD / Furiosa 2026.4, fusion4/Mesh2, isolated allocation with128 GiB host-memory limit; GPU A10080GB; ATOM / Rebel0.11 / Optimum0.11.0.post1, B1/context9216/k1. Pinned KONI revision0172bf7e73977ae213613529f3d2d5400a33f678. Exact commands, dependency hashes and source hashes are in the experiment record/evidence.

• Full2K: 3 finite, bit-exact KV/logit repeats; warm prefill plus host export769.15–769.52ms. KV shape[32,8,2048,128]. This is not HTTP TTFT.
• GPU: native/imported32-token outputs match,64 slots read back exactly. KV/logit relative RMS3.15%/5.87%/2.27% remains.
• ATOM: both importers read back64 slots exactly and produce bit-exact decode logits. Outputs match GPU. Native ATOM diverges at token28: its0.125 winning margin becomes a tie in the imported path. No broad quality-equivalence claim.
• Expanded CPU suite:27passed,2failed. test_kv_cache_generation_matches_hf[1] and [2] fail on unregistered DynamicCache output in AOTAutograd; both reproduce on the unchanged parent.
• Scoped Ruff/diff checks pass. Sensitive scan passes for staged changes; full-repository scan has7 unchanged findings reproduced on parent.
• Full8K:3 finite/bit-exact repeats, warm3.317/2.934s including host export; timing dispersion remains. GPU and ATOM native/imported32-token outputs match,64slots readback each and ATOM importer logits exact. GPU relative RMS K3.30%/V5.85%/logits2.44%. Artifacts in commit a754c26. All owned experiment allocations released.

Compatibility and limits

Research stage-only engine; production engine/SDK defaults unchanged. Single-block ATOM validation does not enable production PD scheduling, prefix caching or concurrency. No true compute/transfer overlap, native-engine speed parity, or completion of the full four-model objective. Failed compile/runtime runs and numerical differences remain in the record.

Related Issue

Related to #221, #219 and #201. Stacked on #220 (exp/219-koni-long).

Generated with Codex; maintainer review and merge required.