[gspark-etri]

Background

The four-model matrix lacked a Gemma RNGD producer. Qwen/Llama equations omit Gemma's offset RMS weights, four residual-path norms, GELU, scaled embeddings, and alternating local/global RoPE and attention. This adds a research producer preserving those semantics across all26layers.

Changes

• Explicit Gemma equations with independent HF layer/cache/causality and full tiny-model checks.
• A checkpoint-pinned RNGD runner and actual-device layer/layout probes.
• Fix compiler sample/runtime strides: transposed sample weights versus contiguous runtime weights caused silent wrong outputs. A matrix probe reproduces137% relativeRMS versus0% with matching layouts; the initial failed full-model run is retained.
• GPU handoff checker now validates absolute cache position and retained local/global state, including readback after cache insertion.
• Research tracking, exact commands and machine-readable failure/success evidence.

Verification

Baseline a754c26; implementation f271dcd; final evidence 90e4a7d. One isolated RNGD, Furiosa2026.4 fusion4/Mesh2,128GiB host-memory limit,4requested/16limited CPUs. GPU A10080GB. Gemma checkpoint5b11413a10db4e486ef16a20101fd028f8f2499c; source and checkpoint hashes in evidence223/summary.json. Embeddings/final norm/head run on CPU; transformer layers on RNGD.

• 128tokens:3finite/bit-exact repeats; warm prefill+export44.83–45.08ms. GPU native/imported32tokens match;52cache slots exact. KV/logit relativeRMS0.826%/0.825%/0.923%.
• 2048tokens:3finite/bit-exact repeats; warm1.894–1.901s. GPU native/imported32tokens match;52slots exact.22local layers retain511positions;4global layers retain2048, with absolute positions verified. Retained KV/logit relativeRMS5.06%/3.32%/1.54%.
• Existing KONI2K checker regression:32tokens match,64slots exact.
• CPU suite38passed,2failed: existing test_kv_cache_generation_matches_hf[1] and [2] DynamicCache/AOTAutograd failures, reproduced on parent in exp: enable partitioned KONI 2K/8K prefill and GPU/ATOM handoff #222. Scoped Ruff/diff checks pass.
• Sensitive scan passes for changed files; full-repository scan retains7 unchanged baseline findings from exp: enable partitioned KONI 2K/8K prefill and GPU/ATOM handoff #222.
• All experiment jobs terminal; owned RNGD allocation released.

Compatibility and limits

Experimental full-prefix producer; no production engine/SDK defaults changed. Fixed output counts ignore EOS and do not establish broad chat quality or numerical identity.8K and experimental Gemma ATOM hybrid-state import follow-up is now in #226, including correction of an invalid native global-attention baseline. Production HTTP/cache/cancellation, genuine compute/transfer overlap and performance optimization remain open. The2K latency requires profiling; no performance-parity claim.

Related Issue

Related to #223 and #201. Stacked on #222 (exp/221-koni-parts).

Generated with Codex; maintainer review and merge required.