[gspark-etri]

Background

The first live Gemma3-1B RNGD→GPU request crashed: the subprocess producer returned all26 model layers in one KV tensor, while the GPU receiver tried to copy it into its22-layer sliding-attention group. Existing grouped Torch producers did not exercise this contract.

Changes

PagedKV.write_span maps dense model-layer KV into receiver groups, preserves existing single-group/recurrent and grouped inputs, and validates group counts and shapes before copying. The subprocess wire format is unchanged. Added nonzero-offset layer-mapping and malformed-input regressions plus dense Gemma PD outputs compared against independent HF greedy beyond the sliding window.

Verification

Code32ce5d5, base273952a (#250); actual failed runtime3e16c33, BF16 Gemma3-1B, RNGD SDK2026.4, GPU Torch2.13.0+cu130, A10080GB, width8192,4slots/context9216/page64/cacheoff. Original failure retained; no completed hardware performance claim.

• RED:3failures reproduced mapping, silent missing group and partial malformed-group write.
• Focused:31passed,29.86s (including existing recurrent/hybrid/disaggregated tests).
• Full CPU:505passed,34skipped,253subtests passed,151.69s; one existing flex-attention warning.
• Commands/results in opt_prime/examples/pd/http_matrix/experiments/2026-10-24-dense-hybrid258.md; evidence258 contains original failure and test logs (trailing whitespace normalized).
• Scoped sensitive scan and diff check clean. Corrected live Gemma32ce5d5 completed24timed batches;20/20nonstream token checks and60/60timed response texts equal unchanged nativeGPU references.84remote prefills,0local,0request errors. Inputs128/8192,outputs32/128,c1/4,3repeats. At8192/output128/c4 median55.019tok/s; no speedup comparison claimed. Raw original failure and corrected artifacts both retained.

Compatibility and limits

Dense spans require uniform head geometry; grouped tensors still support heterogeneous geometry. No default routing, cache policy or transport changes. The recorded hardware matrix now passes. This does not establish arbitrary-prompt quality or all numerical configurations; draft retained for maintainer review. Separate Qwen/KONI/Gemma numerical investigations remain open.

Related Issue

Refs #258, #201. Stacked on #250; maintainer reviews and merges.

[gspark-etri]

Landed on optimus-rm via #283. Closing; the branch is archived locally (experiment records and evidence included).