experiment: partition KONI prefill compilation without shortening context #221

Problem

KONI2K full-prefill compilation fails RMS[2048,4096] lowering. Row-tiled RMS inside the fused graph then fails compiler mapping with memory explosion even at128GiB and one layer per graph (#219).

Hypothesis

Compile normalization, attention and FFN as separate modules. Invoke a512-row RMS module outside the compiled attention/FFN graphs, preserving full2048-token causal attention and all32layers. This removes the problematic fused layout search without dropping prefix context.

Scope

A research partitioned engine for the pinned KONI8B checkpoint; independent CPU reference tests of attention/FFN and actual RNGD compilation/inference. Compare exported KV and32-token GPU/standard ATOM continuation if compilation succeeds. Record all failures, resource limits and timings. No production default, SDK mutation, overlap or performance-parity claim.

Done when

Actual-device evidence establishes whether partitioning allows full2K model execution and handoff, or identifies the next concrete failing operator. A CPU pass alone is insufficient.

Owner

Codex, maintainer review. Related #201,#219,#220.