# Full SigLIP BS3 attempt, 2026-09-13

Current compiler main 2d4e69d, 27 layers, capacity baseline, fused QKV, FP8 scale -4, B1024, unprofiled, eight optimization steps requested. FP32 reference and two resident inferences configured, but execution was never reached.

Default attempt: rejected after 40.743 seconds of plan_package. Geometry-derived per-tile fragments 16742 exceed default 16384. Mid coarse effective peak 550008 bytes; estimated row bytes 148328 (not physical placement).

One diagnostic retry with --exchange-transfer-limit-per-tile 17408: scheduled baseline in 117.559 seconds of plan_package, then rejected because maximum exchange table storage 99104 bytes exceeds per-tile 81920-byte budget. No tensor placement or local optimization reached; no successful executable/hardware result. Production defaults unchanged.

Logs: run.log, relaxed.log. Mid memory estimates: memory/ and relaxed/memory/. This is a failure of the generated baseline, not proof that every BS3 layout is infeasible.