14 days ago
-
Add auto arena mode (80% of available memory, capped at all experts); patch 369ef971
-
Hot/cold expert analysis: skew, profile pinning hit rates; fix hit counting and default pin fraction
-
Upload v2 memory tiers (default and forced row-parallel) with path counters
-
Add v2 row-parallel for gated Flint experts; label all data v1/v2; fix swiglu_limit pass-through
15 days ago
-
Verify EPQ against native llama.cpp; fix swiglu clamp and dump hook; re-run memory tiers
-
Add Flint expert paging port and memory benchmark
