Tune small model expert placement

2026-08-25 00:31:05 -07:00
parent af6ffeeb64
commit 7a17984dbc
5 changed files with 16 additions and 16 deletions
@@ -20,9 +20,9 @@ TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
    grep -q '^PARALLEL=3$' "$ENV_FILE"
 }

-@test "frank-default-small keeps final four expert blocks on CUDA" {
+@test "frank-default-small keeps final five expert blocks on CUDA" {
    grep -q '^OVERRIDE_EXPERT=' "$ENV_FILE"
-    grep -Fq '35|36)\.ffn_' "$ENV_FILE"
+    grep -Fq '34|35)\.ffn_' "$ENV_FILE"
    grep -q '^OVERRIDE_BUFT=CPU$' "$ENV_FILE"
    grep -q '^N_GPU_LAYERS=all$' "$ENV_FILE"
 }