@@ -10,7 +10,8 @@ mkdir -p "$LOG_DIR"
1010
1111CACHE_HOST=" $HOME /.cache/contextbench_repos"
1212HF_HOME_HOST=" $HOME /.cache/huggingface"
13- IMAGE=" treemapper-bench:latest"
13+ IMAGE=" ${IMAGE:- treemapper-bench: latest} "
14+ USE_BAKED_CACHE=" ${USE_BAKED_CACHE:- 1} "
1415
1516MIN_OK_THRESHOLD=550
1617LIMIT=9999
@@ -41,6 +42,7 @@ declare -a CONFIGS=(
4142
4243log () {
4344 printf ' [%s] %s\n' " $( date -u +%FT%TZ) " " $* " | tee -a " $SWEEP_LOG "
45+ return 0
4446}
4547
4648run_one_config () {
@@ -53,6 +55,10 @@ run_one_config() {
5355 local t0
5456 t0=$( date -u +%s)
5557 log " RUN start mode=$mode budget=$budget workers=$workers batch=$batch "
58+ local cache_mount=()
59+ if [[ " $USE_BAKED_CACHE " != " 1" ]]; then
60+ cache_mount=(-v " $CACHE_HOST :/cache/contextbench_repos" )
61+ fi
5662 docker run --rm \
5763 --cpus=" $CPUS " \
5864 --memory=" $MEM " \
@@ -62,7 +68,7 @@ run_one_config() {
6268 -e PYTHONUNBUFFERED=1 \
6369 -e HF_HOME=/cache/huggingface \
6470 -e HF_DATASETS_CACHE=/cache/huggingface/datasets \
65- -v " $CACHE_HOST :/cache/contextbench_repos " \
71+ " ${cache_mount[@]} " \
6672 -v " $HF_HOME_HOST :/cache/huggingface" \
6773 -v " $REPO_ROOT /results:/app/results" \
6874 -v " $REPO_ROOT /benchmarks:/app/benchmarks:ro" \
@@ -81,9 +87,9 @@ count_ok() {
8187 local mode=" $1 "
8288 local budget=" $2 "
8389 local f=" $REPO_ROOT /results/cb_${mode} _n${LIMIT} _b${budget} .json"
84- if [ ! -f " $f " ]; then
90+ if [[ ! -f " $f " ] ]; then
8591 echo 0
86- return
92+ return 0
8793 fi
8894 python3 -c "
8995import json,sys
@@ -106,7 +112,7 @@ for cfg in "${CONFIGS[@]}"; do
106112 mode=$( echo " $cfg " | awk ' {print $1}' )
107113 budget=$( echo " $cfg " | awk ' {print $2}' )
108114 existing_ok=$( count_ok " $mode " " $budget " )
109- if [ " $existing_ok " -ge " $MIN_OK_THRESHOLD " ]; then
115+ if [[ " $existing_ok " -ge " $MIN_OK_THRESHOLD " ] ]; then
110116 log " CFG $mode b=$budget SKIP (existing ok=$existing_ok >= $MIN_OK_THRESHOLD )"
111117 continue
112118 fi
@@ -120,13 +126,13 @@ for cfg in "${CONFIGS[@]}"; do
120126 run_one_config " $mode " " $budget " " $workers " " $batch "
121127 ok_count=$( count_ok " $mode " " $budget " )
122128 log " CFG $mode b=$budget attempt=$attempt RESULT ok=$ok_count threshold=$MIN_OK_THRESHOLD "
123- if [ " $ok_count " -ge " $MIN_OK_THRESHOLD " ]; then
129+ if [[ " $ok_count " -ge " $MIN_OK_THRESHOLD " ] ]; then
124130 log " CFG $mode b=$budget OK (ok=$ok_count , attempt=$attempt )"
125131 break
126132 fi
127133 log " CFG $mode b=$budget BELOW THRESHOLD (ok=$ok_count ) — escalating to next tier"
128134 done
129- if [ " $ok_count " -lt " $MIN_OK_THRESHOLD " ]; then
135+ if [[ " $ok_count " -lt " $MIN_OK_THRESHOLD " ] ]; then
130136 FAILED+=(" $mode b=$budget ok=$ok_count " )
131137 log " CFG $mode b=$budget FAILED ALL TIERS (ok=$ok_count )"
132138 fi
0 commit comments