julian/amd-strix-halo-toolboxes
1
0
Fork 0
Code Issues Pull Requests 2 Actions Packages Projects Releases Wiki Activity

restoring correct llama-bench flags

Browse Source
This commit is contained in:
Donato Capitella
2025-11-17 16:00:10 +00:00
parent de02a53d96
commit 12f057612b
1 changed files with 5 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
benchmark/run_benchmarks.sh
+5 -8
View File
@@ -77,20 +77,17 @@ for MODEL_PATH in "${MODEL_PATHS[@]}"; do
EXTRA_ARGS=( -fa 1 ) EXTRA_ARGS=( -fa 1 )
fi fi
# tune ubatch-size for backend
if [[ "$ENV" == *vulkan* ]]; then
EXTRA_ARGS+=( -ub 512 )
else
EXTRA_ARGS+=( -ub 2048 )
fi
for CTX in default longctx32768; do for CTX in default longctx32768; do
CTX_SUFFIX="" CTX_SUFFIX=""
CTX_ARGS=() CTX_ARGS=()
if [[ "$CTX" == longctx32768 ]]; then if [[ "$CTX" == longctx32768 ]]; then
CTX_SUFFIX="__longctx32768" CTX_SUFFIX="__longctx32768"
# shellcheck disable=SC2206 # intentional word splitting into array
CTX_ARGS=( -p 2048 -n 32 -d 32768 ) CTX_ARGS=( -p 2048 -n 32 -d 32768 )
if [[ "$ENV" == *vulkan* ]]; then
CTX_ARGS+=( -ub 512 )
else
CTX_ARGS+=( -ub 2048 )
fi
fi fi
OUT="$RESULTDIR/${MODEL_NAME}__${ENV}${SUFFIX}${CTX_SUFFIX}.log" OUT="$RESULTDIR/${MODEL_NAME}__${ENV}${SUFFIX}${CTX_SUFFIX}.log"