extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-15 07:32:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

whisper : use flash attention (#2152)

Browse Source 
* whisper : use flash attention in the encoder

* whisper : add kv_pad

* whisper : remove extra backend instance (huh?)

* whisper : use FA for cross-attention

* whisper : use FA for self-attention

* whisper : simplify encoder FA

* whisper : add flash_attn runtime parameter

* scripts : add bench log

* scripts : add M1 Pro bench log
This commit is contained in:
Georgi Gerganov
2024-05-15 09:38:19 +03:00
committed by GitHub
parent 9d5771ae43
commit 7094ea5e75
13 changed files with 657 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
scripts/bench-all.sh
View File

@ -2,7 +2,7 @@
# Helper script to run the bench tool on all models and print the results in share-able format
printf "Usage: ./bench.sh [n_threads] [encoder-only]\n"
printf "Usage: ./bench.sh [n_threads] [encoder-only] [flash-attn]\n"
if [ -z "$1" ]; then
n_threads=4
@ -11,12 +11,19 @@ else
fi
encoder_only=0
if [ -z "$2" ]; then
if [ -z "$2" ] || [ "$2" -eq 0 ]; then
encoder_only=0
else
encoder_only=$2
fi
fattn=""
if [ -z "$3" ] || [ "$3" -eq 0 ]; then
fattn=""
else
fattn="-fa"
fi
models=( \
"tiny" "tiny-q4_0" "tiny-q4_1" "tiny-q5_0" "tiny-q5_1" "tiny-q8_0" \
"base" "base-q4_0" "base-q4_1" "base-q5_0" "base-q5_1" "base-q8_0" \
@ -44,13 +51,19 @@ if [ "$encoder_only" -eq 0 ]; then
printf "\n"
fi
printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "Bch5" "PP" "Commit"
printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"
if [ "$fattn" == "-fa" ]; then
fattn_i=1
else
fattn_i=0
fi
printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "FA" "Enc." "Dec." "Bch5" "PP" "Commit"
printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"
for model in "${models[@]}"; do
# actual run
# store stderr output in a variable in order to parse it later
output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
output=$(./bench -m ./models/ggml-$model.bin -t $n_threads $fattn 2>&1)
ret=$?
# parse the output:
@ -95,6 +108,6 @@ for model in "${models[@]}"; do
commit=$(git rev-parse --short HEAD)
if [ $ret -eq 0 ]; then
printf "| <todo> | <todo> | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
printf "| <todo> | <todo> | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$fattn_i" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
fi
done