llm_mutil_npu / scripts /bench_hccl_adv.sh
xianglarry's picture
Initial C++ aclnn EAGER inference for Qwen3-235B-A22B MoE on Ascend 910 × 16 NPU
4b9fefd
#!/usr/bin/env bash
# bench_hccl_adv.sh — 进阶 HCCL 参数调优
# 在已确定的 ring:200 baseline 上加入 OP_EXPANSION_MODE=AIV 等 knob
set -u
cd "$(dirname "$0")/.."
MODEL="${MODEL_DIR:-/path/to/Qwen3-235B-A22B-Instruct-2507-BF16}"
BIN="./build/qwen3-moe-aclnn"
LAUNCH="./scripts/tp_launch.sh"
TP=16
N_PREDICT=200
N_RUNS=2
PROMPT="The history of artificial intelligence spans several decades and"
VOCAB="tokenizer_data/vocab.bin"
OUT=/tmp/bench_hccl_adv.csv
echo "config,run1,run2,best,median" > $OUT
run_one() {
local name="$1"; shift
# remaining args are env assignments: KEY=VALUE ...
local tgs=()
for r in $(seq 1 $N_RUNS); do
local out
# set env vars for this run
local env_cmd=""
for a in "$@"; do env_cmd="$env_cmd $a"; done
out=$(env HCCL_ALGO=level0:ring HCCL_BUFFSIZE=200 $@ \
${LAUNCH} ${TP} ${BIN} --model-dir "$MODEL" \
--prompt "$PROMPT" --n-predict $N_PREDICT \
--vocab "$VOCAB" --seed 0 2>&1 | grep "decode :" | awk '{print $(NF-2)}')
tgs+=("${out:-0}")
done
local sorted=($(printf '%s\n' "${tgs[@]}" | sort -n))
local best="${sorted[-1]}"
local median="${sorted[$((${#sorted[@]}/2))]}"
echo "$name,${tgs[0]},${tgs[1]},$best,$median" >> $OUT
printf " %-40s %s best=%s median=%s\n" "$name" "${tgs[*]}" "$best" "$median"
}
echo "Adv HCCL bench: baseline ring:200 + additional knobs"
echo "Results → $OUT"
echo ""
run_one "baseline (ring+200 only)"
run_one "+ OP_EXPANSION_MODE=AIV" HCCL_OP_EXPANSION_MODE=AIV
run_one "+ OP_BASE_FFTS_MODE=1" HCCL_OP_BASE_FFTS_MODE_ENABLE=1
run_one "+ OP_EXPANSION=AIV + FFTS=1" HCCL_OP_EXPANSION_MODE=AIV HCCL_OP_BASE_FFTS_MODE_ENABLE=1
run_one "+ OP_EXPANSION=AIV + BUF=256" HCCL_OP_EXPANSION_MODE=AIV HCCL_BUFFSIZE=256
run_one "+ OP_EXPANSION=AIV + BUF=512" HCCL_OP_EXPANSION_MODE=AIV HCCL_BUFFSIZE=512
run_one "+ OP_EXPANSION=AIV + ALGO=fullmesh" HCCL_OP_EXPANSION_MODE=AIV HCCL_ALGO=level0:fullmesh
echo ""
echo "====== Sorted by best TG ======"
(head -1 $OUT; tail -n +2 $OUT | sort -t, -k4 -gr) | column -t -s,