gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Generalize Qwen30 fixed-shape real runner

Browse Source
This commit is contained in:
Gahow Wang
2026-07-17 10:17:21 +08:00
parent 6e8704d525
commit a2cf361ffe
3 changed files with 71 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
runs/frontier-phase-factorial-v0/run_qwen30_prefill_real_config.sh
View File

@@ -6,6 +6,10 @@ OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
TP="${TP:?TP is required}"
MNS="${MNS:?MNS is required}"
RATES="${RATES:-4 8 16 32 64}"
INPUT_TOKENS="${INPUT_TOKENS:-2048}"
OUTPUT_TOKENS="${OUTPUT_TOKENS:-1}"
TPOT_SLO_MS="${TPOT_SLO_MS:-150}"
WARMUP_SECONDS="${WARMUP_SECONDS:-2}"
SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
@@ -35,18 +39,19 @@ if [[ "${#GPU_IDS[@]}" -ne "${TP}" ]]; then
fi
read -r -a RATE_ARRAY <<< "${RATES}"
echo "QWEN30_PREFILL_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 rates=${RATES// /,} rounds=2 requests=64 shape=ISL2048_OSL1 arrivals=uniform prefix=off cuda_graph=runtime_default isolation=fresh_server_per_anchor output=${OUTPUT_ROOT}"
echo "QWEN30_FIXED_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 rates=${RATES// /,} rounds=2 requests=64 shape=ISL${INPUT_TOKENS}_OSL${OUTPUT_TOKENS} ttft_slo=1000+1000*ISL/8000ms tpot_slo=${TPOT_SLO_MS}ms warmup_seconds=${WARMUP_SECONDS} arrivals=uniform prefix=off cuda_graph=runtime_default isolation=fresh_server_per_anchor output=${OUTPUT_ROOT}"
date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
sha256sum run_qwen30_prefill_real_config.sh qwen30_prefill_client.py \
> "${OUTPUT_ROOT}/provenance/source.sha256"
"${VENV_ROOT}/bin/python" - "${TP}" "${MNS}" "${RATES}" \
"${VENV_ROOT}/bin/python" - "${TP}" "${MNS}" "${RATES}" "${INPUT_TOKENS}" "${OUTPUT_TOKENS}" "${TPOT_SLO_MS}" "${WARMUP_SECONDS}" \
> "${OUTPUT_ROOT}/provenance/contract.json" <<'PY'
import importlib.metadata as metadata
import json
import platform
import sys
tp, mns, rates = sys.argv[1:]
tp, mns, rates, input_tokens, output_tokens, tpot_slo_ms, warmup_seconds = sys.argv[1:]
input_tokens = int(input_tokens)
print(json.dumps({
"python": platform.python_version(),
"torch": metadata.version("torch"),
@@ -57,10 +62,11 @@ print(json.dumps({
"rounds": 2,
"requests_per_anchor": 64,
"anchor_isolation": "fresh_server_per_rate_per_round",
"target_rate_warmup_requests": "min(32, max(4, ceil(rate * 2)))",
"input_tokens": 2048,
"output_tokens": 1,
"ttft_slo_ms": 1256.0,
"target_rate_warmup_requests": f"min(32, max(4, ceil(rate * {warmup_seconds})))",
"input_tokens": input_tokens,
"output_tokens": int(output_tokens),
"ttft_slo_ms": 1000.0 + 1000.0 * input_tokens / 8000.0,
"tpot_slo_ms": float(tpot_slo_ms) if int(output_tokens) > 1 else None,
"target_pass_rate": 0.95,
}, indent=2, sort_keys=True))
PY
@@ -112,19 +118,22 @@ for ROUND in 1 2; do
exit 1
fi
WARMUP_REQUESTS="$("${VENV_ROOT}/bin/python" - "${RATE}" <<'PY'
WARMUP_REQUESTS="$("${VENV_ROOT}/bin/python" - "${RATE}" "${WARMUP_SECONDS}" <<'PY'
import math
import sys
print(min(32, max(4, math.ceil(float(sys.argv[1]) * 2.0))))
print(min(32, max(4, math.ceil(float(sys.argv[1]) * float(sys.argv[2])))))
PY
)"
"${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
--served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
--requests "${WARMUP_REQUESTS}" \
--requests "${WARMUP_REQUESTS}" --input-tokens "${INPUT_TOKENS}" \
--output-tokens "${OUTPUT_TOKENS}" --tpot-slo-ms "${TPOT_SLO_MS}" \
--output "${ROUND_ROOT}/results/warmup_${KEY}.json"
"${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
--served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
--requests 64 --output "${ROUND_ROOT}/results/${KEY}.json"
--requests 64 --input-tokens "${INPUT_TOKENS}" \
--output-tokens "${OUTPUT_TOKENS}" --tpot-slo-ms "${TPOT_SLO_MS}" \
--output "${ROUND_ROOT}/results/${KEY}.json"
cleanup
done
done