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

Prepare remaining Qwen30 latency cases

Browse Source
This commit is contained in:
Gahow Wang
2026-07-18 01:00:41 +08:00
parent 65fee8450a
commit 44104bd96e
6 changed files with 650 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
runs/frontier-fidelity-envelope-v1/run_qwen30_exact_trace_real_anchor.sh
View File

@@ -8,10 +8,12 @@ TP="${TP:?TP is required}"
MNS="${MNS:?MNS is required}"
TRACE_LABEL="${TRACE_LABEL:?TRACE_LABEL is required}"
SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
PREFIX_CACHING="${PREFIX_CACHING:-true}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
FLASHINFER_WORKSPACE_BASE="${FLASHINFER_WORKSPACE_BASE:-/tmp/wjh/flashinfer-workspace-vllm020-profiler-v1}"
SERVED_MODEL="qwen3-30b-exact-trace"
EXACT_TRACE_CLIENT="${EXACT_TRACE_CLIENT:-qwen30_exact_trace_client.py}"
SERVER_PID=""
mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/results" \
@@ -37,8 +39,21 @@ if [[ "${#GPU_IDS[@]}" -ne "${TP}" ]]; then
exit 1
fi
case "${PREFIX_CACHING}" in
true)
PREFIX_CACHING_FLAG="--enable-prefix-caching"
;;
false)
PREFIX_CACHING_FLAG="--no-enable-prefix-caching"
;;
*)
echo "ERROR: PREFIX_CACHING must be true or false, got ${PREFIX_CACHING}" >&2
exit 1
;;
esac
REQUEST_COUNT="$(wc -l < "${REQUESTS_FILE}")"
echo "QWEN30_EXACT_TRACE_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 trace=${TRACE_LABEL} requests=${REQUEST_COUNT} source=${REQUESTS_FILE} arrivals=original600s input_output_prompt=exact prefix=on block=16 tpot_slo=150ms flashinfer_workspace=${FLASHINFER_WORKSPACE_BASE} output=${OUTPUT_ROOT} expected_wall=12-35m hard_wall=3600s"
echo "QWEN30_EXACT_TRACE_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 trace=${TRACE_LABEL} requests=${REQUEST_COUNT} source=${REQUESTS_FILE} arrivals=manifest prefix=${PREFIX_CACHING} block=16 metrics=TTFT,TPOT-if-OSL-gt-1,E2E flashinfer_workspace=${FLASHINFER_WORKSPACE_BASE} output=${OUTPUT_ROOT} expected_wall=12-35m hard_wall=3600s"
date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
sha256sum qwen30_exact_trace_client.py run_qwen30_exact_trace_real_anchor.sh \
../frontier-phase-factorial-v0/qwen30_prefill_client.py \
@@ -60,7 +75,7 @@ setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
--host 127.0.0.1 --port "${SERVER_PORT}" --served-model-name "${SERVED_MODEL}" \
--tensor-parallel-size "${TP}" --gpu-memory-utilization 0.92 \
--max-model-len 40960 --max-num-batched-tokens 8192 --max-num-seqs "${MNS}" \
--enable-prefix-caching --enable-chunked-prefill --no-enable-log-requests \
"${PREFIX_CACHING_FLAG}" --enable-chunked-prefill --no-enable-log-requests \
> "${OUTPUT_ROOT}/logs/server.log" 2>&1 &
SERVER_PID=$!
READY=0
@@ -89,7 +104,7 @@ fi
--model-path "${MODEL_ROOT}" --rate 1 --requests 4 --input-tokens 512 \
--output-tokens 1 --output "${OUTPUT_ROOT}/results/warmup.json"
"${VENV_ROOT}/bin/python" qwen30_exact_trace_client.py \
"${VENV_ROOT}/bin/python" "${EXACT_TRACE_CLIENT}" \
--port "${SERVER_PORT}" --requests-file "${REQUESTS_FILE}" \
--output "${OUTPUT_ROOT}/results/result.json" --tpot-slo-ms 150 \
--timeout-seconds 1800