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

Isolate FlashInfer cache for trace replay

Browse Source
This commit is contained in:
Gahow Wang
2026-07-17 12:44:50 +08:00
parent 2aa713f5d6
commit 03aa794448
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
runs/frontier-fidelity-envelope-v1/run_qwen30_exact_trace_real_anchor.sh
View File

@@ -10,10 +10,12 @@ TRACE_LABEL="${TRACE_LABEL:?TRACE_LABEL is required}"
SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}" SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}" 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}" 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" SERVED_MODEL="qwen3-30b-exact-trace"
SERVER_PID="" SERVER_PID=""
mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/results" mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/results" \
"${FLASHINFER_WORKSPACE_BASE}"
exec > >(tee -a "${OUTPUT_ROOT}/logs/controller.log") 2>&1 exec > >(tee -a "${OUTPUT_ROOT}/logs/controller.log") 2>&1
cleanup() { cleanup() {
@@ -36,7 +38,7 @@ if [[ "${#GPU_IDS[@]}" -ne "${TP}" ]]; then
fi fi
REQUEST_COUNT="$(wc -l < "${REQUESTS_FILE}")" REQUEST_COUNT="$(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=${REQUESTS_FILE} arrivals=original600s input_output_prompt=exact prefix=on block=16 tpot_slo=150ms output=${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=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"
date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ" date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
sha256sum qwen30_exact_trace_client.py run_qwen30_exact_trace_real_anchor.sh \ sha256sum qwen30_exact_trace_client.py run_qwen30_exact_trace_real_anchor.sh \
../frontier-phase-factorial-v0/qwen30_prefill_client.py \ ../frontier-phase-factorial-v0/qwen30_prefill_client.py \
@@ -51,6 +53,7 @@ export VLLM_USE_V1=1
export TORCH_CUDA_ARCH_LIST=9.0 export TORCH_CUDA_ARCH_LIST=9.0
export HF_HUB_OFFLINE=1 export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1 export TRANSFORMERS_OFFLINE=1
export FLASHINFER_WORKSPACE_BASE
ulimit -n 65536 ulimit -n 65536
setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \ setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \