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

Isolate Qwen235 DeepGEMM JIT cache

Browse Source
This commit is contained in:
Gahow Wang
2026-07-18 10:57:48 +08:00
parent 28ffb22bee
commit 686b050517
1 changed files with 12 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
runs/frontier-fidelity-envelope-v1/run_qwen235_v020_smoke.sh
View File

@@ -14,6 +14,8 @@ VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-235B-A22B-FP8}"
CLIENT="${CLIENT:?CLIENT is required}"
READY_TIMEOUT_SECONDS="${READY_TIMEOUT_SECONDS:-1800}"
TRTLLM_DG_CACHE_DIR="${TRTLLM_DG_CACHE_DIR:-${OUT}/deep_gemm}"
DEEP_GEMM_JIT_DEBUG="${DEEP_GEMM_JIT_DEBUG:-false}"
SERVED_MODEL="qwen3-235b-v020-smoke"
SERVER_PID=""
@@ -47,8 +49,12 @@ if ! [[ "${READY_TIMEOUT_SECONDS}" =~ ^[1-9][0-9]*$ ]]; then
echo "ERROR: READY_TIMEOUT_SECONDS must be a positive integer" >&2
exit 1
fi
case "${DEEP_GEMM_JIT_DEBUG}" in
true|false) ;;
*) echo "ERROR: DEEP_GEMM_JIT_DEBUG must be true or false" >&2; exit 1 ;;
esac
echo "QWEN235_V020_SMOKE_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 checkpoint=FP8_compute_kv=auto config=TP${TP}_EP${ENABLE_EXPERT_PARALLEL}_MNS${MNS}_MBT${MBT} prefix=false requests=1 shape=2048x1 ready_timeout=${READY_TIMEOUT_SECONDS}s role=compatibility_gate_not_latency_measurement output=${OUT} expected_wall=20-35m expected_gpu_cap=$((${TP} * 35 / 60))_H20h"
echo "QWEN235_V020_SMOKE_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 checkpoint=FP8_compute_kv=auto config=TP${TP}_EP${ENABLE_EXPERT_PARALLEL}_MNS${MNS}_MBT${MBT} prefix=false requests=1 shape=2048x1 ready_timeout=${READY_TIMEOUT_SECONDS}s deep_gemm_cache=${TRTLLM_DG_CACHE_DIR} deep_gemm_jit_debug=${DEEP_GEMM_JIT_DEBUG} role=compatibility_gate_not_latency_measurement output=${OUT} expected_wall=20-35m expected_gpu_cap=$((${TP} * 35 / 60))_H20h"
date -u +START_UTC=%Y-%m-%dT%H:%M:%SZ
sha256sum "${BASH_SOURCE[0]}" "${CLIENT}" "${MODEL_ROOT}/config.json" > "${OUT}/provenance/input.sha256"
"${VENV_ROOT}/bin/vllm" --version > "${OUT}/provenance/vllm.version"
@@ -67,6 +73,11 @@ export TRANSFORMERS_OFFLINE=1
export HOME=/tmp/wjh
export XDG_CACHE_HOME=/tmp/wjh/.cache
export VLLM_CACHE_ROOT=/tmp/wjh/.cache/vllm
export TRTLLM_DG_CACHE_DIR
if [[ "${DEEP_GEMM_JIT_DEBUG}" == true ]]; then
export TRTLLM_DG_JIT_DEBUG=1
fi
mkdir -p "${TRTLLM_DG_CACHE_DIR}"
setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
--host 127.0.0.1 --port "${SERVER_PORT}" --served-model-name "${SERVED_MODEL}" \
--tensor-parallel-size "${TP}" --disable-custom-all-reduce --quantization fp8 \