987f1db4b0
- 实例 A profile 加 --disable-radix-cache（bench 测量纯净） - 新增实例 B profile（kimi3_pro6000_sglang_tp32ep32_instB，.1-.4） - 新增 deploy_dp2.sh 编排脚本（A + B + router --worker-urls） - 新增 README
55 lines
2.9 KiB
Bash
55 lines
2.9 KiB
Bash
# Kimi-K3 SGLang multi-node TP=32 EP=32 deployment profile (4x RTX 6000D).
|
||
# Nodes: 174.1.60.5~8 (rank 0~3), 32x NVIDIA RTX 6000D (85GB, sm_120).
|
||
#
|
||
# 关键点（实测踩坑，勿随意改）:
|
||
# - MoE 后端必须 marlin（K3 的 MXFP4 缩放因子为 uint8，DeepGEMM 只接受 fp32/UE8M0）
|
||
# - RoCE: NCCL_IB_HCA=mlx5_0..3（4 张独立卡, 10.100.21-24/24, RoCEv2 GID index 3）
|
||
# 实测 32-rank 117MB allreduce 2.4ms；勿用 mlx5_bond_0（仅 4.5GB/s）
|
||
# - 容器必须 --ulimit memlock=-1（否则 ibv_create_cq 报 Cannot allocate memory）
|
||
# - 不要设 NCCL_ALGO=TREE（CUDA graph 捕获报 "NCCL error: invalid usage"）
|
||
# - 首次请求含 ~110s Triton JIT 编译，属正常现象，预热一次后回落
|
||
# - flashkda 可选后端（与 triton prefill 性能等价）见 docs/KIMI_K3_DEPLOY.md 附录
|
||
#
|
||
# Model-team only. Ops only run `python -m sskj.bench` against the served URL.
|
||
|
||
PLATFORM=pro6000
|
||
EXPERIMENT=kimi3_pro6000_sglang_tp32ep32
|
||
MODEL_NAME=Kimi-K3
|
||
ENGINE=sglang
|
||
RUNTIME=docker
|
||
DOCKER_IMAGE=lmsysorg/sglang:kimi-k3
|
||
CONTAINER_NAME=${EXPERIMENT}_node${NODE_RANK}
|
||
MODEL_PATH=/data/hf_models/Kimi-K3
|
||
SERVED_MODEL_NAME=kimi-k3
|
||
PORT=30000
|
||
HEALTH_PATH=/health
|
||
HEALTH_HOST=174.1.60.5
|
||
HEALTH_WAIT_S=2400
|
||
CONTAINER_PYTHON=python3
|
||
|
||
# ---- 并行度（固定，勿改）----
|
||
TP=32
|
||
DP=1
|
||
EP_SIZE=32
|
||
|
||
# ---- Multi-node topology (rank order; rank 0 exposes the HTTP API) ----
|
||
NNODES=4
|
||
NODE_HOSTS="174.1.60.5 174.1.60.6 174.1.60.7 174.1.60.8"
|
||
NODE_SSH_USER=root
|
||
LOCAL_NODE_RANK=0
|
||
MASTER_IP=174.1.60.5
|
||
DIST_PORT=20000
|
||
|
||
DEVICE_VARS="CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7"
|
||
ENGINE_ENV="NCCL_SOCKET_IFNAME=bond0 GLOO_SOCKET_IFNAME=bond0 NCCL_IB_HCA=mlx5_0,mlx5_1,mlx5_2,mlx5_3 NCCL_IB_GID_INDEX=3 NCCL_IB_TIMEOUT=22 NCCL_IB_RETRY_CNT=7 NCCL_CUMEM_ENABLE=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0 SGLANG_MOE_FUSED_GATE_RADIX=1"
|
||
|
||
DOCKER_FLAGS="--gpus all --network host --ipc=host --ulimit memlock=-1 --device /dev/infiniband --shm-size 32g --entrypoint ''"
|
||
VOLUMES="${MODEL_PATH}:${MODEL_PATH}:ro"
|
||
PATCH_MOUNTS="/tmp/patch_k3_sm120.py:/tmp/patch_k3_sm120.py:ro"
|
||
|
||
# BOOTSTRAP 在容器内执行: 打 sm_120 补丁 → 按节点 rank 计算 SGLANG_HOST_IP → 启动 sglang。
|
||
# SGLANG_HOST_IP 必须为本节点实际 IP（174.1.60.5~8 = 5 + NODE_RANK）。
|
||
BOOTSTRAP="python3 /tmp/patch_k3_sm120.py && export SGLANG_HOST_IP=\"174.1.60.$((5 + ${NODE_RANK}))\" && exec python3 -m sglang.launch_server ${LAUNCH_ARGS}"
|
||
|
||
LAUNCH_ARGS="--model-path ${MODEL_PATH} --served-model-name ${SERVED_MODEL_NAME} --tp-size ${TP} --ep-size 32 --nnodes ${NNODES} --node-rank ${NODE_RANK} --dist-init-addr ${MASTER_IP}:${DIST_PORT} --trust-remote-code --moe-runner-backend marlin --mem-fraction-static 0.88 --cuda-graph-max-bs-decode 16 --mamba-radix-cache-strategy extra_buffer_lazy --disable-radix-cache --dist-timeout 3600 --mamba-full-memory-ratio 0.36 --host 0.0.0.0 --port ${PORT}"
|