k100AI部署qwen3有相关文档和资源吗?
您好,下文为vllm部署qwen3系列模型的案例:
step1:拉取镜像 https://sourcefind.cn/#/image/dcu/vllm?activeName=overview(拉取最新版本镜像即可)
step2:基于上述镜像构建容器
step3:设置环境变量
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export NCCL_MIN_NCHANNELS=16
export NCCL_MAX_NCHANNELS=16
export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1
针对hycpu环境建议设置绑核操作(VLLM_RANK x_NUMA 变量值需要根据hy-smi --showtopo 自行设置):
export VLLM_NUMA_BIND=1
export VLLM_RANK0_NUMA=0
export VLLM_RANK1_NUMA=1
export VLLM_RANK2_NUMA=2
export VLLM_RANK3_NUMA=3
export VLLM_RANK4_NUMA=4
export VLLM_RANK5_NUMA=5
export VLLM_RANK6_NUMA=6
export VLLM_RANK7_NUMA=7
step4:vllm启动模型服务
235b模型:
vllm serve <model-path>/Qwen3-235B-A22B --trust-remote-code --dtype float16 -tp 8 --max-model-len 32768 --gpu-memory-utilization 0.95
其它模型:
vllm serve <model-path>/Qwen3-32B --trust-remote-code --dtype float16 -tp 4
vllm serve <model-path>/Qwen3-30B-A3B --trust-remote-code --dtype float16 -tp 4
vllm serve <model-path>/Qwen3-14B --trust-remote-code --dtype float16 -tp 2
vllm serve <model-path>/Qwen3-8B --trust-remote-code --dtype float16 -tp 1
您好,下文为vllm部署qwen3系列模型的案例:
step1:拉取镜像 https://sourcefind.cn/#/image/dcu/vllm?activeName=overview(拉取最新版本镜像即可)
step2:基于上述镜像构建容器
step3:设置环境变量
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export NCCL_MIN_NCHANNELS=16
export NCCL_MAX_NCHANNELS=16
export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1
针对hycpu环境建议设置绑核操作(VLLM_RANK x_NUMA 变量值需要根据hy-smi --showtopo 自行设置):
export VLLM_NUMA_BIND=1
export VLLM_RANK0_NUMA=0
export VLLM_RANK1_NUMA=1
export VLLM_RANK2_NUMA=2
export VLLM_RANK3_NUMA=3
export VLLM_RANK4_NUMA=4
export VLLM_RANK5_NUMA=5
export VLLM_RANK6_NUMA=6
export VLLM_RANK7_NUMA=7
step4:vllm启动模型服务
235b模型:
vllm serve <model-path>/Qwen3-235B-A22B --trust-remote-code --dtype float16 -tp 8 --max-model-len 32768 --gpu-memory-utilization 0.95
其它模型:
vllm serve <model-path>/Qwen3-32B --trust-remote-code --dtype float16 -tp 4
vllm serve <model-path>/Qwen3-30B-A3B --trust-remote-code --dtype float16 -tp 4
vllm serve <model-path>/Qwen3-14B --trust-remote-code --dtype float16 -tp 2
vllm serve <model-path>/Qwen3-8B --trust-remote-code --dtype float16 -tp 1
vllm serve <model-path>/Qwen3-235B-A22B-W8A8 --trust-remote-code --dtype float16 -tp 8