{"id":1801,"date":"2026-08-13T15:16:38","date_gmt":"2026-08-13T07:16:38","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1801"},"modified":"2026-08-31T12:30:31","modified_gmt":"2026-08-31T04:30:31","slug":"sglang-%e5%9c%a8%e6%98%87%e8%85%be-910b-%e9%83%a8%e7%bd%b2-qwen3-6-35b-a3b-%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/08\/13\/sglang-%e5%9c%a8%e6%98%87%e8%85%be-910b-%e9%83%a8%e7%bd%b2-qwen3-6-35b-a3b-%e7%ae%80%e8%ae%b0\/","title":{"rendered":"SGLang \u5728\u6607\u817e 910B \u90e8\u7f72 Qwen3.6-35B-A3B \u7b80\u8bb0"},"content":{"rendered":"\n

\u5355\u673a 4 \u5361\u6607\u817e 910B\uff0c\u7528 SGLang \u90e8\u7f72 Qwen3.6-35B-A3B\uff08MoE\uff0c\u6fc0\u6d3b\u53c2\u6570 35B\uff0c\u603b\u53c2\u6570\u7ea6 300B\uff09\uff0cTP=4\uff0c\u4e0a\u4e0b\u6587 256K\u3002\u8bb0\u5f55\u5bb9\u5668\u542f\u52a8\u4e0e\u8c03\u4f18\u811a\u672c\uff0c\u540e\u7eed\u76f4\u63a5\u590d\u7528\u3002<\/p>\n\n\n\n

\u4e00\u3001\u62c9\u53d6\u955c\u50cf<\/h2>\n\n\n\n

\u4f7f\u7528\u5b98\u65b9\u6607\u817e\u9002\u914d\u955c\u50cf\uff0c910B \u5bf9\u5e94 CANN 9.0 \u7248\u672c\uff1a<\/p>\n\n\n\n

export IMAGE=quay.io\/ascend\/sglang:v0.5.16-cann9.0.0-910b\nexport NAME=sglang-qwen36-35b-1<\/code><\/pre>\n\n\n\n
\u4e8c\u3001\u542f\u52a8\u5bb9\u5668<\/h2>\n\n\n\n
\u6607\u817e 910B \u516b\u5757\u5361\u5168\u90e8\u900f\u4f20\u8fdb\u5bb9\u5668\uff0c\u6302\u8f7d DCMI\/HCCL \u9a71\u52a8\u4e0e\u6570\u636e\u76ee\u5f55\u3002\u6ce8\u610f --shm-size<\/code> \u7ed9\u5230 200g\uff0c\u907f\u514d\u5927\u4e0a\u4e0b\u6587\u573a\u666f\u4e0b\u5171\u4eab\u5185\u5b58\u4e0d\u8db3\uff1a<\/p>\n\n\n\n
docker run -itd \\\n --name $NAME \\\n --net=host \\\n --shm-size=200g \\\n --privileged \\\n --device \/dev\/davinci0 \\\n --device \/dev\/davinci1 \\\n --device \/dev\/davinci2 \\\n --device \/dev\/davinci3 \\\n --device \/dev\/davinci4 \\\n --device \/dev\/davinci5 \\\n --device \/dev\/davinci6 \\\n --device \/dev\/davinci7 \\\n --device \/dev\/davinci_manager \\\n --device \/dev\/devmm_svm \\\n --device \/dev\/hisi_hdc \\\n -v \/usr\/local\/dcmi:\/usr\/local\/dcmi \\\n -v \/usr\/local\/Ascend\/driver\/tools\/hccn_tool:\/usr\/local\/Ascend\/driver\/tools\/hccn_tool \\\n -v \/usr\/local\/bin\/npu-smi:\/usr\/local\/bin\/npu-smi \\\n -v \/usr\/local\/Ascend\/driver\/lib64\/:\/usr\/local\/Ascend\/driver\/lib64\/ \\\n -v \/usr\/local\/Ascend\/driver\/version.info:\/usr\/local\/Ascend\/driver\/version.info \\\n -v \/etc\/ascend_install.info:\/etc\/ascend_install.info \\\n -v \/etc\/hccn.conf:\/etc\/hccn.conf \\\n -v \/mnt\/data:\/mnt\/data \\\n -v \/data\/huawei\/:\/data \\\n -it $IMAGE bash<\/code><\/pre>\n\n\n\n
\u4e09\u3001\u7f16\u5199\u8c03\u4f18\u811a\u672c qwen36.sh<\/h2>\n\n\n\n
\u5728\u5bb9\u5668 \/data<\/code> \u4e0b\u7f16\u5199\u542f\u52a8\u811a\u672c\uff0c\u5148\u505a CPU \/ NPU \u8c03\u4f18\uff0c\u518d\u62c9\u8d77\u670d\u52a1\uff1a<\/p>\n\n\n\n
# CPU\u8c03\u4f18\n#echo performance | tee \/sys\/devices\/system\/cpu\/cpu*\/cpufreq\/scaling_governor\nsysctl -w vm.swappiness=0\nsysctl -w kernel.numa_balancing=0\nsysctl -w kernel.sched_migration_cost_ns=50000\nexport ASCEND_VISIBLE_DEVICES=4,5,6,7\nexport ASCEND_RT_VISIBLE_DEVICES=4,5,6,7\nexport SGLANG_SET_CPU_AFFINITY=1\n\n# \u6e05\u9664\u4ee3\u7406\nunset https_proxy http_proxy HTTPS_PROXY HTTP_PROXY\nunset ASCEND_LAUNCH_BLOCKING\n\n# \u6291\u5236torch_npu\u5197\u4f59\u8b66\u544a\nexport TORCH_NPU_TRANSFER_WARNING=0\n\n# CANN\u73af\u5883\nsource \/usr\/local\/Ascend\/ascend-toolkit\/set_env.sh\nsource \/usr\/local\/Ascend\/nnal\/atb\/set_env.sh\n\n# NPU\u8c03\u4f18\u73af\u5883\u53d8\u91cf\nexport STREAMS_PER_DEVICE=32\nexport HCCL_OP_EXPANSION_MODE=AIV\nexport HCCL_SOCKET_IFNAME=lo\nexport GLOO_SOCKET_IFNAME=lo\nexport SGLANG_ENABLE_SPEC_V2=1\nexport SGLANG_ENABLE_OVERLAP_PLAN_STREAM=0\n\n# \u542f\u52a8\u670d\u52a1\npython3 -m sglang.launch_server \\\n --model-path \/data\/models\/Qwen3.6-35B-A3B \\\n --served-model-name Qwen35 \\\n --host 0.0.0.0 \\\n --port 21414 \\\n --tp-size 4 \\\n --context-length 262144 \\\n --reasoning-parser qwen3 \\\n --tool-call-parser qwen3_coder \\\n --device npu \\\n --attention-backend ascend \\\n --disable-radix-cache \\\n --mem-fraction-static 0.70 \\\n --trust-remote-code \\\n --dtype bfloat16 \\\n --enable-prefill-delayer \\\n --prefill-delayer-max-delay-passes 100<\/code><\/pre>\n\n\n\n
\u5176\u4e2d ASCEND_VISIBLE_DEVICES=4,5,6,7<\/code> \u9009\u62e9\u56db\u5f20\u5361\uff0c\u5bf9\u5e94\u5bb9\u5668\u5185\u900f\u4f20\u7684\u8bbe\u5907\uff0c\u6309\u5b9e\u9645\u8d44\u6e90\u81ea\u884c\u8c03\u6574\u3002<\/p>\n\n\n\n
\u56db\u3001\u542f\u52a8\u547d\u4ee4\u53c2\u6570\u89e3\u6790<\/h2>\n\n\n\n
\u53c2\u6570<\/th>\u503c<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
--tp-size<\/code><\/td>4<\/td>\u5f20\u91cf\u5e76\u884c\u5ea6\uff0c\u5bf9\u5e94 4 \u5f20 910B<\/td><\/tr>
--context-length<\/code><\/td>262144<\/td>\u6700\u5927\u4e0a\u4e0b\u6587 256K<\/td><\/tr>
--device npu<\/code><\/td>\u2014<\/td>\u6307\u5b9a\u6607\u817e NPU \u540e\u7aef<\/td><\/tr>
--attention-backend<\/code><\/td>ascend<\/td>\u4f7f\u7528\u6607\u817e attention \u5b9e\u73b0<\/td><\/tr>
--disable-radix-cache<\/code><\/td>\u2014<\/td>\u5173\u95ed radix cache\uff08\u6607\u817e\u573a\u666f\u5e38\u7528\uff09<\/td><\/tr>
--mem-fraction-static<\/code><\/td>0.70<\/td>\u9759\u6001 KV \u7f13\u5b58\u663e\u5b58\u5360\u6bd4<\/td><\/tr>
--dtype<\/code><\/td>bfloat16<\/td>\u6a21\u578b\u7cbe\u5ea6<\/td><\/tr>
--reasoning-parser<\/code><\/td>qwen3<\/td>\u63a8\u7406\u6a21\u578b reasoning \u89e3\u6790<\/td><\/tr>
--tool-call-parser<\/code><\/td>qwen3_coder<\/td>\u5de5\u5177\u8c03\u7528\u89e3\u6790<\/td><\/tr>
--enable-prefill-delayer<\/code><\/td>\u2014<\/td>\u542f\u7528 prefill \u5ef6\u8fdf\uff0c\u914d\u5408 max-delay-passes \u7f13\u89e3\u957f\u8f93\u5165\u65f6\u7684 prefill\/decode \u62a2\u5360<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u8c03\u4f18\u73af\u5883\u53d8\u91cf\u89e3\u6790\uff1a<\/p>\n\n\n\n
  • STREAMS_PER_DEVICE=32<\/code>\uff1a\u63d0\u5347\u5e76\u53d1\u6d41<\/li>
  • HCCL_OP_EXPANSION_MODE=AIV<\/code>\uff1aHCCL \u7b97\u5b50\u6269\u5c55\u6a21\u5f0f<\/li>
  • HCCL_SOCKET_IFNAME=lo<\/code> \/ GLOO_SOCKET_IFNAME=lo<\/code>\uff1a\u591a\u673a\/\u591a\u5361\u8d70\u56de\u73af\u7f51\u5361<\/li>
  • SGLANG_ENABLE_SPEC_V2=1<\/code>\uff1a\u542f\u7528 speculation v2<\/li>
  • SGLANG_ENABLE_OVERLAP_PLAN_STREAM=0<\/code>\uff1a\u5173\u95ed\u8ba1\u5212\u6d41\u91cd\u53e0<\/li><\/ul>\n\n\n\n
    \u4e94\u3001\u6ce8\u610f\u4e8b\u9879<\/h2>\n\n\n\n
    • \u4f7f\u7528\u7684\u5361\u53f7\uff084,5,6,7\uff09\u53ef\u6309\u5b9e\u9645\u8d44\u6e90\u8c03\u6574\uff0c\u9700\u4e0e\u5bb9\u5668\u6302\u8f7d\u7684\u8bbe\u5907\u4e00\u81f4<\/li>
    • --shm-size<\/code> \u573a\u666f\u4e0b\uff0c\u82e5\u51fa\u73b0 OOM \u4f18\u5148\u68c0\u67e5\u5171\u4eab\u5185\u5b58\u5927\u5c0f<\/li>
    • \u82e5\u9700\u591a\u673a\u90e8\u7f72\uff0c\u5c06 HCCL_SOCKET_IFNAME<\/code> \u6539\u4e3a\u5b9e\u9645\u4e1a\u52a1\u7f51\u5361\u5e76\u914d\u7f6e --nnodes<\/code><\/li>
    • \u5927\u4e0a\u4e0b\u6587\u573a\u666f\u5efa\u8bae\u6839\u636e\u663e\u5b58\u4f59\u91cf\u5fae\u8c03 --mem-fraction-static<\/code><\/li><\/ul>\n\n\n\n
      \u4ee5\u4e0a\u4e3a\u5355\u673a 4 \u5361 TP=4 \u90e8\u7f72 Qwen3.6-35B-A3B \u7684\u5b8c\u6574\u7b80\u8bb0\uff0c\u4f9b\u540e\u7eed\u590d\u7528\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
      \u5355\u673a 4 \u5361\u6607\u817e 910B\uff0c\u7528 SGLang \u90e8\u7f72 Qwen3.6-35B-A3B\uff08MoE\uff0c\u6fc0\u6d3b\u53c2\u6570 35B\uff0c\u603b\u53c2\u6570\u7ea6 300B\uff09 …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[25,52,15,66,65,16],"class_list":["post-1801","post","type-post","status-publish","format-standard","hentry","category-uncategorized","tag-910b","tag-docker","tag-npu","tag-qwen3-6","tag-sglang","tag-16"],"jetpack_featured_media_url":"","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1801","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1801"}],"version-history":[{"count":2,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1801\/revisions"}],"predecessor-version":[{"id":1804,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1801\/revisions\/1804"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1801"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1801"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1801"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}