make\r\n<\/code><\/pre>\n\u7f16\u8bd1\u5b8c\u6210\u540e\uff0c\u4f1a\u751f\u6210\u53ef\u5728\u9879\u76ee\u4e2d\u8c03\u7528\u7684\u5185\u6838\u6587\u4ef6\u3002<\/p>\n
5. \u914d\u7f6e\u73af\u5883\u53d8\u91cf<\/h4>\n
\u4e3a\u786e\u4fddDeepEP\u6b63\u786e\u8fd0\u884c\uff0c\u9700\u8bbe\u7f6eNVSHMEM\u76f8\u5173\u53c2\u6570\uff0c\u4f8b\u5982\u865a\u62df\u901a\u9053\u5206\u914d\uff1a<\/p>\n
export NVSHMEM_IB_SL=0 # \u8bbe\u7f6e\u865a\u62df\u901a\u9053\uff0c\u907f\u514d\u6d41\u91cf\u51b2\u7a81\r\n<\/code><\/pre>\n\u5982\u9700\u542f\u7528\u81ea\u9002\u5e94\u8def\u7531\uff08\u4ec5\u9650\u4f4e\u5ef6\u8fdf\u5185\u6838\uff09\uff0c\u53ef\u989d\u5916\u914d\u7f6e\uff1a<\/p>\n
export NVSHMEM_ENABLE_ADAPTIVE_ROUTING=1\r\n<\/code><\/pre>\n6. \u6d4b\u8bd5\u5b89\u88c5<\/h4>\n
\u8fd0\u884c\u63d0\u4f9b\u7684\u6d4b\u8bd5\u811a\u672c\uff0c\u9a8c\u8bc1DeepEP\u529f\u80fd\uff1a<\/p>\n
python tests\/test_low_latency.py\r\n<\/code><\/pre>\n\u82e5\u8f93\u51fa\u663e\u793a\u901a\u4fe1\u6210\u529f\uff0c\u5219\u5b89\u88c5\u5b8c\u6210\u3002<\/p>\n
\u4f7f\u7528\u65b9\u6cd5<\/h3>\n
DeepEP\u4e3b\u8981\u901a\u8fc7\u96c6\u6210\u5230MoE\u6a21\u578b\u7684\u5de5\u4f5c\u6d41\u4e2d\u4f7f\u7528\uff0c\u4ee5\u4e0b\u662f\u4e3b\u8981\u529f\u80fd\u7684\u8be6\u7ec6\u64cd\u4f5c\u6307\u5357\uff1a<\/p>\n
\u529f\u80fd1\uff1a\u8fd0\u884c\u9ad8\u541e\u5410\u91cf\u8bad\u7ec3<\/strong><\/h4>\nDeepEP\u7684\u9ad8\u541e\u5410\u91cf\u5185\u6838\u9002\u7528\u4e8eMoE\u6a21\u578b\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002\u5047\u8bbe\u4f60\u6709\u4e00\u4e2a\u57fa\u4e8eDeepSeek-V3\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u6309\u7167\u4ee5\u4e0b\u6b65\u9aa4\u64cd\u4f5c\uff1a<\/p>\n
\n- \u51c6\u5907\u6a21\u578b\u548c\u6570\u636e<\/strong>\uff1a\u786e\u4fdd\u4f60\u7684MoE\u6a21\u578b\u5df2\u914d\u7f6e\u597d\u4e13\u5bb6\u5e76\u884c\u903b\u8f91\uff0c\u5e76\u51c6\u5907\u597d\u8bad\u7ec3\u6570\u636e\u96c6\u3002<\/li>\n
- \u8c03\u7528DeepEP\u5185\u6838<\/strong>\uff1a\u5728\u8bad\u7ec3\u811a\u672c\u4e2d\u5f15\u5165DeepEP\u7684all-to-all\u901a\u4fe1\u63a5\u53e3\u3002\u4f8b\u5982\uff1a\n
#include \"deep_ep.h\" \r\nvoid moe_train(float* input, float* output, int size) { \r\ndeep_ep_all_to_all(input, output, size, FP8); \r\n}\r\n<\/code><\/pre>\n<\/li>\n- \u914d\u7f6e\u786c\u4ef6<\/strong>\uff1a\u6307\u5b9a\u4f7f\u7528\u7684GPU\u8bbe\u5907\uff0c\u4f8b\u5982\uff1a\n
CUDA_VISIBLE_DEVICES=0,1,2,3 .\/train_script\r\n<\/code><\/pre>\n<\/li>\n- \u8fd0\u884c\u8bad\u7ec3<\/strong>\uff1a\u542f\u52a8\u8bad\u7ec3\u540e\uff0cDeepEP\u4f1a\u81ea\u52a8\u4f18\u5316NVLink\u548cRDMA\u95f4\u7684\u901a\u4fe1\u3002<\/li>\n<\/ol>\n
\u529f\u80fd2\uff1a\u4f4e\u5ef6\u8fdf\u63a8\u7406<\/strong><\/h4>\n\u4f4e\u5ef6\u8fdf\u5185\u6838\u9002\u7528\u4e8e\u5b9e\u65f6\u63a8\u7406\u4efb\u52a1\uff0c\u5982\u5728\u7ebf\u5bf9\u8bdd\u7cfb\u7edf\uff1a<\/p>\n
\n- \u52a0\u8f7d\u6a21\u578b<\/strong>\uff1a\u5c06\u9884\u8bad\u7ec3\u7684MoE\u6a21\u578b\u52a0\u8f7d\u5230\u5185\u5b58\u3002<\/li>\n
- \u8c03\u7528\u63a8\u7406\u5185\u6838<\/strong>\uff1a\u4f7f\u7528\u7eafRDMA\u901a\u4fe1\u63a5\u53e3\u3002\u4f8b\u5982\uff1a\n
#include \"deep_ep.h\" \r\nvoid moe_infer(float* query, float* result, int batch_size) { \r\ndeep_ep_low_latency_all_to_all(query, result, batch_size); \r\n}\r\n<\/code><\/pre>\n<\/li>\n- \u6d4b\u8bd5\u63a8\u7406\u901f\u5ea6<\/strong>\uff1a\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u6d4b\u91cf\u5ef6\u8fdf\uff1a\n
python tests\/test_inference.py --batch_size 128 --hidden_size 7168\r\n<\/code><\/pre>\n\u8f93\u51fa\u5c06\u663e\u793a\u6bcf\u6279\u6b21\u7684\u63a8\u7406\u65f6\u95f4\uff0c\u786e\u4fdd\u6ee1\u8db3\u5b9e\u65f6\u9700\u6c42\u3002<\/li>\n<\/ol>\n
\u529f\u80fd3\uff1aFP8\u8fd0\u7b97\u4f18\u5316<\/strong><\/h4>\nDeepEP\u652f\u6301FP8\u5206\u53d1\uff0c\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\uff1a<\/p>\n
\n- \u542f\u7528FP8\u6a21\u5f0f<\/strong>\uff1a\u5728\u8c03\u7528\u901a\u4fe1\u63a5\u53e3\u65f6\u6307\u5b9a\u6570\u636e\u7c7b\u578b\u4e3aFP8\u3002<\/li>\n
- \u9a8c\u8bc1\u7cbe\u5ea6<\/strong>\uff1a\u8fd0\u884c\u6d4b\u8bd5\u811a\u672c\u5bf9\u6bd4FP8\u4e0eBF16\u7684\u6027\u80fd\u548c\u7cbe\u5ea6\u5dee\u5f02\uff1a\n
python tests\/test_fp8.py\r\n<\/code><\/pre>\n<\/li>\n- \u5e94\u7528\u5230\u751f\u4ea7<\/strong>\uff1a\u5c06FP8\u914d\u7f6e\u96c6\u6210\u5230\u73b0\u6709\u8bad\u7ec3\u6216\u63a8\u7406\u6d41\u7a0b\u4e2d\u3002<\/li>\n<\/ol>\n
\u529f\u80fd4\uff1a\u8d44\u6e90\u63a7\u5236\u4e0e\u4f18\u5316<\/strong><\/h4>\n\u8c03\u8282SM\u6570\u91cf\u4ee5\u9002\u914d\u786c\u4ef6\uff1a<\/p>\n
\n- \u67e5\u770b\u786c\u4ef6SM\u6570\u91cf<\/strong>\uff1a\u4f7f\u7528
nvidia-smi<\/code>\u68c0\u67e5GPU\u7684\u6d41\u5904\u7406\u5668\u6570\u91cf\u3002<\/li>\n- \u8bbe\u7f6eSM\u9650\u5236<\/strong>\uff1a\u5728\u811a\u672c\u4e2d\u6307\u5b9a\uff1a\n
deep_ep_set_sm_limit(32); \/\/ \u9650\u5236\u4e3a32\u4e2aSM\r\n<\/code><\/pre>\n<\/li>\n- \u6d4b\u8bd5\u6027\u80fd<\/strong>\uff1a\u8c03\u6574SM\u6570\u91cf\u540e\u8fd0\u884c\u57fa\u51c6\u6d4b\u8bd5\uff0c\u627e\u5230\u6700\u4f73\u914d\u7f6e\u3002<\/li>\n<\/ol>\n
\u6ce8\u610f\u4e8b\u9879<\/h3>\n\n- \u7f51\u7edc\u914d\u7f6e<\/strong>\uff1aDeepEP\u5728InfiniBand\u7f51\u7edc\u4e0a\u6d4b\u8bd5\u6700\u4f18\uff0cRoCE\u9700\u989d\u5916\u9a8c\u8bc1\u517c\u5bb9\u6027\u3002<\/li>\n
- \u81ea\u9002\u5e94\u8def\u7531<\/strong>\uff1a\u4ec5\u4f4e\u5ef6\u8fdf\u5185\u6838\u652f\u6301\u6b64\u529f\u80fd\uff0c\u666e\u901a\u5185\u6838\u542f\u7528\u53ef\u80fd\u5bfc\u81f4\u6b7b\u9501\u3002<\/li>\n
- \u96c6\u7fa4\u8c03\u4f18<\/strong>\uff1a\u5efa\u8bae\u8fd0\u884c\u6240\u6709\u6d4b\u8bd5\u811a\u672c\uff08\u5982
tests\/<\/code>\u76ee\u5f55\u4e0b\uff09\uff0c\u81ea\u52a8\u8c03\u6574\u914d\u7f6e\u4ee5\u9002\u914d\u4f60\u7684\u96c6\u7fa4\u3002<\/li>\n<\/ul>\n\u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u4f60\u53ef\u4ee5\u5feb\u901f\u4e0a\u624bDeepEP\uff0c\u5145\u5206\u5229\u7528\u5176\u5728MoE\u6a21\u578b\u4e2d\u7684\u901a\u4fe1\u4f18\u5316\u80fd\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
DeepEP\u662f\u7531deepseek-ai\u56e2\u961f\u5f00\u53d1\u7684\u4e00\u6b3e\u5f00\u6e90\u901a\u4fe1\u5e93\uff0c\u4e13\u6ce8\u4e8e\u63d0\u5347Mixture-of-Experts\uff08MoE\uff09\u6a21\u578b\u548c\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u6548\u7387\u3002\u5b83\u901a\u8fc7\u4f18\u5316GPU\u95f4\u7684\u6570\u636e\u4ea4\u6362\uff0c\u63d0\u4f9b\u9ad8\u541e\u5410\u91cf\u548c\u4f4e\u5ef6\u8fdf\u7684\u901a\u4fe1\u652f\u6301\uff0c\u9002\u7528\u4e8e\u5927\u89c4\u6a21\u5206\u5e03\u5f0f…<\/p>\n","protected":false},"author":1,"featured_media":61923,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[],"class_list":["post-26915","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/26915","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/comments?post=26915"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/26915\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/media\/61923"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/media?parent=26915"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/categories?post=26915"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/tags?post=26915"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
![\"DeepEP\uff1a\u4e13\u4e3aMoE\u6a21\u578b\u4f18\u5316\u901a\u4fe1\u6548\u7387\u7684\u5f00\u6e90\u5de5\u5177-1\"](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/f8d56b395f6d1af.png\" "\\"DeepEP\uff1a\u4e13\u4e3aMoE\u6a21\u578b\u4f18\u5316\u901a\u4fe1\u6548\u7387\u7684\u5f00\u6e90\u5de5\u5177-1\\"")
<\/p>\n