{"id":12015,"date":"2026-09-09T01:45:39","date_gmt":"2026-09-08T17:45:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/perplexity-lily-open-source-garden-for-inference-technology\/"},"modified":"2026-09-09T01:45:39","modified_gmt":"2026-09-08T17:45:39","slug":"perplexity-lily-open-source-garden-for-inference-technology","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/perplexity-lily-open-source-garden-for-inference-technology\/","title":{"rendered":"Perplexity \u958b\u6e90 Lily\uff1aMac \u672c\u5730\u63a8\u7406\u5c08\u7528\u63d0\u901f\u5f15\u64ce"},"content":{"rendered":"\n
\"Repository<\/figure>\n\n\n\n

\u7576\u5927\u578b\u6a21\u578b\u958b\u59cb\u8ca0\u8cac\u8655\u7406 Mac \u4e0a\u7684\u79c1\u4eba\u6a94\u6848\u8207\u61c9\u7528\u7a0b\u5f0f\uff0c\u672c\u5730\u63a8\u8ad6\u901f\u5ea6\u5c31\u4e0d\u518d\u53ea\u662f\u958b\u767c\u8005\u5be6\u9a57\u7684\u6307\u6a19\u3002Perplexity \u958b\u6e90\u5605 pplx-garden \u5165\u9762\uff0cLily \u4ee5\u5de5\u5177\u9805\u76ee\u5f62\u5f0f\u8655\u7406 Apple Silicon \u4e0a Qwen3.6-35B-A3B \u7684\u63a8\u8ad6\uff0c\u76ee\u6a19\u4fc2\u4ee4\u63d0\u793a\u8a5e\u8655\u7406\u53ca\u6587\u5b57\u751f\u6210\u66f4\u5feb\uff0c\u4e26\u900f\u904e OpenAI-compatible HTTP API \u4e32\u63a5\u804a\u5929\u6d41\u7a0b\u3002<\/p>\n\n\n\n

Lily \u5514\u4f3c MLX-LM \u822c\u8ffd\u6c42\u652f\u63f4\u591a\u6b3e\u6a21\u578b\uff0c\u800c\u4fc2\u96c6\u4e2d\u670d\u52d9 Qwen3.6-35B-A3B\uff1aRust runtime \u8ca0\u8cac\u8f09\u5165 checkpoint\u3001\u7ba1\u7406 session state \u540c\u751f\u6210\u8ff4\u5708\uff0c\u81ea\u8a02 Metal kernels \u5c31\u8655\u7406\u6a21\u578b\u7279\u5b9a\u904b\u7b97\u3002\u5462\u7a2e\u55ae\u4e00\u9032\u7a0b\u3001\u6a21\u578b\u8207 runtime \u5171\u540c\u5354\u8abf\u5605\u505a\u6cd5\uff0c\u6e1b\u5c11\u901a\u7528 kernel \u5e36\u4f86\u5605\u984d\u5916\u8abf\u5ea6\uff0c\u4ea6\u907f\u958b PyTorch \u548c MLX execution path\u3002<\/p>\n\n\n\n

pplx-garden \u4e0d\u53ea\u662f Mac \u672c\u5730\u63a8\u8ad6\u5de5\u5177\u3002fabric-lib \u63d0\u4f9b RDMA TransferEngine\uff0c\u540c\u6642\u6db5\u84cb P2P MoE dispatch \u8207 combine kernel\uff1bpplx-unigram \u5247\u4fc2\u91dd\u5c0d Unigram tokenizer \u5605 CPU encoder\u3002\u76f8\u95dc\u5167\u5bb9\u4ea6\u5305\u62ec trillion-parameter model \u55ba AWS EFA \u4e0a\u5605\u90e8\u7f72\uff0c\u4ee5\u53ca RL post-training \u6b0a\u91cd\u8f49\u79fb\u3001\u5206\u96e2\u5f0f prefill \u548c decode \u7b49\u7cfb\u7d71\u7814\u7a76\uff0c\u986f\u793a\u5132\u5b58\u5eab\u6db5\u84cb\u7531\u88dd\u7f6e\u7aef\u63a8\u8ad6\u5230\u5206\u6563\u5f0f LLM infrastructure \u5605\u591a\u500b\u74f6\u9838\u3002<\/p>\n\n\n\n

\u9069\u5408\u9700\u8981\u55ba Mac \u8655\u7406\u654f\u611f\u8cc7\u6599\u3001\u53c8\u5e0c\u671b\u4fdd\u7559\u672c\u5730\u751f\u6210\u80fd\u529b\u5605\u958b\u767c\u8005\u53ca\u7814\u7a76\u5718\u968a\uff1b\u9700\u8981 RDMA\u3001MoE \u6e9d\u901a\u6216 tokenizer \u512a\u5316\u5605\u7cfb\u7d71\u5de5\u7a0b\u4eba\u54e1\u4ea6\u53ef\u53c3\u8003\u76f8\u95dc\u5143\u4ef6\u3002\u73fe\u6709\u8cc7\u6599\u96c6\u4e2d\u8b1b\u8ff0 Lily \u5206\u5225\u91cf\u5ea6 prefill throughput \u540c decode throughput\u3002<\/p>\n\n\n\n

\u91cd\u9ede\u6458\u8981\uff1a
\n– Lily \u5c08\u70ba Apple Silicon \u8207 Qwen3.6-35B-A3B \u8a2d\u8a08\uff0c\u652f\u63f4 OpenAI-compatible HTTP API\u3002
\n– \u4ee5 Rust runtime \u914d\u5408\u81ea\u8a02 Metal kernels\uff0c\u5206\u958b\u8655\u7406 prefill \u540c decode\u3002
\n– \u4e0d\u7d93 PyTorch \u6216 MLX execution path\uff0c\u4ee3\u50f9\u4fc2\u6a21\u578b\u53ca\u786c\u4ef6\u652f\u63f4\u7bc4\u570d\u8f03\u5c08\u9580\u3002
\n– fabric-lib \u540c p2p-all-to-all \u9762\u5411 RDMA\u3001MoE dispatch \u8207 combine \u7b49\u5206\u6563\u5f0f\u63a8\u8ad6\u554f\u984c\u3002
\n– \u6548\u80fd\u61c9\u6309\u88dd\u7f6e\u3001\u4e0a\u4e0b\u6587\u53ca\u751f\u6210\u8ca0\u8f09\u5be6\u6e2c\uff0c\u4e0d\u80fd\u53ea\u4f9d\u8cf4\u5b98\u65b9\u5b9a\u4f4d\u4f5c\u6bd4\u8f03\u3002<\/p>\n\n\n\n

\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Perplexity \u63a8\u51fa\u91dd\u5c0d Apple Silicon \u8207 Qwen3.6-35B-A3B \u7684\u672c\u5730\u63a8\u8ad6\u5f15\u64ce Lily\uff0c\u7e5e\u904e PyTorch \u8207 MLX\uff0c\u4ee5 Rust \u548c\u81ea\u8a02 Metal kernels \u6539\u5584\u9810\u586b\u5145\u53ca\u89e3\u78bc\u6548\u7387\u3002<\/p>\n","protected":false},"author":8,"featured_media":12014,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,185,153,137,38,166,152,150,161,76,127,187,199],"tags":[],"class_list":["post-12015","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-openai","category-api","category-38","category-mac","category-python","category-150","category-161","category-76","category-127","category-187","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12015","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=12015"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12015\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12014"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12015"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12015"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12015"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}