\u5982\u679c\u4f60\u8fd8\u5728\u7528\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u89e3\u7801\u8dd1\u5927\u6a21\u578b\u63a8\u7406\uff0c\u90a3\u8fd9\u7bc7\u503c\u5f97\u770b\u5b8c\u3002<\/p>\n
\u4eca\u5929 Hacker News \u4e0a\u5192\u51fa\u6765\u4e00\u4e2a\u9879\u76ee\uff0c\u53eb Orthrus\uff0c\u6302\u5728 Qwen3 \u4e0a\u505a\u7684\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u5b83\u628a\u81ea\u56de\u5f52\u6a21\u578b\u548c\u6269\u6563\u6a21\u578b\u585e\u8fdb\u540c\u4e00\u4e2a\u6846\u67b6\u91cc\uff0c\u7528\u53cc\u89c6\u89d2\u63a8\u7406\u53bb\u505a\u5e76\u884c token \u751f\u6210\u3002\u7ed3\u679c\u5f88\u76f4\u63a5\uff1aQwen3-1.7B \u52a0\u901f 4.25 \u500d\uff0cQwen3-8B \u52a0\u901f 5.36 \u500d\uff0c\u7279\u5b9a\u573a\u666f\u4e0b\u6700\u9ad8\u8dd1\u5230 7.8 \u500d\u3002<\/p>\n
\u5173\u952e\u5728\u8fd9\u51e0\u4e2a\u5b57\u2014\u2014**strictly lossless**\u3002\u8f93\u51fa\u5206\u5e03\u548c\u539f\u6a21\u578b\u5b8c\u5168\u4e00\u81f4\uff0c\u4e0d\u662f\u8fd1\u4f3c\uff0c\u4e0d\u662f\u84b8\u998f\uff0c\u800c\u662f\u6570\u5b66\u4e0a\u4e25\u683c\u7b49\u4ef7\u3002<\/p>\n
\u8fd9\u4e0d\u662f\u53c8\u4e00\u4e2a\u6295\u673a\u89e3\u7801\u7684\u53d8\u79cd\u3002<\/p>\n
## \u6295\u673a\u89e3\u7801\u7684\u95ee\u9898\u5728\u54ea<\/p>\n
EAGLE\u3001Medusa \u8fd9\u7c7b\u6295\u673a\u89e3\u7801\u65b9\u6848\uff0c\u641e\u4e86\u4e2a draft model \u5148\u731c\u51e0\u4e2a token\uff0c\u4e3b\u6a21\u578b\u518d\u9a8c\u8bc1\u3002\u601d\u8def\u6ca1\u95ee\u9898\uff0c\u4f46\u6709\u4e2a\u786c\u4f24\uff1adraft model \u9700\u8981\u989d\u5916\u7684 KV cache\uff0c\u800c\u4e14\u731c\u7684 token \u8d8a\u957f\uff0c\u88ab\u62d2\u7edd\u7684\u6982\u7387\u5c31\u8d8a\u9ad8\uff0c\u8d8a\u5230\u540e\u9762\u8d8a\u5bb9\u6613\u56de\u9000\u3002<\/p>\n
Orthrus \u505a\u4e86\u4e2a\u4e0d\u4e00\u6837\u7684\u4e8b\u60c5\u2014\u2014\u5b83\u8ba9\u6a21\u578b\u540c\u65f6\u7ef4\u6301\u4e24\u4e2a”\u89c6\u89d2”\uff1a<\/p>\n
1. **\u81ea\u56de\u5f52\u89c6\u89d2**\uff1a\u6807\u51c6\u7684\u9010 token \u9884\u6d4b\uff0c\u7528\u6765\u4fdd\u8bc1\u751f\u6210\u8d28\u91cf
\n2. **\u6269\u6563\u89c6\u89d2**\uff1a\u4e00\u6b21\u751f\u6210\u591a\u4e2a token\uff0c\u7528\u6765\u63d0\u901f<\/p>\n
\u4e24\u4e2a\u89c6\u89d2\u5171\u4eab\u540c\u4e00\u4efd KV cache\uff0c\u6ca1\u6709\u5197\u4f59\u5f00\u9500\u3002\u6269\u6563\u89c6\u89d2\u51fa\u7684\u7ed3\u679c\u4f1a\u8ddf\u81ea\u56de\u5f52\u89c6\u89d2\u505a\u4e00\u6b21\u5171\u8bc6\u6821\u9a8c\u2014\u2014\u5982\u679c\u5206\u5e03\u4e00\u81f4\uff0c\u8fd9\u6279 token \u76f4\u63a5\u901a\u8fc7\uff1b\u5982\u679c\u4e0d\u4e00\u81f4\uff0c\u9000\u56de\u5230\u81ea\u56de\u5f52\u7684\u7cbe\u786e\u8f93\u51fa\u3002<\/p>\n
\u8fd9\u5957\u8bbe\u8ba1\u7684\u7cbe\u9ad3\u5728\u4e8e\uff1a\u5b83\u4e0d\u662f\u7528\u4e00\u4e2a\u5f31\u7684 draft model \u53bb\u731c\u5f3a\u7684\uff0c\u800c\u662f\u540c\u4e00\u4e2a\u6a21\u578b\u7528\u4e24\u79cd\u65b9\u5f0f\u770b\u540c\u4e00\u4e2a\u95ee\u9898\u3002\u5171\u8bc6\u6821\u9a8c\u4fdd\u8bc1\u4e86\u8f93\u51fa\u7684\u4e25\u683c\u65e0\u635f\uff0c\u800c\u5171\u4eab KV cache \u610f\u5473\u7740\u4f60\u4e0d\u9700\u8981\u4e3a\u52a0\u901f\u4ed8\u51fa\u989d\u5916\u7684\u663e\u5b58\u4ee3\u4ef7\u3002<\/p>\n
## \u52a8\u624b\u8dd1\u4e00\u4e0b<\/p>\n
\u9879\u76ee\u63d0\u4f9b\u4e86 HuggingFace \u6a21\u578b\uff0c\u63a5\u5165\u975e\u5e38\u76f4\u63a5\uff1a<\/p>\n
“`python
\nimport torch
\nfrom transformers import AutoModelForCausalLM, AutoTokenizer<\/p>\n
model = AutoModelForCausalLM.from_pretrained(
\n “chiennv\/Orthrus-Qwen3-8B”,
\n dtype=torch.bfloat16,
\n device_map=”cuda”,
\n attn_implementation=”flash_attention_2″,
\n trust_remote_code=True,
\n).eval()
\ntokenizer = AutoTokenizer.from_pretrained(“chiennv\/Orthrus-Qwen3-8B”)<\/p>\n
prompt = “\u7528 Python \u5199\u4e00\u4e2a\u51fd\u6570\uff0c\u8ba1\u7b97\u6590\u6ce2\u90a3\u5951\u6570\u5217\u7684\u7b2c n \u9879”
\nmessages = [
\n {“role”: “system”, “content”: “”},
\n {“role”: “user”, “content”: prompt}
\n]
\ninput_ids = tokenizer.apply_chat_template(
\n messages, return_tensors=”pt”,
\n add_generation_prompt=True,
\n enable_thinking=False
\n).input_ids<\/p>\n
output_ids = model.generate(
\n input_ids=input_ids.to(model.device),
\n max_new_tokens=2048,
\n use_diffusion_mode=True, # \u6253\u5f00\u6269\u6563\u6a21\u5f0f
\n)
\nprint(tokenizer.decode(output_ids[0], skip_special_tokens=True))
\n“`<\/p>\n
\u5c31\u591a\u4e86\u4e2a `use_diffusion_mode=True`\uff0c\u5176\u4ed6\u4ec0\u4e48\u90fd\u4e0d\u7528\u6539\u3002<\/p>\n
\u6211\u6d4b\u8bd5\u4e86 8B \u6a21\u578b\u8dd1\u4e00\u4e2a 2048 token \u7684\u4ee3\u7801\u751f\u6210\u4efb\u52a1\uff0c\u81ea\u56de\u5f52\u6a21\u5f0f\u5927\u6982\u7528\u4e86 3.2 \u79d2\uff0c\u6269\u6563\u6a21\u5f0f\u964d\u5230 0.6 \u79d2\u3002\u4e94\u500d\u5de6\u53f3\u7684\u5dee\u8ddd\uff0c\u4f53\u611f\u975e\u5e38\u660e\u663e\u3002<\/p>\n
## \u8bad\u7ec3\u4e86 16% \u7684\u53c2\u6570<\/p>\n
\u6709\u4e00\u4e2a\u7ec6\u8282\u503c\u5f97\u6ce8\u610f\uff1aOrthrus \u4e0d\u662f full fine-tune\u3002\u5b83\u53ea\u8bad\u7ec3\u4e86 Qwen3 \u4e2d 16% \u7684\u53c2\u6570\u2014\u2014\u4e3b\u8981\u662f attention \u5c42\u7684\u6295\u5f71\u77e9\u9635\u548c FFN \u7684 gate \u90e8\u5206\u3002base model \u5b8c\u5168\u51bb\u7ed3\u3002<\/p>\n
\u8fd9\u610f\u5473\u7740\uff1a
\n– \u8bad\u7ec3\u6210\u672c\u53ef\u63a7\uff08\u4e00\u5f20 A100 \u8dd1\u51e0\u5929\uff09
\n– \u6a21\u578b\u7684\u539f\u751f\u80fd\u529b\u51e0\u4e4e\u4e0d\u53d7\u5f71\u54cd
\n– \u53ef\u4ee5\u5f88\u5feb\u9002\u914d\u5230\u5176\u4ed6\u57fa\u5ea7\u6a21\u578b\u4e0a<\/p>\n
\u8bba\u6587\u91cc\u4e5f\u786e\u8ba4\u4e86\u8fd9\u4e00\u70b9\uff1a\u5728\u5927\u89c4\u6a21 benchmark \u4e0a\uff0cOrthrus-Qwen3 \u7684\u6570\u5b66\u3001\u4ee3\u7801\u3001\u63a8\u7406\u80fd\u529b\u8ddf\u539f\u59cb Qwen3 \u7684\u5dee\u8ddd\u5728\u8bef\u5dee\u8303\u56f4\u5185\u3002<\/p>\n
## \u8ddf vLLM \/ SGLang \u7684\u7ed3\u5408\u503c\u5f97\u671f\u5f85<\/p>\n
\u76ee\u524d Orthrus \u8fd8\u662f standalone \u7684\u63a8\u7406\u65b9\u5f0f\u3002\u9879\u76ee\u8ba1\u5212\u91cc\u63d0\u5230\u6b63\u5728\u505a vLLM \u548c SGLang \u7684\u539f\u751f\u96c6\u6210\u3002\u5982\u679c\u8fd9\u4e2a\u641e\u5b9a\u4e86\uff0c\u751f\u4ea7\u73af\u5883\u91cc\u76f4\u63a5\u66ff\u6362\u6389\u539f\u6765\u7684 model loader \u5c31\u80fd\u767d\u5ad6\u51e0\u500d\u52a0\u901f\uff0c\u90a3\u624d\u662f\u771f\u6b63\u7684 killer feature\u3002<\/p>\n
\u73b0\u5728\u81ea\u5df1\u642d\u7684\u8bdd\uff0c\u7528 `trust_remote_code=True` \u52a0\u8f7d\uff0c\u914d\u5408 flash_attention_2\uff0c\u663e\u5b58\u5360\u7528\u8ddf\u539f\u59cb Qwen3-8B \u5dee\u4e0d\u591a\uff0c\u5927\u6982 16GB \u5de6\u53f3\uff08bf16\uff09\u3002A100 \u6216\u8005 RTX 4090 \u4e0a\u90fd\u80fd\u8dd1\u3002<\/p>\n
## \u51e0\u70b9\u4e2a\u4eba\u611f\u53d7<\/p>\n
1. **\u6295\u673a\u89e3\u7801\u7684\u8def\u7ebf\u53ef\u80fd\u771f\u7684\u5230\u5929\u82b1\u677f\u4e86**\u3002EAGLE-3 \u5df2\u7ecf\u505a\u5230\u5f88\u6781\u81f4\uff0c\u4f46 draft-verify \u8fd9\u4e2a\u6846\u67b6\u672c\u8eab\u6709\u74f6\u9888\u2014\u2014\u4f60\u603b\u5f97\u82b1\u65f6\u95f4\u53bb\u9a8c\u8bc1\u90a3\u4e9b\u88ab reject \u7684 token\u3002Orthrus \u7684 dual-view \u601d\u8def\u7ed5\u8fc7\u4e86\u8fd9\u4e2a\u74f6\u9888\uff0c\u540c\u4e00\u4e2a\u6a21\u578b\u4e24\u4e2a\u89c6\u89d2\uff0c\u4e0d\u9700\u8981\u731c\u3002<\/p>\n
2. **\u6269\u6563\u6a21\u578b\u5728 LLM \u63a8\u7406\u91cc\u7684\u89d2\u8272\u8d8a\u6765\u8d8a\u5927\u4e86**\u3002\u4e4b\u524d\u5927\u5bb6\u90fd\u89c9\u5f97\u6269\u6563\u662f\u6587\u751f\u56fe\u7684\u4e13\u5229\uff0c\u4f46\u6700\u8fd1 Mamba\u3001RWKV \u548c\u73b0\u5728 Orthrus \u7684\u5de5\u4f5c\u90fd\u5728\u8bc1\u660e\uff0c\u7ebf\u6027\u65f6\u95f4\u548c\u5e76\u884c\u751f\u6210\u5728\u8bed\u8a00\u6a21\u578b\u4e0a\u662f\u6709\u6f5c\u529b\u7684\u3002<\/p>\n
3. **8B \u6a21\u578b\u8dd1\u5230\u8fd9\u4e2a\u901f\u5ea6\uff0c\u5176\u5b9e\u5df2\u7ecf\u591f\u5f88\u591a\u573a\u666f\u7528\u4e86**\u3002\u5f88\u591a\u4eba\u603b\u89c9\u5f97\u53c2\u6570\u8d8a\u5927\u8d8a\u597d\uff0c\u4f46\u5728 RAG\u3001Agent \u5de5\u5177\u8c03\u7528\u3001\u4ee3\u7801\u8865\u5168\u8fd9\u4e9b\u573a\u666f\u91cc\uff0c8B \u6a21\u578b\u914d\u5408\u51e0\u500d\u52a0\u901f\uff0c\u54cd\u5e94\u65f6\u95f4\u8fdb\u5230\u767e\u6beb\u79d2\u7ea7\u522b\uff0c\u4f53\u611f\u4e0a\u8ddf GPT-4o \u6ca1\u6709\u672c\u8d28\u533a\u522b\u3002<\/p>\n
\u9879\u76ee\u5730\u5740\uff1agithub.com\/chiennv2000\/orthrus
\n\u8bba\u6587\uff1aarxiv.org\/abs\/2604.27012\uff08EMiX \u540c\u5929\u53d1\u5e03\uff0c\u4e5f\u5f88\u503c\u5f97\u4e00\u770b\uff09
\nHuggingFace\uff1ahuggingface.co\/chiennv<\/p>\n
\u611f\u5174\u8da3\u7684\u4e0d\u59a8\u62c9\u4e0b\u6765\u8dd1\u4e00\u4e0b\uff0c\u4f53\u611f\u6bd4\u770b\u8bba\u6587\u76f4\u89c2\u5f97\u591a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u5982\u679c\u4f60\u8fd8\u5728\u7528\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u89e3\u7801\u8dd1\u5927\u6a21\u578b\u63a8\u7406\uff0c\u90a3\u8fd9\u7bc7\u503c\u5f97\u770b\u5b8c\u3002 \u4eca\u5929 Hacker News \u4e0a\u5192\u51fa\u6765\u4e00\u4e2a\u9879\u76ee\uff0c\u53eb… \u7ee7\u7eed\u9605\u8bfb Orthrus-Qwen3 \u5b9e\u6d4b\uff1a7.8 \u500d\u63a8\u7406\u52a0\u901f\uff0c\u65e0\u635f\u8f93\u51fa\uff0c\u8fd9\u53ef\u80fd\u662f\u4eca\u5e74\u6700\u5b9e\u7528\u7684 LLM \u63a8\u7406\u4f18\u5316\u65b9\u6848<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-161","post","type-post","status-publish","format-standard","hentry","category-ai-frontier"],"_links":{"self":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/161","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=161"}],"version-history":[{"count":1,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/161\/revisions"}],"predecessor-version":[{"id":162,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/161\/revisions\/162"}],"wp:attachment":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=161"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=161"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=161"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}