{"id":4597,"date":"2026-07-23T06:00:33","date_gmt":"2026-07-22T22:00:33","guid":{"rendered":"http:\/\/viplao.com\/index.php\/2026\/07\/23\/%e7%83%ad%e9%97%a8%e6%8a%80%e8%83%bd-vol-045-deepspec-%e6%8e%a8%e7%90%86%e6%8f%90%e9%80%9f2-3%e5%80%8d\/"},"modified":"2026-07-23T06:00:33","modified_gmt":"2026-07-22T22:00:33","slug":"%e7%83%ad%e9%97%a8%e6%8a%80%e8%83%bd-vol-045-deepspec-%e6%8e%a8%e7%90%86%e6%8f%90%e9%80%9f2-3%e5%80%8d","status":"publish","type":"post","link":"http:\/\/viplao.com\/index.php\/2026\/07\/23\/%e7%83%ad%e9%97%a8%e6%8a%80%e8%83%bd-vol-045-deepspec-%e6%8e%a8%e7%90%86%e6%8f%90%e9%80%9f2-3%e5%80%8d\/","title":{"rendered":"\u70ed\u95e8\u6280\u80fd Vol.045 | DeepSpec \u2014 \u63a8\u7406\u63d0\u901f2-3\u500d"},"content":{"rendered":"
\n

<\/p>\n

\n
\u6bcf\u65e5\u70ed\u95e8\u6280\u80fd\u5206\u4eab<\/div>\n
<\/div>\n
DeepSpec \u2014 \u63a8\u7406\u63d0\u901f2-3\u500d<\/div>\n
<\/div>\n
DeepSeek \u5f00\u6e90\u7684\u63a8\u6d4b\u89e3\u7801\u5168\u6808\u5de5\u5177\u5e93<\/div>\n
<\/div>\n
\n 3 \u5927\u7b97\u6cd5 + 9 \u4e2a\u5f00\u7bb1\u5373\u7528 Checkpoint \u00b7 \u8ba9 Qwen3 \/ Gemma 4 \u63a8\u7406\u5feb 2-3 \u500d<\/span>\n <\/div>\n<\/div>\n

<\/p>\n

<\/div>\n
\n
\u5927\u6a21\u578b\u63a8\u7406\u5360 LLM \u670d\u52a1 80% \u7b97\u529b\u9884\u7b97,\u81ea\u56de\u5f52\u4e00\u6b21\u53ea\u80fd\u751f\u6210 1 \u4e2a token,GPU \u5229\u7528\u7387\u957f\u671f\u4f4e\u4e8e 20%\u3002DeepSeek \u628a\u7ebf\u4e0a\u751f\u4ea7\u540c\u6b3e<\/strong>\u7684\u63a8\u6d4b\u89e3\u7801\u5168\u6808\u6d41\u6c34\u7ebf(DSpark+DFlash+Eagle3)\u5f00\u6e90,\u76f4\u63a5\u7ed9 Qwen3 \/ Gemma 4 \u7b49\u5f00\u6e90\u6a21\u578b\u63d0\u901f 2-3 \u500d<\/strong>\u3002<\/div>\n<\/div>\n

<\/p>\n

<\/div>\n\n\n
\n
~6.7K<\/div>\n
<\/div>\n
Stars<\/div>\n<\/td>\n
\n
3<\/div>\n
<\/div>\n
\u7b97\u6cd5<\/div>\n<\/td>\n
\n
9<\/div>\n
<\/div>\n
\u9884\u8bad\u7ec3\u6743\u91cd<\/div>\n<\/td>\n
\n
85%<\/div>\n
<\/div>\n
\u7ebf\u4e0a\u63d0\u901f<\/div>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\ud83d\udca1 \u4e24\u79cd\u89e3\u7801\u65b9\u5f0f\u5bf9\u6bd4<\/div>\n
<\/div>\n\n\n
\n
\ud83d\ude29 \u666e\u901a\u81ea\u56de\u5f52<\/div>\n
<\/div>\n
\n \u2717 \u4e00\u6b21\u751f\u6210 1 \u4e2a token,\u4e0b\u4e00\u4e2a\u5fc5\u987b\u7b49
\n \u2717 GPU \u7b97\u529b\u5229\u7528\u7387\u957f\u671f\u4f4e\u4e8e 20%
\n \u2717 \u63a8\u7406\u5360 LLM \u670d\u52a1 80% \u7b97\u529b\u6210\u672c
\n \u2717 \u7528\u6237\u54cd\u5e94\u5ef6\u8fdf\u968f\u8f93\u51fa\u957f\u5ea6\u7ebf\u6027\u589e\u957f\n <\/div>\n<\/td>\n
<\/td>\n\n
\ud83d\ude80 DeepSpec \u63a8\u6d4b\u89e3\u7801<\/div>\n
<\/div>\n
\n \u2713 \u5c0f\u6a21\u578b draft 7 \u4e2a token,\u5927\u6a21\u578b\u4e00\u6b21 verify
\n \u2713 \u547d\u4e2d\u7387 78-82%,\u7aef\u5230\u7aef\u52a0\u901f 2-3 \u500d
\n \u2713 9 \u4e2a\u9884\u8bad\u7ec3\u6743\u91cd\u5f00\u7bb1\u5373\u7528,MIT \u534f\u8bae
\n \u2713 \u9ad8\u5e76\u53d1\u573a\u666f\u541e\u5410\u91cf\u66b4\u6da8 661%\n <\/div>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\u2728 \u4e09\u5927\u7b97\u6cd5\u4e00\u7ad9\u96c6\u6210<\/div>\n
<\/div>\n
\u4e0d\u7528\u5206\u522b\u627e\u4e09\u5957\u4ed3\u5e93\u8e29\u4e09\u5957\u4f9d\u8d56<\/div>\n
<\/div>\n\n\n
\n
\n
\u26a1<\/div>\n
<\/div>\n
DSpark<\/div>\n
<\/div>\n
DeepSeek 2026 \u81ea\u7814,\u534a\u81ea\u56de\u5f52 + Confidence-Scheduled \u8c03\u5ea6 + Markov head\u3002\u7ebf\u4e0a\u751f\u4ea7\u540c\u6b3e<\/strong>,\u9ad8\u5e76\u53d1\u573a\u666f\u541e\u5410\u66b4\u6da8 661%\u3002<\/div>\n<\/p><\/div>\n<\/td>\n
\n
\n
\ud83e\uddca<\/div>\n
<\/div>\n
DFlash<\/div>\n
<\/div>\n
MIT z-lab \u51fa\u54c1,\u5e76\u884c\u5757\u6269\u6563(Block Diffusion)\u3002\u5355\u5361\u672c\u5730\u3001\u4f4e\u5e76\u53d1\u573a\u666f\u7684\u9996\u9009<\/strong>,\u5b9e\u73b0\u7b80\u5355\u3001\u663e\u5b58\u53cb\u597d\u3002<\/div>\n<\/p><\/div>\n<\/td>\n
\n
\n
\ud83e\udd85<\/div>\n
<\/div>\n
Eagle3<\/div>\n
<\/div>\n
SGLang \u56e2\u961f\u81ea\u56de\u5f52 + \u6811\u72b6 attention \u6295\u673a\u89e3\u7801\u3002\u591a\u5019\u9009\u63a2\u7d22\u3001\u7075\u6d3b\u5ea6\u4f18\u5148,\u751f\u6001\u6700\u6210\u719f<\/strong>\u7684\u65b9\u6848\u3002<\/div>\n<\/p><\/div>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\ud83d\udce6 9 \u4e2a\u5f00\u7bb1\u5373\u7528 Checkpoint<\/div>\n
<\/div>\n
4 \u4e2a\u76ee\u6807\u6a21\u578b \u00d7 3 \u79cd\u7b97\u6cd5,\u76f4\u63a5\u62c9\u6743\u91cd\u5c31\u80fd\u7528<\/div>\n
<\/div>\n\n\n\n
\n
Qwen3-4B<\/div>\n
<\/div>\n
Eagle3 \u00b7 DFlash \u00b7 DSpark<\/div>\n<\/td>\n
\n
Qwen3-8B<\/div>\n
<\/div>\n
Eagle3 \u00b7 DFlash \u00b7 DSpark<\/div>\n<\/td>\n
\n
Qwen3-14B<\/div>\n
<\/div>\n
Eagle3 \u00b7 DFlash \u00b7 DSpark<\/div>\n<\/td>\n<\/tr>\n
\n
Gemma-4-12B-it<\/div>\n
<\/div>\n
Eagle3 \u00b7 DFlash \u00b7 DSpark<\/div>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\ud83d\ude80 5 \u5206\u949f\u7ed9 Qwen3 \u52a0\u901f 2 \u500d<\/div>\n
<\/div>\n
\u4e0d\u7528\u8bad\u7ec3,\u76f4\u63a5\u62c9\u73b0\u6210 checkpoint \u8d70 vLLM<\/div>\n
<\/div>\n\n\n
\n\n\n\n\n
\n
STEP 1<\/div>\n<\/td>\n
\n
\u62c9\u53d6\u9884\u8bad\u7ec3 draft \u6a21\u578b<\/div>\n
<\/div>\n
\u4ece HuggingFace \u62c9 DSpark \u9002\u914d Qwen3-8B \u7684\u8349\u7a3f\u6743\u91cd,\u8df3\u8fc7\u8bad\u7ec3\u76f4\u63a5\u7528\u3002<\/div>\n
<\/div>\n
snapshot_download(repo_id=”deepseek-ai\/dspark_qwen3_8b_block7″, local_dir=”.\/draft_model”)<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
\n\n\n\n\n
\n
STEP 2<\/div>\n<\/td>\n
\n
\u8dd1\u8bc4\u6d4b\u770b\u52a0\u901f\u6548\u679c<\/div>\n
<\/div>\n
9 \u4e2a\u57fa\u51c6(gsm8k\/math500\/humaneval \u7b49)\u7ed9\u51fa\u63a5\u53d7\u7387 + \u52a0\u901f\u6bd4<\/strong>,\u6bd4\u5982 gsm8k \u63a5\u53d7\u7387 78% \u52a0\u901f 2.3x\u3002<\/div>\n
<\/div>\n
bash scripts\/eval\/eval.sh # \u5217\u51fa 9 \u4e2a\u57fa\u51c6\u7684\u63a5\u53d7\u7387\/\u52a0\u901f\u6bd4<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
\n\n\n
\n
STEP 3<\/div>\n<\/td>\n
\n
\u63a5\u5165 vLLM \u4e0a\u7ebf\u670d\u52a1<\/div>\n
<\/div>\n
vLLM nightly build \u5df2\u5408\u5e76 DSpark \u652f\u6301(PR #46995),\u4e00\u6761\u547d\u4ee4\u542f\u52a8\u52a0\u901f\u670d\u52a1\u3002<\/div>\n
<\/div>\n
vllm serve Qwen\/Qwen3-8B –speculative-model deepseek-ai\/dspark_qwen3_8b_block7 –num-speculative-tokens 5<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\ud83d\udcca \u7efc\u5408\u8bc4\u7ea7<\/div>\n
<\/div>\n
\u57fa\u4e8e\u5de5\u7a0b\u4ef7\u503c\u3001\u53ef\u590d\u73b0\u6027\u548c\u751f\u6001\u6210\u719f\u5ea6\u8bc4\u4f30<\/div>\n
<\/div>\n\n\n
\n\n\n\n\n
\n S<\/span>\n <\/td>\n\n
\u751f\u4ea7\u53ef\u4fe1\u5ea6<\/div>\n
<\/div>\n
DSpark \u5df2\u5728 DeepSeek-V4 \u4e07\u5361\u96c6\u7fa4\u7ebf\u4e0a\u8dd1\u901a,\u4e0d\u662f demo<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
\n\n\n\n\n
\n S<\/span>\n <\/td>\n\n
\u5f00\u7bb1\u5373\u7528\u5ea6<\/div>\n
<\/div>\n
9 \u4e2a HF checkpoint + vLLM \u4e00\u884c\u547d\u4ee4\u63a5\u5165,\u65e0\u9700\u81ea\u5df1\u8bad\u7ec3<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
\n\n\n\n\n
\n A+<\/span>\n <\/td>\n\n
\u7b97\u6cd5\u5b8c\u6574\u5ea6<\/div>\n
<\/div>\n
DSpark\/DFlash\/Eagle3 \u4e00\u7ad9\u96c6\u6210,\u4e0d\u7528\u5206\u522b\u8e29\u4e09\u5957\u4f9d\u8d56<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n
<\/td>\n<\/tr>\n
\n\n\n
\n B+<\/span>\n <\/td>\n\n
\u786c\u4ef6\u95e8\u69db<\/div>\n
<\/div>\n
\u9ed8\u8ba4\u914d\u7f6e\u8981 8\u00d7A100\/H100,\u5c0f\u56e2\u961f\u9700\u51cf\u5c11 GPU \u6216\u76f4\u63a5\u62c9 checkpoint<\/div>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n<\/table>\n

<\/p>\n

<\/div>\n
\n
\ud83c\udfaf \u9002\u5408\u8c01?<\/div>\n
<\/div>\n
\u81ea\u6258\u7ba1\u5927\u6a21\u578b\u60f3\u538b\u63a8\u7406\u6210\u672c\u7684\u4f01\u4e1a\/\u56e2\u961f;\u60f3\u7ed9 Qwen3 \/ Gemma 4 \u52a0\u901f 2-3 \u500d\u7684\u5de5\u7a0b\u5e08;\u7814\u7a76\u63a8\u6d4b\u89e3\u7801\u539f\u7406\u9700\u8981\u53ef\u590d\u73b0 pipeline \u7684\u5b66\u751f\/\u7814\u7a76\u5458;\u4ee5\u53ca\u6240\u6709\u5728\u4e3a”GPU \u7b97\u529b\u8d35\/\u54cd\u5e94\u6162”\u5934\u75bc\u7684 LLM \u670d\u52a1\u65b9\u3002<\/div>\n<\/div>\n

<\/p>\n

<\/div>\n
\n
\u5148\u628a GPU \u7b97\u529b\u6253 5 \u6298,\u518d\u8c08\u66f4\u5927\u7684\u6a21\u578b<\/div>\n
<\/div>\n
DSpark \u5df2\u7ecf\u5728 DeepSeek \u7ebf\u4e0a\u4e07\u5361\u96c6\u7fa4\u8dd1\u901a,\u4e0d\u662f PPT \u52a0\u901f<\/div>\n
<\/div>\n
github.com\/deepseek-ai\/DeepSpec<\/div>\n<\/div>\n

<\/p>\n

\n
\u5982\u679c\u89c9\u5f97\u6709\u7528,\u8f6c\u53d1\u7ed9\u88ab GPU \u8d26\u5355\u523a\u75db\u7684\u670b\u53cb<\/strong><\/div>\n
<\/div>\n
\u597d\u7684\u5185\u5bb9\u4e0d\u8be5\u53ea\u505c\u5728\u6d4f\u89c8\u5668\u6807\u7b7e\u9875\u91cc,\u5b83\u5e94\u8be5\u8d70\u8fdb\u4e00\u4e2a\u4eba\u7684\u884c\u52a8\u3002<\/div>\n
<\/div>\n
\u26a0\ufe0f \u6280\u80fd\u6765\u6e90\u4e8e\u4e92\u8054\u7f51,\u4f7f\u7528\u524d\u8bf7\u5148\u6838\u5b9e\u5176\u5b89\u5168\u6027<\/div>\n
<\/div>\n
\u70ed\u95e8\u6280\u80fd Vol.045 \u00b7 \u4f5c\u8005:\u4e00\u679d\u503c\u5343\u91d1 \u00b7 2026-07-23<\/div>\n<\/div>\n<\/div>\n
\n

\u8d44\u6e90\u94fe\u63a5<\/strong>
GitHub: https:\/\/github.com\/deepseek-ai\/DeepSpec<\/a>
\u534f\u8bae: MIT<\/p>\n<\/section>\n","protected":false},"excerpt":{"rendered":"

\u6bcf\u65e5\u70ed\u95e8\u6280\u80fd\u5206\u4eab DeepSpec \u2014 \u63a8\u7406\u63d0\u901f2-3\u500d DeepSeek \u5f00\u6e90\u7684\u63a8\u6d4b\u89e3\u7801\u5168\u6808\u5de5\u5177\u5e93… \u7ee7\u7eed\u9605\u8bfb \u70ed\u95e8\u6280\u80fd Vol.045 | DeepSpec \u2014 \u63a8\u7406\u63d0\u901f2-3\u500d<\/span><\/i><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[49],"tags":[],"views":9,"_links":{"self":[{"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/posts\/4597"}],"collection":[{"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/comments?post=4597"}],"version-history":[{"count":0,"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/posts\/4597\/revisions"}],"wp:attachment":[{"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/media?parent=4597"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/categories?post=4597"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/viplao.com\/index.php\/wp-json\/wp\/v2\/tags?post=4597"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}