[{"data":1,"prerenderedAt":2020},["ShallowReactive",2],{"post-\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab":3,"archive-public-posts":1514},{"id":4,"title":5,"body":6,"categories":1491,"cover":1493,"coverAlt":1494,"date":1495,"description":1496,"draft":1497,"extension":1498,"haloPublished":1499,"haloSync":1499,"hidden":1497,"legacy":1497,"meta":1500,"navigation":1499,"path":1501,"published":1497,"readingTime":1502,"seo":1503,"series":1502,"seriesOrder":1502,"stem":1504,"tags":1505,"__hash__":1513},"posts\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab.md","从 Mac 到单卡 GPU：Qwen2.5-VL 部署、GRPO 与计数 LoRA 实践",{"type":7,"value":8,"toc":1471},"minimark",[9,13,16,24,29,83,86,89,97,101,104,114,117,128,131,137,140,143,146,203,206,216,220,223,273,280,292,295,298,334,337,340,344,351,354,401,404,407,414,418,421,437,440,447,494,500,506,509,512,515,648,651,655,658,694,697,703,709,712,715,721,724,731,738,800,803,806,809,812,869,876,879,890,909,912,916,919,926,929,933,936,939,989,999,1005,1008,1011,1072,1075,1079,1082,1085,1088,1095,1207,1214,1220,1224,1227,1230,1233,1237,1240,1243,1270,1273,1279,1309,1313,1334,1341,1344,1347,1350,1373,1376,1379,1382,1464,1467],[10,11,12],"p",{},"这次实验从一个很具体的问题开始：M2 Pro、16 GB 统一内存，能不能在本机跑 Qwen2.5-VL？跑起来之后，我又想把图文问答、推理性能测量和强化学习训练接成一条完整流程。",[10,14,15],{},"本机的 MLX 4-bit 服务和云端的 SGLang BF16 服务都能进行图文问答。最初 3 步、随后 30 步 GRPO 都完成了真实参数更新和训练前后评测，但没有观察到可靠的质量提升。后续改用有监督语言 LoRA，并根据独立压力测试补充训练数据，合成计数任务才有了明确改善。",[10,17,18,19,23],{},"本文最初发布于 2026 年 8 月 19 日，于 ",[20,21,22],"strong",{},"2026 年 8 月 21 日补充更新","。部署、推理优化和最初 3 步 GRPO 的记录保留，后半部分补充 30 步 GRPO、SFT 扩展、压力测试及两轮定向补训。后续提升不能归因于 GRPO，也不能当成真实照片计数能力的证明。",[25,26,28],"h2",{"id":27},"两个环境各自解决什么问题","两个环境，各自解决什么问题",[30,31,32,51],"table",{},[33,34,35],"thead",{},[36,37,38,42,45,48],"tr",{},[39,40,41],"th",{},"环境",[39,43,44],{},"模型与精度",[39,46,47],{},"后端",[39,49,50],{},"本轮用途",[52,53,54,69],"tbody",{},[36,55,56,60,63,66],{},[57,58,59],"td",{},"MacBook Pro，M2 Pro，16 GB 统一内存",[57,61,62],{},"Qwen2.5-VL-3B，社区 MLX 4-bit 转换版",[57,64,65],{},"MLX-VLM 0.7.6",[57,67,68],{},"本地图文问答、批量调用、低并发测量",[36,70,71,74,77,80],{},[57,72,73],{},"云端单卡，设备标称 RTX 4080 SUPER，报告显存约 32 GiB",[57,75,76],{},"原始 Qwen2.5-VL-3B-Instruct，BF16",[57,78,79],{},"SGLang 0.5.5",[57,81,82],{},"CUDA 推理、调参、verl 0.6.1 多模态 GRPO",[10,84,85],{},"云端设备报告了 32760 MiB 显存，这里按实际环境描述，不把它当作标准零售显卡的规格。训练环境固定为 Python 3.12、PyTorch 2.8.0+cu128、Transformers 4.57.1；Mac 的 MLX 环境单独维护。",[10,87,88],{},"两边硬件、精度和后端都不同，性能数字用于说明各自的行为。直接拿它们计算“哪个推理框架快几倍”，无法排除这些变量。",[10,90,91,92,96],{},"本机 4-bit 权重约 3.07 GB，但运行内存还要容纳视觉计算、KV Cache 和临时张量。第一次测试能读出 ",[93,94,95],"code",{},"LOCAL TEST 42","，也能识别颜色，却把红色正方形称为矩形。接口返回、页面可用和形状判断正确，需要分别检查。我保留了这条错误，没有为了让验收通过而改判定标准。",[25,98,100],{"id":99},"一次图文回答怎样走到首个-token","一次图文回答怎样走到首个 token",[10,102,103],{},"Qwen2.5-VL 的输入链路可以概括为：",[105,106,112],"pre",{"className":107,"code":109,"language":110,"meta":111},[108],"language-text","图片与问题\n  → 图片预处理、视觉编码\n  → 视觉 token 与文本 token 组成语言模型输入\n  → Prefill：处理输入，建立各层 KV 状态\n  → 首个输出 token\n  → Decode：复用历史 KV，逐步生成后续 token\n","text","",[93,113,109],{"__ignoreMap":111},[10,115,116],{},"Prefill 处理已经知道的输入序列；Decode 则根据历史输入和已生成内容继续预测。每步 Decode 会增加新位置的 K\u002FV，避免重新计算整个历史前缀。图片越大，processor 往往生成越多视觉 token，视觉编码、Prefill 和缓存占用也会随之变化。",[10,118,119,120,123,124,127],{},"实测 processor 对 448×336 图片输出 ",[93,121,122],{},"image_grid_thw=[1,24,32]","，空间合并系数为 2，因此得到 ",[93,125,126],{},"1×24×32÷2²=192"," 个视觉 token。加上文本和模板后，主性能实验的输入总长度是 335 tokens。",[10,129,130],{},"KV Cache 保存注意力中间状态。普通全注意力、各层结构相同且 KV 未量化时，可作如下估算：",[105,132,135],{"className":133,"code":134,"language":110,"meta":111},[108],"KV bytes ≈ 2 × L × B × T × Hkv × Dhead × bytes_per_element\n",[93,136,134],{"__ignoreMap":111},[10,138,139],{},"其中 L 是层数，B 是活跃序列数，T 是每条序列长度，Hkv 是 KV 头数。对本次模型的 36 层、2 个 KV 头、每头 128 维和 2-byte KV，每 token、每序列约占 36 KiB；4096 tokens 约占 144 MiB。这个估算不含权重、视觉激活、临时张量和缓存池空闲空间。GQA 应使用 KV 头数，不能把 16 个 Query 头直接代入；4-bit 权重量化也不代表 KV 自动变成 4-bit。",[25,141,142],{"id":142},"先把测量口径固定下来",[10,144,145],{},"我用流式接口记录首个非空文本片段，同时保留实际输出 token 数和停止原因。几个指标的含义如下：",[30,147,148,158],{},[33,149,150],{},[36,151,152,155],{},[39,153,154],{},"指标",[39,156,157],{},"本次口径",[52,159,160,168,176,184,192],{},[36,161,162,165],{},[57,163,164],{},"TTFT，首字延迟",[57,166,167],{},"HTTP 请求发出至首个非空文本 delta，不计只有 role 的片段",[36,169,170,173],{},[57,171,172],{},"完整响应延迟",[57,174,175],{},"请求发出至流结束，包含后续 Decode",[36,177,178,181],{},[57,179,180],{},"整组输出吞吐",[57,182,183],{},"该组成功请求的实际输出 token 总数，除以整组墙钟时间",[36,185,186,189],{},[57,187,188],{},"Mac 内存峰值",[57,190,191],{},"MLX allocator 的进程峰值，十进制 GB",[36,193,194,197],{},[57,195,196],{},"GPU 显存采样峰值",[57,198,199,202],{},[93,200,201],{},"nvidia-smi"," 每 0.2 秒采样的设备占用最大值，MiB",[10,204,205],{},"TTFT 包含请求提交后的排队、传输、服务端图像处理、视觉编码和语言模型 Prefill；客户端读图、缩放和编码在计时前完成，不能把 TTFT 全部归为纯 Prefill 计算。GPU 性能测量客户端运行在云端机器本机，不包含 Mac 到云端的网络耗时。",[10,207,208,211,212,215],{},[93,209,210],{},"max_tokens=128"," 只规定上限，模型可能提前结束。性能实验要比较相同输出负载：主实验检查实际输出均达到上限，九格调参进一步使用 ",[93,213,214],{},"min_tokens=max_tokens=128","。这类回答会被长度限制截断，适合测计算负载；质量评测则允许模型正常结束，保留完整回答。",[25,217,219],{"id":218},"并发提高了吞吐也增加了等待","并发提高了吞吐，也增加了等待",[10,221,222],{},"GPU 主实验固定 448×336 图片、335-token 输入、128-token 输出，服务端活跃序列上限为 4，关闭前缀缓存。改变客户端并发后得到：",[30,224,225,238],{},[33,226,227],{},[36,228,229,233,236],{},[39,230,232],{"align":231},"right","客户端并发",[39,234,235],{"align":231},"TTFT p50",[39,237,180],{"align":231},[52,239,240,251,262],{},[36,241,242,245,248],{},[57,243,244],{"align":231},"1",[57,246,247],{"align":231},"67 ms",[57,249,250],{"align":231},"97.43 tokens\u002Fs",[36,252,253,256,259],{},[57,254,255],{"align":231},"2",[57,257,258],{"align":231},"107 ms",[57,260,261],{"align":231},"168.00 tokens\u002Fs",[36,263,264,267,270],{},[57,265,266],{"align":231},"4",[57,268,269],{"align":231},"166 ms",[57,271,272],{"align":231},"341.41 tokens\u002Fs",[10,274,275],{},[276,277],"img",{"alt":278,"src":279},"GPU 并发从 1 增至 4 时，输出吞吐与首字延迟同时增加的实测图","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-gpu-concurrency-20261010.webp",[10,281,282,283,291],{},"图 1：本轮 GPU 并发实验。主实验七个场景共 48 次测量；并发 4 场景有 12 个请求，其他场景各 6 个，预热不计入统计。",[284,285,290],"a",{"href":279,"target":286,"rel":287},"_blank",[288,289],"noopener","noreferrer","查看大图","。",[10,293,294],{},"Mac 上也观察到同样的取舍：并发 1、2、4 时，吞吐为 57.36、73.80、81.01 tokens\u002Fs，TTFT p50 为 0.676、1.183、2.453 秒。交互问答往往更在意等待时间，批量离线任务则可能更在意总吞吐，选参数时需要先确定目标。",[10,296,297],{},"客户端并发 C 和服务端活跃批大小 B 是两个设置。为了检查排队的影响，我固定 C=4，只改变 GPU 服务的活跃上限：",[30,299,300,311],{},[33,301,302],{},[36,303,304,307,309],{},[39,305,306],{"align":231},"活跃序列上限",[39,308,235],{"align":231},[39,310,180],{"align":231},[52,312,313,324],{},[36,314,315,318,321],{},[57,316,317],{"align":231},"B=1",[57,319,320],{"align":231},"3.931 秒",[57,322,323],{"align":231},"98.44 tokens\u002Fs",[36,325,326,329,332],{},[57,327,328],{"align":231},"B=4",[57,330,331],{"align":231},"0.166 秒",[57,333,272],{"align":231},[10,335,336],{},"这组负载下，B=4 的吞吐约为 B=1 的 3.47 倍。C=4、B=1 时，四个请求并没有同时在 GPU 上推进，后续请求需要排队。连续批处理允许完成的请求退出、等待请求加入，提高活跃槽位的利用率，但它不会保证所有请求都获得相同收益。",[10,338,339],{},"分辨率和输出长度也会改变负载。GPU 实验中，图片从 224×168 放大到 896×672，输入长度从 191 增至 911 tokens，TTFT p50 从 53 增至 164 ms；输出上限从 64 增至 256 tokens，完整响应 p50 从 0.681 增至 2.577 秒。同一张图放大不会新增真实细节，这组实验用于观察计算成本。",[25,341,343],{"id":342},"前缀缓存需要命中证据","前缀缓存，需要命中证据",[10,345,346,347,350],{},"KV Cache 支持当前请求继续 Decode；前缀缓存则复用此前请求相同输入前缀的 KV。SGLang 的 Radix Cache 用树组织可共享的前缀。使用 ",[93,348,349],{},"--disable-radix-cache"," 关闭前缀复用后，当前请求仍然需要并使用自己的 KV Cache。",[10,352,353],{},"我用相同图片和相同的 1741-token 图文前缀，分别测五对冷、暖请求：每对先清理前缀缓存，再发冷请求，随后立即重复输入。",[30,355,356,372],{},[33,357,358],{},[36,359,360,363,366,369],{},[39,361,362],{},"配置",[39,364,365],{"align":231},"冷请求 TTFT p50",[39,367,368],{"align":231},"重复请求 TTFT p50",[39,370,371],{"align":231},"重复请求 cached tokens",[52,373,374,388],{},[36,375,376,379,382,385],{},[57,377,378],{},"前缀缓存开启",[57,380,381],{"align":231},"194 ms",[57,383,384],{"align":231},"75 ms",[57,386,387],{"align":231},"1740",[36,389,390,393,396,398],{},[57,391,392],{},"前缀缓存关闭",[57,394,395],{"align":231},"177 ms",[57,397,395],{"align":231},[57,399,400],{"align":231},"0",[10,402,403],{},"开启组的暖请求明确报告复用了 1740 个 token。关闭组没有同等的延迟下降，因此这里同时有缓存命中和时间差的证据。只看到第二次更快，还可能是编译、内核或其他预热效果。",[10,405,406],{},"SGLang 0.5.5 会省略零缓存量的详情字段。探针核对版本、cache-report 和开关状态后，才按该版本语义把这种省略解释为零，并保留原始 usage。",[10,408,409,410,413],{},"还有一处容易误读：视觉特征缓存复用图像 embedding，与语言模型前缀缓存属于不同层。MLX-VLM 虽然包含通用 ",[93,411,412],{},"VisionFeatureCache","，当前 Qwen2.5-VL 路径并不直接读写它。配置一个视觉缓存容量，不能据此宣称重复图片已经命中。",[25,415,417],{"id":416},"sglang-九格调参再独立复测","SGLang 九格调参，再独立复测",[10,419,420],{},"接下来扫描两个参数：",[422,423,424,431],"ul",{},[425,426,427,430],"li",{},[93,428,429],{},"chunked_prefill_size","：256、512、1024。",[425,432,433,436],{},[93,434,435],{},"mem_fraction_static","：0.55、0.70、0.80。",[10,438,439],{},"这一轮固定 1735-token 长图文输入、C4\u002FB4、原始 BF16 模型、前缀缓存关闭和 128-token 实际输出。每组独立重启服务，先运行 4 条预热，再测 12 条请求。首次尝试出现提前 EOS，我单独归档了这些结果，重新固定输出长度后再比较。",[10,441,442,443,446],{},"筛选规则在执行前写定：排除失败或输出不一致的组合，要求吞吐至少达到基线的 95%，再比较 TTFT p95；p95 接近时优先选较小的显存池。选出的 ",[93,444,445],{},"chunk1024 \u002F mem0.55","，还要和基线分别重新启动，各测一组 12 请求：",[30,448,449,464],{},[33,450,451],{},[36,452,453,456,459,461],{},[39,454,455],{},"独立复测配置",[39,457,458],{"align":231},"TTFT p95",[39,460,180],{"align":231},[39,462,463],{"align":231},"设备显存采样峰值",[52,465,466,480],{},[36,467,468,471,474,477],{},[57,469,470],{},"基线：chunk512 \u002F mem0.70",[57,472,473],{"align":231},"664 ms",[57,475,476],{"align":231},"252.27 tokens\u002Fs",[57,478,479],{"align":231},"23243 MiB",[36,481,482,485,488,491],{},[57,483,484],{},"候选：chunk1024 \u002F mem0.55",[57,486,487],{"align":231},"595 ms",[57,489,490],{"align":231},"260.34 tokens\u002Fs",[57,492,493],{"align":231},"18533 MiB",[10,495,496],{},[276,497],{"alt":498,"src":499},"SGLang 独立复测中首字延迟、输出吞吐和设备显存采样峰值的对照图","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-sglang-confirmation-20261010.webp",[10,501,502,503,291],{},"图 2：独立确认数据。候选 TTFT p95 下降约 10.5%，吞吐增加约 3.2%，采样显存峰值减少 4710 MiB，约 4.60 GiB。九格加两组确认共 132 次正式测量，全部成功。",[284,504,290],{"href":499,"target":286,"rel":505},[288,289],[10,507,508],{},"显存池缩小后，服务报告的 KV 容量也从 424083 降至 287664 tokens。本轮 C4 负载没有触及容量限制，所以减少闲置池占用更划算；更高并发或更长输入需要重新测量。显存采样包含模型加载、预热和预分配池，可能漏过瞬时峰值；每组只有 12 个请求，p95 也不能当作生产 SLA。",[10,510,511],{},"这轮长输入与前面的 335-token 并发实验不同，不能把 260 和 341 tokens\u002Fs 当成同一个负载的前后对照。",[10,513,514],{},"最终交互服务沿用原始 BF16 模型，使用 TP1、B4、chunk1024、mem0.55，并开启前缀缓存。固定 SGLang 0.5.5 环境准备好后，关键启动参数是：",[105,516,520],{"className":517,"code":518,"language":519,"meta":111,"style":111},"language-bash shiki shiki-themes github-dark-dimmed github-light houston","python -m sglang.launch_server \\\n  --model-path .\u002Fmodels\u002FQwen2.5-VL-3B-Instruct \\\n  --host 127.0.0.1 --port 30000 \\\n  --dtype bfloat16 --context-length 4096 \\\n  --tp-size 1 --max-running-requests 4 \\\n  --chunked-prefill-size 1024 --mem-fraction-static 0.55 \\\n  --attention-backend triton --mm-attention-backend sdpa \\\n  --enable-metrics --enable-cache-report\n","bash",[93,521,522,543,554,571,588,605,622,639],{"__ignoreMap":111},[523,524,527,531,535,539],"span",{"class":525,"line":526},"line",1,[523,528,530],{"class":529},"sLsJe","python",[523,532,534],{"class":533},"sgENW"," -m",[523,536,538],{"class":537},"sWqY2"," sglang.launch_server",[523,540,542],{"class":541},"sxpgZ"," \\\n",[523,544,546,549,552],{"class":525,"line":545},2,[523,547,548],{"class":533},"  --model-path",[523,550,551],{"class":537}," .\u002Fmodels\u002FQwen2.5-VL-3B-Instruct",[523,553,542],{"class":541},[523,555,557,560,563,566,569],{"class":525,"line":556},3,[523,558,559],{"class":533},"  --host",[523,561,562],{"class":533}," 127.0.0.1",[523,564,565],{"class":533}," --port",[523,567,568],{"class":533}," 30000",[523,570,542],{"class":541},[523,572,574,577,580,583,586],{"class":525,"line":573},4,[523,575,576],{"class":533},"  --dtype",[523,578,579],{"class":537}," bfloat16",[523,581,582],{"class":533}," --context-length",[523,584,585],{"class":533}," 4096",[523,587,542],{"class":541},[523,589,591,594,597,600,603],{"class":525,"line":590},5,[523,592,593],{"class":533},"  --tp-size",[523,595,596],{"class":533}," 1",[523,598,599],{"class":533}," --max-running-requests",[523,601,602],{"class":533}," 4",[523,604,542],{"class":541},[523,606,608,611,614,617,620],{"class":525,"line":607},6,[523,609,610],{"class":533},"  --chunked-prefill-size",[523,612,613],{"class":533}," 1024",[523,615,616],{"class":533}," --mem-fraction-static",[523,618,619],{"class":533}," 0.55",[523,621,542],{"class":541},[523,623,625,628,631,634,637],{"class":525,"line":624},7,[523,626,627],{"class":533},"  --attention-backend",[523,629,630],{"class":537}," triton",[523,632,633],{"class":533}," --mm-attention-backend",[523,635,636],{"class":537}," sdpa",[523,638,542],{"class":541},[523,640,642,645],{"class":525,"line":641},8,[523,643,644],{"class":533},"  --enable-metrics",[523,646,647],{"class":533}," --enable-cache-report\n",[10,649,650],{},"模型路径需要指向原始 HF 权重。本次设备不支持 Hopper 专用的 FA3，语言注意力使用 Triton、视觉注意力使用 SDPA。上述配置对应当前实验，换版本或换硬件后应重新核对可用参数与实际服务配置。",[25,652,654],{"id":653},"用看图计数跑通多模态-grpo","用看图计数跑通多模态 GRPO",[10,656,657],{},"我选择了一个能用规则评分的任务：数出图片中的红色圆形和蓝色正方形，按固定 JSON 返回结果。其他颜色和形状作为干扰项。",[105,659,663],{"className":660,"code":661,"language":662,"meta":111,"style":111},"language-json shiki shiki-themes github-dark-dimmed github-light houston","{\"red_circles\": 3, \"blue_squares\": 2}\n","json",[93,664,665],{"__ignoreMap":111},[523,666,667,671,675,678,681,684,687,689,691],{"class":525,"line":526},[523,668,670],{"class":669},"sWURj","{",[523,672,674],{"class":673},"sjXwz","\"red_circles\"",[523,676,677],{"class":669},": ",[523,679,680],{"class":533},"3",[523,682,683],{"class":669},", ",[523,685,686],{"class":673},"\"blue_squares\"",[523,688,677],{"class":669},[523,690,255],{"class":533},[523,692,693],{"class":669},"}\n",[10,695,696],{},"这只是输出格式示例。训练集包含 128 张合成图，最初留出集为 32 张；图片以 bytes 写入 Parquet，避免依赖本机绝对路径。真值进入奖励字段，不进入图片或提示词。",[10,698,699],{},[276,700],{"alt":701,"src":702},"用输入图形卡、四张采样卡和参数旋钮表达采样、比较与小幅更新的概念插画","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-grpo-study-20261010.webp",[10,704,705,706,291],{},"图 3：原创概念插画，AI 辅助绘制。卡片和旋钮只表达采样与更新，不对应真实回答、奖励值或能力增益。",[284,707,290],{"href":702,"target":286,"rel":708},[288,289],[10,710,711],{},"每个图片问题采样 4 个回答，temperature 为 1.0。每步有 4 个不同问题，因此产生 16 条回答。奖励规则很简单：格式合法且两项都对得 1，只对一项得 0.5，格式不合法或两项都错得 0。格式校验还会拒绝重复键、额外字段、布尔值和尾部解释。",[10,713,714],{},"GRPO 比较同一个问题的四个回答。固定版本的组内优势可写为：",[105,716,719],{"className":717,"code":718,"language":110,"meta":111},[108],"A_i = (r_i - mean(r_1, ..., r_4)) \u002F (std(r_1, ..., r_4) + 1e-6)\n",[93,720,718],{"__ignoreMap":111},[10,722,723],{},"如果一组奖励完全相同，归一化后的优势为零；保存一个 checkpoint 本身不足以证明获得了有效学习信号。我同时检查奖励差异、梯度和权重变化。",[10,725,726,727,730],{},"训练使用原始 BF16 权重，冻结视觉塔，采用 FSDP2、gradient checkpointing、参数与优化器 CPU offload，学习率为 ",[93,728,729],{},"1e-6","，不加载 critic 或 KL reference。总共只跑 3 步，实际采样 12 组图片问题，目标是核验整个流程。",[10,732,733,734,737],{},"第一次反向传播完成后，AdamW 的 optimizer step 显存不足。将 ",[93,735,736],{},"foreach=False"," 后，逐个矩阵更新减少了这段临时张量的峰值，训练才完整跑完。CPU offload 也不意味着整个优化器更新都在 CPU 上执行。",[30,739,740,750],{},[33,741,742],{},[36,743,744,747],{},[39,745,746],{},"核验项目",[39,748,749],{},"实际结果",[52,751,752,760,768,776,784,792],{},[36,753,754,757],{},[57,755,756],{},"参数更新",[57,758,759],{},"完成 3 步，48 条训练采样",[36,761,762,765],{},[57,763,764],{},"同问题组内奖励差异",[57,766,767],{},"12 组中有 9 组",[36,769,770,773],{},[57,771,772],{},"三步梯度范数",[57,774,775],{},"2.328125、3.78125、2.5",[36,777,778,781],{},[57,779,780],{},"完整 HF 导出",[57,782,783],{},"824 个 tensor 的名称、shape、dtype 与值逐一核对",[36,785,786,789],{},[57,787,788],{},"语言层变化",[57,790,791],{},"434 个语言 tensor 中 392 个改变，最大绝对差约 5.72×10^-6",[36,793,794,797],{},[57,795,796],{},"视觉层冻结",[57,798,799],{},"390 个视觉 tensor 全部保持原值",[10,801,802],{},"这些证据说明采样、奖励、反向传播、更新、导出和重新加载都执行了。保存的 checkpoint 只含模型，没有 Adam 状态，适合作为权重产物，但不能保证精确恢复优化器进度。",[25,804,805],{"id":805},"训练后到底有没有变好",[10,807,808],{},"训练程序最初的验证得到了 7\u002F32 全对。后续受控评测的服务与解码条件不同，这个数字不能直接当作前测基线。为了比较效果，我把原始模型和 step3 导出模型分别独立启动，用完全相同的输入和实际服务控制评测。",[10,810,811],{},"两份模型都采用 SGLang 0.5.5、BF16、TP1\u002FB1、chunk2048 \u002F mem0.60、seed42、贪心解码和 64-token 输出上限，关闭前缀缓存、overlap 与 CUDA graph，开启 deterministic inference；仅模型路径不同。计数评测包含原 32 张留出图和另外 128 张新图，按文件和 RGB 像素哈希排除与训练集重复。",[30,813,814,826],{},[33,815,816],{},[36,817,818,820,823],{},[39,819,154],{},[39,821,822],{"align":231},"原始 BF16",[39,824,825],{"align":231},"GRPO step3",[52,827,828,838,849,859],{},[36,829,830,833,836],{},[57,831,832],{},"160 张计数图两项全对率",[57,834,835],{"align":231},"22\u002F160，13.75%",[57,837,835],{"align":231},[36,839,840,843,846],{},[57,841,842],{},"160 张计数图平均奖励",[57,844,845],{"align":231},"0.384375",[57,847,848],{"align":231},"0.38125",[36,850,851,854,857],{},[57,852,853],{},"计数 JSON 格式合法率",[57,855,856],{"align":231},"100%",[57,858,856],{"align":231},[36,860,861,864,867],{},[57,862,863],{},"TextVQA64，官方方法 soft score",[57,865,866],{"align":231},"81.25%",[57,868,866],{"align":231},[10,870,871,872,875],{},"计数平均奖励差为 -0.003125，配对图片 bootstrap 的 95% 区间是 ",[93,873,874],{},"[-0.015625, 0.00625]","，包含零。6 张图的计数答案改变，但全对与否的逐图判定全部不变。这里没有观察到提升证据，也不足以确定整体退化。",[10,877,878],{},"真实图片从 TextVQA v0.5.1 validation 的固定镜像分片中，按预测前固定的 SHA 排序选取 64 张不同图片，每图一题、10 个人类参考答案。输入只含图片和问题，不传额外 OCR tokens 或参考答案；评分采用固定的官方 MMF EvalAI 规范化和留一标注者 soft score。",[10,880,881,882,885,886,889],{},"两份模型在这 64 题上逐题同分，单模型分数的图片 bootstrap 95% 区间为 ",[93,883,884],{},"[72.1875%, 89.53125%]","。这是分片小样本，不能当成官方 5000 题的全量成绩，预训练是否见过图片也未知。配对差区间为 ",[93,887,888],{},"[0,0]","，仅表示本样本评分差全部为零。",[10,891,892,893,896,897,900,901,904,905,908],{},"错误复核也很有用：有把 ",[93,894,895],{},"BIG NORM"," 读成 ",[93,898,899],{},"BIG NORTH"," 的 OCR 错误，也有目标选择失配。还有回答包含正确文字、但答案范围不符合参考的情况；例如单位 ",[93,902,903],{},"200ml"," 和 ",[93,906,907],{},"200 ml"," 在固定评分规则下并不相同。我保留完整回答，不在看到结果后改评分规则。",[10,910,911],{},"计数前后共 320 个请求，真实图片前后共 128 个请求，448 个评测请求全部成功。HTTP 成功率和任务得分分别统计，能避免把“没有报错”写成“模型答对了”。",[25,913,915],{"id":914},"后续-30-步-grpo更新了参数仍未观察到可靠提升","后续 30 步 GRPO：更新了参数，仍未观察到可靠提升",[10,917,918],{},"从固定的原始 BF16 权重和全新 AdamW 状态重新开始，完成 30 个优化步。120 张不同训练图各采样 4 条回答，共 480 条；梯度非零，导出权重也确实发生改变。迁移到新机器后核对权重哈希，重新完成两个模型各 352 张、共 704 次推理，不混用迁移前后的不完整结果。",[10,920,921,922,925],{},"新增、训练前冻结的 128 张留出图上，两项计数都正确的图片从 22 张变成 23 张，但平均奖励从 0.4453125 降至 0.44140625，奖励差的配对 bootstrap 95% 区间为 ",[93,923,924],{},"[-0.0234375, 0.015625]","，包含零。重复 TextVQA64 的 soft score 仍为 81.25% → 81.25%。",[10,927,928],{},"这次实验说明训练链路和参数更新有效，不说明任务质量改善。后来的 SFT 使用不同训练方法、数据量和推理口径，不能把其结果写成“30 步 GRPO 带来的提升”。",[25,930,932],{"id":931},"换成有监督-lora先建立能改善的基线","换成有监督 LoRA：先建立能改善的基线",[10,934,935],{},"小规模试训观察到改善后，扩展到 768 张新训练图，覆盖红圆、蓝方各 0–7 的组合和渲染变体。冻结视觉与基座参数，仅训练语言侧 LoRA，rank 为 8、alpha 为 16；基座和计算使用 BF16，训练参数与优化器状态使用 FP32。",[10,937,938],{},"学习率为 2e-4，共 6 轮、1152 个优化步。96 张独立开发图在第 2、4、6 轮候选中选出第 4 轮，最终测试不参与选权。所有比较都采用同一 Transformers 后端和 448×448 图像处理。",[30,940,941,954],{},[33,942,943],{},[36,944,945,948,951],{},[39,946,947],{},"独立合成测试",[39,949,950],{"align":231},"原始基座",[39,952,953],{"align":231},"开发集选出的 LoRA",[52,955,956,967,978],{},[36,957,958,961,964],{},[57,959,960],{},"原渲染器新图 128 张，两项计数都正确",[57,962,963],{"align":231},"19\u002F128，14.84%",[57,965,966],{"align":231},"118\u002F128，92.19%",[36,968,969,972,975],{},[57,970,971],{},"渲染变体新图 128 张，两项计数都正确",[57,973,974],{"align":231},"17\u002F128，13.28%",[57,976,977],{"align":231},"113\u002F128，88.28%",[36,979,980,983,986],{},[57,981,982],{},"重复 TextVQA64，soft score",[57,984,985],{"align":231},"70.00%",[57,987,988],{"align":231},"70.625%",[10,990,991,992,904,995,998],{},"前两组奖励增量的配对 95% 区间分别为 ",[93,993,994],{},"[0.476562, 0.597656]",[93,996,997],{},"[0.519531, 0.660156]","，均高于零。1056 次生成完成，7 项预先冻结的验收条件通过。",[10,1000,1001,1004],{},[20,1002,1003],{},"这里的 TextVQA 70% 与前文 SGLang 的 81.25% 不是同一推理和图像处理口径，不能拼成一条训练前后曲线。"," TextVQA64 也只是重复使用的回归保护集，不是新增的完整基准测试。",[25,1006,1007],{"id":1007},"压力测试暴露了什么",[10,1009,1010],{},"固定上述适配器，不再训练，另生成四组各 64 张新图，比较原始基座和适配器，共 512 次推理。结果没有同分布测试那么整齐：",[30,1012,1013,1026],{},[33,1014,1015],{},[36,1016,1017,1020,1023],{},[39,1018,1019],{},"压力场景",[39,1021,1022],{"align":231},"原始基座全对数",[39,1024,1025],{"align":231},"SFT 适配器全对数",[52,1027,1028,1039,1050,1061],{},[36,1029,1030,1033,1036],{},[57,1031,1032],{},"散布排列",[57,1034,1035],{"align":231},"6\u002F64",[57,1037,1038],{"align":231},"56\u002F64",[36,1040,1041,1044,1047],{},[57,1042,1043],{},"密集小物体",[57,1045,1046],{"align":231},"0\u002F64",[57,1048,1049],{"align":231},"48\u002F64",[36,1051,1052,1055,1058],{},[57,1053,1054],{},"空心形状",[57,1056,1057],{"align":231},"7\u002F64",[57,1059,1060],{"align":231},"31\u002F64",[36,1062,1063,1066,1069],{},[57,1064,1065],{},"数量 8–12",[57,1067,1068],{"align":231},"2\u002F64",[57,1070,1071],{"align":231},"1\u002F64",[10,1073,1074],{},"对象都可见且不重叠，但空心形状明显更难；训练仅覆盖 0–7 时，8–12 没有改善。高数量样本中经常回答 7，这与训练范围有关是一个合理假设，却不是因果证明。不能凭常规图的 92.19% 就推断模型已经学会普适计数。",[25,1076,1078],{"id":1077},"两轮定向补训第一次没达标第二次才通过","两轮定向补训：第一次没达标，第二次才通过",[10,1080,1081],{},"第一轮从扩展 SFT 的第 4 轮适配器继续训练。新增 1014 张图，覆盖两个计数各 0–12 的 169 种组合，并回放 266 张旧训练图，总计 1280 张。学习率 1e-4，训练 6 轮、1920 步，192 张独立开发图选择第 4 轮。",[10,1083,1084],{},"这一轮有改善，但散布排列全对率仅 76.56%、空心 73.44%、8–12 为 53.125%，共 4 项冻结条件未通过。候选、日志和失败结果全部保留，没有降低阈值，也没有用最终测试去挑第 6 轮权重。",[10,1086,1087],{},"第二轮从第一轮开发集选出的适配器继续训练，新增 1280 张图：676 张均衡散布图、400 张高数量组合图和 204 张空心图；再回放第一轮全部 1280 张训练图。学习率降至 5e-5，训练 3 轮、1920 步，独立开发集从第 1、2、3 轮选择第 2 轮。",[10,1089,1090,1091,1094],{},"两个补训轮次使用各自新生成的开发与最终测试图。旧压力测试和第一轮最终测试图没有加入第二轮训练。",[20,1092,1093],{},"下表两列都来自第二轮同一批新测试图：before 是第一轮适配器，不是原始基座。"," 不同轮次测试图和数量范围有变化，不能横向拼接各轮分数。",[30,1096,1097,1113],{},[33,1098,1099],{},[36,1100,1101,1104,1107,1110],{},[39,1102,1103],{},"第二轮独立测试，每组 64 张",[39,1105,1106],{"align":231},"第一轮适配器",[39,1108,1109],{"align":231},"最终适配器",[39,1111,1112],{},"部分奖励增量的 95% 配对区间",[52,1114,1115,1130,1145,1161,1177,1193],{},[36,1116,1117,1120,1123,1125],{},[57,1118,1119],{},"常规排列，0–7",[57,1121,1122],{"align":231},"63\u002F64，98.44%",[57,1124,1122],{"align":231},[57,1126,1127],{},[523,1128,1129],{},"-0.0234, 0.0234",[36,1131,1132,1135,1138,1140],{},[57,1133,1134],{},"渲染变体，0–7",[57,1136,1137],{"align":231},"64\u002F64，100.00%",[57,1139,1137],{"align":231},[57,1141,1142],{},[523,1143,1144],{},"0.0000, 0.0000",[36,1146,1147,1150,1153,1156],{},[57,1148,1149],{},"散布排列，0–12",[57,1151,1152],{"align":231},"52\u002F64，81.25%",[57,1154,1155],{"align":231},"58\u002F64，90.625%",[57,1157,1158],{},[523,1159,1160],{},"-0.0078, 0.1016",[36,1162,1163,1166,1169,1172],{},[57,1164,1165],{},"密集小物体，0–12",[57,1167,1168],{"align":231},"51\u002F64，79.69%",[57,1170,1171],{"align":231},"60\u002F64，93.75%",[57,1173,1174],{},[523,1175,1176],{},"0.0312, 0.1641",[36,1178,1179,1182,1185,1188],{},[57,1180,1181],{},"空心形状，0–12",[57,1183,1184],{"align":231},"40\u002F64，62.50%",[57,1186,1187],{"align":231},"55\u002F64，85.94%",[57,1189,1190],{},[523,1191,1192],{},"0.0625, 0.2031",[36,1194,1195,1197,1200,1202],{},[57,1196,1065],{},[57,1198,1199],{"align":231},"32\u002F64，50.00%",[57,1201,1187],{"align":231},[57,1203,1204],{},[523,1205,1206],{},"0.1094, 0.2500",[10,1208,1209,1210,1213],{},"“全对”要求红圆与蓝方两个数量同时正确；部分奖励是两个字段正确率的均值。表中的区间对应",[20,1211,1212],{},"部分奖励差","，不是全对率差。散布组分数上升，但奖励区间仍包含零，不能声称这组已证明统计显著改善。",[10,1215,1216],{},[276,1217],{"alt":1218,"src":1219},"第二轮定向补训的配对测试结果","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-counting-repair-v2-results-20261011.png",[1221,1222,1223],"h3",{"id":1223},"验收规则没有事后修改",[10,1225,1226],{},"第二轮沿用第一轮预先冻结的 15 项条件：常规、渲染变体、散布、密集、空心、8–12 的全对率下限依次为 85%、80%、80%、70%、75%、60%；空心和 8–12 的部分奖励至少 0.8，且奖励差区间下界高于零；其余四组部分奖励及 TextVQA 的下降不能超过 5 个百分点。",[10,1228,1229],{},"最终 15 项全部通过。重复 TextVQA64 的同口径 soft score 从 70.625% 变为 70.15625%，下降 0.46875 个百分点，在预先设定的容许范围内。通过回归门槛不等于完全没有退化，更不等于通用能力提升。",[10,1231,1232],{},"第二轮共完成 1600 次生成：训练图诊断、独立最终测试的前后比较，以及三个候选的开发集评分；其中最终比较每个模型包含 384 张合成图和 64 张 TextVQA 图。训练与评测约 83.3 分钟。两轮补训合计执行 3840 个优化步、3200 次评测生成。",[1221,1234,1236],{"id":1235},"参数数据和调用路径都做了复核","参数、数据和调用路径都做了复核",[10,1238,1239],{},"独立审计复算图片哈希与数据隔离、场景标签、对象可见性和非重叠、训练覆盖、梯度与监督掩码、实际 FP32 权重变化、开发集选权规则，以及最终适配器是否与选中候选一致。配对评测的输入张量哈希也逐图核对。",[10,1241,1242],{},"相关 CPU 测试 54 项通过，历史证据复核 8 个部分通过。另用独立命令实际加载最终适配器，测试一张 11 个红圆、12 个蓝方的图，返回：",[105,1244,1246],{"className":660,"code":1245,"language":662,"meta":111,"style":111},"{\"blue_squares\": 12, \"red_circles\": 11}\n",[93,1247,1248],{"__ignoreMap":111},[523,1249,1250,1252,1254,1256,1259,1261,1263,1265,1268],{"class":525,"line":526},[523,1251,670],{"class":669},[523,1253,686],{"class":673},[523,1255,677],{"class":669},[523,1257,1258],{"class":533},"12",[523,1260,683],{"class":669},[523,1262,674],{"class":673},[523,1264,677],{"class":669},[523,1266,1267],{"class":533},"11",[523,1269,693],{"class":669},[10,1271,1272],{},"回答、生成 token IDs 和输入张量哈希与原评测一致。这张图已用于评测，所以只算调用一致性验证，不算新增独立成绩。",[10,1274,1275],{},[276,1276],{"alt":1277,"src":1278},"最终适配器的独立调用一致性示例","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-counting-repair-v2-parity-20261011.png",[10,1280,1281,1282,1285,1286,1289,1290,1289,1293,1289,1296,1289,1299,1289,1302,904,1305,1308],{},"最终产物是约 57.2 MiB 的语言 LoRA，保存在实验项目的 ",[93,1283,1284],{},"verl_lab\u002Foutputs\u002Fcounting-repair-v2\u002Fselected\u002Fadapter\u002F","，没有导出完整合并模型。原始证据保留在同级目录的 ",[93,1287,1288],{},"plan.json","、",[93,1291,1292],{},"before.jsonl",[93,1294,1295],{},"after.jsonl",[93,1297,1298],{},"training.jsonl",[93,1300,1301],{},"dev-epoch*.jsonl",[93,1303,1304],{},"audit.json",[93,1306,1307],{},"completion.json","。这些是本地实验路径，不是公开下载地址。",[1221,1310,1312],{"id":1311},"目前能说什么不能说什么","目前能说什么，不能说什么",[422,1314,1315,1318,1321,1328,1331],{},[425,1316,1317],{},"可以说：固定测试协议下，合成、可见、无重叠几何形状的计数有改善，最终达到预设操作性门槛。",[425,1319,1320],{},"不能说：GRPO 已有效提升；有明确改善的是后续有监督 LoRA。",[425,1322,1323,1324,1327],{},"不能说：8–12 是最终模型的数值外推。补训已经覆盖 0–12，内部数据分组名虽沿用 ",[93,1325,1326],{},"extrapolate","，实际是训练范围内的新图片泛化。",[425,1329,1330],{},"不能说：真实照片、遮挡、重叠或超过 12 的数量已经验证。TextVQA 回归集不能替代真实照片计数评测。",[425,1332,1333],{},"不能说：单次训练结果证明稳定收益。只有一个训练随机种子，bootstrap 重采样的是配对图片，不是多次训练运行。",[10,1335,1336,1337,1340],{},"新的 LoRA 已可通过独立命令调用，但",[20,1338,1339],{},"未替换原网页默认模型，也不能直接用于旧 Mac 4-bit MLX 模型","。实验结束后释放了 GPU 显存与训练锁；云实例没有关机，释放显存不等于停止计费。",[10,1342,1343],{},"下一步应先补真实照片计数评测，再测重叠、遮挡、超过 12 的数量，并扩大独立样本、做多训练种子复现，而不是继续刷同一类合成图。",[25,1345,1346],{"id":1346},"这轮实验留下的做法",[10,1348,1349],{},"现在回看，最有用的记录并不是单独一个 token\u002Fs，而是能说明这个数字怎样得到的条件：模型与精度、输入长度、实际输出长度、并发与活跃上限、缓存状态、预热和计时范围。",[10,1351,1352,1353,1289,1356,1289,1359,1362,1363,1289,1366,1289,1369,1372],{},"读源码时，我沿着 SGLang 的 ",[93,1354,1355],{},"get_next_batch_to_run()",[93,1357,1358],{},"get_new_batch_prefill()",[93,1360,1361],{},"update_running_batch()"," 看请求怎样进入和退出批次，再沿着 Radix Cache 的 ",[93,1364,1365],{},"match_prefix()",[93,1367,1368],{},"insert()",[93,1370,1371],{},"evict()"," 看状态怎样复用。这样能把 C4\u002FB1 的排队和缓存冷暖差异对应到实际实现。",[10,1374,1375],{},"本轮只使用单卡，TP=2 没有实测。M2 Pro 的 19 个 GPU 核心属于一个设备，也不等于 19 张卡。要研究张量并行的收益，需要多设备和通信测量。",[10,1377,1378],{},"后续补训印证了按失败类型补数据的价值，也暴露了同分布高分的局限。训练流程跑通、权重发生变化、特定测试达到门槛和真实场景可用，是四件不同的事。原 3 步和 30 步 GRPO 的无可靠提升结果，与后续 SFT 的合成计数改善，应当一起保留。",[25,1380,1381],{"id":1381},"参考资料与固定版本",[422,1383,1384,1398,1410,1433,1451],{},[425,1385,1386,1392,1393,291],{},[284,1387,1391],{"href":1388,"rel":1389},"https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2.5-vl\u002F",[1390],"nofollow","Qwen2.5-VL 官方介绍","与",[284,1394,1397],{"href":1395,"rel":1396},"https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2.5-VL-3B-Instruct\u002Ftree\u002F66285546d2b821cf421d4f5eb2576359d3770cd3",[1390],"本次原始 3B 模型 revision",[425,1399,1400,1392,1405,291],{},[284,1401,1404],{"href":1402,"rel":1403},"https:\u002F\u002Fgithub.com\u002FBlaizzy\u002Fmlx-vlm\u002Ftree\u002F1dcc142dbe2c579d1de7638aaac263f7e4e11734",[1390],"MLX-VLM 固定 commit",[284,1406,1409],{"href":1407,"rel":1408},"https:\u002F\u002Fhuggingface.co\u002Fmlx-community\u002FQwen2.5-VL-3B-Instruct-4bit\u002Ftree\u002F46d4cf06a06ffc1a766c214174f9cbed2f45bcab",[1390],"MLX 4-bit 模型 revision",[425,1411,1412,1417,1418,1289,1423,1289,1428,291],{},[284,1413,1416],{"href":1414,"rel":1415},"https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\u002Ftree\u002F0c006b8809cd99e1f95926401a2823dd952641c8",[1390],"SGLang 0.5.5 固定 commit","：",[284,1419,1422],{"href":1420,"rel":1421},"https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\u002Fblob\u002F0c006b8809cd99e1f95926401a2823dd952641c8\u002Fpython\u002Fsglang\u002Fsrt\u002Fmanagers\u002Fscheduler.py",[1390],"调度器",[284,1424,1427],{"href":1425,"rel":1426},"https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\u002Fblob\u002F0c006b8809cd99e1f95926401a2823dd952641c8\u002Fpython\u002Fsglang\u002Fsrt\u002Fmem_cache\u002Fradix_cache.py",[1390],"Radix Cache",[284,1429,1432],{"href":1430,"rel":1431},"https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\u002Fblob\u002F0c006b8809cd99e1f95926401a2823dd952641c8\u002Fpython\u002Fsglang\u002Fsrt\u002Fserver_args.py",[1390],"启动参数",[425,1434,1435,1417,1440,1289,1445,1450],{},[284,1436,1439],{"href":1437,"rel":1438},"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl\u002Ftree\u002Fd62da4950573d7a4b7ef2362337952e7ab59e78d",[1390],"verl 0.6.1 固定 commit",[284,1441,1444],{"href":1442,"rel":1443},"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl\u002Fblob\u002Fd62da4950573d7a4b7ef2362337952e7ab59e78d\u002Fexamples\u002Fgrpo_trainer\u002Frun_qwen2_5_vl-7b-sglang.sh",[1390],"官方多模态 GRPO 示例",[284,1446,1449],{"href":1447,"rel":1448},"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl\u002Fblob\u002Fd62da4950573d7a4b7ef2362337952e7ab59e78d\u002Fverl\u002Futils\u002Fdataset\u002Frl_dataset.py",[1390],"数据集实现","。官方示例使用 7B；本文实际实验使用 3B 和上述单卡配置。",[425,1452,1453,1458,1459,291],{},[284,1454,1457],{"href":1455,"rel":1456},"https:\u002F\u002Ftextvqa.org\u002Fdataset\u002F",[1390],"TextVQA 官方数据与任务说明","及",[284,1460,1463],{"href":1461,"rel":1462},"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fmmf\u002Fblob\u002F72f898ec35af86567423c7a019dd1dc5175d1a5a\u002Fmmf\u002Futils\u002Fm4c_evaluators.py",[1390],"MMF TextVQA 评分实现",[10,1465,1466],{},"封面与正文概念插画由 AI 辅助生成；性能图和新增计数结果图按实测数据绘制，调用示例使用实际测试图和输出。",[1468,1469,1470],"style",{},"html pre.shiki code .sLsJe, html code.shiki .sLsJe{--shiki-default:#F69D50;--shiki-light:#6F42C1;--shiki-cyber:#00DAEF}html pre.shiki code .sgENW, html code.shiki .sgENW{--shiki-default:#6CB6FF;--shiki-light:#005CC5;--shiki-cyber:#FFD493}html pre.shiki code .sWqY2, html code.shiki .sWqY2{--shiki-default:#96D0FF;--shiki-light:#032F62;--shiki-cyber:#FFD493}html pre.shiki code .sxpgZ, html code.shiki .sxpgZ{--shiki-default:#F47067;--shiki-light:#005CC5;--shiki-cyber:#EEF0F9}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .cyber .shiki span {color: var(--shiki-cyber);background: var(--shiki-cyber-bg);font-style: var(--shiki-cyber-font-style);font-weight: var(--shiki-cyber-font-weight);text-decoration: var(--shiki-cyber-text-decoration);}html.cyber .shiki span {color: var(--shiki-cyber);background: var(--shiki-cyber-bg);font-style: var(--shiki-cyber-font-style);font-weight: var(--shiki-cyber-font-weight);text-decoration: var(--shiki-cyber-text-decoration);}html pre.shiki code .sWURj, html code.shiki .sWURj{--shiki-default:#ADBAC7;--shiki-light:#24292E;--shiki-cyber:#EEF0F9}html pre.shiki code .sjXwz, html code.shiki .sjXwz{--shiki-default:#8DDB8C;--shiki-light:#005CC5;--shiki-cyber:#4BF3C8}",{"title":111,"searchDepth":545,"depth":556,"links":1472},[1473,1474,1475,1476,1477,1478,1479,1480,1481,1482,1483,1484,1489,1490],{"id":27,"depth":545,"text":28},{"id":99,"depth":545,"text":100},{"id":142,"depth":545,"text":142},{"id":218,"depth":545,"text":219},{"id":342,"depth":545,"text":343},{"id":416,"depth":545,"text":417},{"id":653,"depth":545,"text":654},{"id":805,"depth":545,"text":805},{"id":914,"depth":545,"text":915},{"id":931,"depth":545,"text":932},{"id":1007,"depth":545,"text":1007},{"id":1077,"depth":545,"text":1078,"children":1485},[1486,1487,1488],{"id":1223,"depth":556,"text":1223},{"id":1235,"depth":556,"text":1236},{"id":1311,"depth":556,"text":1312},{"id":1346,"depth":545,"text":1346},{"id":1381,"depth":545,"text":1381},[1492],"AI","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-lab-cover-20261010.webp","笔记本与单卡计算模块通过图像和 token 光轨连接的实验台插画","2026-08-19","在 M2 Pro 和云端单卡上部署 Qwen2.5-VL-3B，实测连续批处理、前缀缓存与 SGLang 参数调优，再用 verl 完成多模态 GRPO 更新和同条件前后评测，记录有效优化、兼容性问题与未观察到提升的结果。",false,"md",true,{},"\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab",null,{"title":5,"description":1496},"posts\u002Fqwen25-vl-sglang-verl-single-gpu-lab",[1506,1507,1508,1509,1510,1511,1512],"Qwen2.5-VL","多模态","SGLang","verl","GRPO","推理优化","MLX","i8vCR0dmVtDOqLB-Ar8rTPSWiKjdk0D-Sv5ePI81EPg",[1515,1533,1549,1563,1576,1590,1605,1620,1633,1645,1659,1671,1685,1698,1711,1725,1742,1757,1768,1780,1794,1807,1821,1833,1847,1859,1874,1888,1902,1913,1926,1943,1948,1961,1973,1985,1998,2011],{"slug":1516,"path":1517,"title":1518,"date":1519,"tags":1520,"description":1526,"cover":1527,"coverAlt":1528,"excerpt":1529,"draft":1497,"hidden":1497,"published":1499,"readingTime":641,"series":1530,"seriesOrder":641,"codeUrl":1531,"attachmentUrl":1532},"cs336-cross-entropy-adamw-training-loop","\u002Fposts\u002Fcs336-cross-entropy-adamw-training-loop","Stanford CS336 学习记录 08：数值稳定交叉熵、AdamW、余弦调度与完整训练闭环","2026-10-27",[1521,1522,1523,1524,1525],"CS336","PyTorch","AdamW","CrossEntropy","训练闭环","用 LogSumExp 消去交叉熵中的指数溢出，手写解耦权重衰减的 AdamW 优化器、全局 L2 梯度裁剪、Warmup 余弦退火调度、memmap 数据采样、Checkpoint 序列化与 Top-p 解码，完成 CS336 Assignment 1 全部 48 项测试。","\u002Fimages\u002Fposts\u002Fcs336-cross-entropy-adamw-training-loop\u002Fcs336-cross-entropy-adamw-training-loop-cover.webp","滑动窗口采样、余弦退火曲线、双动量优化轨迹与检查点归档构成的训练闭环 AI 辅助概念插画","作为 Stanford CS336 Assignment 1 系列的收官篇，我们已经拥有了分词器（bpe.py、tokenizer.py）和完整的语言模型架构（transformer.py）。要让模型在语料上真正训练并生成文本，还需","Stanford CS336","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002Fb5c93ad0cc73178b2b2a62e518a0900da47a0fb7","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-08-adamw-training-loop-20261027.zip",{"slug":1534,"path":1535,"title":1536,"date":1537,"tags":1538,"description":1545,"cover":1546,"coverAlt":1547,"excerpt":1548,"draft":1497,"hidden":1497,"published":1499,"readingTime":641},"editcache-video-dit-inference-acceleration","\u002Fposts\u002Feditcache-video-dit-inference-acceleration","EditCache：视频 Diffusion Transformer (DiT) 局部编辑的跨步特征复用与区域重算加速","2026-10-25",[1539,1540,1541,1542,1543,1544],"扩散模型","视频生成","AI Infra","推理加速","GPU","论文解读","拆解视频扩散模型（Wan2.1-VACE）局部编辑中的三重算力冗余，通过可重入 Latent、跨步 Block 隐状态缓存复用、带 Halo 边界的局部区域重算与自适应推理起点，在单卡 H100 上实现 2.46× 模型计算加速与 1.49× 严格同步端到端加速。","\u002Fimages\u002Fposts\u002Feditcache-video-dit-inference-acceleration\u002Feditcache-dit-cover-20261025.webp","视频扩散 Transformer 时空三维体素网格中冻结背景缓存与暖橙色局部编辑重算通道的等距技术插画","在基于 Diffusion Transformer（DiT）的视频编辑与局部重绘（Video Inpainting \u002F Local Editing）任务中，用户的典型工作流并不是一次性生成最终成片，而是针对同一段源视频反复调整掩码区",{"slug":1550,"path":1551,"title":1552,"date":1537,"tags":1553,"description":1558,"cover":1559,"coverAlt":1560,"excerpt":1561,"draft":1497,"hidden":1497,"published":1499,"readingTime":641,"series":1530,"seriesOrder":624,"codeUrl":1531,"attachmentUrl":1562},"cs336-bpe-tokenizer-training-encoding","\u002Fposts\u002Fcs336-bpe-tokenizer-training-encoding","Stanford CS336 学习记录 07：Byte-Level BPE 分词器训练、倒排索引加速与流式编码",[1521,1554,1555,1556,1557],"Tokenizer","BPE","Tiktoken","算法优化","实现 GPT-2 正则预分词、特殊 token 硬边界切分、基于 pair-to-word 倒排索引的增量 BPE 合并算法（\u003C 0.25s 通过速度基准），以及与 tiktoken 完全对齐的 Tokenizer。","\u002Fimages\u002Fposts\u002Fcs336-bpe-tokenizer-training-encoding\u002Fcs336-bpe-tokenizer-training-encoding-cover.webp","UTF-8 字节块经过正则切分闸口后沿合并树聚合成子词单元的 AI 辅助概念插画","前面六篇完成了 Transformer 语言模型本身的全部神经网络组件。要让模型真正处理自然语言文本，还需要打通入口处的 ByteLevel BPE（BytePair Encoding）分词器。这一篇记录 cs336basics\u002Fb","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-07-bpe-tokenizer-20261025.zip",{"slug":1564,"path":1565,"title":1566,"date":1567,"tags":1568,"description":1572,"cover":1573,"coverAlt":1574,"excerpt":1575,"draft":1497,"hidden":1497,"published":1499,"readingTime":641},"in-situ-autoguidance-editable-scientific-illustration","\u002Fposts\u002Fin-situ-autoguidance-editable-scientific-illustration","从原位自校正引导（In-situ Autoguidance）到可编辑科研插图生成（AutoFigure-Edit）","2026-10-24",[1539,1507,1569,1570,1571,1544],"Agent","ICML","ACL","串联两项生成式模型与多模态智能体研究：如何通过推理期局部随机微扰在单模型内构造免训练扩散自引导向量（ICML 2025 Workshop 第一作者，arXiv:2510.17136），以及如何通过五阶段视觉-符号混合流水线生成原生可编辑 SVG 科研插图（ACL 2026 Demo，FigureBench 83.0% 胜率）。","\u002Fimages\u002Fposts\u002Fin-situ-autoguidance-editable-scientific-illustration\u002Fautoguidance-autofigure-cover-20261024.webp","左侧扩散流形自校正轨迹矢量场与右侧五阶段可编辑 SVG 科研插图分层网格相衔接的等距技术插画","生成式视觉模型正在经历从“生成高保真像素（Pixel Synthesis）”向“可控结构化生成（Controllable & Editable Generation）”的演进。在这一进程中，我作为第一作者或核心贡献者参与了两项分别处",{"slug":1577,"path":1578,"title":1579,"date":1580,"tags":1581,"description":1585,"cover":1586,"coverAlt":1587,"excerpt":1588,"draft":1497,"hidden":1497,"published":1499,"readingTime":641,"series":1530,"seriesOrder":607,"codeUrl":1531,"attachmentUrl":1589},"cs336-pre-norm-transformer-block-lm","\u002Fposts\u002Fcs336-pre-norm-transformer-block-lm","Stanford CS336 学习记录 06：Pre-Norm Transformer Block 与完整语言模型装配","2026-10-23",[1521,1522,1582,1583,1584],"Transformer","Pre-Norm","FLOPs","将 RMSNorm、带 RoPE 的因果多头自注意力和 SwiGLU 组装成 Pre-Norm Transformer Block，搭建完整 TransformerLM，并核算 GPT-2 XL 的参数量与前向 FLOPs。","\u002Fimages\u002Fposts\u002Fcs336-pre-norm-transformer-block-lm\u002Fcs336-pre-norm-transformer-block-lm-cover.webp","干净残差流主干穿过多个 Pre-Norm 子层并汇聚到词表投影层的 AI 辅助概念插画","前五篇把 Linear、Embedding、RMSNorm、SwiGLU、RotaryPositionalEmbedding 和因果 MultiHeadSelfAttention 全部单独打通了。这一篇进入 Assignment 1","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-06-transformer-block-lm-20261023.zip",{"slug":1591,"path":1592,"title":1593,"date":1580,"tags":1594,"description":1600,"cover":1601,"coverAlt":1602,"excerpt":1603,"draft":1497,"hidden":1497,"published":1499,"readingTime":1604},"xgrammar-concurrent-cache-generation-fix","\u002Fposts\u002Fxgrammar-concurrent-cache-generation-fix","XGrammar 结构化解码缓存剖析：Unlimited 模式并发 Clear 字节记账竞态与代际修复（PR #1011）",[1541,1595,1596,1597,1598,1599],"C++","XGrammar","结构化解码","并发编程","源码分析","拆解 XGrammar GrammarCompiler 的双层缓存架构与 ThreadSafeLRUCache 锁分离实现，定位 Unlimited 模式下 Clear() 字节残留与在途编译（In-flight Computation）跨代污染根因，给出代际号 + 读写锁原子修复与零 sleep 确定性并发 GoogleTest 验证。","\u002Fimages\u002Fposts\u002Fxgrammar-concurrent-cache-generation-fix\u002Fxgrammar-cache-fix-cover-20261022.webp","双层磨砂玻璃流道经暖橙色代际闸门过滤后汇入晶格缓存矩阵的等距技术插画","在大语言模型服务（如 SGLang、vLLM）的结构化输出场景中，JSON Schema 或 EBNF 语法需要先编译为词表级掩码缓存（CompiledGrammar），才能在每个 Decode 步以微秒级开销生成合法 token ",14,{"slug":1606,"path":1607,"title":1608,"date":1609,"tags":1610,"description":1616,"cover":1617,"coverAlt":1618,"excerpt":1619,"draft":1497,"hidden":1497,"published":1499,"readingTime":1604},"multi-constraint-agent-routing-dag-eval","\u002Fposts\u002Fmulti-constraint-agent-routing-dag-eval","多约束生产级 Agent 系统实践：两阶段工具路由、DAG 局部重规划与轨迹级评测闭环","2026-10-22",[1569,1611,1612,1613,1614,1615],"LLM","工具调用","上下文工程","DPO","强化学习","围绕跨境酒旅等强耦合多约束业务场景，拆解向量 Top-K 召回 + LLM 精排 + 跨语言实体链接与 Schema 门控（96%+ 调用合法率）、基于状态机的 DAG 局部子图重规划、Observation 冗余裁剪（节省 40% Token），以及从线上失败恢复轨迹构造 SFT\u002FDPO 偏好对的评测与训练闭环。","\u002Fimages\u002Fposts\u002Fmulti-constraint-agent-routing-dag-eval\u002Fagent-dag-routing-cover-20261020.webp","候选模块经漏斗滤网进入双分支有向无环图轨道并在受阻节点触发暖橙色局部旁路重连的等距技术插画","在玩具级 Demo 里，几个工具加上一段 ReAct 循环就能跑通大多数单步问答。但当 Agent 进入跨境酒旅、复杂交易或研发沙箱等真实业务生产环境时，单靠线性 ReAct 循环很快会撞上四堵墙： 1. 工具空间与参数耦合爆炸：候",{"slug":1621,"path":1622,"title":1623,"date":1624,"tags":1625,"description":1628,"cover":1629,"coverAlt":1630,"excerpt":1631,"draft":1497,"hidden":1497,"published":1499,"readingTime":1632},"llm-mrd-distillation-auditable-review-agents","\u002Fposts\u002Fllm-mrd-distillation-auditable-review-agents","多视角推理蒸馏（LLM-MRD）与可审计生成式评审系统（DeepReviewer 2.0）深度拆解","2026-10-21",[1626,1507,1506,1569,1627,1544],"知识蒸馏","Post-training","对比大模型落地的两条互补路径：如何用 Qwen2.5-VL 离线多视角分析结合 MLP 残差校准与温度缩放 KL 蒸馏，将推理能力压入 358M 学生模型实现 17.2 ms 独立推理（LLM-MRD，DASFAA 2026 Oral）；以及如何通过主张-证据-风险台账、PASA 可比性门控与导出预算校验构建可审计的同行评审 Agent（DeepReviewer 2.0）。","\u002Fimages\u002Fposts\u002Fllm-mrd-distillation-auditable-review-agents\u002Fmrd-review-distill-cover-20261016.webp","三块半透明几何棱镜发出平行光束穿过中央校准环汇聚至右侧高密度晶体模块的等距技术插画","把大语言模型或视觉语言模型（VLM）的深度推理能力用到实际系统中，通常会面临一个两难选择： 如果场景受限于毫秒级在线延迟（如高吞吐多模态内容审核与虚假信息检测）：直接在线调用 7B～72B 的多模态大模型做长链条思维链（CoT）生成",18,{"slug":1634,"path":1635,"title":1636,"date":1624,"tags":1637,"description":1640,"cover":1641,"coverAlt":1642,"excerpt":1643,"draft":1497,"hidden":1497,"published":1499,"readingTime":624,"series":1530,"seriesOrder":590,"codeUrl":1531,"attachmentUrl":1644},"cs336-rope-rotary-position-embedding","\u002Fposts\u002Fcs336-rope-rotary-position-embedding","Stanford CS336 学习记录 05：RoPE 旋转位置编码与带位置的多头注意力",[1521,1522,1582,1638,1639],"RoPE","Attention","从二维平面正交旋转推导 RoPE 的奇偶特征配对公式，理解为什么内积只依赖相对位置差 (i - j)，并用 register_buffer 预计算 cos\u002Fsin 缓存接入因果多头自注意力。","\u002Fimages\u002Fposts\u002Fcs336-rope-rotary-position-embedding\u002Fcs336-rope-rotary-position-embedding-cover.webp","二维特征平面中的旋转向量对与预计算余弦正弦网格的 AI 辅助概念插画","上一篇完成了不含位置编码的因果多头自注意力。如果只看那一套矩阵乘法，把输入序列前两个 token 交换顺序（并在因果 mask 允许的范围内看同一个后缀 query 对它们的点积），点积本身并不知道哪个 key 离当前 query ","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-05-rope-mha-20261021.zip",{"slug":1646,"path":1647,"title":1648,"date":1649,"tags":1650,"description":1655,"cover":1656,"coverAlt":1657,"excerpt":1658,"draft":1497,"hidden":1497,"published":1499,"readingTime":641},"llm-gateway-semantic-cache-dynamic-batching","\u002Fposts\u002Fllm-gateway-semantic-cache-dynamic-batching","高并发 LLM 推理网关架构实践：两级精确\u002F语义缓存、50ms 有界动态批处理与尾部延迟治理","2026-10-19",[1541,1651,1652,1653,1654,47],"LLM 推理","分布式系统","Redis","FastAPI","深度拆解生产级大模型推理网关 LLM-Gateway 的核心系统设计：基于 SHA-256 与 Redis 向量余弦相似度的两级精确\u002F语义缓存、50ms 有界等待动态批处理调度器（Dynamic Batching）、令牌桶多租户限流与 Prometheus 全链路可观测性，将命中请求 P99 延迟从 5,307 ms 压降至 17 ms。","\u002Fimages\u002Fposts\u002Fllm-gateway-semantic-cache-dynamic-batching\u002Fllm-gateway-cover-20261019.webp","高并发请求流依次穿过精确哈希滤网、向量球面语义缓存层与 50ms 动态批处理时间窗拼装器的等距系统架构插画","在将大语言模型接入高并发业务系统时，如果让上游微服务直接裸连底层推理引擎（vLLM \u002F SGLang 或外部模型 API），系统很快会暴露出三个典型的生产痛点： 1. 高频重复与近义查询消耗巨量算力：大量用户请求在语义上高度等价（例",{"slug":1660,"path":1661,"title":1662,"date":1649,"tags":1663,"description":1665,"cover":1666,"coverAlt":1667,"excerpt":1668,"draft":1497,"hidden":1497,"published":1499,"readingTime":1604,"series":1530,"seriesOrder":573,"codeUrl":1669,"attachmentUrl":1670},"cs336-multihead-attention-shapes","\u002Fposts\u002Fcs336-multihead-attention-shapes","Stanford CS336 学习记录 04：多头注意力怎样拆头、计算和合头",[1521,1522,1582,1639,1664],"张量形状","沿着 (2, 5, 64) 的输入追踪 QKV 投影、reshape、transpose、共享因果 mask 和合头，理解每个头仍处理完整序列，以及最后为什么还需要输出 Linear。","\u002Fimages\u002Fposts\u002Fcs336-multihead-attention-shapes\u002Fcs336-multihead-attention-shapes-cover.webp","宽向量束分成多条并行特征轨道后重新合并的 AI 辅助概念插画","上一篇完成了缩放点积注意力。那个函数接收已经准备好的 Q、K、V，计算匹配分数，应用 mask，再用权重汇总 V。这一篇继续把它放进一个带参数的多头自注意力模块。 这次最需要看清的是“拆头”。代码里只是几行 reshape 和 tr","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002F31363223498fafd1528c550a51efbe0aed231736","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-04-multihead-attention-20261019.zip",{"slug":1672,"path":1673,"title":1674,"date":1675,"tags":1676,"description":1680,"cover":1681,"coverAlt":1682,"excerpt":1683,"draft":1497,"hidden":1497,"published":1499,"readingTime":1684},"qwen25-sft-dpo-alignment-tax-ablation","\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation","大模型后训练偏好对齐实录：Qwen2.5 SFT 多维消融、动态参考 DPO 与对齐税（Alignment Tax）量化分析","2026-10-18",[1627,1614,1677,1678,1615,1679],"SFT","LoRA","Qwen2.5","基于 Qwen2.5 完整复现 Instruct SFT → Bradley-Terry DPO 两阶段后训练管线：通过 3 随机种子消融 LoRA Rank、学习率与数据规模（GSM8K +12.73pp），实现免双倍权重的动态 Adapter 切换 DPO 训练，并定量揭示偏好对齐对确定性数学推理的对齐税效应。","\u002Fimages\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation\u002Fqwen25-sft-dpo-cover-20261018.webp","左侧监督微调轨迹流经中央 Bradley-Terry 偏好分流天平分离出正负样本通道并连接右侧对齐税监控仪表盘的等距技术插画","在大语言模型后训练（PostTraining）管线中，监督微调（SFT）负责教会模型任务格式与领域推理模式，而直接偏好优化（DPO, Direct Preference Optimization）则利用成对偏好数据 $(x, yw,",9,{"slug":1686,"path":1687,"title":1688,"date":1675,"tags":1689,"description":1691,"cover":1692,"coverAlt":1693,"excerpt":1694,"draft":1497,"hidden":1497,"published":1499,"readingTime":1695,"series":1530,"seriesOrder":556,"codeUrl":1696,"attachmentUrl":1697},"cs336-softmax-attention-mask","\u002Fposts\u002Fcs336-softmax-attention-mask","Stanford CS336 学习记录 03：Softmax 的数值稳定与 Attention Mask",[1521,1522,1582,1690,1639],"Softmax","从减去最大值为何不改变 Softmax 开始，理清函数与有参数层的区别，再沿着 Q、K、V 的形状理解缩放点积注意力，以及 mask 为什么必须填负无穷。","\u002Fimages\u002Fposts\u002Fcs336-softmax-attention-mask\u002Fcs336-softmax-attention-mask-cover.webp","注意力分数、权重与因果遮罩的 AI 辅助概念插画","上一篇写到 SwiGLU，接下来进入 Attention。真正开始相乘之前，先做了 Softmax：它把一组分数变成总和为 1 的权重，后面用这些权重汇总不同 token 提供的信息。 这段学习里，有两个很具体的问题值得记下来。一个",11,"https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002Fa39eb684710817f7f389626423fcb26ca0f6d629","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-03-softmax-sdpa-20261018.zip",{"slug":1699,"path":1700,"title":1701,"date":1702,"tags":1703,"description":1706,"cover":1707,"coverAlt":1708,"excerpt":1709,"draft":1497,"hidden":1497,"published":1499,"readingTime":1684,"series":1530,"seriesOrder":545,"codeUrl":1696,"attachmentUrl":1710},"cs336-rmsnorm-swiglu","\u002Fposts\u002Fcs336-rmsnorm-swiglu","Stanford CS336 学习记录 02：RMSNorm 的广播与 SwiGLU 的门控","2026-10-16",[1521,1522,1582,1704,1705],"RMSNorm","SwiGLU","用一个 token 向量理解均方根归一化，拆开 dim=-1 与 keepdim=True 的作用，再沿着两条分支读懂 SiLU、逐元素乘法和 SwiGLU。","\u002Fimages\u002Fposts\u002Fcs336-rmsnorm-swiglu\u002Fcs336-rmsnorm-swiglu-cover.webp","特征尺度整理和双分支门控的 AI 辅助概念插画","上一篇把 Linear 和 Embedding 的参数形状理顺了。继续往下，最先遇到的是 RMSNorm：公式不算长，但把它落到一批 token 上时，dim=1、keepdim=True 和广播必须一起理解。 然后是 SiLU 和","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-02-rmsnorm-swiglu-20261016.zip",{"slug":1712,"path":1713,"title":1714,"date":1715,"tags":1716,"description":1719,"cover":1720,"coverAlt":1721,"excerpt":1722,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1530,"seriesOrder":526,"codeUrl":1696,"attachmentUrl":1724},"cs336-linear-embedding-shapes","\u002Fposts\u002Fcs336-linear-embedding-shapes","Stanford CS336 学习记录 01：从 Linear、Embedding 搞清张量形状","2026-10-14",[1521,1522,1582,1717,1718],"Linear","Embedding","从 PyTorch 类、参数与 forward 的关系开始，理解 Linear 为什么保存两行三列的权重、计算时为什么要转置，再区分 Embedding 查表与线性变换。","\u002Fimages\u002Fposts\u002Fcs336-linear-embedding-shapes\u002Fcs336-linear-embedding-shapes-cover.webp","token 方块、向量表与矩阵变换的 AI 辅助概念插画","开始沿着 Stanford CS336 的 Assignment 1 一步步实现。BPE 之前已经接触过，这轮先跳过分词器，从 Transformer 的基础层开始。这里记录的是当时写代码时真正卡住的问题，以及对应测试能确认到的范围",10,"\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-01-linear-embedding-20261014.zip",{"slug":1726,"path":1727,"title":1728,"date":1729,"tags":1730,"description":1735,"cover":1736,"coverAlt":1737,"excerpt":1738,"draft":1497,"hidden":1497,"published":1499,"readingTime":1739,"series":1731,"seriesOrder":1604,"codeUrl":1740,"attachmentUrl":1741},"cmu-15-779-kv-cache-granularity-study","\u002Fposts\u002Fcmu-15-779-kv-cache-granularity-study","CMU 15-779 学习记录 14：KV 页大小、共享边界与 CPU 实现成本","2026-10-12",[1731,1732,1733,1734],"CMU 15-779","KV cache","PagedAttention","系统实验","用 210 组 CPU 配置比较连续与分页 KV cache，观察 31\u002F32-token 共享前缀的页边界，分别记录活跃槽位、池容量、复制计数和完整调用耗时。","\u002Fimages\u002Fposts\u002Fcmu-15-779-kv-serving-speculative-sampling\u002Fcmu-serving-cover-20261019.webp","浅色手绘分页存储抽屉与共享前缀卡片，旁边连接不同长度的请求队列","个人课程项目中的 CPU 实验记录，延续记录 11 的分页 KV 实现。已完成 210 组配置与 1470 次计时，输出均通过连续存储参考对照；本篇不代表正式作业提交、整门课程或项目全部交付完成。 同样是共享 31 个 token，",12,"https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F200faf569eccf94fb125e65110bb5d14094d454a","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fkv-cache-granularity-20261019.zip",{"slug":1743,"path":1744,"title":1745,"date":1746,"tags":1747,"description":1750,"cover":1751,"coverAlt":1752,"excerpt":1753,"draft":1497,"hidden":1497,"published":1499,"readingTime":590,"series":1731,"seriesOrder":1754,"codeUrl":1755,"attachmentUrl":1756},"cmu-15-779-moe-routing-capacity","\u002Fposts\u002Fcmu-15-779-moe-routing-capacity","CMU 15-779 学习记录 13：MoE 路由、稀疏分发与容量溢出","2026-10-09",[1731,1748,1749],"MoE","Routing","用 CPU top-2 MoE 检查稀疏分发的输出与梯度，统计 expert-token 行数，并保留容量丢弃会改变输出的反例。","\u002Fimages\u002Fposts\u002Fcmu-15-779-moe-routing-capacity\u002Fcmu-moe-cover-20261019.webp","浅色手绘分拣器将小珠分配到四个专家工作屋，队列呈现负载不均","第 18 讲的个人配套实验。36 号脚本串行运行 CPU 专家，检查 topk 路由后的分发与合并；没有训练 router，也没有执行 GPU grouped GEMM 或跨设备 alltoall。 上一节记录低秩适配，这次转向每个",13,"https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002Fc4851e0cedadf13af7c421867e83c8ff57244fa0","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcpu-parallel-serving-lora-moe-20261019.zip",{"slug":1758,"path":1759,"title":1760,"date":1761,"tags":1762,"description":1764,"cover":1765,"coverAlt":1766,"excerpt":1767,"draft":1497,"hidden":1497,"published":1499,"readingTime":590,"series":1731,"seriesOrder":1739,"codeUrl":1755,"attachmentUrl":1756},"cmu-15-779-lora-uniform-quantization","\u002Fposts\u002Fcmu-15-779-lora-uniform-quantization","CMU 15-779 学习记录 12：LoRA 首步梯度、权重合并与均匀 4-bit 打包","2026-10-06",[1731,1678,1763],"Quantization","用一次 CPU SGD 更新检查 LoRA 初始化、冻结基座与合并等价，再分别记录均匀 4-bit 存储字节数和量化误差。","\u002Fimages\u002Fposts\u002Fcmu-15-779-lora-uniform-quantization\u002Fcmu-lora-cover-20261019.webp","浅色手绘的大计算板、两块细长适配板和成对打包的小方块","第 17 讲的个人配套实验，35 号脚本实际执行一次 CPU LoRA 更新，并实现均匀 4bit 打包。实验检查机制与数值误差，不评估微调任务效果，也没有实现完整 QLoRA。 这次把两个问题分开记录：低秩分支是否能合并进原权重，",{"slug":1769,"path":1770,"title":1771,"date":1772,"tags":1773,"description":1777,"cover":1736,"coverAlt":1778,"excerpt":1779,"draft":1497,"hidden":1497,"published":1499,"readingTime":624,"series":1731,"seriesOrder":1695,"codeUrl":1755,"attachmentUrl":1756},"cmu-15-779-kv-serving-speculative-sampling","\u002Fposts\u002Fcmu-15-779-kv-serving-speculative-sampling","CMU 15-779 学习记录 11：分页 KV、连续批处理与推测采样","2026-10-03",[1731,1774,1775,1776],"LLM Serving","KV Cache","Speculative Decoding","在 CPU 上验证 KV 页共享和写时复制，用单位成本模型检查连续批处理，再用条件概率小模型核对推测采样的接受与补采样。","浅色手绘的共享卡片抽屉与不断补入小珠子的请求托盘","第 15、16 讲的个人配套实验。33 号实现 CPU 分页 KV 存储，37 号模拟请求准入，34 号运行小概率模型的推测采样。它们分别验证存储、调度和分布问题，没有测量真实 LLM 服务吞吐或 GPU 延迟。 记录 08已经对照",{"slug":1781,"path":1782,"title":1783,"date":1784,"tags":1785,"description":1789,"cover":1790,"coverAlt":1791,"excerpt":1792,"draft":1497,"hidden":1497,"published":1499,"readingTime":1793},"qwen3-mid-post-train-mlx","\u002Fposts\u002Fqwen3-mid-post-train-mlx","在 M2 Pro 上完成 Qwen3 两阶段训练：Mid Train、SFT 与可复核评估","2026-09-30 11:19:10",[1611,1786,1512,1678,1787,1677,1788],"Qwen3","持续预训练","Apple Silicon","用 Qwen3 0.6B Base 和 MLX 完成领域持续训练与监督微调，记录数据划分、聊天模板、损失掩码、数值稳定性修复和独立测试结果。","\u002Fimages\u002Fposts\u002Fqwen3-mid-post-train-mlx\u002Fqwen3-mlx-cover-20261018.webp","Apple Silicon 统一内存芯片底座支撑两阶段持续预训练与带精确 Loss 掩码滤网的监督微调桥接插画","在一台 M2 Pro、16 GB 内存的 Mac 上，能否把一个 Base 语言模型接续训练成输出 JSON 的运维事件分级助手？这次项目完成了领域持续训练、监督微调、检查点选择、重新加载和独立测试，并保留了权重、日志和数据哈希。 ",16,{"slug":1795,"path":1796,"title":1797,"date":1798,"tags":1799,"description":1803,"cover":1804,"coverAlt":1805,"excerpt":1806,"draft":1497,"hidden":1497,"published":1499,"readingTime":607,"series":1731,"seriesOrder":1723,"codeUrl":1755,"attachmentUrl":1756},"cmu-15-779-cpu-parallelism-pipeline","\u002Fposts\u002Fcmu-15-779-cpu-parallelism-pipeline","CMU 15-779 学习记录 10：加权梯度、状态分片与同步流水线","2026-09-30",[1731,1800,1801,1802],"Parallel Training","Gloo","Pipeline","用双 CPU 进程 Gloo 检查不等数据分片的梯度，推导持久状态字节数，再用依赖模型比较 GPipe 与同步 1F1B 的激活存活。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cpu-parallelism-pipeline\u002Fcmu-parallel-cover-20261019.webp","浅色手绘的四阶段桌面流水线与按数量汇总的托盘","第 12 至 14 讲的个人配套实验，脚本编号 31、32 不对应官方作业。Gloo 实际运行两个 CPU 进程；张量切分、状态字节数和流水线分别按下文的数学或成本模型解释。真实 GPU 实测按学习安排跳过。 上一节检查单个计算图的",{"slug":1808,"path":1809,"title":1810,"date":1811,"tags":1812,"description":1815,"cover":1816,"coverAlt":1817,"excerpt":1818,"draft":1497,"hidden":1497,"published":1499,"readingTime":624,"series":1731,"seriesOrder":1684,"codeUrl":1819,"attachmentUrl":1820},"cmu-15-779-cpu-compiler-optimizations","\u002Fposts\u002Fcmu-15-779-cpu-compiler-optimizations","CMU 15-779 学习记录 09：CPU 编译优化、合法重写与候选搜索","2026-09-27",[1731,1813,1814,1522],"ML Compiler","TorchInductor","用 CPU 实验验证 BatchNorm 折叠、RMSNorm 重排与非线性反例，测量分块候选，并运行 TorchInductor，分别记录首调用与预热后的时间。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cpu-compiler-optimizations\u002Fcmu-cpu-compiler-cover-20261019.webp","淡紫计算积木经过薄荷绿折叠模具变成浅蓝积木，前方摆放四种候选托盘的浅色手绘插画","编译章节的个人配套实验，编号 30 不对应官方作业。图重写、分块候选和 TorchInductor 均在 CPU 上运行；真实 GPU 实验按学习安排跳过。 上一节记录了分块 Attention 和 KV cache。这次转到编译优","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F9261bea6aedf31cdccbbab2541ca693fc97b3743","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcpu-compiler-optimizations-20261019.zip",{"slug":1822,"path":1823,"title":1824,"date":1825,"tags":1826,"description":1828,"cover":1829,"coverAlt":1830,"excerpt":1831,"draft":1497,"hidden":1497,"published":1499,"readingTime":1739,"series":1731,"seriesOrder":641,"codeUrl":1740,"attachmentUrl":1832},"cmu-15-779-attention-online-softmax","\u002Fposts\u002Fcmu-15-779-attention-online-softmax","CMU 15-779 学习记录 08：分块 Attention、在线 Softmax 与 KV cache","2026-09-24",[1731,1639,1690,1827,1522],"FlashAttention","在 CPU 上对照完整 Attention、在线 Softmax 分块实现与 PyTorch SDPA，检查因果掩码、浮点误差和分数块大小，理解 FlashAttention 的一部分算法基础。","\u002Fimages\u002Fposts\u002Fcmu-15-779-attention-online-softmax\u002Fcmu-online-attention-cover-20261016.webp","浅紫色卡片分批进入绿色计算托盘，旁边三个状态容器保存累积量的浅色手绘插画","第 5 讲 Transformers & Attention 的个人配套实验。28\u002F29 号是 CPU 单头前向与 KV cache；后续补入 38 号 CUDA 多头合成张量的显式 SDPA MATH\u002FFLASH 对照。没有手写 ","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcuda-triton-sdpa-20261019.zip",{"slug":1834,"path":1835,"title":1836,"date":1837,"tags":1838,"description":1843,"cover":1844,"coverAlt":1845,"excerpt":1846,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1731,"seriesOrder":624,"codeUrl":1740,"attachmentUrl":1832},"cmu-15-779-matmul-tiling-tree-reduction","\u002Fposts\u002Fcmu-15-779-matmul-tiling-tree-reduction","CMU 15-779 学习记录 07：矩阵分块、数据复用与树形归约","2026-09-21",[1731,1839,1840,1841,1842],"CUDA","矩阵乘法","数据复用","归约","用 CPU 模型统计矩阵分块的输入读取次数，再比较串行求和和树形归约的依赖深度，记录补零、同步与浮点加法顺序的边界。","\u002Fimages\u002Fposts\u002Fcmu-15-779-matmul-tiling-tree-reduction\u002Fcmu-matmul-reduction-cover-20261016.webp","矩阵积木在共享托盘中复用，旁边的小珠沿树形管道逐层合并的浅色手绘插画","GPU Architecture & CUDA Programming 的个人配套实验。首次记录用 CPU 模型检查数据复用与依赖深度；后续补入单卡 GPU 原生 CUDA 矩阵验证，以及 Triton 两阶段归约和矩阵乘法实测。 ",{"slug":1848,"path":1849,"title":1850,"date":1851,"tags":1852,"description":1855,"cover":1856,"coverAlt":1857,"excerpt":1858,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1731,"seriesOrder":607,"codeUrl":1740,"attachmentUrl":1832},"cmu-15-779-cuda-thread-indexing","\u002Fposts\u002Fcmu-15-779-cuda-thread-indexing","CMU 15-779 学习记录 06：CUDA 线程索引与边界检查","2026-09-18",[1731,1839,1543,1853,1854],"线程索引","边界检查","用 CPU 演示 CUDA 的 block、thread 与全局下标：10 个元素如何分给 3 个线程块，为什么要跳过越界线程，以及这些验证能说明什么。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cuda-thread-indexing\u002Fcmu-cuda-indexing-cover-20261015.webp","分组的浅紫色计算方块与末端边界挡板组成线程分配教具的浅色插画","GPU Architecture & CUDA Programming 的个人配套练习。首次记录为 CPU 索引演示；后续在单卡 GPU 环境补入原生 CUDA 编译、运行与 memcheck，以及 Triton 向量加法实测。 上",{"slug":1860,"path":1861,"title":1862,"date":1863,"tags":1864,"description":1868,"cover":1869,"coverAlt":1870,"excerpt":1871,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1731,"seriesOrder":590,"codeUrl":1872,"attachmentUrl":1873},"cmu-15-779-profiler-operator-batching","\u002Fposts\u002Fcmu-15-779-profiler-operator-batching","CMU 15-779 学习记录 05：Profiler、算子调用与批量推理","2026-09-15",[1731,1522,1865,1866,1867],"Profiler","批量推理","性能分析","对照 ML Systems 基础，用 Profiler 拆开两层网络的执行过程，再固定 1000 个样本扫描 batch size，观察算子调用次数、尾批和 CPU 耗时。","\u002Fimages\u002Fposts\u002Fcmu-15-779-profiler-operator-batching\u002Fcmu-pytorch-profiling-cover-20261015.webp","放大镜观察线性层与 ReLU 模块，旁边面板呈现执行片段的浅色技术教具插画","本篇对照课程官网的 Introduction to ML Systems 整理，是个人配套 CPU 实验，不是课程官方作业。脚本编号是练习顺序，不对应课程讲次。 前几篇一直在补 PyTorch 基础。到了这一步，我在学习聊天里问：现","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002Fcecf5110d8a95161bb5763eadfca94cd7a0cc50f","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-profiler-batching-20261015.zip",{"slug":1875,"path":1876,"title":1877,"date":1878,"tags":1879,"description":1882,"cover":1883,"coverAlt":1884,"excerpt":1885,"draft":1497,"hidden":1497,"published":1499,"readingTime":1684,"series":1731,"seriesOrder":573,"codeUrl":1886,"attachmentUrl":1887},"cmu-15-779-pytorch-linear-relu-absolute-value","\u002Fposts\u002Fcmu-15-779-pytorch-linear-relu-absolute-value","CMU 15-779 学习记录 04：多输出线性层、ReLU 与绝对值网络","2026-09-12",[1731,1522,1840,1880,1881],"ReLU","神经网络","把 nn.Linear 对应到矩阵乘法，弄清多个输出怎样计算，再逐层拆开一个绝对值网络，比较手动设置参数与训练得到的结果。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-linear-relu-absolute-value\u002Fcmu-pytorch-absolute-cover-20261015.webp","双分支经过折角模块后汇入紫色 V 形曲线面板的浅色技术教具插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 接着前一篇的批量推理与矩阵乘法，我继续把 model(x) 拆开看：一层里有多个输出时，每个数是怎么算出来的？两","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F4cc63904a3a583da4a2122c246a14b43aac38528","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-linear-relu-abs-20261015.zip",{"slug":1889,"path":1890,"title":1891,"date":1892,"tags":1893,"description":1896,"cover":1897,"coverAlt":1898,"excerpt":1899,"draft":1497,"hidden":1497,"published":1499,"readingTime":1684,"series":1731,"seriesOrder":556,"codeUrl":1900,"attachmentUrl":1901},"cmu-15-779-pytorch-dataloader-checkpoints-batching","\u002Fposts\u002Fcmu-15-779-pytorch-dataloader-checkpoints-batching","CMU 15-779 学习记录 03：DataLoader、模型保存与批量推理","2026-09-09",[1731,1522,1894,1866,1895],"DataLoader","模型保存","用 DataLoader 分批训练平方函数，保存并加载模型参数，再比较 1000 个样本与 1 个样本的推理耗时，最后把批量计算对应到矩阵乘法。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-dataloader-checkpoints-batching\u002Fcmu-pytorch-batching-cover-20261015.webp","样本卡片按批次进入计算模块，旁边配有参数存储盒和输出矩阵面板的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 接着前一篇的平方函数实验，这一节继续练习分批训练、保存参数和独立预测。最后做了一次计时：同一个小网络，逐个处理 ","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-batching-20261015.zip",{"slug":1903,"path":1904,"title":1905,"date":1906,"tags":1907,"description":1908,"cover":1909,"coverAlt":1910,"excerpt":1911,"draft":1497,"hidden":1497,"published":1499,"readingTime":1684,"series":1731,"seriesOrder":545,"codeUrl":1900,"attachmentUrl":1912},"cmu-15-779-pytorch-linear-module-relu","\u002Fposts\u002Fcmu-15-779-pytorch-linear-module-relu","CMU 15-779 学习记录 02：nn.Linear、张量形状与 ReLU","2026-09-06",[1731,1522,1881,1880],"从带偏置的线性回归开始，弄清 nn.Linear 和 nn.Module 的分工，再用平方函数观察线性模型的限制、ReLU 的作用和测试点的选择。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-linear-module-relu\u002Fcmu-pytorch-relu-cover-20261015.webp","输入经过两层神经网络和 ReLU 折角模块，输出曲线拟合结果的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 上一篇里，我用一个参数拟合了 $y=2x$。这一篇继续顺着往下写，先加上偏置，再把手写参数换成 nn.Linea","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-models-20261015.zip",{"slug":1914,"path":1915,"title":1916,"date":1917,"tags":1918,"description":1921,"cover":1922,"coverAlt":1923,"excerpt":1924,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1731,"seriesOrder":526,"codeUrl":1900,"attachmentUrl":1925},"cmu-15-779-pytorch-autograd-linear-regression","\u002Fposts\u002Fcmu-15-779-pytorch-autograd-linear-regression","CMU 15-779 学习记录 01：PyTorch 自动求导与线性回归","2026-09-03",[1731,1522,1919,1920],"自动求导","线性回归","从自动求导和手动梯度下降开始，比较学习率，再用 SGD 训练线性回归。记录 w * 4 写错后停在 1.25 的原因和修正结果。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-autograd-linear-regression\u002Fcmu-pytorch-autograd-cover-20261015.webp","梯度曲面上的小球沿路径下降，前侧面板展示散点与拟合直线的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 开始学习 CMU 15779 时，我会 Python，但还没用过 PyTorch。这次先在 Mac 上用 CPU","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-basics-20261013.zip",{"slug":1927,"path":1928,"title":1929,"date":1930,"tags":1931,"description":1937,"cover":1938,"coverAlt":1939,"excerpt":1940,"draft":1497,"hidden":1497,"published":1499,"readingTime":1754,"codeUrl":1941,"attachmentUrl":1942},"embodied-pickplace-lab-robustness","\u002Fposts\u002Fembodied-pickplace-lab-robustness","Pickplace Lab 项目复盘：抓取成功率之外的时延与失败分析","2026-08-29",[1932,1933,1934,1522,1935,1936],"具身智能","模仿学习","MuJoCo","机器人","鲁棒性","从 MuJoCo 抓取搬运闭环出发，记录 22,100 个评测回合。动作平滑在自建任务的 50 ms 延迟下有效，换到 Panda Lift 却没有支持同样的收益。","\u002Fimages\u002Fposts\u002Fembodied-pickplace-lab-robustness\u002Fpickplace-grasp-cover-20261016.webp","MuJoCo 实际渲染的笛卡尔夹爪抓取方块场景，桌面上标出放置目标","Pickplace Lab 从一个简单任务开始：让夹爪抓起方块，搬到指定位置，放下，再退开。先把仿真、示范采集、训练和闭环评测接起来，随后再问：干净条件下成功的策略，换一种执行条件，还能不能完成任务？ 最值得讨论的结果出现在指令延迟","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fembodied-pickplace-lab\u002Ftree\u002Fv0.1.0","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fembodied-pickplace-lab\u002Freleases\u002Fdownload\u002Fv0.1.0\u002Fpickplace-lab-research-v1.zip",{"slug":1944,"path":1501,"title":5,"date":1495,"tags":1945,"description":1496,"cover":1493,"coverAlt":1494,"excerpt":1946,"draft":1497,"hidden":1497,"published":1499,"readingTime":1947},"qwen25-vl-sglang-verl-single-gpu-lab",[1506,1507,1508,1509,1510,1511,1512],"这次实验从一个很具体的问题开始：M2 Pro、16 GB 统一内存，能不能在本机跑 Qwen2.5VL？跑起来之后，我又想把图文问答、推理性能测量和强化学习训练接成一条完整流程。 本机的 MLX 4bit 服务和云端的 SGLang",29,{"slug":1949,"path":1950,"title":1951,"date":1952,"tags":1953,"description":1957,"cover":1958,"coverAlt":1959,"excerpt":1960,"draft":1497,"hidden":1497,"published":1499,"readingTime":1754},"cycleresearcher-post-training","\u002Fposts\u002Fcycleresearcher-post-training","从 DeepReviewer 2.0 到 CycleResearcher：科研 Agent 如何真正做 Post-training","2026-08-07 17:33:05",[1611,1627,1569,1954,1955,1956],"SimPO","Reward Model","ICLR","区分推理时 Agent 编排与模型后训练，并拆解 CycleResearcher 的领域 SFT、生成式 Reward Model、Iterative SimPO 和评测风险。","\u002Fimages\u002Fposts\u002Fcycleresearcher-post-training\u002Fcycleresearcher-cover-20261018.webp","策略生成模块产出多份候选文稿经多维评审棱镜排序后通过偏好梯度反馈弧闭环优化的等距技术插画","看到一个系统会检索、规划、反思和调用工具，不能据此判断它做过 Posttraining。Agent 描述的是模型在运行时如何完成任务；Posttraining 描述的是预训练之后，是否继续用监督数据、偏好数据或奖励信号更新模型参数。",{"slug":1962,"path":1963,"title":1964,"date":1965,"tags":1966,"description":1968,"cover":1969,"coverAlt":1970,"excerpt":1971,"draft":1497,"hidden":1497,"published":1499,"readingTime":1604,"series":1972,"seriesOrder":526},"cuda-fused-rmsnorm-kernel","\u002Fposts\u002Fcuda-fused-rmsnorm-kernel","从 PyTorch 到 CUDA Kernel：融合 RMSNorm 在 RTX 4090D 上加速 4.27 倍","2026-07-22 18:16:50",[1839,1541,1543,1522,1651,1967],"Kernel 优化","从零实现 PyTorch CUDA 扩展，拆解 RMSNorm 的访存瓶颈、warp 归约、向量化加载、算子融合与冷热缓存 benchmark。","\u002Fimages\u002Fposts\u002Fcuda-fused-rmsnorm-kernel\u002Fcuda-rmsnorm-cover-20261018.webp","融合残差加法与向量化 float4 访存经 Warp Shuffle 蝴蝶树归约输出归一化张量的等距硬件架构插画","为了准备 AI Infra 岗位，我做了一个尽量小而完整的 CUDA 项目：不追求堆很多算子，而是选一个真实的 LLM 推理路径，把实现、正确性、benchmark 和性能解释全部做扎实。 项目实现了两个 PyTorch CUDA ","AI Infra 实战",{"slug":1974,"path":1975,"title":1976,"date":1977,"tags":1978,"description":1980,"cover":1981,"coverAlt":1982,"excerpt":1983,"draft":1497,"hidden":1497,"published":1499,"readingTime":641,"series":1984,"seriesOrder":526},"agent-memory","\u002Fposts\u002Fagent-memory","Agent 上下文工程与分层记忆架构：从 Observation 剪枝、滚动摘要到向量召回闭环","2026-04-24",[1611,1569,1979,1541],"RAG","结合生产级多约束 Agent 与超 100K Token 科研智能体实践，系统拆解长程会话中的上下文膨胀与注意力稀释问题：涵盖 Observation 结构化剪枝（降低约 40% Token 开销）、滚动工作摘要、向量检索长期记忆及 Prefix Cache 友好的上下文组装。","\u002Fimages\u002Fposts\u002Fagent-memory\u002Fagent-memory-cover-20261018.webp","底层高噪数据经剪枝滤网进入中期滑动窗口与金色状态台账并由顶层向量星座按需召回的三层记忆架构插画","在构建多步工具调用（ToolUse）与长周期自主规划（Autonomous Planning）的 Agent 系统时，上下文管理（Context Engineering）往往比单纯改写 System Prompt 更决定线上稳定性。","Agent 上下文工程",{"slug":1986,"path":1987,"title":1988,"date":1989,"tags":1990,"description":1993,"cover":1994,"coverAlt":1995,"excerpt":1996,"draft":1497,"hidden":1497,"published":1499,"readingTime":590,"series":1997,"seriesOrder":545},"backend-stack-deep-dive","\u002Fposts\u002Fbackend-stack-deep-dive","高并发 AI 平台后端基础设施实践：FastAPI 异步网关、SQLAlchemy 2.0 Async、PostgreSQL 与 Docker 沙箱编排","2026-04-07",[47,1652,1654,1991,1992,1541],"PostgreSQL","Docker","结合 DeepScientist 生产级 AI 科研平台实践，拆解 FastAPI ASGI 异步事件循环、SQLAlchemy 2.0 + AsyncPG 连接池调优、PostgreSQL + MinIO 存储分层与 Docker SDK 动态沙箱容器编排的职责边界与核心设计。","\u002Fimages\u002Fposts\u002Fbackend-stack-deep-dive\u002Fbackend-stack-cover-20261018.webp","ASGI 异步事件循环涡轮将非阻塞协程流分发至 PostgreSQL 连接池、MinIO 存储矩阵与动态 Docker 沙箱的等距架构插画","结合 DeepScientist 生产级 AI 科研平台（服务 7,000+ 用户）的实际工程落地，本文系统拆解支撑高并发流式推理与容器沙箱调度的后端基础设施栈——FastAPI（ASGI 异步网关 + SSE 流式推送）、SQLA","DeepScientist 工程复盘",{"slug":1999,"path":2000,"title":2001,"date":2002,"tags":2003,"description":2006,"cover":2007,"coverAlt":2008,"excerpt":2009,"draft":1497,"hidden":1497,"published":1499,"readingTime":1695,"series":2010,"seriesOrder":526},"minicode-source-analysis","\u002Fposts\u002Fminicode-source-analysis","MiniCode 架构与工具评测深度拆解：开源终端 Agentic Coding 引擎（1k+ Stars）的 ReAct、MCP 与 30 项基准验证","2026-04-06",[1569,2004,2005,1599,1541],"TypeScript","MCP","深度剖析开源终端 AI 编程智能体 MiniCode（1k+ GitHub Stars）的核心架构：从零重型依赖的 ReAct 规划循环、声明式工具注册表与权限拦截层、MCP 协议帧自动协商，到 30 项可执行 Tool-Use 评测基准（86.7% 完成率、100% 错误恢复率）的失败分布归因与闭环优化。","\u002Fimages\u002Fposts\u002Fminicode-source-analysis\u002Fminicode-cover-20261018.webp","终端编程智能体控制台被 ReAct 状态环、Unified Diff 补丁预览门与 MCP 协议插槽环绕的等距技术插画","MiniCode 是我作为核心开发者参与构建的开源终端 Agentic Coding 引擎（目前已在 GitHub 获得 1,000+ Stars）。项目的核心设计目标是：在不引入庞大运行时框架的前提下，用约 5,000 行高内聚的","AI 项目源码导读",{"slug":2012,"path":2013,"title":2014,"date":2002,"tags":2015,"description":2016,"cover":2017,"coverAlt":2018,"excerpt":2019,"draft":1497,"hidden":1497,"published":1499,"readingTime":1723,"series":1997,"seriesOrder":526},"deepscientist-tech-stack","\u002Fposts\u002Fdeepscientist-tech-stack","DeepScientist 架构与运行时全解析：超 100K 上下文科研智能体、MCP 沙箱与高并发流式系统",[1569,1541,2005,1992,1654,1652],"深度拆解服务 7,000+ 用户的科研自主智能体平台 DeepScientist：从一题一仓库（Quest）持久化状态机、MCP 工具栈（memory \u002F artifact \u002F bash_exec）、Docker SDK 动态容器编排，到 SSE + WebSocket 双流通信与 PostgreSQL \u002F Redis \u002F Prometheus 生产级基础设施。","\u002Fimages\u002Fposts\u002Fdeepscientist-tech-stack\u002Fdeepscientist-cover-20261018.webp","科研智能体编排核心通过 memory、artifact 与隔离容器沙箱三座协议桥连接双流工作台的等距架构插画","DeepScientist（deepscientist.cc）是我们在西湖大学自然语言处理实验室研发的自主科学发现与科研管理平台，目前累计服务 7,000+ 科研用户。与单轮对话或纯文本文献问答工具不同，科研智能体需要在数小时乃至数",1795088944397]