{"id":110032,"date":"2026-09-25T18:17:06","date_gmt":"2026-09-25T10:17:06","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/110032.html"},"modified":"2026-09-25T18:17:06","modified_gmt":"2026-09-25T10:17:06","slug":"day-19%c2%b71-vit%e8%a7%86%e8%a7%89%e7%bc%96%e7%a0%81%ef%bc%9a%e5%9b%be%e7%89%87%e5%a6%82%e4%bd%95%e5%8f%98%e6%88%90%e8%a7%86%e8%a7%89token","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/110032.html","title":{"rendered":"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token"},"content":{"rendered":"

> \u4e0a\u4e00\u7bc7：\u00a018\u00b73 \u300aMRoPE\u2014\u20143D \u4f4d\u7f6e\u7f16\u7801\u7ed9\u89c6\u89c9\u7559\u7684\u5e2d\u4f4d\u300b｜ \u4e0b\u4e00\u7bc7：\u00a0\u00a019\u00b71 \u300aViT\u89c6\u89c9\u7f16\u7801：\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token\u300b<\/p>\n

\u4e00\u53e5\u8bdd\u5bfc\u8bfb：ViT \u89c6\u89c9\u7f16\u7801\u6d41\u7a0b：\u56fe\u7247\u50cf\u7d20\u7ecf\u9884\u5904\u7406\u3001PatchEmbed\u300124 \u5c42\u81ea\u6ce8\u610f\u529b\u4e0e\u5408\u5e76\u5c42\u53d8\u6210\u89c6\u89c9 token；\u677f\u7aef 64\u00d764 \u5408\u6210\u56fe\u5b9e\u6d4b 4 \u4e2a\u89c6\u89c9 token\u3001\u7f16\u7801\u8017\u65f6 295.3 ms\u3002<\/p>\n

\u771f\u673a\u5b9e\u6d4b\u901a\u8fc7：\u672c\u6587\u5b9e\u9a8c\u5df2\u5728 RK3588 \u677f\u7aef\u5b9e\u6d4b\u5b8c\u6210（2026-09；\u65b9\u6cd5\u5b66\u4e0e\u539f\u59cb\u8bb0\u5f55\u89c1\u4ed3\u5e93 docs \u4e0e\u300a\u5b9e\u9a8c\u811a\u672c\u300b\u76ee\u5f55）<\/p>\n

Day 18 \u6536\u5c3e\u8bb2\u4e86 MRoPE：\u89c6\u89c9 token \u5728\u8fdb\u5165 LLM \u524d\u4f1a\u62ff\u5230 (pos_t, pos_h, pos_w) \u4e09\u5957\u5750\u6807\u3002\u8fd9\u7bc7\u56de\u7b54\u201c\u8fd9\u4e9b\u89c6\u89c9 token \u4ece\u54ea\u6765\u201d\u7684\u7b2c\u4e00\u534a：\u56fe\u7247\u50cf\u7d20 \u2192 ViT \u7f16\u7801 \u2192 \u89c6\u89c9\u7279\u5f81\u3002\u5f15\u64ce\u7684\u89c6\u89c9\u5854\u5728 src\/model\/vllm_vision.c，\u677f\u7aef（RK3588 + Qwen3-VL-2B \u771f\u6743\u91cd）\u5b9e\u6d4b：64\u00d764 \u5408\u6210\u56fe\u7ecf ViT 24 \u5c42\u5f97\u5230 4 \u4e2a\u89c6\u89c9 token，\u7f16\u7801\u8017\u65f6 295.3 ms\u3002<\/p>\n

1. \u77e5\u8bc6\u70b9：\u89c6\u89c9\u5854\u628a\u56fe\u7247\u201c\u88c1\u201d\u6210\u4ec0\u4e48<\/h3>\n

1.1 \u591a\u6a21\u6001 LLM \u7684\u56fe\u7247\u8fdb\u6cd5<\/h4>\n

Qwen3-VL \u7684\u89c6\u89c9\u5854\u662f\u4e00\u6761\u72ec\u7acb\u7684 ViT（Vision Transformer），\u4e0e LLM \u5206\u5f00\u8bad\u7ec3\u6743\u91cd\u3001\u5206\u5f00\u91cf\u5316：<\/p>\n

RGB \u56fe\u7247 \u2192 resize\/\u5f52\u4e00\u5316（preprocess）
\n \u2192 PatchEmbed（patch=16，\u628a\u56fe\u5207\u6210 16\u00d716 \u7684\u5c0f\u5757\u5d4c\u5165）
\n \u2192 \u53ef\u5b66\u4e60\u4f4d\u7f6e\u5d4c\u5165 + \u89c6\u89c9 RoPE
\n \u2192 ViT 24 \u5c42\u81ea\u6ce8\u610f\u529b（DeepStack \u5728\u6307\u5b9a\u5c42\u989d\u5916\u62bd\u7279\u5f81）
\n \u2192 2\u00d72 \u7a7a\u95f4\u5408\u5e76（spatial_merge）
\n \u2192 Merger \u4e24\u5c42 MLP \u6295\u5f71\u5230 LLM \u9690\u7a7a\u95f4
\n \u2192 \u6bcf\u5f20\u56fe\u5f97\u5230 n \u4e2a\u201c\u89c6\u89c9 token\u201d（LLM \u7ef4\u5ea6） <\/p>\n

\u56fe\u7247\u5206\u8fa8\u7387\u51b3\u5b9a token \u6570：\u4e00\u5f20 64\u00d764 \u7684\u56fe \u2192 64\/16 \u00d7 64\/16 = 4\u00d74 = 16 \u4e2a patch \u2192 2\u00d72 \u5408\u5e76 \u2192 4 \u4e2a\u89c6\u89c9 token\u3002\u89c6\u89c9 token \u6570\u968f\u56fe\u53d8（\u4e0d\u662f\u56fa\u5b9a\u6570\u91cf），\u8fd9\u6b63\u662f MRoPE \u7f51\u683c (g_h, g_w) \u7684\u6765\u6e90\u3002<\/p>\n

1.2 \u5b9e\u6d4b\u7684 2B \u89c6\u89c9\u914d\u7f6e<\/h4>\n

\u677f\u7aef\u52a0\u8f7d\u65e5\u5fd7（\u4e0d\u662f\u5934\u6587\u4ef6\u91cc\u7684\u201c27 \u5c42\u201d\u5ba3\u4f20\u6ce8\u91ca，\u662f config.json \u9a71\u52a8\u3001manifest \u4f50\u8bc1\u7684\u771f\u5b9e\u503c）：<\/p>\n

[ST] Vision: depth=24 hidden=1024 heads=16 ffn=4096 patch=16 temporal=2 merge=2 out=2048
\n[ST] DeepStack layers: 5 11 17
\n[ST] MRoPE sections: 24 20 20
\nConfig: dim=2048 layers=28 heads=16 kv_heads=8 hd=128 ffn=6144 vocab=151936 <\/p>\n

\u6ce8\u610f：2B \u7684 ViT \u662f 24 \u5c42（vllm_vision.h \u5934\u6ce8\u91ca\u5199\u201c27-layer ViT\u201d，\u90a3\u662f 8B \u7684\u53e3\u5f84），\u89c6\u89c9\u9690\u5bbd 1024\u3001patch 16\u30012\u00d72 \u5408\u5e76\u3001temporal=2（\u89c6\u9891\u6cbf\u65f6\u95f4\u8f74\u4e5f\u505a patch）\u3002DeepStack \u5728\u5c42 5\/11\/17 \u5404\u62bd\u4e00\u7ec4\u4e2d\u95f4\u7279\u5f81\u3002\u751f\u4ea7 model.vqf（2,331,316,232 B，flags=0x63）\u91cc\u80fd\u770b\u5230\u5bf9\u5e94\u7684 27 \u4e2a v_* \u5f20\u91cf：4 \u4e2a Q8 \u5927\u77e9\u9635（v_q8_qkv\/proj\/fc1\/fc2，\u5404 24 \u5c42）+ patch\/pos \u5d4c\u5165 + \u504f\u7f6e\/\u5f52\u4e00 + merger 6 \u4e2a + deepstack 6 \u4e2a\u3002<\/p>\n

2. \u5bf9\u5e94\u4ee3\u7801：encode \u7684\u516b\u6b65\u6d41\u6c34<\/h3>\n

st_vision_encode_image（vllm_vision.c 1144 \u884c\u8d77）\u6309 8 \u6b65\u8d70\u5b8c\u5168\u6d41\u7a0b：<\/p>\n

\/* Step 1: Preprocess (resize + normalize) *\/ preprocess_image(…) \/\/ 1144-1174
\n\/* Step 2: Patch embedding *\/ patch_embed_image(…) \/\/ 1176-1179
\n\/* Step 3: Interpolated position embedding + vision RoPE table *\/
\n vision_pos_encode(…) \/\/ 1181-1183
\n\/* Step 5: ViT forward with DeepStack *\/
\nfor (int l = 0; l < cfg->vis_depth; l++) { \/\/ 1194-1214
\n vit_block_forward(vis, l, n_patches);
\n if (cfg->vis_ds_idx[d] == l) { \/\/ DeepStack \u5c42 5\/11\/17
\n vision_ds_merge(…); \/\/ \u62bd\u4e2d\u95f4\u7279\u5f81 \u2192 ds_features
\n }
\n}
\n\/* Step 6: Spatial merge of final ViT output *\/ spatial_merge(…) \/\/ 1217-1226
\n\/* Step 7: Main merger *\/ merger_forward(…) \/\/ 1228-1236
\n\/* Step 8: Visual token output = main merged tokens only *\/ \/\/ 1238+ <\/p>\n

\u914d\u5957\u7684\u6743\u91cd\u7ed3\u6784\u5728 include\/model\/vllm_vision.h：STVisionWeights（patch_embed\u3001pos_embed\u3001attn_qkv\/proj\u3001norm1\/2\u3001mlp_fc1\/fc2\u3001merger_*\u3001ds_*）\u4e0e STVisionState（\u8fd0\u884c\u65f6\u7f13\u51b2：pixel_values\u3001hidden_states\u3001rot_cos\/sin\u3001ds_features）\u3002\u8981\u76ef\u7684\u8f6c\u6362\u5173\u7cfb\u6709\u4e24\u6761：<\/p>\n

  • \u5c42\u6570\u7531 config \u9a71\u52a8：vis_depth\/vis_hidden\/vis_heads\/vis_patch\/vis_merge \u5168\u90e8\u6765\u81ea config.json，\u4ee3\u7801\u6ca1\u6709\u5199\u6b7b 8B \u6570\u503c\u2014\u2014\u6240\u4ee5\u540c\u4e00\u4e2a\u5f15\u64ce\u80fd\u8dd1 2B（24 \u5c42）\u4e5f\u80fd\u8dd1 8B（27 \u5c42）；<\/li>\n
  • Q8_0 \u6743\u91cd\u53ea\u8bfb：ViT \u7684 QKV\/O\/fc1\/fc2 \u52a0\u8f7d\u540e\u91cf\u5316\u4e3a Q8_0 \u5e76\u91ca\u653e F32 \u526f\u672c（st_vision_weights_free_f32，vllm_vision.h 153-159 \u884c\u6ce8\u91ca：RK3588 16GB \u677f\u5b50\u4e0a\u8fd9\u6b65\u80fd\u7701\u7ea6 1.65GB \u5185\u5b58）\u2014\u2014\u4e0e\u6587\u672c\u6743\u91cd\u540c\u4e00\u4e2a\u201c\u63a8\u7406\u53ea\u8bfb\u91cf\u5316\u526f\u672c\u201d\u7eaa\u5f8b\u3002<\/li>\n

    3. \u6539\u52a8\u540e\u679c：\u677f\u7aef\u8dd1\u4e00\u6b21\u5b8c\u6574\u56fe\u7247 \u2192 \u7f16\u7801 \u2192 \u751f\u6210<\/h3>\n

    \u5b9e\u6d4b\u53e3\u5f84：RK3588（Orange Pi 5 Plus）\/ aarch64 \/ Release \/ 2026-09-07\u3002\u65b9\u6cd5：–multimodal \u81ea\u6d4b（\u9ed8\u8ba4 64\u00d764 \u5408\u6210\u6e10\u53d8\u56fe），\u6a21\u578b\u76ee\u5f55\u7528 bind-mount \u628a 2B \u6302\u5230\u81ea\u6d4b\u9ed8\u8ba4\u5bfb\u627e\u7684\u8def\u5f84\u4e0b\u8fd0\u884c\u3002<\/p>\n

    [ST-Q8] Estimated memory: 4.5 GB (F32 token + Q8_0 + Q4_0 nibble weights)
    \nWeights loading: 20.5 sec
    \n[TOK] Loaded 151669 tokens from vocab.bin (max_len=256)
    \n Tokenizer: 151669 tokens, vision IDs: start=151652 end=151653 img=151655 vid=151656<\/p>\n

    — Image 1\/1: synthetic (64×64) —
    \n[VIS] Image encoded: 64×64 \u2192 16 patches \u2192 4 visual tokens
    \n 64×64 -> 4 patches -> 4 visual tokens (grid 2×2)
    \n [VIS] Encode: 295.3 ms
    \n Prompt: 17 tokens (10 text, 6 visual)
    \n Prefill: 513ms (33.16 tok\/s)
    \n Decode: 60 tokens in 2974ms (TPOT=50ms, 20.2 tok\/s)
    \n Total: TTFT=513ms E2E=3486ms<\/p>\n

    — Video Analysis —
    \n[VIS] Video encoded: 4 frames 64×64 \u2192 32 patches \u2192 8 visual tokens
    \n Video 4 frames 64×64 -> 8 visual tokens
    \n [VIS] Encode: 628.5 ms <\/p>\n

    \u4e09\u4ef6\u4e8b\u503c\u5f97\u62c6\u5f00\u8bfb：<\/p>\n

  • token \u6570\u968f\u201c\u5c42\u201d\u8d70：\u5355\u56fe 64\u00d764 \u2192 16 patch \u2192 \u5408\u5e76 \u2192 4 token；4 \u5e27\u89c6\u9891\u56e0\u4e3a temporal=2 \u5148\u6cbf\u65f6\u95f4\u8f74\u4e24\u4e24\u5f52\u5e76（4 \u5e27 \u2192 2 \u7ec4），\u518d\u5404\u81ea 4\u00d74=16 patch \u2192 2\u00d716=32 patch \u2192 \u5408\u5e76 \u2192 2\u00d74 = 8 token\u3002temporal patch \u662f\u201c\u89c6\u9891 token \u53d8\u5c11\u201d\u7684\u5173\u952e（\u5e27\u95f4\u4fe1\u606f\u5728 patch \u9636\u6bb5\u5148\u878d\u5408）\u3002<\/li>\n
  • encode \u662f\u677f\u8f7d\u5b9e\u6570：64\u00d764 \u7f16\u7801 295.3 ms\u30014 \u5e27\u89c6\u9891 628.5 ms\u2014\u2014\u8fd9\u5c31\u662f\u201c\u89c6\u89c9\u5854\u5728\u8fb9\u7f18\u677f\u5b50\u4e0a\u7b97\u4e00\u6b21\u201d\u7684\u6210\u672c\u91cf\u7ea7，\u4e5f\u662f 19-2 \u91cc\u5ba2\u6237\u7aef\u9884\u7f16\u7801\u534f\u8bae\u5b58\u5728\u7684\u7406\u7531\u3002<\/li>\n
  • \u8bda\u5b9e\u8fb9\u754c：\u8be5\u81ea\u6d4b\u7684\u6a21\u578b\u76ee\u5f55\u540d\u4e0e\u63d0\u793a\u8bcd\u662f\u6309 8B \u5199\u7684，\u6211\u4eec\u7528 bind-mount \u628a 2B \u6302\u8fdb\u53bb\u8dd1\u901a；2B \u7684 lm_head \u662f tied（safetensors \u91cc\u6ca1\u6709\u72ec\u7acb lm_head.weight），\u6240\u4ee5\u89e3\u7801\u6587\u672c\u6bb5\u4e0d\u4f5c\u4e3a\u8d28\u91cf\u7ed3\u8bba\u2014\u2014\u672c\u6587\u5f15\u7528\u7684 encode\/grid\/\u65f6\u5e8f\u6570\u5b57\u6765\u81ea\u7f16\u7801\u4e0e prefill \u6bb5，\u5747\u4e0e serve \u90e8\u7f72\u8def\u5f84（19-2）\u4ea4\u53c9\u9a8c\u8bc1\u4e00\u81f4\u3002\u89c6\u89c9\u7f16\u7801\u540e\u7684\u771f\u6b63\u201c\u770b\u56fe\u56de\u7b54\u201d\u8d28\u91cf\u9a8c\u8bc1，\u8bf7\u770b 19-2 \u7528 model.vqf \u90e8\u7f72\u8dd1\u7684 serve \u7ed3\u679c\u3002<\/li>\n

    4. \u5b66\u5458\u8c03\u8bd5\u4efb\u52a1<\/h3>\n
      \n
    • A \u6863（\u677f\u7aef\u52a8\u624b）：\n
    • \u590d\u523b\u672c\u6587：\u8dd1 –multimodal（\u9700\u8981\u628a 2B \u6216 8B \u6a21\u578b\u76ee\u5f55\u653e\u5230\u4ee3\u7801\u67e5\u627e\u7684 Modl\/\u5343\u95ee3_VL_8B_Instruct\/ \u4e0b，\u6216\u81ea\u5df1 bind-mount），\u6838\u5bf9\u201c16 patches \u2192 4 visual tokens \/ Encode \u2248 295ms \/ grid 2×2\u201d；<\/li>\n
    • \u5f00 VLLM_MMDBG=1 \u91cd\u8dd1\u4e00\u6b21，\u89c2\u5bdf [VIS] preprocess\/patch_embed\/pos_embed\/vit_all\/merge \u5404\u6b65\u9aa4\u7684\u7d2f\u8ba1\u8017\u65f6\u2014\u2014\u54ea\u4e00\u6b65\u662f\u70ed\u70b9？（\u8fd9\u662f\u7ed9 Day 28 \u6027\u80fd\u7bc7\u7559\u7684\u4f0f\u7b14）<\/li>\n
    • \u7528 manifest_vqf.py（Day 17）\u6838\u5bf9 model.vqf \u91cc 27 \u4e2a v_* \u5f20\u91cf\u7684\u884c\u6570\u4e0e [ST] Vision: \u4e00\u884c\u5bf9\u5f97\u4e0a（qkv\/proj\/fc1\/fc2 \u90fd\u662f 24 \u884c）\u3002<\/li>\n<\/li>\n
    • B \u6863（\u7eaf\u8bfb\u6e90\u7801）：\u8bfb vllm_vision.c 1144\u20131240，\u56de\u7b54：\u2460 \u4e3a\u4ec0\u4e48 encode \u7684\u8f93\u51fa\u53ea\u7b97\u201cmain merged tokens\u201d\u800c DeepStack \u7279\u5f81\u662f\u53e6\u5b58\u5230 ds_features（\u63d0\u793a：vllm_vision.h 124 \u884c\u6ce8\u91ca\u2014\u2014\u5b83\u8981\u6ce8\u5165\u5230 LLM \u7684\u54ea\u4e9b\u5c42，\u8fdb\u5165 prefill \u7684\u54ea\u4e2a\u73af\u8282）？\u2461 \u82e5\u628a vis_merge \u4ece 2 \u6539\u6210 1，\u540c\u6837 64\u00d764 \u4f1a\u4ea7\u51fa\u591a\u5c11 token？\u6539\u6210 4 \u5462？\u2462 \u82e5 config.json \u7684 vis_depth \u4e0e model.vqf \u91cc v_q8_qkv \u7684\u5b9e\u9645\u5c42\u6570\u4e0d\u4e00\u81f4，\u52a0\u8f7d\u6216\u524d\u5411\u4f1a\u5728\u54ea\u4e00\u6b65\u5148\u51fa\u9519（\u8bfb st_vision_load_weights \u7684\u5206\u5c42\u5faa\u73af）？<\/li>\n<\/ul>\n

      \u9884\u671f\u8f93\u51fa：\u4e00\u5f20\u4f60\u81ea\u5df1\u677f\u5b50\u7684\u201c\u56fe\u7247\u5c3a\u5bf8 \u2192 patch \u6570 \u2192 \u5408\u5e76\u540e token \u6570 \u2192 encode ms\u201d\u5c0f\u8868，\u5e76\u80fd\u8bf4\u51fa ViT \u5404\u6b65\u9aa4\u91cc\u6700\u8017\u65f6\u7684\u4e00\u73af\u3002<\/p>\n

      \u6536\u5c3e<\/h3>\n
        \n
      • \u672c\u7bc7\u6e90\u7801\u70b9\u540d：vllm_vision.c（encode_image \u516b\u6b65 1144\u20131240\u3001vit_block_forward\/merger）\u3001vllm_vision.h（\u6743\u91cd\/\u8fd0\u884c\u65f6\u7ed3\u6784）\u3001vllm_safetensors.c（\u89c6\u89c9\u5f20\u91cf\u52a0\u8f7d\u4e0e Q8 \u91cf\u5316）<\/li>\n
      • \u5f00\u6e90\u4ed3\u5e93：Kestrel-LLM (Gitee)（\u6e90\u7801\u53ef\u5f97\u53cc\u8bb8\u53ef：\u5b66\u4e60 \/ \u5b66\u672f\u7814\u7a76\u514d\u8d39）<\/li>\n
      • \u4e0b\u7bc7\u9884\u544a：\u677f\u8f7d ViT \u8981 295ms，\u90a3\u201c\u628a\u7f16\u7801\u632a\u5230\u5ba2\u6237\u7aef\u201d\u4f1a\u600e\u6837？19-2 \u62c6 vision_tokens \u5ba2\u6237\u7aef\u9884\u7f16\u7801\u534f\u8bae：\u56fe\u7247 base64 \u4e0a\u884c\u7684\u677f\u7aef ViT \u8def\u5f84\u3001\u7f13\u5b58\u547d\u4e2d\u8def\u5f84\u3001\u4ee5\u53ca\u201c\u5ba2\u6237\u7aef\u81ea\u5df1\u7b97\u597d\u7279\u5f81\u76f4\u63a5\u585e\u8fdb\u6765\u201d\u7684\u7b2c\u4e09\u6761\u8def\u3002<\/li>\n
      • \u5173\u952e\u8bcd：ViT\u3001\u89c6\u89c9 token\u3001\u89c6\u89c9\u5854\u3001\u591a\u6a21\u6001\u3001PatchEmbed\u3001DeepStack\u3001MRoPE\u3001RK3588<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

        > \u4e0a\u4e00\u7bc7：\u00a018\u00b73 \u300aMRoPE\u2014\u20143D \u4f4d\u7f6e\u7f16\u7801\u7ed9\u89c6\u89c9\u7559\u7684\u5e2d\u4f4d\u300b｜ \u4e0b\u4e00\u7bc7：\u00a0\u00a019\u00b71 \u300aViT\u89c6\u89c9\u7f16\u7801：\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token\u300b
        \n\u4e00\u53e5\u8bdd\u5bfc\u8bfb：ViT \u89c6\u89c9\u7f16\u7801\u6d41\u7a0b：\u56fe\u7247\u50cf\u7d20\u7ecf\u9884\u5904\u7406\u3001PatchEmbed\u300124 \u5c42\u81ea\u6ce8\u610f\u529b\u4e0e\u5408\u5e76\u5c42\u53d8\u6210\u89c6\u89c9 token；\u677f\u7aef 64\u00d764 \u5408\u6210\u56fe\u5b9e\u6d4b 4 \u4e2a\u89c6\u89c9 token\u3001\u7f16\u7801\u8017\u65f6 295.3 ms\u3002 \u771f\u673a\u5b9e\u6d4b\u901a\u8fc7：\u672c\u6587<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[174,280,50,427,492],"topic":[],"class_list":["post-110032","post","type-post","status-publish","format-standard","hentry","category-server","tag-c","tag-harmonyos","tag-50","tag-427","tag-492"],"yoast_head":"\nDay 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/110032.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"> \u4e0a\u4e00\u7bc7：\u00a018\u00b73 \u300aMRoPE\u2014\u20143D \u4f4d\u7f6e\u7f16\u7801\u7ed9\u89c6\u89c9\u7559\u7684\u5e2d\u4f4d\u300b｜ \u4e0b\u4e00\u7bc7：\u00a0\u00a019\u00b71 \u300aViT\u89c6\u89c9\u7f16\u7801：\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token\u300b \u4e00\u53e5\u8bdd\u5bfc\u8bfb：ViT \u89c6\u89c9\u7f16\u7801\u6d41\u7a0b：\u56fe\u7247\u50cf\u7d20\u7ecf\u9884\u5904\u7406\u3001PatchEmbed\u300124 \u5c42\u81ea\u6ce8\u610f\u529b\u4e0e\u5408\u5e76\u5c42\u53d8\u6210\u89c6\u89c9 token；\u677f\u7aef 64\u00d764 \u5408\u6210\u56fe\u5b9e\u6d4b 4 \u4e2a\u89c6\u89c9 token\u3001\u7f16\u7801\u8017\u65f6 295.3 ms\u3002 \u771f\u673a\u5b9e\u6d4b\u901a\u8fc7：\u672c\u6587\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/110032.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-25T10:17:06+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/110032.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/110032.html\",\"name\":\"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-25T10:17:06+00:00\",\"dateModified\":\"2026-09-25T10:17:06+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/110032.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/110032.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/110032.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/110032.html","og_locale":"zh_CN","og_type":"article","og_title":"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"> \u4e0a\u4e00\u7bc7：\u00a018\u00b73 \u300aMRoPE\u2014\u20143D \u4f4d\u7f6e\u7f16\u7801\u7ed9\u89c6\u89c9\u7559\u7684\u5e2d\u4f4d\u300b｜ \u4e0b\u4e00\u7bc7：\u00a0\u00a019\u00b71 \u300aViT\u89c6\u89c9\u7f16\u7801：\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token\u300b \u4e00\u53e5\u8bdd\u5bfc\u8bfb：ViT \u89c6\u89c9\u7f16\u7801\u6d41\u7a0b：\u56fe\u7247\u50cf\u7d20\u7ecf\u9884\u5904\u7406\u3001PatchEmbed\u300124 \u5c42\u81ea\u6ce8\u610f\u529b\u4e0e\u5408\u5e76\u5c42\u53d8\u6210\u89c6\u89c9 token；\u677f\u7aef 64\u00d764 \u5408\u6210\u56fe\u5b9e\u6d4b 4 \u4e2a\u89c6\u89c9 token\u3001\u7f16\u7801\u8017\u65f6 295.3 ms\u3002 \u771f\u673a\u5b9e\u6d4b\u901a\u8fc7：\u672c\u6587","og_url":"https:\/\/www.wsisp.com\/helps\/110032.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-25T10:17:06+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/110032.html","url":"https:\/\/www.wsisp.com\/helps\/110032.html","name":"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-25T10:17:06+00:00","dateModified":"2026-09-25T10:17:06+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/110032.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/110032.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/110032.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Day 19\u00b71 ViT\u89c6\u89c9\u7f16\u7801\uff1a\u56fe\u7247\u5982\u4f55\u53d8\u6210\u89c6\u89c9token"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/110032","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=110032"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/110032\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=110032"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=110032"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=110032"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=110032"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}