{"id":120648,"date":"2026-10-31T23:48:24","date_gmt":"2026-10-31T15:48:24","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/120648.html"},"modified":"2026-10-31T23:48:24","modified_gmt":"2026-10-31T15:48:24","slug":"%e4%bb%8e-1024-%e5%9d%97-gpu-%e5%88%b0-558-%e4%b8%87%e7%be%8e%e5%85%83%ef%bc%9a%e6%8b%86%e8%a7%a3-deepseek-v3-%e7%9a%84%e7%ae%97%e6%b3%95-%e7%b3%bb%e7%bb%9f%e5%8d%8f%e5%90%8c%e8%ae%be%e8%ae%a1","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/120648.html","title":{"rendered":"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1"},"content":{"rendered":"

\u6458\u8981：\u4e00\u4e2a 671B \u53c2\u6570\u7684\u6a21\u578b，\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b 37B，\u8bad\u7ec3\u5168\u7a0b 278.8 \u4e07 H800 GPU \u5c0f\u65f6，\u6ca1\u6709\u51fa\u73b0\u4e00\u6b21\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0\u2014\u2014DeepSeek-V3 \u7528\u4e00\u7ec4\u8fd1\u4e4e\u201c\u53cd\u5e38\u8bc6\u201d\u7684\u6570\u5b57，\u628a\u5927\u6a21\u578b\u8bad\u7ec3\u4ece\u201c\u5806\u7b97\u529b\u201d\u7684\u53d9\u4e8b\u4e2d\u62fd\u4e86\u51fa\u6765\u3002<\/span>\u672c\u6587\u4e0d\u6ee1\u8db3\u4e8e\u5217\u4e3e MLA\u3001MoE\u3001MTP \u8fd9\u4e9b\u540d\u8bcd，\u800c\u662f\u6cbf\u7740\u201c\u89e3\u8026\u201d\u8fd9\u6761\u4e3b\u7ebf，\u9010\u5c42\u62c6\u89e3\u6bcf\u4e00\u4e2a\u8bbe\u8ba1\u51b3\u7b56\u80cc\u540e\u7684\u7ea6\u675f\u3001\u53d6\u820d\u4e0e\u5de5\u7a0b\u5b9e\u73b0：\u4e3a\u4ec0\u4e48 MLA \u8981\u628a KV \u7f13\u5b58\u538b\u8fdb\u4f4e\u79e9\u7a7a\u95f4，\u4e3a\u4ec0\u4e48 MoE \u8981\u628a\u4e13\u5bb6\u5207\u5230 256 \u4e2a\u518d\u914d\u4e0a top-1 \u8def\u7531，\u4e3a\u4ec0\u4e48 MTP \u4e0d\u662f\u63a8\u7406\u52a0\u901f\u5668\u800c\u9996\u5148\u662f\u8bad\u7ec3\u4fe1\u53f7\u589e\u7a20\u5668，\u4e3a\u4ec0\u4e48 DualPipe \u80fd\u628a\u901a\u4fe1\u201c\u85cf\u201d\u8fdb\u8ba1\u7b97\u91cc\u3002\u6700\u540e，\u56de\u5230 1024 \u7a0b\u5e8f\u5458\u8282\u7684\u8bed\u5883：\u5f53\u201c\u9ad8\u6548\u201d\u6210\u4e3a AI \u5de5\u7a0b\u7684\u786c\u7ea6\u675f，\u6211\u4eec\u4ece DeepSeek-V3 \u8eab\u4e0a\u80fd\u6284\u8d70\u7684\u4e0d\u53ea\u662f\u67d0\u4e2a\u6a21\u5757，\u800c\u662f\u4e00\u79cd\u5728\u8d44\u6e90\u8fb9\u754c\u4e0a\u505a\u8bbe\u8ba1\u7684\u65b9\u6cd5\u8bba\u3002<\/span><\/p>\n

\"\"<\/p>\n

1. \u4e00\u4e2a\u8ba9\u7845\u8c37\u91cd\u65b0\u7b97\u8d26\u7684\u6570\u5b57<\/h3>\n

2024 \u5e74\u5e95，DeepSeek-V3 \u7684\u6280\u672f\u62a5\u544a\u7ed9\u51fa\u4e86\u4e00\u4e2a\u5728\u5f53\u65f6\u8fd1\u4e4e\u523a\u773c\u7684\u6570\u5b57：671B \u603b\u53c2\u6570，278.8 \u4e07 H800 GPU \u5c0f\u65f6，\u6309 $2\/\u5361\u65f6\u4f30\u7b97\u7ea6 558 \u4e07\u7f8e\u5143\u3002\u4f5c\u4e3a\u5bf9\u7167，\u540c\u7b49\u53c2\u6570\u89c4\u6a21\u7684\u7a20\u5bc6\u6a21\u578b，\u8bad\u7ec3\u6210\u672c\u901a\u5e38\u662f\u8fd9\u4e2a\u6570\u5b57\u7684\u6570\u500d\u5230\u6570\u5341\u500d\u3002<\/span><\/p>\n

\"\"<\/p>\n

\u8fd9\u4e2a\u6210\u672c\u4e0d\u662f\u9760\u201c\u5c11\u8bad\u201d\u6362\u6765\u7684\u300214.8T token \u7684\u8bad\u7ec3\u6570\u636e，2048 \u5f20 H800 \u8dd1\u7ea6\u4e24\u4e2a\u6708，\u5168\u7a0b\u6ca1\u6709\u51fa\u73b0\u4e00\u6b21\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0，\u6ca1\u6709\u505a\u8fc7\u4e00\u6b21\u56de\u6eda\u2014\u2014\u5728\u5927\u89c4\u6a21 MoE \u8bad\u7ec3\u4e2d，\u8fd9\u672c\u8eab\u5c31\u662f\u4e00\u4e2a\u7cfb\u7edf\u7a33\u5b9a\u6027\u7684\u5de5\u7a0b\u58f0\u660e\u3002<\/p>\n

\u4f46\u771f\u6b63\u503c\u5f97\u6280\u672f\u4eba\u5458\u53cd\u590d\u5480\u56bc\u7684，\u4e0d\u662f\u6210\u672c\u6570\u5b57\u672c\u8eab，\u800c\u662f\u8fd9\u4e2a\u6570\u5b57\u5f97\u4ee5\u6210\u7acb\u7684\u7ed3\u6784\u6027\u524d\u63d0：DeepSeek-V3 \u5728\u56db\u4e2a\u5f7c\u6b64\u8026\u5408\u7684\u5c42\u9762\u4e0a，\u5206\u522b\u505a\u4e86\u4e00\u6b21\u201c\u89e3\u8026\u201d\u3002<\/p>\n\n\u89e3\u8026\u7ef4\u5ea6\u6838\u5fc3\u673a\u5236\u89e3\u51b3\u7684\u95ee\u9898<\/tr>\n\n\n\n\n\n
\u5bb9\u91cf vs \u8ba1\u7b97<\/td>\n\u7ec6\u7c92\u5ea6 MoE，671B \u603b\u53c2 \/ 37B \u6fc0\u6d3b<\/td>\n\u53c2\u6570\u91cf\u51b3\u5b9a\u77e5\u8bc6\u5bb9\u91cf，\u4f46\u6bcf token \u7b97\u529b\u4e0d\u518d\u968f\u53c2\u6570\u91cf\u7ebf\u6027\u589e\u957f<\/td>\n<\/tr>\n
\u5185\u5b58 vs \u7cbe\u5ea6<\/td>\nMLA \u4f4e\u79e9 KV \u538b\u7f29 + FP8 \u6df7\u5408\u7cbe\u5ea6<\/td>\nKV \u7f13\u5b58\u4ece 32768 \u5143\u7d20\/ token\/ \u5c42\u538b\u5230 576，\u663e\u5b58\u5e26\u5bbd\u4e0d\u518d\u662f\u63a8\u7406\u74f6\u9888<\/td>\n<\/tr>\n
\u8bad\u7ec3\u4fe1\u53f7 vs \u6b65\u6570<\/td>\n\u4e32\u884c MTP（\u591a token \u9884\u6d4b）<\/td>\n\u6bcf\u4e2a\u8bad\u7ec3\u4f4d\u7f6e\u4ea7\u751f\u591a\u500d\u68af\u5ea6\u4fe1\u53f7，\u6837\u672c\u6548\u7387\u4e0d\u518d\u53d7\u9650\u4e8e\u5355\u6b65\u5355 token \u76d1\u7763<\/td>\n<\/tr>\n
\u901a\u4fe1 vs \u8ba1\u7b97<\/td>\nDualPipe \u53cc\u5411\u6d41\u6c34\u7ebf<\/td>\n\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\u7684 all-to-all \u901a\u4fe1\u88ab\u201c\u85cf\u201d\u8fdb\u524d\u5411 \/ \u53cd\u5411\u8ba1\u7b97\u4e2d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u8fd9\u56db\u4e2a\u89e3\u8026\u4e0d\u662f\u5b64\u7acb\u7684\u4f18\u5316，\u800c\u662f\u4e00\u5957\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1：\u7b97\u6cd5\u5c42\u9762\u7684\u7a00\u758f\u6027\u9009\u62e9，\u5fc5\u987b\u7531\u7cfb\u7edf\u5c42\u9762\u7684\u901a\u4fe1\u3001\u663e\u5b58\u3001\u5e76\u884c\u7b56\u7565\u6765\u627f\u63a5；\u7cfb\u7edf\u5c42\u9762\u7684\u786c\u4ef6\u7ea6\u675f，\u53cd\u8fc7\u6765\u7ea6\u675f\u4e86\u7b97\u6cd5\u53ef\u4ee5\u8d70\u591a\u8fdc\u3002\u4e0b\u9762\u9010\u5c42\u62c6\u5f00\u3002<\/p>\n

\"\"<\/p>\n

2. MoE：\u628a\u201c\u53c2\u6570\u662f\u80fd\u529b，\u4e5f\u662f\u7a0e\u201d\u8fd9\u6761\u516c\u5f0f\u62c6\u6389<\/h3>\n

2.1 \u7a20\u5bc6 Transformer \u7684\u7b97\u529b\u56f0\u5c40<\/h4>\n

\u6807\u51c6\u7a20\u5bc6 Transformer \u4e2d，\u6bcf\u4e2a token \u90fd\u8981\u6d41\u8fc7\u5168\u90e8\u53c2\u6570\u3002FFN \u5c42\u7684\u8ba1\u7b97\u91cf\u6b63\u6bd4\u4e8e\u53c2\u6570\u603b\u91cf，\u6a21\u578b\u53d8\u5927 N \u500d，\u6bcf token \u7684\u8bad\u7ec3\u548c\u63a8\u7406\u7b97\u529b\u4e5f\u8fd1\u4f3c\u6da8 N \u500d\u3002GPT-3 \u7684 175B \u53c2\u6570\u5df2\u7ecf\u8ba9\u8fd9\u79cd\u7ebf\u6027\u5173\u7cfb\u663e\u5f97\u6602\u8d35，\u800c\u5230\u4e86 671B \u91cf\u7ea7，\u5982\u679c\u4ecd\u7136\u7a20\u5bc6，\u6bcf token \u7684\u6fc0\u6d3b\u53c2\u6570\u5c31\u662f 671B\u2014\u2014\u8bad\u7ec3\u6210\u672c\u548c\u63a8\u7406\u5ef6\u8fdf\u90fd\u4f1a\u53d8\u5f97\u96be\u4ee5\u627f\u53d7\u3002<\/p>\n

MoE \u7684\u7834\u5c40\u601d\u8def\u6781\u5176\u7b80\u6d01：\u53c2\u6570\u53ef\u4ee5\u5f88\u591a，\u4f46\u6bcf\u4e2a token \u53ea\u8d70\u5176\u4e2d\u4e00\u5c0f\u90e8\u5206\u3002\u628a Transformer \u4e2d\u7684 FFN \u5c42\u66ff\u6362\u4e3a\u4e00\u7ec4\u5e76\u884c\u7684\u201c\u4e13\u5bb6\u201d\u7f51\u7edc，\u6bcf\u4e2a\u4e13\u5bb6\u672c\u8d28\u4e0a\u662f\u72ec\u7acb\u7684 FFN，\u8def\u7531\u5668\u4e3a\u6bcf\u4e2a token \u9009\u62e9\u5c11\u6570\u4e13\u5bb6\u6fc0\u6d3b，\u5176\u4f59\u4e13\u5bb6\u4fdd\u6301\u4f11\u7720\u3002<\/span>\u4e8e\u662f\u201c\u53c2\u6570\u91cf\u201d\u548c\u201c\u6bcf token \u7b97\u529b\u201d\u88ab\u62c6\u6210\u4e86\u4e24\u4e2a\u53ef\u4ee5\u72ec\u7acb\u8c03\u8282\u7684\u65cb\u94ae：\u524d\u8005\u51b3\u5b9a\u77e5\u8bc6\u5bb9\u91cf，\u540e\u8005\u51b3\u5b9a\u5b9e\u9645\u5f00\u9500\u3002<\/p>\n

\"\"<\/p>\n

2.2 DeepSeek-V3 \u7684 MoE：256 \u4e2a\u8def\u7531\u4e13\u5bb6 + 1 \u4e2a\u5171\u4eab\u4e13\u5bb6<\/h4>\n

DeepSeek-V3 \u7684\u6bcf\u4e2a MoE \u5c42\u5305\u542b\u00a0256 \u4e2a\u8def\u7531\u4e13\u5bb6\u548c 1 \u4e2a\u5171\u4eab\u4e13\u5bb6，\u603b\u5171 61 \u5c42 Transformer，\u9664\u524d\u4e09\u5c42\u5916 FFN \u5168\u90e8\u66ff\u6362\u4e3a MoE\u3002\u5171\u4eab\u4e13\u5bb6\u59cb\u7ec8\u6fc0\u6d3b，\u627f\u62c5\u901a\u7528\u3001\u9ad8\u9891\u7684\u77e5\u8bc6；\u8def\u7531\u4e13\u5bb6\u5219\u88ab\u7ec6\u7c92\u5ea6\u5207\u5206，\u6bcf\u4e2a\u4e13\u5bb6\u7684 FFN \u7ef4\u5ea6\u53ea\u6709 2048，\u8d1f\u8d23\u4e13\u7528\u3001\u7a00\u758f\u7684\u80fd\u529b\u3002<\/p>\n

\u8fd9\u91cc\u6709\u4e24\u4e2a\u5bb9\u6613\u88ab\u5ffd\u7565\u7684\u8bbe\u8ba1\u76f4\u89c9\u3002<\/p>\n

\u7b2c\u4e00，\u4e3a\u4ec0\u4e48\u8981\u201c\u7ec6\u7c92\u5ea6\u201d？\u00a0\u4f20\u7edf\u7684 MoE（\u5982 Mixtral 8\u00d77B）\u4f7f\u7528\u5c11\u91cf\u5927\u4e13\u5bb6，\u6bcf\u4e2a\u4e13\u5bb6\u53c2\u6570\u91cf\u5f88\u5927，\u8def\u7531\u7684\u9009\u62e9\u7a7a\u95f4\u5176\u5b9e\u6709\u9650\u3002DeepSeek-V3 \u628a\u4e13\u5bb6\u5207\u5c0f\u3001\u6570\u91cf\u62c9\u591a，\u76f8\u5f53\u4e8e\u628a\u8def\u7531\u5668\u7684\u201c\u51b3\u7b56\u5206\u8fa8\u7387\u201d\u63d0\u9ad8\u4e86：\u540c\u6837\u6fc0\u6d3b 37B \u53c2\u6570，\u4ece 8 \u4e2a\u5927\u4e13\u5bb6\u4e2d\u9009，\u548c\u4ece 256 \u4e2a\u5c0f\u4e13\u5bb6\u4e2d\u9009，\u540e\u8005\u80fd\u7ec4\u5408\u51fa\u66f4\u7cbe\u7ec6\u7684\u4e13\u5bb6\u5206\u5de5\u3002\u4ee3\u4ef7\u662f\u8def\u7531\u548c\u8c03\u5ea6\u66f4\u590d\u6742，\u4f46\u8fd9\u6b63\u597d\u7531\u7cfb\u7edf\u5c42\u6765\u6d88\u5316\u3002<\/p>\n

\u7b2c\u4e8c，\u4e3a\u4ec0\u4e48\u8981 top-1 \u8def\u7531？\u00a0\u6bcf\u4e2a token \u7ecf Sigmoid \u95e8\u63a7\u6253\u5206\u540e，\u53ea\u6fc0\u6d3b\u5206\u6570\u6700\u9ad8\u7684\u00a01 \u4e2a\u8def\u7531\u4e13\u5bb6\u52a0\u4e0a\u5171\u4eab\u4e13\u5bb6，\u6bd4 DeepSeek-V2 \u7684 top-2 \u66f4\u6fc0\u8fdb，\u628a\u6bcf token \u6fc0\u6d3b\u53c2\u6570\u538b\u5230 37B\u3002top-1 \u7684\u6536\u76ca\u662f\u8ba1\u7b97\u6700\u7701，\u4f46\u98ce\u9669\u662f\u8def\u7531\u574d\u7f29\u2014\u2014\u8def\u7531\u5668\u503e\u5411\u4e8e\u628a\u6240\u6709 token \u4e22\u7ed9\u5c11\u6570\u4e13\u5bb6，\u5176\u4f59\u4e13\u5bb6\u201c\u997f\u6b7b\u201d\u3002DeepSeek-V3 \u7684\u5e94\u5bf9\u4e0d\u662f\u52a0\u8f85\u52a9\u635f\u5931，\u800c\u662f\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861：\u7ed9\u6bcf\u4e2a\u4e13\u5bb6\u7ef4\u62a4\u4e00\u4e2a bias \u9879，\u6839\u636e\u5b9e\u65f6\u8d1f\u8f7d\u52a8\u6001\u8c03\u6574，\u8ba9\u8def\u7531\u6982\u7387\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u81ea\u53d1\u8d8b\u4e8e\u5747\u8861，\u800c\u4e0d\u662f\u7528\u989d\u5916\u7684 loss \u9879\u53bb\u201c\u60e9\u7f5a\u201d\u4e0d\u5747\u8861\u2014\u2014\u540e\u8005\u4f1a\u5728\u4e3b\u4efb\u52a1\u6027\u80fd\u4e0a\u5f15\u5165\u9000\u5316\u3002\u4e0b\u9762\u8fd9\u6bb5\u95e8\u63a7\u4ee3\u7801\u662f\u8fd9\u4e00\u673a\u5236\u6700\u76f4\u63a5\u7684\u4f53\u73b0：<\/p>\n

class MoEGate(nn.Module):
\n def __init__(self, hidden_size, n_routed_experts, num_experts_per_tok,
\n n_group, topk_group, routed_scaling_factor=1.0):
\n super().__init__()
\n self.top_k = num_experts_per_tok
\n self.n_group = n_group
\n self.topk_group = topk_group
\n self.weight = nn.Parameter(torch.empty(n_routed_experts, hidden_size))
\n # bias \u662f buffer，\u4e0d\u662f parameter：\u7531\u8d1f\u8f7d\u7edf\u8ba1\u66f4\u65b0，\u4e0d\u8d70\u68af\u5ea6
\n self.register_buffer("e_score_correction_bias",
\n torch.zeros(n_routed_experts))<\/p>\n

def forward(self, hidden_states):
\n bsz, seq_len, h = hidden_states.shape
\n hidden_states = hidden_states.view(-1, h)<\/p>\n

# \u5173\u952e：\u7528 sigmoid \u800c\u975e softmax \u6253\u5206
\n logits = F.linear(hidden_states.float(), self.weight.float())
\n scores = logits.sigmoid()<\/p>\n

# bias \u53ea\u53c2\u4e0e\u201c\u9009\u62e9\u201d，\u4e0d\u53c2\u4e0e\u201c\u52a0\u6743\u201d
\n scores_for_choice = scores + self.e_score_correction_bias.unsqueeze(0)
\n group_scores = (scores_for_choice
\n .view(bsz * seq_len, self.n_group, -1)
\n .topk(2, dim=-1)[0].sum(dim=-1))
\n group_idx = torch.topk(group_scores, k=self.topk_group,
\n dim=-1, sorted=False)[1]
\n group_mask = torch.zeros_like(group_scores)
\n group_mask.scatter_(1, group_idx, 1)
\n score_mask = (group_mask.unsqueeze(-1)
\n .expand(bsz * seq_len, self.n_group,
\n self.n_routed_experts \/\/ self.n_group)
\n .reshape(bsz * seq_len, -1))
\n tmp_scores = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
\n _, topk_idx = torch.topk(tmp_scores, k=self.top_k, dim=-1, sorted=False)<\/p>\n

# \u6743\u91cd\u6765\u81ea\u672a\u52a0 bias \u7684\u539f\u59cb scores
\n topk_weight = scores.gather(1, topk_idx)
\n if self.top_k > 1:
\n denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
\n topk_weight = topk_weight \/ denominator
\n topk_weight = topk_weight * self.routed_scaling_factor
\n return topk_idx, topk_weight <\/p>\n

2.3 \u4e13\u5bb6\u5e76\u884c\u4e0e all-to-all：MoE \u771f\u6b63\u7684\u5de5\u7a0b\u95e8\u69db<\/h4>\n

MoE \u5728\u6570\u5b66\u4e0a\u662f\u6f02\u4eae\u7684，\u4f46\u5728\u7cfb\u7edf\u4e0a\u6709\u4e00\u4e2a\u786c\u9aa8\u5934：256 \u4e2a\u4e13\u5bb6\u4e0d\u53ef\u80fd\u5168\u90e8\u585e\u8fdb\u4e00\u5f20 GPU\u3002\u91c7\u7528\u4e13\u5bb6\u5e76\u884c\u540e，\u4e13\u5bb6\u88ab\u5206\u5e03\u5230\u591a\u5f20\u5361\u4e0a，\u4e00\u4e2a token \u5982\u679c\u88ab\u8def\u7531\u5230\u8fdc\u7aef\u5361\u4e0a\u7684\u4e13\u5bb6，\u5c31\u9700\u8981\u4e00\u6b21\u8de8\u5361 all-to-all \u901a\u4fe1\u3002\u5728 DeepSeek-V3 \u7684\u8bad\u7ec3\u914d\u7f6e\u4e2d，\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\u5e26\u6765\u7684\u901a\u4fe1\u5f00\u9500\u4e00\u5ea6\u4f7f\u8ba1\u7b97-\u901a\u4fe1\u6bd4\u63a5\u8fd1\u00a01:1\u2014\u2014\u4e5f\u5c31\u662f\u8bf4，GPU \u6709\u5c06\u8fd1\u4e00\u534a\u7684\u65f6\u95f4\u5728\u7b49\u901a\u4fe1\u3002<\/p>\n

all-to-all \u5728\u4ee3\u7801\u91cc\u7684\u5f62\u6001\u5e76\u4e0d\u590d\u6742，\u4f46\u9b54\u9b3c\u5728\u7ec6\u8282\u91cc\u3002\u4e0b\u9762\u662f\u4ece\u63a8\u7406 \/ \u8bad\u7ec3\u8def\u5f84\u4e2d\u6458\u51fa\u7684\u6838\u5fc3\u9aa8\u67b6，\u5b83\u5c55\u793a\u4e86 token \u5982\u4f55\u6309\u4e13\u5bb6\u5206\u7ec4\u3001\u8de8 rank \u4ea4\u6362\u3001\u8ba1\u7b97\u540e\u518d\u6536\u56de：<\/p>\n

# \u6bcf\u4e2a token \u88ab\u8def\u7531\u5230\u7684\u4e13\u5bb6\u8ba1\u6570
\ncnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
\ncnts.scatter_(1, topk_ids, 1)
\ntokens_per_expert = cnts.sum(dim=0)
\nidxs = topk_ids.view(-1).argsort()
\nsorted_tokens = x[idxs \/\/ topk_ids.shape[1]]<\/p>\n

if self.ep_size > 1:
\n # \u5148\u4ea4\u6362\u201c\u6bcf\u4e2a rank \u6709\u591a\u5c11 token \u8981\u53bb\u54ea\u4e9b\u4e13\u5bb6\u201d\u7684\u5143\u4fe1\u606f
\n tokens_per_ep_rank = tokens_per_expert.view(self.ep_size, -1).sum(dim=1)
\n tokens_per_expert_group = tokens_per_expert.new_empty(tokens_per_expert.shape[0])
\n dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)<\/p>\n

output_splits = (tokens_per_expert_group.view(self.ep_size, -1)
\n .sum(1).cpu().numpy().tolist())
\n gathered_tokens = sorted_tokens.new_empty(
\n tokens_per_expert_group.sum(dim=0).cpu().item(), sorted_tokens.shape[1])
\n input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()<\/p>\n

# \u771f\u6b63\u7684 token \u4ea4\u6362
\n dist.all_to_all(
\n list(gathered_tokens.split(output_splits)),
\n list(sorted_tokens.split(input_split_sizes)),
\n )<\/p>\n

# \u672c\u5730\u4e13\u5bb6\u8ba1\u7b97\u540e，\u518d\u53cd\u5411 all_to_all \u6536\u56de <\/p>\n

\u8fd9\u5c31\u5f15\u51fa\u4e86\u540e\u9762\u7684 DualPipe\u3002\u4f46\u5728\u6b64\u4e4b\u524d，\u8fd8\u6709\u4e00\u4e2a\u540c\u6837\u7d27\u8feb\u7684\u74f6\u9888：\u63a8\u7406\u4fa7\u7684 KV \u7f13\u5b58\u3002<\/p>\n

3. MLA：\u628a KV \u7f13\u5b58\u4ece 32768 \u538b\u5230 576<\/h3>\n

\"\"<\/p>\n

3.1 KV \u7f13\u5b58\u4e3a\u4ec0\u4e48\u662f\u63a8\u7406\u7684\u7b2c\u4e00\u74f6\u9888<\/h4>\n

\u81ea\u56de\u5f52\u751f\u6210\u4e2d，\u6bcf\u751f\u6210\u4e00\u4e2a\u65b0 token，\u90fd\u9700\u8981\u7528\u4e4b\u524d\u6240\u6709 token \u7684 Key \u548c Value \u8ba1\u7b97\u6ce8\u610f\u529b\u3002\u4e3a\u4e86\u907f\u514d\u91cd\u590d\u8ba1\u7b97，K \u548c V \u4f1a\u88ab\u7f13\u5b58\u4e0b\u6765\u3002\u5bf9\u4e8e\u4e00\u4e2a\u6709\u00a0LL\u00a0\u5c42\u3001HH\u00a0\u4e2a\u6ce8\u610f\u529b\u5934\u3001\u5934\u7ef4\u5ea6\u00a0d_{head}dhead\u200b\u00a0\u7684\u6a21\u578b，KV \u7f13\u5b58\u7684\u5185\u5b58\u9700\u6c42\u662f：<\/p>\n

\\\\text{Memory}_{KV} = 2 \\\\times L \\\\times H \\\\times d_{head} \\\\times T \\\\times \\\\text{bytes}MemoryKV\u200b=2\u00d7L\u00d7H\u00d7dhead\u200b\u00d7T\u00d7bytes<\/p>\n

\u4ee5 GPT-3 \u91cf\u7ea7（96 \u5c42\u300196 \u5934\u3001128 \u7ef4\u3001\u5e8f\u5217\u957f\u5ea6 2048\u3001FP16）\u4e3a\u4f8b，KV \u7f13\u5b58\u7ea6\u00a09.6 GB，\u5373\u4f7f\u5728\u9ad8\u7aef GPU \u4e0a\u4e5f\u53ea\u591f\u540c\u65f6\u670d\u52a1\u5c11\u6570\u51e0\u4e2a\u7528\u6237\u3002\u5728\u957f\u4e0a\u4e0b\u6587\u548c\u9ad8\u5e76\u53d1\u573a\u666f\u4e0b，\u9650\u5236\u90e8\u7f72\u5bc6\u5ea6\u7684\u5f80\u5f80\u4e0d\u662f\u7b97\u529b，\u800c\u662f\u663e\u5b58\u5e26\u5bbd\u3002<\/p>\n

\u6807\u51c6 MHA \u7684 KV \u7f13\u5b58\u5927\u5c0f\u4e0e\u5934\u6570\u00a0HH\u00a0\u548c\u5934\u7ef4\u5ea6\u00a0d_{head}dhead\u200b\u00a0\u6210\u6b63\u6bd4\u3002GQA \u901a\u8fc7\u8ba9\u591a\u4e2a Query \u5934\u5171\u4eab\u4e00\u7ec4 K\/V \u6765\u538b\u7f29，\u4f46\u538b\u7f29\u7684\u7c92\u5ea6\u662f\u201c\u5934\u7ea7\u201d\u7684\u3002MLA \u8d70\u4e86\u4e00\u6761\u66f4\u6839\u672c\u7684\u8def\u7ebf：\u4e0d\u5728\u5934\u7ea7\u5171\u4eab，\u800c\u662f\u5728\u4f4e\u79e9\u7a7a\u95f4\u91cc\u538b\u7f29\u3002<\/p>\n

3.2 \u4f4e\u79e9\u538b\u7f29\u7684\u6570\u5b66\u5f62\u5f0f<\/h4>\n

MLA \u7684\u6838\u5fc3\u89c2\u5bdf\u6765\u81ea\u4e00\u4e2a\u7ecf\u9a8c\u4e8b\u5b9e：\u4f20\u7edf\u6ce8\u610f\u529b\u6743\u91cd\u77e9\u9635\u5b58\u5728\u663e\u8457\u7684\u4f4e\u79e9\u7279\u6027，\u5178\u578b\u79e9\u7ea6\u4e3a\u9690\u85cf\u5c42\u7ef4\u5ea6\u7684 15%\u201320%\u3002\u65e2\u7136 K\/V \u7684\u6709\u6548\u4fe1\u606f\u96c6\u4e2d\u5728\u4e00\u4e2a\u4f4e\u7ef4\u5b50\u7a7a\u95f4\u91cc，\u5c31\u6ca1\u6709\u5fc5\u8981\u5b58\u50a8\u5b8c\u6574\u7684\u00a0d_{head}dhead\u200b\u00a0\u7ef4\u8868\u793a，\u800c\u662f\u5148\u538b\u7f29\u5230\u4f4e\u7ef4\u6f5c\u5728\u5411\u91cf，\u9700\u8981\u65f6\u518d\u89e3\u538b\u7f29\u3002<\/p>\n

\u5177\u4f53\u6765\u8bf4，MLA \u4e0d\u76f4\u63a5\u7f13\u5b58 K \u548c V，\u800c\u662f\u7f13\u5b58\u4e00\u4e2a\u4f4e\u79e9\u538b\u7f29\u540e\u7684\u9690\u5411\u91cf：<\/p>\n

c_{KV} = \\\\text{RMSNorm}(W_{dkv} \\\\cdot x)cKV\u200b=RMSNorm(Wdkv\u200b\u22c5x)<\/p>\n

\u5176\u4e2d\u00a0W_{dkv}Wdkv\u200b\u00a0\u662f\u4e0b\u6295\u5f71\u77e9\u9635，d_{kv}dkv\u200b\u00a0\u662f\u4f4e\u79e9\u7ef4\u5ea6\u3002\u5728\u6ce8\u610f\u529b\u8ba1\u7b97\u65f6，\u518d\u901a\u8fc7\u4e0a\u6295\u5f71\u6062\u590d：<\/p>\n

K_{content} = W_{uk} \\\\cdot c_{KV}, \\\\quad V = W_{uv} \\\\cdot c_{KV}Kcontent\u200b=Wuk\u200b\u22c5cKV\u200b,V=Wuv\u200b\u22c5cKV\u200b<\/p>\n

\u53ea\u7f13\u5b58\u00a0c_{KV}cKV\u200b，\u800c\u975e\u5b8c\u6574\u7684 K \u548c V\u3002KV \u7f13\u5b58\u7684\u7f29\u51cf\u56e0\u5b50\u4e3a\u00a0(2 \\\\times d_{head}) \/ d_{kv}(2\u00d7dhead\u200b)\/dkv\u200b\u3002-26<\/p>\n

\u771f\u6b63\u60ca\u4eba\u7684\u662f\u538b\u7f29\u7684\u7edd\u5bf9\u91cf\u7ea7\u3002\u4ee5 DeepSeek-V3 \u7684\u91cf\u7ea7，\u6bcf token \u6bcf\u5c42\u7684 KV \u7f13\u5b58\u5143\u7d20\u6570\u4ece\u6807\u51c6 MHA \u7684\u00a032768 \u538b\u7f29\u5230 576，\u76f8\u5dee\u00a056.89 \u500d\u3002\u8fd9 576 \u7ef4\u7531\u4e24\u90e8\u5206\u7ec4\u6210：\u4e00\u4e2a\u00a0512 \u7ef4\u7684\u538b\u7f29\u9690\u5411\u91cf，\u52a0\u4e0a\u00a064 \u7ef4\u5171\u4eab\u7684 RoPE \u952e\u5206\u91cf\u3002RoPE \u5206\u91cf\u9700\u8981\u4fdd\u7559\u662f\u4e3a\u4e86\u7ef4\u6301\u4f4d\u7f6e\u7f16\u7801\u7684\u5b8c\u6574\u6027，\u4f46\u5b83\u53ea\u6709 64 \u7ef4，\u5360\u6bd4\u8f83\u5c0f\u3002<\/p>\n

\u8fd9\u4e2a\u538b\u7f29\u6bd4\u610f\u5473\u7740\u4ec0\u4e48？\u5728\u76f8\u540c\u7684 H800 \u663e\u5b58\u4e0b，MLA \u5141\u8bb8\u7684\u5e76\u53d1\u5e8f\u5217\u6570\u6bd4\u6807\u51c6 MHA \u9ad8\u51fa\u63a5\u8fd1\u4e00\u4e2a\u6570\u91cf\u7ea7\u3002\u63a8\u7406\u6210\u672c\u7684\u51b3\u5b9a\u6743\u4ece\u201c\u663e\u5b58\u80fd\u88c5\u591a\u5c11 KV\u201d\u56de\u5230\u4e86\u201c\u7b97\u529b\u80fd\u8dd1\u591a\u5c11 FLOPs\u201d\u2014\u2014\u800c\u540e\u8005\u6070\u6070\u662f GPU \u66f4\u5145\u88d5\u7684\u8d44\u6e90\u3002<\/p>\n

3.3 MLA \u4e0e MoE \u7684\u914d\u5408\u5173\u7cfb<\/h4>\n

MLA \u548c MoE \u4e0d\u662f\u4e24\u4e2a\u72ec\u7acb\u7684\u201c\u4eae\u70b9\u529f\u80fd\u201d，\u5b83\u4eec\u5728\u63a8\u7406\u8def\u5f84\u4e0a\u5f62\u6210\u4e92\u8865\u3002MoE \u628a\u6bcf token \u7684\u8ba1\u7b97\u91cf\u538b\u5230 37B \u53c2\u6570\u7684\u89c4\u6a21，MLA \u628a\u6bcf token \u7684\u5185\u5b58\u8db3\u8ff9\u538b\u5230 576 \u5143\u7d20\/\u5c42\u3002\u4e24\u8005\u53e0\u52a0\u7684\u6548\u679c\u662f：\u4e00\u4e2a 671B \u603b\u53c2\u6570\u7684\u6a21\u578b，\u5728\u63a8\u7406\u65f6\u65e2\u80fd\u4ee5\u63a5\u8fd1\u4e2d\u7b49\u7a20\u5bc6\u6a21\u578b\u7684\u7b97\u529b\u8fd0\u884c，\u53c8\u80fd\u4ee5\u63a5\u8fd1\u5c0f\u6a21\u578b\u7684\u663e\u5b58\u5e26\u5bbd\u8fd0\u884c\u3002\u5982\u679c\u6ca1\u6709 MLA，MoE \u7701\u4e0b\u6765\u7684\u8ba1\u7b97\u4f18\u52bf\u4f1a\u88ab KV \u7f13\u5b58\u7684\u5185\u5b58\u74f6\u9888\u5403\u6389；\u5982\u679c\u6ca1\u6709 MoE，MLA \u7701\u4e0b\u6765\u7684\u663e\u5b58\u53c8\u4f1a\u88ab\u7a20\u5bc6 FFN \u7684\u8ba1\u7b97\u91cf\u6df9\u6ca1\u3002<\/p>\n

\u8fd9\u79cd\u6210\u5bf9\u8bbe\u8ba1\u662f DeepSeek-V3 \u67b6\u6784\u601d\u7ef4\u7684\u4e00\u4e2a\u7f29\u5f71：\u4e0d\u8981\u5b64\u7acb\u5730\u4f18\u5316\u4e00\u4e2a\u6307\u6807，\u800c\u8981\u540c\u65f6\u770b\u8ba1\u7b97\u548c\u5185\u5b58\u4e24\u6761\u66f2\u7ebf，\u8ba9\u5b83\u4eec\u5728\u540c\u4e00\u4e2a\u5de5\u4f5c\u70b9\u4e0a\u76f8\u9047\u3002<\/p>\n

4. MTP：\u8bad\u7ec3\u4fe1\u53f7\u589e\u7a20，\u800c\u4e0d\u662f\u63a8\u7406\u52a0\u901f\u5668<\/h3>\n

\"\"<\/p>\n

4.1 \u6807\u51c6\u81ea\u56de\u5f52\u8bad\u7ec3\u7684\u4fe1\u53f7\u74f6\u9888<\/h4>\n

\u6807\u51c6\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d，\u6bcf\u4e2a\u4f4d\u7f6e\u53ea\u9884\u6d4b\u4e0b\u4e00\u4e2a token，\u635f\u5931\u51fd\u6570\u53ea\u5728\u00a0t_{i+1}ti+1\u200b\u00a0\u8fd9\u4e00\u4e2a\u4f4d\u7f6e\u4ea7\u751f\u68af\u5ea6\u3002\u5bf9\u4e8e\u4e00\u6bb5\u957f\u5ea6\u4e3a\u00a0TT\u00a0\u7684\u6587\u672c，\u8bad\u7ec3\u4fe1\u53f7\u662f\u201c\u5355\u70b9\u201d\u7684：\u6a21\u578b\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u53ea\u88ab\u544a\u77e5\u201c\u4e0b\u4e00\u4e2a\u8bcd\u662f\u4ec0\u4e48\u201d，\u800c\u201c\u4e0b\u4e0b\u4e2a\u8bcd\u3001\u4e0b\u4e0b\u4e0b\u4e2a\u8bcd\u201d\u7684\u4fe1\u606f\u6ca1\u6709\u76f4\u63a5\u8fdb\u5165\u68af\u5ea6\u3002<\/p>\n

MTP（Multi-Token Prediction）\u6539\u53d8\u4e86\u8fd9\u4e00\u70b9：\u8ba9\u6bcf\u4e2a\u4f4d\u7f6e\u540c\u65f6\u6216\u4e32\u884c\u5730\u9884\u6d4b\u672a\u6765\u00a0nn\u00a0\u4e2a token，\u4f7f\u76d1\u7763\u4fe1\u53f7\u4ece\u5355\u70b9\u53d8\u6210\u4e00\u6bb5\u7a20\u5bc6\u7684\u4fe1\u53f7\u5e26\u3002\u540c\u4e00\u6bb5\u6587\u672c\u4ea7\u751f\u00a0nn\u00a0\u500d\u7684 next-k \u4fe1\u53f7，\u5173\u952e\u8f6c\u79fb（choice point）\u7684\u540e\u679c\u4f1a\u51fa\u73b0\u5728\u591a\u4e2a\u9884\u6d4b\u5934\u7684 loss \u4e2d，\u4ece\u800c\u9690\u5f0f\u63d0\u9ad8\u6a21\u578b\u5bf9\u91cd\u8981\u51b3\u7b56\u7684\u654f\u611f\u5ea6\u3002<\/p>\n

4.2 DeepSeek-V3 \u7684\u4e32\u884c MTP \u4e0e\u56e0\u679c\u94fe\u4fdd\u6301<\/h4>\n

\"\"<\/p>\n

MTP \u6709\u4e24\u6761\u6280\u672f\u8def\u7ebf\u3002Gloeckle \u7b49\u4eba\u7684\u5e76\u884c MTP \u4f7f\u7528\u5171\u4eab trunk \u52a0\u591a\u4e2a\u72ec\u7acb\u8f93\u51fa\u5934，\u5404\u5934\u6761\u4ef6\u72ec\u7acb\u9884\u6d4b\u672a\u6765\u4e0d\u540c\u4f4d\u7f6e\u7684 token，t_{t+2}tt+2\u200b\u00a0\u7684\u5934\u770b\u4e0d\u5230\u00a0t_{t+1}tt+1\u200b\u00a0\u7684\u8f93\u51fa\u3002DeepSeek-V3 \u9009\u62e9\u4e86\u53e6\u4e00\u6761\u8def：\u4e32\u884c MTP，\u4fdd\u6301\u5b8c\u6574\u7684\u56e0\u679c\u94fe，\u8ba9\u7b2c\u00a0kk\u00a0\u6b65\u7684\u9884\u6d4b\u4f9d\u8d56\u524d\u4e00\u6b65\u7684\u7ed3\u679c，\u800c\u4e0d\u662f\u5404\u4e2a\u5934\u5404\u81ea\u4e3a\u6218\u3002<\/p>\n

\u8fd9\u4e2a\u9009\u62e9\u7684\u5173\u952e\u5728\u4e8e\u4e3b\u76ee\u6807\u4e0d\u540c\u3002\u5e76\u884c MTP \u7684\u989d\u5916\u5934\u5728\u63a8\u7406\u65f6\u53ef\u4ee5\u62ff\u6765\u505a speculative decoding，\u4f46 DeepSeek-V3 \u7684 MTP \u9996\u5148\u662f\u8bad\u7ec3\u76ee\u6807：\u8bad\u7ec3\u589e\u76ca\u662f\u4e3b\u76ee\u6807，\u63a8\u7406\u65f6\u65e2\u53ef\u4ee5\u4e22\u5f03\u989d\u5916\u7684\u9884\u6d4b\u6a21\u5757，\u4e5f\u53ef\u4ee5\u4fdd\u7559\u4e0b\u6765\u5f53 drafter\u3002\u4e32\u884c\u7ed3\u6784\u4fdd\u6301\u4e86\u56e0\u679c\u6027，\u4f7f\u5f97\u8bad\u7ec3\u4fe1\u53f7\u589e\u7a20\u7684\u8fc7\u7a0b\u672c\u8eab\u5c31\u5728\u6559\u6a21\u578b\u201c\u4e3a\u540e\u9762\u51e0\u6b65\u9884\u7559\u4fe1\u606f\u201d，\u800c\u4e0d\u662f\u7b80\u5355\u5730\u5e76\u884c\u62df\u5408\u591a\u4e2a\u72ec\u7acb\u76ee\u6807\u3002<\/p>\n

\u4e32\u884c MTP \u7684\u5b9e\u73b0\u96c6\u4e2d\u5728\u8f93\u5165\u5e8f\u5217\u7684\u79fb\u4f4d\u548c\u9010\u6df1\u5ea6\u7684 MTP \u5757\u4e0a\u3002\u4e0b\u9762\u8fd9\u6bb5\u00a0roll_mtp_sequence\u00a0\u7684\u8bed\u4e49，\u662f\u7406\u89e3 MTP \u8bad\u7ec3\u6570\u636e\u6d41\u7684\u6700\u5c0f\u5165\u53e3：<\/p>\n

# MTP \u8f93\u5165 \/ \u6807\u7b7e\u6309 depth \u79fb\u4f4d（\u793a\u610f\u903b\u8f91）
\n# \u4e3b\u6a21\u578b: \u8f93\u5165 [t0, t1, t2, t3, t4, t5]
\n# \u6807\u7b7e [t1, t2, t3, t4, t5, t6]
\n#
\n# MTP depth 1: \u8f93\u5165 [t1, t2, t3, t4, t5, FILL]
\n# \u6807\u7b7e [t2, t3, t4, t5, t6, IGNORE]
\n#
\n# MTP depth 2: \u8f93\u5165 [t2, t3, t4, t5, FILL, FILL]
\n# \u6807\u7b7e [t3, t4, t5, t6, IGNORE, IGNORE]<\/p>\n

def roll_mtp_sequence(tokens, positions, shift):
\n """\u6cbf\u5e8f\u5217\u7ef4\u79fb\u4f4d；\u8de8 document \u8fb9\u754c\u7684\u4f4d\u7f6e\u586b\u5145 IGNORE\/FILL\u3002"""
\n shifted = tokens[…, shift:]
\n # packed \u5e8f\u5217\u4e2d，position \u91cd\u7f6e\u5904\u89c6\u4e3a\u8fb9\u754c
\n valid_mask = (positions[…, shift:] > positions[…, :-shift])
\n return shifted, valid_mask <\/p>\n

\u4ee3\u4ef7\u662f\u8bad\u7ec3\u65f6\u7684\u8ba1\u7b97\u56fe\u66f4\u957f\u3001\u663e\u5b58\u7ba1\u7406\u66f4\u590d\u6742\u3002\u4f46 DeepSeek-V3 \u7684\u7cfb\u7edf\u5c42\u518d\u6b21\u627f\u63a5\u4e86\u8fd9\u4e2a\u4ee3\u4ef7\u2014\u2014FP8 \u6df7\u5408\u7cbe\u5ea6\u628a\u6fc0\u6d3b\u503c\u548c\u6743\u91cd\u7684\u5b58\u50a8\u538b\u5230 8 \u4f4d，DualPipe \u628a\u989d\u5916\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u91cd\u53e0\u8d77\u6765\u3002\u7b97\u6cd5\u63d0\u51fa\u9700\u6c42，\u7cfb\u7edf\u8d1f\u8d23\u4e70\u5355，\u8fd9\u6761\u5206\u5de5\u7ebf\u5728 MTP \u4e0a\u4f53\u73b0\u5f97\u5c24\u4e3a\u6e05\u6670\u3002<\/p>\n

\"\"<\/p>\n

4.3 MTP \u7684\u771f\u5b9e\u6536\u76ca\u8fb9\u754c<\/h4>\n

\u9700\u8981\u8bda\u5b9e\u5730\u8bf4，MTP \u5e76\u4e0d\u662f\u4e00\u4e2a\u201c\u514d\u8d39\u5348\u9910\u201d\u3002\u5b83\u5728\u5c0f\u6a21\u578b\u4e0a\u7684\u6536\u76ca\u5e76\u4e0d\u7a33\u5b9a，\u5728 13B \u91cf\u7ea7\u624d\u5f00\u59cb\u660e\u663e\u62c9\u5f00\u5dee\u8ddd\u3002\u6362\u53e5\u8bdd\u8bf4，MTP \u662f\u4e00\u79cd\u201c\u5927\u6a21\u578b\u624d\u5212\u7b97\u201d\u7684\u8bad\u7ec3\u6280\u5de7：\u6a21\u578b\u8d8a\u5c0f，\u989d\u5916\u9884\u6d4b\u5934\u5e26\u6765\u7684 overhead \u76f8\u5bf9\u4e8e\u4e3b\u4efb\u52a1\u6536\u76ca\u7684\u5360\u6bd4\u8d8a\u9ad8；\u6a21\u578b\u8d8a\u5927，\u4fe1\u53f7\u589e\u7a20\u7684\u6536\u76ca\u8d8a\u80fd\u8986\u76d6\u7cfb\u7edf\u6210\u672c\u3002DeepSeek-V3 \u5728 671B \u91cf\u7ea7\u4e0a\u4f7f\u7528 MTP，\u662f\u5728\u6b63\u786e\u7684\u89c4\u6a21\u4e0a\u505a\u4e86\u6b63\u786e\u7684\u9009\u62e9\u3002<\/p>\n

5. \u7cfb\u7edf\u5c42：FP8\u3001DualPipe \u4e0e\u201c\u628a\u901a\u4fe1\u85cf\u8d77\u6765\u201d<\/h3>\n

\"\"<\/p>\n

5.1 FP8 \u6df7\u5408\u7cbe\u5ea6：\u663e\u5b58\u548c\u5e26\u5bbd\u7684\u53cc\u91cd\u677e\u7ed1<\/h4>\n

FP8 \u8bad\u7ec3\u7684\u6838\u5fc3\u4ef7\u503c\u4e0d\u662f\u201c\u7b97\u5f97\u5feb\u201d，\u800c\u662f\u5728\u540c\u6837\u7684 GPU \u4e0a\u88c5\u4e0b\u66f4\u591a\u4e1c\u897f\u3002DeepSeek-V3 \u5728\u8bad\u7ec3\u4e2d\u4f7f\u7528 FP8 \u7cbe\u5ea6\u7684\u8ba1\u7b97\u548c\u5b58\u50a8，\u5927\u5e45\u964d\u4f4e\u4e86 GPU \u5185\u5b58\u9700\u6c42\u548c\u5b58\u50a8\u5e26\u5bbd\u538b\u529b\u3002\u5728\u63a8\u7406\u4fa7，FP8 \u6a21\u5f0f\u4e0b\u7684\u663e\u5b58\u5360\u7528\u53ea\u6709 BF16 \u7684\u4e00\u534a，\u4f7f\u5f97\u5728\u6709\u9650\u663e\u5b58\u4e0b\u90e8\u7f72 671B \u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n

@triton.jit
\ndef blockwise_fp8_gemm_kernel(a_ptr, a_s_ptr, b_ptr, b_s_ptr, y_ptr, …):
\n # \u6bcf\u4e2a program \u5904\u7406\u4e00\u5757\u8f93\u51fa；a\/b \u4ee5 FP8 \u5b58\u50a8，
\n # \u5bf9\u5e94\u7684 scale \u4ee5 block \u7c92\u5ea6\u5b58\u50a8\u5e76\u5728\u7d2f\u52a0\u65f6\u9010\u5757\u5e94\u7528
\n …
\n # \u6982\u5ff5\u4e0a：
\n # a_fp8_block * a_scale_block -> \u53cd\u91cf\u5316\u540e\u7684 a_block
\n # b_fp8_block * b_scale_block -> \u53cd\u91cf\u5316\u540e\u7684 b_block
\n # \u7d2f\u52a0\u4f7f\u7528 FP32，\u8f93\u51fa\u518d\u6309\u9700\u91cf\u5316 <\/p>\n

\u4f46 FP8 \u8bad\u7ec3\u7684\u96be\u70b9\u4e0d\u5728\u683c\u5f0f\u672c\u8eab，\u800c\u5728\u6570\u503c\u7a33\u5b9a\u6027\u3002\u5927\u89c4\u6a21 MoE \u8bad\u7ec3\u4e2d，\u4e13\u5bb6\u6fc0\u6d3b\u503c\u7684\u52a8\u6001\u8303\u56f4\u5dee\u5f02\u5f88\u5927，\u76f4\u63a5\u7528 FP8 \u5f88\u5bb9\u6613\u5728\u67d0\u4e2a\u4e13\u5bb6\u4e0a\u51fa\u73b0\u6ea2\u51fa\u6216\u4e0b\u6ea2，\u8fdb\u800c\u5f15\u53d1 loss \u5c16\u5cf0\u3002DeepSeek-V3 \u5168\u7a0b\u6ca1\u6709\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0，\u8bf4\u660e FP8 \u7684\u7f29\u653e\u7b56\u7565\u548c\u6570\u503c\u8fb9\u754c\u7ba1\u7406\u4e0e\u6a21\u578b\u67b6\u6784\u662f\u534f\u540c\u8bbe\u8ba1\u7684，\u800c\u4e0d\u662f\u4e8b\u540e\u5957\u4e0a\u53bb\u7684\u91cf\u5316\u65b9\u6848\u3002<\/p>\n

5.2 DualPipe：\u5f53\u901a\u4fe1\u5f00\u9500\u63a5\u8fd1 1:1 \u65f6<\/h4>\n

\u56de\u5230\u7b2c 2 \u8282\u7559\u4e0b\u7684\u95ee\u9898：\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\u7684 all-to-all \u901a\u4fe1\u4f7f\u8ba1\u7b97-\u901a\u4fe1\u6bd4\u63a5\u8fd1\u00a01:1，\u8fd9\u610f\u5473\u7740\u5982\u679c\u901a\u4fe1\u548c\u8ba1\u7b97\u4e32\u884c\u6267\u884c，GPU \u6709\u4e00\u534a\u65f6\u95f4\u5728\u7a7a\u8f6c\u3002<\/p>\n

DualPipe \u7684\u89e3\u6cd5\u4e0d\u662f\u201c\u51cf\u5c11\u901a\u4fe1\u91cf\u201d，\u800c\u662f\u628a\u901a\u4fe1\u85cf\u8fdb\u8ba1\u7b97\u91cc\u3002\u5b83\u6784\u5efa\u53cc\u5411\u6570\u636e\u6d41\u901a\u9053，\u8ba9\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\u5404\u81ea\u62e5\u6709\u72ec\u7acb\u7684\u7ba1\u9053，\u8ba1\u7b97\u4efb\u52a1\u548c\u901a\u4fe1\u4efb\u52a1\u5728\u4e24\u6761\u7ba1\u9053\u4e0a\u5e76\u884c\u63a8\u8fdb，\u5b9e\u73b0\u524d\u5411 \/ \u53cd\u5411\u8ba1\u7b97-\u901a\u4fe1\u9636\u6bb5\u7684\u5b8c\u5168\u91cd\u53e0，\u540c\u65f6\u51cf\u5c11\u6d41\u6c34\u7ebf\u6c14\u6ce1（\u8bbe\u5907\u7a7a\u95f2\u7b49\u5f85）\u3002\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u7684\u4f9d\u8d56\u5206\u6790，\u53ef\u5e76\u884c\u8ba1\u7b97\u7684\u68af\u5ea6\u8282\u70b9\u88ab\u63d0\u524d\u8c03\u5ea6，\u901a\u4fe1\u4e0e\u8ba1\u7b97\u7684\u91cd\u53e0\u7387\u63d0\u5347\u5230 90% \u4ee5\u4e0a\u3002<\/p>\n

\u8fd9\u91cc\u7684\u8bbe\u8ba1\u54f2\u5b66\u503c\u5f97\u5355\u72ec\u62ce\u51fa\u6765：\u5f53\u901a\u4fe1\u4e0d\u53ef\u907f\u514d\u65f6，\u4e0d\u8981\u53bb\u548c\u901a\u4fe1\u201c\u5bf9\u6297\u201d，\u800c\u8981\u91cd\u65b0\u6392\u5217\u8ba1\u7b97\u56fe，\u8ba9\u901a\u4fe1\u53d1\u751f\u5728\u8ba1\u7b97\u7684\u9634\u5f71\u91cc\u3002DualPipe \u6ca1\u6709\u53d1\u660e\u66f4\u5feb\u7684\u7f51\u7edc，\u5b83\u6539\u53d8\u7684\u662f\u8ba1\u7b97\u548c\u901a\u4fe1\u7684\u65f6\u95f4\u5173\u7cfb\u3002<\/p>\n

5.3 \u5e76\u884c\u7b56\u7565\u7684\u6574\u4f53\u62fc\u56fe<\/h4>\n

\u628a\u8fd9\u4e9b\u7cfb\u7edf\u7ec4\u4ef6\u653e\u56de\u5168\u5c40，DeepSeek-V3 \u7684\u5e76\u884c\u7b56\u7565\u662f\u4e00\u5f20\u7cbe\u5fc3\u914d\u6bd4\u7684\u62fc\u56fe：16 \u8def\u6d41\u6c34\u7ebf\u5e76\u884c\u300164 \u8def\u4e13\u5bb6\u5e76\u884c，\u8de8 8 \u4e2a\u7269\u7406\u8282\u70b9，\u6570\u636e\u5e76\u884c\u91c7\u7528 ZeRO-1，\u523b\u610f\u907f\u5f00\u4e86\u901a\u4fe1\u5f00\u9500\u8f83\u5927\u7684\u5f20\u91cf\u5e76\u884c\u3002\u8fd9\u4e2a\u7ec4\u5408\u4e0d\u662f\u201c\u80fd\u7528\u7684\u6700\u5927\u503c\u201d，\u800c\u662f\u5728\u7ed9\u5b9a\u7f51\u7edc\u62d3\u6251（\u8282\u70b9\u5185 NVLink\u3001\u8282\u70b9\u95f4 InfiniBand）\u548c\u6a21\u578b\u7ed3\u6784\u4e0b\u7684\u7ea6\u675f\u6700\u4f18\u89e3：\u4e13\u5bb6\u5e76\u884c\u5fc5\u987b\u8de8\u8282\u70b9，\u6240\u4ee5\u7528 DualPipe \u91cd\u53e0\u901a\u4fe1；\u5f20\u91cf\u5e76\u884c\u901a\u4fe1\u592a\u8d35，\u6240\u4ee5\u4e0d\u7528；\u6d41\u6c34\u7ebf\u5e76\u884c\u7528\u6765\u5207\u5c42，\u6570\u636e\u5e76\u884c\u7528\u6765\u6269\u5c55\u541e\u5410\u3002\u6bcf\u4e00\u4e2a\u9009\u62e9\u90fd\u5728\u56de\u7b54\u540c\u4e00\u4e2a\u95ee\u9898：\u5728\u8fd9\u4e2a\u786c\u4ef6\u8fb9\u754c\u4e0a，\u54ea\u79cd\u5e76\u884c\u7ef4\u5ea6\u7684\u8fb9\u9645\u6210\u672c\u6700\u4f4e？<\/p>\n

6. \u56de\u5230 1024：\u9ad8\u6548\u521b\u4f5c\u7684\u65b9\u6cd5\u8bba<\/h3>\n

\"\"<\/p>\n

DeepSeek-V3 \u7684\u6280\u672f\u62c6\u89e3\u5230\u8fd9\u91cc，\u53ef\u4ee5\u6536\u675f\u5230\u4e00\u6761\u5bf9\u5f00\u53d1\u8005\u771f\u6b63\u53ef\u8fc1\u79fb\u7684\u4e3b\u7ebf\u4e0a\u3002<\/p>\n

1024 \u7a0b\u5e8f\u5458\u8282\u7684\u6d3b\u52a8\u4e3b\u9898\u662f\u201c\u9ad8\u6548\u521b\u4f5c\u4e4b\u9053\u201d\u3002\u5728 AI \u5de5\u7a0b\u8bed\u5883\u4e0b，\u201c\u9ad8\u6548\u201d\u5f88\u5bb9\u6613\u88ab\u7b80\u5316\u4e3a\u201c\u7528\u66f4\u5c11\u7684\u5361\u201d\u6216\u201c\u82b1\u66f4\u5c11\u7684\u94b1\u201d\u3002\u4f46 DeepSeek-V3 \u7ed9\u51fa\u7684\u542f\u793a\u6070\u6070\u76f8\u53cd：\u9ad8\u6548\u4e0d\u662f\u67d0\u4e00\u4e2a\u5355\u70b9\u4f18\u5316\u7684\u7ed3\u679c，\u800c\u662f\u4e00\u7ec4\u8bbe\u8ba1\u51b3\u7b56\u5728\u591a\u4e2a\u5c42\u9762\u4e0a\u5f7c\u6b64\u5bf9\u9f50\u4e4b\u540e\u6d8c\u73b0\u51fa\u6765\u7684\u7cfb\u7edf\u5c5e\u6027\u3002<\/p>\n

MoE \u7684\u7a00\u758f\u6027\u8981\u6c42\u7cfb\u7edf\u5c42\u6709\u4e13\u5bb6\u5e76\u884c\u548c all-to-all \u4f18\u5316\u6765\u627f\u63a5；MLA \u7684\u4f4e\u79e9\u538b\u7f29\u8981\u6c42\u63a8\u7406\u5f15\u64ce\u6709\u5bf9\u5e94\u7684 kernel \u548c\u7f13\u5b58\u7ba1\u7406\u6765\u5151\u73b0；MTP \u7684\u4fe1\u53f7\u589e\u7a20\u8981\u6c42 FP8 \u548c DualPipe \u628a\u989d\u5916\u7684\u8ba1\u7b97\u4e0e\u901a\u4fe1\u6210\u672c\u5438\u6536\u6389\u3002\u5982\u679c\u53ea\u6284\u5176\u4e2d\u4e00\u4e2a\u6a21\u5757，\u800c\u4e0d\u540c\u65f6\u8c03\u6574\u5b83\u6240\u4f9d\u8d56\u7684\u7cfb\u7edf\u5c42，\u6536\u76ca\u4f1a\u88ab\u5176\u4ed6\u5c42\u9762\u7684\u74f6\u9888\u5403\u6389\u3002<\/p>\n

\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u201c\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1\u201d\u6bd4\u201c\u8ffd\u9010\u5355\u4e2a SOTA \u6a21\u5757\u201d\u66f4\u96be，\u4f46\u4e5f\u66f4\u6709\u957f\u671f\u4ef7\u503c\u3002\u771f\u6b63\u7684\u6280\u672f\u5224\u65ad\u529b，\u4e0d\u5728\u4e8e\u77e5\u9053\u54ea\u4e2a\u673a\u5236\u5728\u8bba\u6587\u91cc\u6700\u4eae\u773c，\u800c\u5728\u4e8e\u770b\u6e05\u4e00\u4e2a\u673a\u5236\u5728\u5b8c\u6574\u7684\u8ba1\u7b97-\u5185\u5b58-\u901a\u4fe1\u7ea6\u675f\u4e0b，\u5230\u5e95\u80fd\u91ca\u653e\u591a\u5c11\u3001\u53c8\u9700\u8981\u4ec0\u4e48\u6765\u627f\u63a5\u3002<\/p>\n

278.8 \u4e07 H800 \u5c0f\u65f6\u548c 558 \u4e07\u7f8e\u5143，\u662f\u8fd9\u79cd\u5224\u65ad\u529b\u7684\u4e00\u6b21\u96c6\u4e2d\u5151\u73b0\u3002\u5b83\u63d0\u9192\u6211\u4eec：\u5728\u7b97\u529b\u4ecd\u7136\u7a00\u7f3a\u7684\u5e74\u4ee3，\u6700\u6709\u6760\u6746\u7387\u7684\u521b\u4f5c\u4e0d\u662f\u5199\u66f4\u591a\u7684\u4ee3\u7801，\u800c\u662f\u5728\u7ea6\u675f\u7684\u8fb9\u754c\u4e0a，\u627e\u5230\u90a3\u4e2a\u8ba9\u591a\u4e2a\u7ef4\u5ea6\u540c\u65f6\u677e\u52a8\u7684\u5de5\u4f5c\u70b9\u3002\u8fd9\u5927\u6982\u4e5f\u662f 1024 \u8fd9\u4e2a\u7531 0 \u548c 1 \u6784\u6210\u7684\u8282\u65e5，\u5728 AI \u65f6\u4ee3\u4ecd\u7136\u503c\u5f97\u88ab\u53cd\u590d\u63d0\u8d77\u7684\u539f\u56e0\u3002<\/p>\n

\"\"<\/p>\n

\u53c2\u8003\u6587\u732e\u4e0e\u5ef6\u4f38\u9605\u8bfb<\/p>\n

  • \n

    DeepSeek-AI,\u00a0DeepSeek-V3 Technical Report, arXiv:2412.19437.<\/p>\n<\/li>\n

  • \n

    MLCommons,\u00a0DeepSeek-V3: A Large-Scale MoE Pretraining Benchmark for MLPerf Training v6.0, 2026.<\/p>\n<\/li>\n

  • \n

    Multi-Token Prediction：\u4ece\u8bad\u7ec3\u76ee\u6807\u5230\u63a8\u7406 drafter，llm-infra-atlas.<\/p>\n<\/li>\n

  • \n

    deepseek-v3-moe-mla：671B \u53c2\u6570\u53ea\u6fc0\u6d3b 37B，CSDN，2026.<\/p>\n<\/li>\n

  • \n

    vLLM vs SGLang vs DeepInfer：2026 \u5e74\u90e8\u7f72 DeepSeek \u63a8\u7406\u6846\u67b6\u600e\u4e48\u9009，2026.<\/p>\n<\/li>\n

  • \n

    DeepSeek | DeepSeek-V3 \u6b63\u5f0f\u53d1\u5e03<\/p>\n<\/li>\n","protected":false},"excerpt":{"rendered":"

    \u6458\u8981：\u4e00\u4e2a 671B \u53c2\u6570\u7684\u6a21\u578b，\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b 37B，\u8bad\u7ec3\u5168\u7a0b 278.8 \u4e07 H800 GPU \u5c0f\u65f6，\u6ca1\u6709\u51fa\u73b0\u4e00\u6b21\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0\u2014\u2014DeepSeek-V3 \u7528\u4e00\u7ec4\u8fd1\u4e4e\u201c\u53cd\u5e38\u8bc6\u201d\u7684\u6570\u5b57，\u628a\u5927\u6a21\u578b\u8bad\u7ec3\u4ece\u201c\u5806\u7b97\u529b\u201d\u7684\u53d9\u4e8b\u4e2d\u62fd\u4e86\u51fa\u6765\u3002\u672c\u6587\u4e0d\u6ee1\u8db3\u4e8e\u5217\u4e3e MLA\u3001MoE\u3001MTP \u8fd9\u4e9b\u540d\u8bcd，\u800c\u662f\u6cbf\u7740\u201c\u89e3\u8026\u201d\u8fd9\u6761\u4e3b\u7ebf，\u9010\u5c42\u62c6\u89e3\u6bcf\u4e00\u4e2a\u8bbe\u8ba1\u51b3\u7b56\u80cc\u540e\u7684\u7ea6\u675f\u3001\u53d6\u820d\u4e0e\u5de5\u7a0b\u5b9e\u73b0&<\/p>\n","protected":false},"author":2,"featured_media":120637,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[13758,68,50,207,86],"topic":[],"class_list":["post-120648","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-deepseek-v3","tag-deepseek","tag-50","tag-207","tag-86"],"yoast_head":"\n\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \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\/120648.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6458\u8981：\u4e00\u4e2a 671B \u53c2\u6570\u7684\u6a21\u578b，\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b 37B，\u8bad\u7ec3\u5168\u7a0b 278.8 \u4e07 H800 GPU \u5c0f\u65f6，\u6ca1\u6709\u51fa\u73b0\u4e00\u6b21\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0\u2014\u2014DeepSeek-V3 \u7528\u4e00\u7ec4\u8fd1\u4e4e\u201c\u53cd\u5e38\u8bc6\u201d\u7684\u6570\u5b57，\u628a\u5927\u6a21\u578b\u8bad\u7ec3\u4ece\u201c\u5806\u7b97\u529b\u201d\u7684\u53d9\u4e8b\u4e2d\u62fd\u4e86\u51fa\u6765\u3002\u672c\u6587\u4e0d\u6ee1\u8db3\u4e8e\u5217\u4e3e MLA\u3001MoE\u3001MTP \u8fd9\u4e9b\u540d\u8bcd，\u800c\u662f\u6cbf\u7740\u201c\u89e3\u8026\u201d\u8fd9\u6761\u4e3b\u7ebf，\u9010\u5c42\u62c6\u89e3\u6bcf\u4e00\u4e2a\u8bbe\u8ba1\u51b3\u7b56\u80cc\u540e\u7684\u7ea6\u675f\u3001\u53d6\u820d\u4e0e\u5de5\u7a0b\u5b9e\u73b0&\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/120648.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-31T15:48:24+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261031154806-6ae60db66596b.gif\" \/>\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=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/120648.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/120648.html\",\"name\":\"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-31T15:48:24+00:00\",\"dateModified\":\"2026-10-31T15:48:24+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/120648.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/120648.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/120648.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1\"}]},{\"@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":"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \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\/120648.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6458\u8981：\u4e00\u4e2a 671B \u53c2\u6570\u7684\u6a21\u578b，\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b 37B，\u8bad\u7ec3\u5168\u7a0b 278.8 \u4e07 H800 GPU \u5c0f\u65f6，\u6ca1\u6709\u51fa\u73b0\u4e00\u6b21\u4e0d\u53ef\u6062\u590d\u7684 loss \u5c16\u5cf0\u2014\u2014DeepSeek-V3 \u7528\u4e00\u7ec4\u8fd1\u4e4e\u201c\u53cd\u5e38\u8bc6\u201d\u7684\u6570\u5b57，\u628a\u5927\u6a21\u578b\u8bad\u7ec3\u4ece\u201c\u5806\u7b97\u529b\u201d\u7684\u53d9\u4e8b\u4e2d\u62fd\u4e86\u51fa\u6765\u3002\u672c\u6587\u4e0d\u6ee1\u8db3\u4e8e\u5217\u4e3e MLA\u3001MoE\u3001MTP \u8fd9\u4e9b\u540d\u8bcd，\u800c\u662f\u6cbf\u7740\u201c\u89e3\u8026\u201d\u8fd9\u6761\u4e3b\u7ebf，\u9010\u5c42\u62c6\u89e3\u6bcf\u4e00\u4e2a\u8bbe\u8ba1\u51b3\u7b56\u80cc\u540e\u7684\u7ea6\u675f\u3001\u53d6\u820d\u4e0e\u5de5\u7a0b\u5b9e\u73b0&","og_url":"https:\/\/www.wsisp.com\/helps\/120648.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-31T15:48:24+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261031154806-6ae60db66596b.gif"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/120648.html","url":"https:\/\/www.wsisp.com\/helps\/120648.html","name":"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-31T15:48:24+00:00","dateModified":"2026-10-31T15:48:24+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/120648.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/120648.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/120648.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4ece 1024 \u5757 GPU \u5230 558 \u4e07\u7f8e\u5143\uff1a\u62c6\u89e3 DeepSeek-V3 \u7684\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1"}]},{"@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\/120648","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=120648"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/120648\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/120637"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=120648"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=120648"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=120648"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=120648"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}