{"id":81444,"date":"2026-07-25T02:01:12","date_gmt":"2026-07-24T18:01:12","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/81444.html"},"modified":"2026-07-25T02:01:12","modified_gmt":"2026-07-24T18:01:12","slug":"%e9%b2%b2%e9%b9%8fcpu-8%e5%8d%a1910a-npu%e6%9c%8d%e5%8a%a1%e5%99%a8%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e9%83%a8%e7%bd%b2%e6%b5%8b%e8%af%95%e6%8a%a5%e5%91%8a","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/81444.html","title":{"rendered":"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a"},"content":{"rendered":"

\u00a0<\/p>\n\n\n\n\n\n\n\n
\n

\u62a5\u544a\u7f16\u53f7<\/p>\n<\/td>\n

\n

RPT-910A-20260531-002<\/p>\n<\/td>\n<\/tr>\n

\n

\u6d4b\u8bd5\u65e5\u671f<\/p>\n<\/td>\n

\n

2026\u5e745\u670830\u65e5-31\u65e5<\/p>\n<\/td>\n<\/tr>\n

\n

\u00a0<\/p>\n<\/td>\n

\n

\u00a0<\/p>\n<\/td>\n<\/tr>\n

\n

\u670d\u52a1\u5668<\/p>\n<\/td>\n

\n

bms-ytcs910a-260529 (\u9cb2\u9e4fCPU + 8\u5361910A NPU)<\/p>\n<\/td>\n<\/tr>\n

\n

\u62a5\u544a\u7248\u672c<\/p>\n<\/td>\n

\n

V2.0 (\u7efc\u5408\u7248)<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u00a0<\/p>\n

\u4e00\u3001\u6d4b\u8bd5\u6982\u8ff0<\/span><\/span><\/h2>\n

\u672c\u6b21\u6d4b\u8bd5\u5386\u65f6\u7ea612\u5c0f\u65f6，\u7cfb\u7edf\u6027\u5730\u9a8c\u8bc1\u4e86\u5728\u534e\u4e3a\u6607\u817e910A NPU\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u670d\u52a1\u7684\u53ef\u884c\u6027\u3002<\/span><\/p>\n

\u6d4b\u8bd5\u5148\u540e\u5c1d\u8bd5\u4e86\u4e09\u79cd\u4e3b\u6d41\u63a8\u7406\u6846\u67b6：<\/span><\/p>\n

    \n
  • vLLM + vLLM-Ascend：\u57fa\u4e8ePyTorch\u751f\u6001\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6，\u652f\u6301\u8fde\u7eed\u6279\u5904\u7406\u548cPagedAttention\u3002\u6700\u7ec8\u56e0910A\u786c\u4ef6\u7b97\u5b50\u5e93\u7f3a\u5931（SwiGlu\u7b49\u5173\u952e\u7b97\u5b50\u4e0d\u652f\u6301）\u800c\u5931\u8d25\u3002<\/li>\n
  • llama.cpp + CANN\u540e\u7aef：\u7eafC++\u5b9e\u73b0\u7684\u8f7b\u91cf\u7ea7\u63a8\u7406\u6846\u67b6\u3002\u56e0CANN 9.1.0-beta.1\u4e0ellama.cpp\u6700\u65b0\u4ee3\u7801\u5934\u6587\u4ef6\u4e0d\u517c\u5bb9（aclnn_fused_infer_attention_score_v2.h\u7f3a\u5931），\u4e14GCC 7.3.0\u4e0d\u652f\u6301ARM NEON x2\/x4 intrinsics\u800c\u7f16\u8bd1\u5931\u8d25\u3002<\/li>\n
  • PyTorch Native + torch_npu + transformers：\u6700\u7ec8\u6210\u529f\u65b9\u6848\u3002\u57fa\u4e8e\u534e\u4e3a\u539f\u751fPyTorch NPU\u540e\u7aef，\u901a\u8fc7transformers\u7684device_map="auto"\u5b9e\u73b0\u591a\u5361\u6a21\u578b\u5e76\u884c，\u6210\u529f\u90e8\u7f72Qwen3-32B Dense\u6a21\u578b\u3002<\/li>\n<\/ul>\n

    \u00a0<\/p>\n

    \u4e8c\u3001\u6d4b\u8bd5\u73af\u5883<\/span><\/span><\/h2>\n

    2.1 \u786c\u4ef6\u73af\u5883<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n\n
    \n

    \u9879\u76ee<\/p>\n<\/td>\n

    		\n

    \u914d\u7f6e<\/p>\n<\/td>\n<\/tr>\n

    \n

    \u670d\u52a1\u5668\u578b\u53f7<\/p>\n<\/td>\n

    		\n

    \u9cb2\u9e4fCPU + 8\u5361910A NPU<\/p>\n<\/td>\n<\/tr>\n

    \n

    CPU\u67b6\u6784<\/p>\n<\/td>\n

    		\n

    ARM aarch64<\/p>\n<\/td>\n<\/tr>\n

    \n

    NPU\u578b\u53f7<\/p>\n<\/td>\n

    		\n

    \u6607\u817e910PremiumA \u00d7 8\u5361<\/p>\n<\/td>\n<\/tr>\n

    \n

    NPU\u663e\u5b58<\/p>\n<\/td>\n

    		\n

    \u5355\u536132GB HBM2（\u603b\u91cf256GB）<\/p>\n<\/td>\n<\/tr>\n

    \n

    NPU\u9a71\u52a8<\/p>\n<\/td>\n

    		\n

    25.5.0（\u5347\u7ea7\u540e）<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u00a0<\/p>\n

    2.2 \u8f6f\u4ef6\u73af\u5883（\u6700\u7ec8\u6210\u529f\u65b9\u6848）<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n\n\n\n\n
    \n

    \u7ec4\u4ef6<\/p>\n<\/td>\n

    		\n

    \u7248\u672c<\/p>\n<\/td>\n

    		\n

    \u72b6\u6001<\/p>\n<\/td>\n<\/tr>\n

    \n

    \u64cd\u4f5c\u7cfb\u7edf<\/p>\n<\/td>\n

    		\n

    EulerOS 2.0 (SP8) \/ Ubuntu 22.04<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    CANN<\/p>\n<\/td>\n

    		\n

    8.5.0（\u5bb9\u5668\u5185）<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    Python<\/p>\n<\/td>\n

    		\n

    3.11.6（\u5bb9\u5668\u5185）<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    PyTorch<\/p>\n<\/td>\n

    		\n

    2.1.0+cpu<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    torch_npu<\/p>\n<\/td>\n

    		\n

    2.1.0.post18.dev20251112<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    transformers<\/p>\n<\/td>\n

    		\n

    4.51.0（\u6700\u7ec8\u7a33\u5b9a\u7248）<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    accelerate<\/p>\n<\/td>\n

    		\n

    0.30.0<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n

    \n

    Flask<\/p>\n<\/td>\n

    		\n

    \u6700\u65b0\u7248<\/p>\n<\/td>\n

    		\n

    \u2705<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u00a0<\/p>\n

    \u4e09\u3001\u6d4b\u8bd5\u76ee\u6807\u4e0e\u6a21\u578b\u77e9\u9635<\/span><\/span><\/h2>\n

    \u6d4b\u8bd5\u76ee\u6807：\u5728910A\u670d\u52a1\u5668\u4e0a\u627e\u5230\u80fd\u591f\u7a33\u5b9a\u8fd0\u884c\u3001\u4e14\u53c2\u6570\u89c4\u6a21\u5c3d\u53ef\u80fd\u5927\u7684\u5f00\u6e90\u5927\u8bed\u8a00\u6a21\u578b，\u5e76\u9a8c\u8bc1\u5176\u591a\u5361\u5e76\u884c\u63a8\u7406\u80fd\u529b\u3002<\/span><\/p>\n

    3.1 \u6d4b\u8bd5\u6a21\u578b\u77e9\u9635<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n\n\n\n
    \n

    \u6a21\u578b<\/p>\n<\/td>\n

    		\n

    \u67b6\u6784<\/p>\n<\/td>\n

    		\n

    \u53c2\u6570\u91cf<\/p>\n<\/td>\n

    		\n

    \u5c42\u6570<\/p>\n<\/td>\n

    		\n

    \u6a21\u6001<\/p>\n<\/td>\n

    		\n

    \u6d4b\u8bd5\u7ed3\u679c<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3.6-27B<\/p>\n<\/td>\n

    		\n

    Dense<\/p>\n<\/td>\n

    		\n

    27B<\/p>\n<\/td>\n

    		\n

    64\u5c42<\/p>\n<\/td>\n

    		\n

    \u591a\u6a21\u6001(\u6587\u672c+\u56fe\u50cf+\u89c6\u9891)<\/p>\n<\/td>\n

    		\n

    \u274c transformers\u521d\u59cb\u5316\u5361\u4f4f（vision\u6a21\u5757）<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3-30B-A3B<\/p>\n<\/td>\n

    		\n

    MoE<\/p>\n<\/td>\n

    		\n

    30B\u603b\/3B\u6fc0\u6d3b<\/p>\n<\/td>\n

    		\n

    48\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u274c \u52a0\u8f7d\u6210\u529f，\u63a8\u7406\u5361\u4f4f（MoE\u8def\u7531\u7b97\u5b50\u4e0d\u652f\u6301）<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3.5-35B-A3B<\/p>\n<\/td>\n

    		\n

    MoE<\/p>\n<\/td>\n

    		\n

    35B\u603b\/3B\u6fc0\u6d3b<\/p>\n<\/td>\n

    		\n

    48\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u274c transformers 4.51.0\u4e0d\u652f\u6301qwen3_5_moe\u67b6\u6784<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3.5-27B<\/p>\n<\/td>\n

    		\n

    Dense<\/p>\n<\/td>\n

    		\n

    27B<\/p>\n<\/td>\n

    		\n

    64\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u274c transformers 4.51.0\u4e0d\u652f\u6301qwen3_5\u67b6\u6784<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen2.5-14B-Instruct<\/p>\n<\/td>\n

    		\n

    Dense<\/p>\n<\/td>\n

    		\n

    14B<\/p>\n<\/td>\n

    		\n

    48\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u2705 2\u5361\/4\u5361\/8\u5361\u5747\u9a8c\u8bc1\u6210\u529f<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3-32B<\/p>\n<\/td>\n

    		\n

    Dense<\/p>\n<\/td>\n

    		\n

    32B<\/p>\n<\/td>\n

    		\n

    64\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u2705 8\u5361\u9a8c\u8bc1\u6210\u529f（\u6700\u7ec8\u751f\u4ea7\u65b9\u6848）<\/p>\n<\/td>\n<\/tr>\n

    \n

    Qwen3-32B<\/p>\n<\/td>\n

    		\n

    Dense<\/p>\n<\/td>\n

    		\n

    32B<\/p>\n<\/td>\n

    		\n

    64\u5c42<\/p>\n<\/td>\n

    		\n

    \u7eaf\u6587\u672c<\/p>\n<\/td>\n

    		\n

    \u274c 4\u5361OOM（910A\u8fde\u7eed\u5185\u5b58\u9650\u5236）<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u56db\u3001\u5404\u65b9\u6848\u8be6\u7ec6\u6d4b\u8bd5\u8fc7\u7a0b<\/span><\/span><\/h2>\n

    4.1 vLLM-Ascend\u65b9\u6848（\u7ea610\u5c0f\u65f6，\u6700\u7ec8\u5931\u8d25）<\/span><\/span><\/h3>\n

    vLLM-Ascend\u662f\u534e\u4e3a\u4e0e\u793e\u533a\u5408\u4f5c\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6，\u5b98\u65b9\u660e\u786e\u6807\u6ce8910A"\u6682\u65e0\u8ba1\u5212"\u652f\u6301\u3002\u4f46\u57fa\u4e8e\u6280\u672f\u63a2\u7d22\u76ee\u7684，\u4ecd\u8fdb\u884c\u4e86\u5b8c\u6574\u5c1d\u8bd5\u3002<\/span><\/p>\n

      \n
    • CANN\u5b89\u88c5：CANN 9.0.0\u5b89\u88c5\u56e0\u9a71\u52a8\u517c\u5bb9\u6027\u68c0\u67e5\u5931\u8d25；CANN 9.1.0-beta.1\u901a\u8fc7RPM\u5305\u5b89\u88c5\u6210\u529f，\u4f46\u7f3a\u5c11ATB\/NNAL\u52a0\u901f\u5e93\u3002<\/li>\n
    • Python\u73af\u5883：\u5bbf\u4e3b\u673aPython 3.7.0\u592a\u8001，\u7f16\u8bd1\u5b89\u88c5Python 3.10.13，\u521b\u5efa\u865a\u62df\u73af\u5883vllm-910a-env\u3002<\/li>\n
    • torch_npu\u9002\u914d：\u7ecf\u53862.1.0\u21922.4.0\u21922.5.1\u4e09\u4e2a\u7248\u672c\u8fed\u4ee3，\u6700\u7ec82.5.1\u914d\u5408\u5347\u7ea7\u540e\u7684\u9a71\u52a8\u624d\u80fd\u6b63\u5e38\u5de5\u4f5c\u3002<\/li>\n
    • vLLM\u7f16\u8bd1：vLLM 0.7.3\u6e90\u7801\u5b89\u88c5，\u4fee\u6539setup.py\u786c\u7f16\u7801\u7248\u672c\u53f7\u7ed5\u8fc7Git\u68c0\u6d4b；\u8bbe\u7f6eVLLM_TARGET_DEVICE=empty\u3002<\/li>\n
    • vLLM-Ascend\u5b89\u88c5：pip install vllm-ascend==0.7.3，\u81ea\u52a8\u5b89\u88c5torch 2.5.1\u3001torch-npu 2.5.1\u7b49\u4f9d\u8d56\u3002<\/li>\n
    • \u6a21\u578b\u914d\u7f6e\u95ee\u9898：Qwen3.6-27B\u7684text_config\u4e3adict\u7c7b\u578b，vLLM\u7684get_hf_text_config\u65ad\u8a00\u5931\u8d25；\u6253\u8865\u4e01\u517c\u5bb9\u3002<\/li>\n
    • tokenizer\u95ee\u9898：transformers 5.9.0\u79fb\u9664\u4e86all_special_tokens_extended，vLLM 0.7.3\u8fd8\u5728\u4f7f\u7528；\u6253\u8865\u4e01\u517c\u5bb9\u3002<\/li>\n
    • ATB\/NNAL\u7f3a\u5931：libatb.so\u7f3a\u5931，\u5b89\u88c5NNAL 9.1.0-beta.1\u5931\u8d25；\u521b\u5efa\u8f6f\u94fe\u63a5\u3001\u8bbe\u7f6eLD_PRELOAD\u7b49\u5747\u672a\u89e3\u51b3\u3002<\/li>\n
    • \u81f4\u547d\u9519\u8bef：SwiGlu\u7b97\u5b50：aclnnSwiGlu failed: socVersion [ascend910] does not support opType [SwiGlu]\u3002910A\u786c\u4ef6CANN 9.1.0-beta.1\u7b97\u5b50\u5e93\u4e0d\u652f\u6301SwiGlu，\u8fd9\u662fvLLM-Ascend\u4f9d\u8d56ATB\u7b97\u5b50\u52a0\u901f\u5e93\u7684\u56fa\u6709\u7f3a\u9677，\u65e0\u6cd5\u7ed5\u8fc7\u3002<\/li>\n<\/ul>\n

      \u7ed3\u8bba：vLLM-Ascend 0.7.3\u5b98\u65b9\u8981\u6c42CANN 8.1.RC1，\u4e0e910A\u7684CANN 9.1.0-beta.1\u4e0d\u5339\u914d，\u4e14910A\u7b97\u5b50\u5e93\u7f3a\u5c11SwiGlu\u7b49\u5173\u952e\u7b97\u5b50\u3002<\/span><\/p>\n

      4.2 llama.cpp + CANN\u540e\u7aef\u65b9\u6848（\u7f16\u8bd1\u5931\u8d25）<\/span><\/span><\/h3>\n
        \n
      • CMake\u914d\u7f6e：CANN 9.1.0-beta.1\u8bc6\u522b\u6b63\u5e38，SOC_VERSION=Ascend910PremiumA，CANN\u540e\u7aef\u5df2\u5305\u542b\u3002<\/li>\n
      • \u7f16\u8bd1\u9519\u8bef1：aclnn_fused_infer_attention_score_v2.h\u5934\u6587\u4ef6\u7f3a\u5931\u3002llama.cpp\u6700\u65b0\u4ee3\u7801(b9260)\u9488\u5bf9CANN 8.5.T63\u5f00\u53d1，CANN 9.1.0 beta\u7684ACLNN API\u6709\u53d8\u5316\u3002<\/li>\n
      • \u7f16\u8bd1\u9519\u8bef2：GCC 7.3.0\u4e0d\u652f\u6301ARM NEON x2\/x4 intrinsics（vld1q_s16_x2\u7b49），\u9700\u8981GCC 9+\u3002<\/li>\n
      • \u964d\u7ea7\u5c1d\u8bd5：\u5c1d\u8bd5\u56de\u9000\u5230b8547\u7248\u672c，\u4f46910A\u517c\u5bb9\u6027\u4ecd\u4e0d\u660e\u786e\u3002<\/li>\n<\/ul>\n

        \u7ed3\u8bba：llama.cpp CANN\u540e\u7aef\u5728910A + CANN 9.1.0-beta.1 + GCC 7.3.0\u73af\u5883\u4e0b\u7f16\u8bd1\u5931\u8d25，\u9700\u8981\u5347\u7ea7GCC\u6216\u7b49\u5f85\u5b98\u65b9\u9002\u914d\u3002<\/span><\/p>\n

        4.3 PyTorch Native + torch_npu + transformers\u65b9\u6848（\u6210\u529f）<\/span><\/span><\/h3>\n

        \u57fa\u4e8e\u6210\u90fd\u667a\u7b97\u4e2d\u5fc3\u63d0\u4f9b\u7684mindie_910a:2.3.0 Docker\u955c\u50cf，\u5185\u7f6eCANN 8.5.0 + PyTorch 2.1.0 + torch_npu 2.1.0，\u4e0e\u9a71\u52a825.5.0\u5b8c\u7f8e\u517c\u5bb9\u3002<\/span><\/p>\n

          \n
        • \u955c\u50cf\u52a0\u8f7d：docker load -i mindie_230_910a.tar \u2192 mindie_910a:2.3.0<\/li>\n
        • \u5bb9\u5668\u542f\u52a8：\u6302\u8f7dNPU\u8bbe\u5907\u3001\u9a71\u52a8\u3001\u65e5\u5fd7\u76ee\u5f55，-v \/data:\/home\u6620\u5c04\u6a21\u578b\u76ee\u5f55<\/li>\n
        • PyTorch\u5347\u7ea7：\u5bb9\u5668\u5185PyTorch 2.1.0\u4e0etransformers 5.2.0\u4e0d\u517c\u5bb9，\u964d\u7ea7transformers\u52304.41.2\/4.51.0<\/li>\n
        • accelerate\u95ee\u9898：transformers 5.2.0\u7684accelerate\u68c0\u6d4b\u903b\u8f91\u6709bug，\u624b\u52a8\u521b\u5efaaccelerate.py stub\u6587\u4ef6\u7ed5\u8fc7<\/li>\n
        • device_map\u5206\u914d：\u624b\u52a8\u6784\u5efadevice_map\u5b57\u5178，\u5c0664\u5c42\u5747\u5300\u5206\u914d\u5230\u5404NPU，embed_tokens\u548clm_head\u5fc5\u987b\u5728\u540c\u4e00\u8bbe\u5907（\u5171\u4eab\u6743\u91cd）<\/li>\n
        • 910A\u5173\u952e\u9650\u5236：do_sample=False（greedy decoding）：910A\u7684MultinomialWithReplacement AICPU\u7b97\u5b50\u6709bug，\u5fc5\u987b\u7981\u7528temperature\/top_p\/top_k\u91c7\u6837<\/li>\n
        • attention\u5b9e\u73b0：attn_implementation="eager"：\u907f\u514dsliding window attention\u5728sdpa\u4e0b\u7684\u517c\u5bb9\u6027\u95ee\u9898<\/li>\n<\/ul>\n

          \u4e94\u3001\u5361\u6570\u914d\u7f6e\u5c1d\u8bd5\u8bb0\u5f55<\/span><\/span><\/h2>\n

          \u9488\u5bf9Qwen3-32B Dense\u6a21\u578b（64\u5c42，5120 hidden_size，FP16\u7ea664GB\u6743\u91cd），\u7cfb\u7edf\u6027\u5730\u6d4b\u8bd5\u4e861\u5361\/2\u5361\/4\u5361\/8\u5361\u914d\u7f6e：<\/span><\/p>\n\n\n\n\n\n\n\n\n
          \n

          \u5361\u6570<\/p>\n<\/td>\n

          		\n

          \u6bcf\u5361\u5c42\u6570<\/p>\n<\/td>\n

          		\n

          \u6bcf\u5361\u6743\u91cd<\/p>\n<\/td>\n

          		\n

          \u603bHBM<\/p>\n<\/td>\n

          		\n

          \u7ed3\u679c<\/p>\n<\/td>\n

          		\n

          \u539f\u56e0<\/p>\n<\/td>\n<\/tr>\n

          \n

          1\u5361<\/p>\n<\/td>\n

          		\n

          64\u5c42<\/p>\n<\/td>\n

          		\n

          ~64GB<\/p>\n<\/td>\n

          		\n

          32GB<\/p>\n<\/td>\n

          		\n

          \u274c OOM<\/p>\n<\/td>\n

          		\n

          \u6743\u91cd\u8fdc\u8d85\u5355\u5361\u5bb9\u91cf<\/p>\n<\/td>\n<\/tr>\n

          \n

          2\u5361<\/p>\n<\/td>\n

          		\n

          32\u5c42<\/p>\n<\/td>\n

          		\n

          ~32GB<\/p>\n<\/td>\n

          		\n

          64GB<\/p>\n<\/td>\n

          		\n

          \u274c OOM<\/p>\n<\/td>\n

          		\n

          \u6bcf\u536132GB > \u5b9e\u9645\u53ef\u7528\u8fde\u7eed\u5185\u5b58<\/p>\n<\/td>\n<\/tr>\n

          \n

          4\u5361<\/p>\n<\/td>\n

          		\n

          16\u5c42<\/p>\n<\/td>\n

          		\n

          ~16GB<\/p>\n<\/td>\n

          		\n

          128GB<\/p>\n<\/td>\n

          		\n

          \u274c OOM\/\u788e\u7247<\/p>\n<\/td>\n

          		\n

          910A PyTorch NPU\u5206\u914d\u5668\u65e0\u6cd5\u5206\u914d16GB\u8fde\u7eed\u5757，max_split_size_mb\u53c2\u6570\u65e0\u6cd5\u89e3\u51b3<\/p>\n<\/td>\n<\/tr>\n

          \n

          6\u5361<\/p>\n<\/td>\n

          		\n

          10-11\u5c42<\/p>\n<\/td>\n

          		\n

          ~10-11GB<\/p>\n<\/td>\n

          		\n

          192GB<\/p>\n<\/td>\n

          		\n

          \u23f8 \u672a\u6d4b\u8bd5<\/p>\n<\/td>\n

          		\n

          \u7406\u8bba\u53ef\u884c，\u75592\u5361\u7ed9\u5176\u4ed6\u4efb\u52a1<\/p>\n<\/td>\n<\/tr>\n

          \n

          8\u5361<\/p>\n<\/td>\n

          		\n

          8\u5c42<\/p>\n<\/td>\n

          		\n

          ~8GB<\/p>\n<\/td>\n

          		\n

          256GB<\/p>\n<\/td>\n

          		\n

          \u2705 \u6210\u529f<\/p>\n<\/td>\n

          		\n

          \u6bcf\u53618GB\u6743\u91cd，\u4f59\u91cf24GB\u7528\u4e8eKV Cache\u548c\u5e76\u53d1<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \u00a0<\/p>\n

          \u5173\u952e\u53d1\u73b0：910A\u5355\u5361\u867d\u7136\u6807\u79f032GB HBM，\u4f46PyTorch NPU\u540e\u7aef\u5b9e\u9645\u53ef\u7528\u7684\u8fde\u7eed\u5185\u5b58\u5757\u7ea6\u4e3a15-16GB（\u53d7\u5206\u914d\u5668\u9650\u5236）\u3002\u56e0\u6b644\u5361\u6bcf\u536116\u5c42（\u7ea616GB）\u7684\u914d\u7f6e\u4f1a\u89e6\u53d1OOM，\u800c8\u5361\u6bcf\u53618\u5c42（\u7ea68GB）\u5219\u5b8c\u5168\u5728\u5b89\u5168\u8303\u56f4\u5185\u3002<\/span><\/p>\n

          \u00a0<\/p>\n

          \u516d\u3001\u5173\u952e\u95ee\u9898\u8e29\u5751\u8bb0\u5f55<\/span><\/span><\/h2>\n

          \u95ee\u98981：transformers\u7248\u672c\u517c\u5bb9\u6027
          \ntransformers 5.2.0\u8981\u6c42PyTorch \u22652.4，\u4f46\u5bb9\u5668\u5185PyTorch 2.1.0\u4e0eCANN 8.5.0\u7ed1\u5b9a，\u5347\u7ea7PyTorch\u4f1a\u7834\u574fNPU\u517c\u5bb9\u6027\u3002\u89e3\u51b3\u65b9\u6848：\u964d\u7ea7transformers\u52304.51.0（\u652f\u6301PyTorch 2.1.0\u4e14\u652f\u6301Qwen3\u67b6\u6784）\u3002<\/span><\/p>\n

          \u95ee\u98982：accelerate\u6a21\u5757\u7f3a\u5931
          \ntransformers 5.2.0\u7684device_map\u68c0\u6d4b\u903b\u8f91\u6709bug，\u5373\u4f7f\u5b89\u88c5\u4e86accelerate 0.34.2\u4e5f\u62a5\u9519"requires accelerate"\u3002\u89e3\u51b3\u65b9\u6848：\u624b\u52a8\u521b\u5efaaccelerate.py stub\u6587\u4ef6，\u6ce8\u5165is_accelerate_available=True\u548ccheck_and_set_device_map\u51fd\u6570\u3002<\/span><\/p>\n

          \u95ee\u98983：tied parameters\u8bbe\u5907\u5206\u914d
          \nlm_head.weight\u548cmodel.embed_tokens.weight\u5171\u4eab\u6743\u91cd，device_map\u4e2d\u5fc5\u987b\u5206\u914d\u5230\u540c\u4e00NPU\u8bbe\u5907，\u5426\u5219\u4f1a\u62a5"Tied parameters are on different devices"\u9519\u8bef\u3002<\/span><\/p>\n

          \u95ee\u98984：bfloat16\u4e0d\u652f\u6301
          \nQwen3\u7cfb\u5217\u6a21\u578bconfig\u9ed8\u8ba4torch_dtype=bfloat16，910A\u786c\u4ef6\u4e0d\u652f\u6301bfloat16\u3002\u89e3\u51b3\u65b9\u6848：\u5f3a\u5236config.torch_dtype=torch.float16，\u5e76\u5728\u52a0\u8f7d\u540e\u904d\u5386\u53c2\u6570\u8f6c\u6362\u6b8b\u7559bfloat16\u3002<\/span><\/p>\n

          \u95ee\u98985：MultinomialWithReplacement AICPU\u7b97\u5b50bug
          \n910A\u7684CANN 8.5.0\u4e2dMultinomialWithReplacement\u7b97\u5b50\u5b9e\u73b0\u6709\u7f3a\u9677，do_sample=True\u65f6\u89e6\u53d1ERR99999 UNKNOWN application exception\u3002\u89e3\u51b3\u65b9\u6848：\u5f3a\u5236do_sample=False（greedy decoding），\u7981\u7528temperature\/top_p\/top_k\u3002<\/span><\/p>\n

          \u95ee\u98986：Sliding Window Attention
          \nQwen3\u4f7f\u7528sliding window attention，transformers\u7684sdpa\u5b9e\u73b0\u4e0d\u652f\u6301\u3002\u89e3\u51b3\u65b9\u6848：attn_implementation="eager"，\u4f7f\u7528\u7eafPyTorch\u5b9e\u73b0\u3002<\/span><\/p>\n

          \u95ee\u98987：MoE\u6a21\u578b\u8def\u7531\u7b97\u5b50\u4e0d\u652f\u6301
          \nQwen3-30B-A3B（MoE）\u52a0\u8f7d\u6210\u529f\u4f46\u63a8\u7406\u5361\u4f4f，910A\u4e0d\u652f\u6301MoE\u7684topk\/gate\u8def\u7531\u7b97\u5b50\u3002\u89e3\u51b3\u65b9\u6848：\u653e\u5f03MoE\u6a21\u578b，\u6539\u7528Dense\u6a21\u578b\u3002<\/span><\/p>\n

          \u95ee\u98988：\u591a\u6a21\u6001\u6a21\u578b\u521d\u59cb\u5316\u5361\u4f4f
          \nQwen3.6-27B\u542bvision encoder，transformers\u521d\u59cb\u5316\u65f6\u6784\u5efa\u56fe\u50cf\u5904\u7406\u6a21\u5757\u8017\u65f6\u6781\u957f（30\u5206\u949f+\u65e0\u8f93\u51fa）\u3002\u89e3\u51b3\u65b9\u6848：\u653e\u5f03\u591a\u6a21\u6001\u6a21\u578b，\u6539\u7528\u7eaf\u6587\u672c\u6a21\u578b\u3002<\/span><\/p>\n

          \u95ee\u98989：Docker\u955c\u50cf\u6587\u4ef6\u635f\u574f
          \nmindie_230_910a.tar\u4f20\u8f93\u8fc7\u7a0b\u4e2d\u635f\u574f（unexpected EOF），\u91cd\u65b0\u4e0b\u8f7d\u5b8c\u6574\u955c\u50cf\u540e\u89e3\u51b3\u3002<\/span><\/p>\n

          \u95ee\u989810：\u9a71\u52a8\u7248\u672c\u4e0d\u5339\u914d
          \n\u521d\u59cb\u9a71\u52a823.0.0\u4e0eCANN 8.5.0\u4e0d\u517c\u5bb9，\u5bfc\u81f4AICPU ops\u52a0\u8f7d\u5931\u8d25（507033\/E39006）\u3002\u89e3\u51b3\u65b9\u6848：\u5347\u7ea7\u9a71\u52a8\u523025.5.0\u3002<\/span><\/p>\n

          \u00a0<\/p>\n

          \u4e03\u3001\u6700\u7ec8\u6210\u529f\u65b9\u6848：Qwen3-32B 8\u5361\u90e8\u7f72<\/span><\/span><\/h2>\n

          7.1 \u90e8\u7f72\u67b6\u6784<\/span><\/span><\/h3>\n

          \u91c7\u7528Flask + PyTorch Native + torch_npu + transformers\u67b6\u6784，8\u5361\u6a21\u578b\u5e76\u884c（Model Parallelism），\u6bcf\u5361\u627f\u8f7d8\u5c42Transformer\u3002<\/span><\/p>\n\n\n\n\n\n\n\n
          \n

          \u63a8\u7406\u6846\u67b6<\/p>\n<\/td>\n

          		\n

          PyTorch Native + torch_npu<\/p>\n<\/td>\n<\/tr>\n

          \n

          API\u5c42<\/p>\n<\/td>\n

          		\n

          Flask (threaded=True)<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u5e76\u884c\u7b56\u7565<\/p>\n<\/td>\n

          		\n

          \u6a21\u578b\u5e76\u884c（Model Parallelism）<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u5361\u6570<\/p>\n<\/td>\n

          		\n

          8\u5361（NPU 0-7）<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u6bcf\u5361\u8d1f\u8f7d<\/p>\n<\/td>\n

          		\n

          8\u5c42 + \u90e8\u5206\u5171\u4eab\u6743\u91cd<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          7.2 \u5173\u952e\u914d\u7f6e\u53c2\u6570<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n\n\n
          \n

          \u6a21\u578b\u8def\u5f84<\/p>\n<\/td>\n

          		\n

          \/home\/models\/Qwen3-32B<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u6570\u636e\u7c7b\u578b<\/p>\n<\/td>\n

          		\n

          torch.float16（910A\u4e0d\u652f\u6301bfloat16）<\/p>\n<\/td>\n<\/tr>\n

          \n

          device_map<\/p>\n<\/td>\n

          		\n

          \u624b\u52a8\u5206\u914d，8\u5361\u54048\u5c42<\/p>\n<\/td>\n<\/tr>\n

          \n

          attention\u5b9e\u73b0<\/p>\n<\/td>\n

          		\n

          eager（\u907f\u514dsliding window\u4e0d\u517c\u5bb9）<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u91c7\u6837\u7b56\u7565<\/p>\n<\/td>\n

          		\n

          do_sample=False（greedy decoding）<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u6700\u5927\u4e0a\u4e0b\u6587<\/p>\n<\/td>\n

          		\n

          512-4096 tokens（\u53d7KV Cache\u663e\u5b58\u9650\u5236）<\/p>\n<\/td>\n<\/tr>\n

          \n

          API\u7aef\u53e3<\/p>\n<\/td>\n

          		\n

          8000<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          7.3 \u670d\u52a1\u542f\u52a8\u547d\u4ee4<\/span><\/span><\/h3>\n

          export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
          \ncd \/home && python qwen_server_32b_8card.py<\/span><\/p>\n

          7.4 API\u6d4b\u8bd5\u793a\u4f8b<\/span><\/span><\/h3>\n

          curl -H "Content-Type: application\/json" -d '{
          \n\u00a0\u00a0"model": "Qwen3-32B",
          \n\u00a0\u00a0"messages": [
          \n\u00a0\u00a0\u00a0\u00a0{"role": "system", "content": "\u4f60\u662f\u4e00\u4e2a\u6709\u5e2e\u52a9\u7684\u52a9\u624b"},
          \n\u00a0\u00a0\u00a0\u00a0{"role": "user", "content": "\u8bf7\u7528\u4e00\u53e5\u8bdd\u4ecb\u7ecd\u81ea\u5df1"}
          \n\u00a0\u00a0],
          \n\u00a0\u00a0"max_tokens": 128
          \n}' http:\/\/localhost:8000\/v1\/chat\/completions<\/span><\/p>\n

          \u516b\u3001\u7ed3\u8bba\u4e0e\u5efa\u8bae<\/span><\/span><\/h2>\n

          8.1 \u6d4b\u8bd5\u7ed3\u8bba<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n\n\n
          \n

          vLLM-Ascend\u65b9\u6848<\/p>\n<\/td>\n

          		\n

          \u274c \u4e0d\u53ef\u884c<\/p>\n<\/td>\n

          		\n

          910A\u7b97\u5b50\u5e93\u7f3a\u5c11SwiGlu\u7b49\u5173\u952e\u7b97\u5b50，vLLM-Ascend\u5b98\u65b9\u660e\u786e\u4e0d\u652f\u6301910A<\/p>\n<\/td>\n<\/tr>\n

          \n

          llama.cpp\u65b9\u6848<\/p>\n<\/td>\n

          		\n

          \u274c \u4e0d\u53ef\u884c<\/p>\n<\/td>\n

          		\n

          CANN 9.1.0-beta.1\u4e0ellama.cpp\u5934\u6587\u4ef6\u4e0d\u517c\u5bb9，GCC 7.3.0\u4e0d\u652f\u6301NEON x2\/x4<\/p>\n<\/td>\n<\/tr>\n

          \n

          PyTorch Native\u65b9\u6848<\/p>\n<\/td>\n

          		\n

          \u2705 \u53ef\u884c<\/p>\n<\/td>\n

          		\n

          \u57fa\u4e8etransformers + torch_npu，8\u5361\u6a21\u578b\u5e76\u884c，\u6210\u529f\u90e8\u7f72Qwen3-32B<\/p>\n<\/td>\n<\/tr>\n

          \n

          MoE\u6a21\u578b<\/p>\n<\/td>\n

          		\n

          \u274c \u4e0d\u652f\u6301<\/p>\n<\/td>\n

          		\n

          910A\u4e0d\u652f\u6301MoE\u8def\u7531\u7b97\u5b50（topk\/gate），Qwen3-30B-A3B\/Qwen3.5-35B-A3B\u5747\u5931\u8d25<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u591a\u6a21\u6001\u6a21\u578b<\/p>\n<\/td>\n

          		\n

          \u274c \u4e0d\u652f\u6301<\/p>\n<\/td>\n

          		\n

          Qwen3.6-27B\u542bvision encoder，transformers\u521d\u59cb\u5316\u5361\u4f4f<\/p>\n<\/td>\n<\/tr>\n

          \n

          Dense\u6a21\u578b<\/p>\n<\/td>\n

          		\n

          \u2705 \u652f\u6301<\/p>\n<\/td>\n

          		\n

          Qwen2.5-14B\u548cQwen3-32B Dense\u67b6\u6784\u9a8c\u8bc1\u6210\u529f<\/p>\n<\/td>\n<\/tr>\n

          \n

          \u00a0<\/p>\n<\/td>\n

          		\n

          \u00a0<\/p>\n<\/td>\n

          		\n

          \u00a0<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          8.2 \u5173\u952e\u9650\u5236<\/span><\/span><\/h3>\n
            \n
          • \u91c7\u6837\u9650\u5236：\u5fc5\u987b\u7981\u7528do_sample（temperature\/top_p\/top_k\u65e0\u6548），\u53ea\u80fd\u4f7f\u7528greedy decoding<\/li>\n
          • \u7cbe\u5ea6\u9650\u5236：\u4ec5\u652f\u6301float16，\u4e0d\u652f\u6301bfloat16\/FP8\/INT8\u91cf\u5316<\/li>\n
          • \u5e76\u53d1\u9650\u5236：Flask threaded\u6a21\u5f0f，\u771f\u5b9e\u5e76\u53d1\u80fd\u529b\u7ea61-2 req\/s（NPU\u7b97\u529b\u74f6\u9888）<\/li>\n
          • \u4e0a\u4e0b\u6587\u9650\u5236：max_tokens\u5efa\u8bae\u22644096，\u66f4\u5927\u4e0a\u4e0b\u6587\u53ef\u80fd\u89e6\u53d1OOM<\/li>\n
          • \u6a21\u578b\u9650\u5236：\u4ec5\u652f\u6301\u7eaf\u6587\u672cDense\u6a21\u578b，\u4e0d\u652f\u6301MoE\/\u591a\u6a21\u6001<\/li>\n<\/ul>\n

            8.3 \u540e\u7eed\u4f18\u5316\u5efa\u8bae<\/span><\/span><\/h3>\n
              \n
            • \u6027\u80fd\u4f18\u5316：\u4f7f\u7528gunicorn\u66ff\u4ee3Flask\u5f00\u53d1\u670d\u52a1\u5668，\u6216\u6539\u7528\u5f02\u6b65\u6846\u67b6（\u5982FastAPI + uvicorn）\u63d0\u5347\u5e76\u53d1\u80fd\u529b<\/li>\n
            • \u663e\u5b58\u4f18\u5316：\u63a2\u7d22910A\u539f\u751fINT8\u91cf\u5316（\u901a\u8fc7\u534e\u4e3aCloudMatrix-Infer\u6216NPU-INT8\u5e93），\u53ef\u5c06\u6a21\u578b\u538b\u7f29\u81f31\/4，\u652f\u6301\u66f4\u5927\u4e0a\u4e0b\u6587<\/li>\n
            • \u6a21\u578b\u5347\u7ea7：\u8ddf\u8e2atransformers 4.57.1+\u5bf9Qwen3.5\u7cfb\u5217\u7684\u652f\u6301，\u4f46\u5347\u7ea7\u9700\u8c28\u614e\u9a8c\u8bc1PyTorch\u517c\u5bb9\u6027<\/li>\n
            • \u591a\u5361\u4f18\u5316：\u6d4b\u8bd56\u5361\u914d\u7f6e（\u75592\u5361\u7ed9\u5176\u4ed6\u4efb\u52a1），\u9a8c\u8bc1Qwen3-32B\u57286\u5361\u4e0a\u7684\u7a33\u5b9a\u6027<\/li>\n
            • \u751f\u4ea7\u90e8\u7f72：\u521b\u5efasystemd\u670d\u52a1\u6587\u4ef6，\u5b9e\u73b0\u5f00\u673a\u81ea\u542f；\u914d\u7f6eNginx\u53cd\u5411\u4ee3\u7406，\u589e\u52a0API\u9650\u6d41\u548c\u8ba4\u8bc1<\/li>\n
            • \u5f02\u6784\u96c6\u7fa4：\u8003\u8651\u4f7f\u7528GPUStack\u7b49\u4e2d\u95f4\u5c42\u7edf\u4e00\u7ba1\u7406910A（PyTorch Native）\u4e0e5090\/4090（vLLM）\u5f02\u6784\u96c6\u7fa4<\/li>\n<\/ul>\n

              \u4e5d\u3001\u9644\u5f55<\/span><\/span><\/h2>\n

              9.1 \u5173\u952e\u547d\u4ee4\u8bb0\u5f55<\/span><\/span><\/h3>\n

              # \u542f\u52a8\u5bb9\u5668
              \ndocker run -it –ipc=host –name=mindie –shm-size=80G –net=host –privileged=true \\\\
              \n\u00a0\u00a0–device=\/dev\/davinci_manager \\\\
              \n\u00a0\u00a0–device=\/dev\/devmm_svm \\\\
              \n\u00a0\u00a0–device=\/dev\/hisi_hdc \\\\
              \n\u00a0\u00a0-v \/usr\/local\/Ascend\/driver:\/usr\/local\/Ascend\/driver \\\\
              \n\u00a0\u00a0-v \/usr\/local\/Ascend\/add-ons\/:\/usr\/local\/Ascend\/add-ons\/ \\\\
              \n\u00a0\u00a0-v \/usr\/local\/sbin\/npu-smi:\/usr\/local\/sbin\/npu-smi \\\\
              \n\u00a0\u00a0-v \/var\/log\/npu\/slog\/:\/var\/log\/npu\/slog \\\\
              \n\u00a0\u00a0-v \/data:\/home \\\\
              \n\u00a0\u00a0mindie_910a:2.3.0 \/bin\/bash<\/p>\n

              # \u9a8c\u8bc1NPU
              \nnpu-smi info<\/p>\n

              # \u9a8c\u8bc1PyTorch NPU
              \npython -c "import torch; import torch_npu; print(torch.randn(2,2).npu())"<\/p>\n

              # \u542f\u52a8\u670d\u52a1
              \nexport ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
              \ncd \/home && python qwen_server_32b_8card.py<\/p>\n

              # \u6d4b\u8bd5API
              \ncurl -H "Content-Type: application\/json" -d '{
              \n\u00a0\u00a0"model": "Qwen3-32B",
              \n\u00a0\u00a0"messages": [{"role":"user","content":"\u4f60\u597d"}],
              \n\u00a0\u00a0"max_tokens": 128
              \n}' http:\/\/localhost:8000\/v1\/chat\/completions<\/span><\/p>\n

              9.2 \u7248\u672c\u517c\u5bb9\u6027\u77e9\u9635<\/span><\/span><\/h3>\n\n\n\n\n\n\n\n
              \n

              \u7ec4\u4ef6<\/p>\n<\/td>\n

              		\n

              vLLM\u65b9\u6848<\/p>\n<\/td>\n

              		\n

              llama.cpp\u65b9\u6848<\/p>\n<\/td>\n

              		\n

              PyTorch Native\u65b9\u6848<\/p>\n<\/td>\n

              		\n

              \u5907\u6ce8<\/p>\n<\/td>\n<\/tr>\n

              \n

              CANN<\/p>\n<\/td>\n

              		\n

              9.1.0-beta.1<\/p>\n<\/td>\n

              		\n

              9.1.0-beta.1<\/p>\n<\/td>\n

              		\n

              8.5.0（\u5bb9\u5668\u5185）<\/p>\n<\/td>\n

              		\n

              910A\u9700CANN 8.5+<\/p>\n<\/td>\n<\/tr>\n

              \n

              PyTorch<\/p>\n<\/td>\n

              		\n

              2.5.1<\/p>\n<\/td>\n

              		\n

              N\/A<\/p>\n<\/td>\n

              		\n

              2.1.0<\/p>\n<\/td>\n

              		\n

              \u5bb9\u5668\u51852.1.0\u7a33\u5b9a<\/p>\n<\/td>\n<\/tr>\n

              \n

              transformers<\/p>\n<\/td>\n

              		\n

              5.9.0<\/p>\n<\/td>\n

              		\n

              N\/A<\/p>\n<\/td>\n

              		\n

              4.51.0<\/p>\n<\/td>\n

              		\n

              4.51.0\u652f\u6301Qwen3<\/p>\n<\/td>\n<\/tr>\n

              \n

              \u6a21\u578b\u683c\u5f0f<\/p>\n<\/td>\n

              		\n

              HuggingFace<\/p>\n<\/td>\n

              		\n

              GGUF<\/p>\n<\/td>\n

              		\n

              HuggingFace<\/p>\n<\/td>\n

              		\n

              Native\u65b9\u6848\u6700\u7075\u6d3b<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

              9.3 \u53c2\u8003\u6587\u6863<\/span><\/span><\/h3>\n
                \n
              • vLLM-Ascend\u5b98\u65b9\u6587\u6863：https:\/\/vllm-ascend.readthedocs.io\/<\/li>\n
              • llama.cpp CANN\u540e\u7aef\u6587\u6863：https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/docs\/backend\/CANN.md<\/li>\n
              • \u534e\u4e3aCANN\u6587\u6863：https:\/\/www.hiascend.com\/software\/cann<\/li>\n
              • Qwen\u6a21\u578b\u6587\u6863：https:\/\/qwen.readthedocs.io\/<\/li>\n<\/ul>\n

                \u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"

                \u62a5\u544a\u7f16\u53f7RPT-910A-20260531-002\u6d4b\u8bd5\u65e5\u671f2026\u5e745\u670830\u65e5-31\u65e5\u670d\u52a1\u5668bms-ytcs910a-260529 (\u9cb2\u9e4fCPU 8\u5361910A NPU)\u62a5\u544a\u7248\u672cV2.0 (\u7efc\u5408\u7248)\u4e00\u3001\u6d4b\u8bd5\u6982\u8ff0\u672c\u6b21\u6d4b\u8bd5\u5386\u65f6\u7ea612\u5c0f\u65f6，\u7cfb\u7edf\u6027\u5730\u9a8c\u8bc1\u4e86\u5728\u534e\u4e3a\u6607\u817e910A NPU\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u670d\u52a1\u7684\u53ef\u884c\u6027\u3002\u6d4b\u8bd5\u5148\u540e\u5c1d\u8bd5\u4e86\u4e09\u79cd\u4e3b\u6d41\u63a8\u7406\u6846\u67b6：vLLM vLLM-Ascend：\u57fa\u4e8ePyTorch\u751f\u6001\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6，\u652f<\/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":[43,51,44],"topic":[],"class_list":["post-81444","post","type-post","status-publish","format-standard","hentry","category-server","tag-43","tag-51","tag-44"],"yoast_head":"\n\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \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\/81444.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u62a5\u544a\u7f16\u53f7RPT-910A-20260531-002\u6d4b\u8bd5\u65e5\u671f2026\u5e745\u670830\u65e5-31\u65e5\u670d\u52a1\u5668bms-ytcs910a-260529 (\u9cb2\u9e4fCPU 8\u5361910A NPU)\u62a5\u544a\u7248\u672cV2.0 (\u7efc\u5408\u7248)\u4e00\u3001\u6d4b\u8bd5\u6982\u8ff0\u672c\u6b21\u6d4b\u8bd5\u5386\u65f6\u7ea612\u5c0f\u65f6，\u7cfb\u7edf\u6027\u5730\u9a8c\u8bc1\u4e86\u5728\u534e\u4e3a\u6607\u817e910A NPU\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u670d\u52a1\u7684\u53ef\u884c\u6027\u3002\u6d4b\u8bd5\u5148\u540e\u5c1d\u8bd5\u4e86\u4e09\u79cd\u4e3b\u6d41\u63a8\u7406\u6846\u67b6：vLLM vLLM-Ascend：\u57fa\u4e8ePyTorch\u751f\u6001\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6，\u652f\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/81444.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-24T18:01:12+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=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/81444.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/81444.html\",\"name\":\"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-24T18:01:12+00:00\",\"dateModified\":\"2026-07-24T18:01:12+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/81444.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/81444.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/81444.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a\"}]},{\"@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":"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \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\/81444.html","og_locale":"zh_CN","og_type":"article","og_title":"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u62a5\u544a\u7f16\u53f7RPT-910A-20260531-002\u6d4b\u8bd5\u65e5\u671f2026\u5e745\u670830\u65e5-31\u65e5\u670d\u52a1\u5668bms-ytcs910a-260529 (\u9cb2\u9e4fCPU 8\u5361910A NPU)\u62a5\u544a\u7248\u672cV2.0 (\u7efc\u5408\u7248)\u4e00\u3001\u6d4b\u8bd5\u6982\u8ff0\u672c\u6b21\u6d4b\u8bd5\u5386\u65f6\u7ea612\u5c0f\u65f6，\u7cfb\u7edf\u6027\u5730\u9a8c\u8bc1\u4e86\u5728\u534e\u4e3a\u6607\u817e910A NPU\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u670d\u52a1\u7684\u53ef\u884c\u6027\u3002\u6d4b\u8bd5\u5148\u540e\u5c1d\u8bd5\u4e86\u4e09\u79cd\u4e3b\u6d41\u63a8\u7406\u6846\u67b6：vLLM vLLM-Ascend：\u57fa\u4e8ePyTorch\u751f\u6001\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6，\u652f","og_url":"https:\/\/www.wsisp.com\/helps\/81444.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-24T18:01:12+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/81444.html","url":"https:\/\/www.wsisp.com\/helps\/81444.html","name":"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-24T18:01:12+00:00","dateModified":"2026-07-24T18:01:12+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/81444.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/81444.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/81444.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u9cb2\u9e4fCPU + 8\u5361910A NPU\u670d\u52a1\u5668\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u90e8\u7f72\u6d4b\u8bd5\u62a5\u544a"}]},{"@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\/81444","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=81444"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/81444\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=81444"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=81444"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=81444"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=81444"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}