{"id":1237,"date":"2025-12-11T18:56:00","date_gmt":"2025-12-11T10:56:00","guid":{"rendered":"http:\/\/www.o6s.net\/?p=1237"},"modified":"2025-12-11T18:56:00","modified_gmt":"2025-12-11T10:56:00","slug":"llama-b7356-bin-win-cuda-12-4-x64-llama-server%e5%8f%82%e6%95%b0%e5%8f%8a%e4%bd%bf%e7%94%a8%e8%af%b4%e6%98%8e","status":"publish","type":"post","link":"https:\/\/www.o6s.net\/index.php\/2025\/12\/11\/llama-b7356-bin-win-cuda-12-4-x64-llama-server%e5%8f%82%e6%95%b0%e5%8f%8a%e4%bd%bf%e7%94%a8%e8%af%b4%e6%98%8e\/","title":{"rendered":"llama-b7356-bin-win-cuda-12.4-x64-llama-server\u53c2\u6570\u53ca\u4f7f\u7528\u8bf4\u660e"},"content":{"rendered":"

—– \u5e38\u89c1\u53c2\u6570 —–<\/p>\n

-h, –help, –usage \u6253\u5370\u7528\u6cd5\u5e76\u9000\u51fa
\n–version \u663e\u793a\u7248\u672c\u548c\u6784\u5efa\u4fe1\u606f
\n-cl, –cache-list \u663e\u793a\u7f13\u5b58\u4e2d\u7684\u6a21\u578b\u5217\u8868
\n–completion-bash \u6253\u5370\u53ef\u7528\u4e8e\u6e90\u4ee3\u7801\u7684bash\u8865\u5168\u811a\u672c
\n–verbose-prompt \u5728\u751f\u6210\u524d\u6253\u5370\u8be6\u7ec6\u7684\u63d0\u793a\u4fe1\u606f\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n-t, –threads N \u751f\u6210\u671f\u95f4\u4f7f\u7528\u7684CPU\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a-1\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_THREADS\uff09
\n-tb, –threads-batch N \u6279\u5904\u7406\u548c\u63d0\u793a\u5904\u7406\u671f\u95f4\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a
\n\u4e0e–threads\u76f8\u540c\uff09
\n-C, –cpu-mask M CPU\u4eb2\u548c\u6027\u63a9\u7801\uff1a\u4efb\u610f\u957f\u5ea6\u7684\u5341\u516d\u8fdb\u5236\u3002\u4e0ecpu-range\u4e92\u8865
\n\uff08\u9ed8\u8ba4\uff1a””\uff09
\n-Cr, –cpu-range lo-hi \u7528\u4e8e\u4eb2\u548c\u6027\u7684CPU\u8303\u56f4\u3002\u4e0e–cpu-mask\u4e92\u8865
\n–cpu-strict <0|1> \u4f7f\u7528\u4e25\u683c\u7684CPU\u653e\u7f6e\uff08\u9ed8\u8ba4\uff1a0\uff09
\n–prio N \u8bbe\u7f6e\u8fdb\u7a0b\/\u7ebf\u7a0b\u4f18\u5148\u7ea7\uff1alow(-1), normal(0), medium(1), high(2),
\nrealtime(3)\uff08\u9ed8\u8ba4\uff1a0\uff09
\n–poll <0…100> \u4f7f\u7528\u8f6e\u8be2\u7ea7\u522b\u7b49\u5f85\u5de5\u4f5c\uff080 – \u4e0d\u8f6e\u8be2\uff0c\u9ed8\u8ba4\uff1a50\uff09
\n-Cb, –cpu-mask-batch M \u6279\u5904\u7406\u7684CPU\u4eb2\u548c\u6027\u63a9\u7801\uff1a\u4efb\u610f\u957f\u5ea6\u7684\u5341\u516d\u8fdb\u5236\u3002\u4e0ecpu-range-batch\u4e92\u8865
\n\uff08\u9ed8\u8ba4\uff1a\u4e0e–cpu-mask\u76f8\u540c\uff09
\n-Crb, –cpu-range-batch lo-hi \u6279\u5904\u7406\u7684CPU\u4eb2\u548c\u6027\u8303\u56f4\u3002\u4e0e–cpu-mask-batch\u4e92\u8865
\n–cpu-strict-batch <0|1> \u4f7f\u7528\u4e25\u683c\u7684CPU\u653e\u7f6e\uff08\u9ed8\u8ba4\uff1a\u4e0e–cpu-strict\u76f8\u540c\uff09
\n–prio-batch N \u8bbe\u7f6e\u8fdb\u7a0b\/\u7ebf\u7a0b\u4f18\u5148\u7ea7\uff1a0-normal, 1-medium, 2-high, 3-realtime
\n\uff08\u9ed8\u8ba4\uff1a0\uff09
\n–poll-batch <0|1> \u4f7f\u7528\u8f6e\u8be2\u7b49\u5f85\u5de5\u4f5c\uff08\u9ed8\u8ba4\uff1a\u4e0e–poll\u76f8\u540c\uff09
\n-c, –ctx-size N \u63d0\u793a\u4e0a\u4e0b\u6587\u7684\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a4096\uff0c0 = \u4ece\u6a21\u578b\u52a0\u8f7d\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CTX_SIZE\uff09
\n-n, –predict, –n-predict N \u9884\u6d4b\u7684token\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a-1\uff0c-1 = \u65e0\u9650\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_PREDICT\uff09
\n-b, –batch-size N \u903b\u8f91\u6700\u5927\u6279\u5904\u7406\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a2048\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_BATCH\uff09
\n-ub, –ubatch-size N \u7269\u7406\u6700\u5927\u6279\u5904\u7406\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a512\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_UBATCH\uff09
\n–keep N \u4ece\u521d\u59cb\u63d0\u793a\u4e2d\u4fdd\u7559\u7684token\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a0\uff0c-1 =
\n\u5168\u90e8\uff09
\n–swa-full \u4f7f\u7528\u5b8c\u6574\u5927\u5c0f\u7684SWA\u7f13\u5b58\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n[(\u66f4\u591a\u4fe1\u606f)](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/13194#issuecomment-2868343055)
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SWA_FULL\uff09
\n–kv-unified, -kvu \u4e3a\u6240\u6709\u5e8f\u5217\u7684KV\u7f13\u5b58\u4f7f\u7528\u5355\u4e00\u7edf\u4e00\u7684KV\u7f13\u51b2\u533a
\n\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n[(\u66f4\u591a\u4fe1\u606f)](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/14363)
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_KV_UNIFIED\uff09
\n-fa, –flash-attn [on|off|auto] \u8bbe\u7f6eFlash Attention\u4f7f\u7528\uff08’on’, ‘off’, \u6216 ‘auto’, \u9ed8\u8ba4\uff1a’auto’\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_FLASH_ATTN\uff09
\n–no-perf \u7981\u7528\u5185\u90e8libllama\u6027\u80fd\u8ba1\u65f6\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_PERF\uff09
\n-e, –escape \u5904\u7406\u8f6c\u4e49\u5e8f\u5217\uff08\\n, \\r, \\t, \\’, \\”, \\\\\uff09\uff08\u9ed8\u8ba4\uff1atrue\uff09
\n–no-escape \u4e0d\u5904\u7406\u8f6c\u4e49\u5e8f\u5217
\n–rope-scaling {none,linear,yarn} RoPE\u9891\u7387\u7f29\u653e\u65b9\u6cd5\uff0c\u9ed8\u8ba4\u4e3alinear\u9664\u975e\u6a21\u578b\u6307\u5b9a
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_SCALING_TYPE\uff09
\n–rope-scale N RoPE\u4e0a\u4e0b\u6587\u7f29\u653e\u56e0\u5b50\uff0c\u5c06\u4e0a\u4e0b\u6587\u6269\u5c55N\u500d
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_SCALE\uff09
\n–rope-freq-base N RoPE\u57fa\u7840\u9891\u7387\uff0c\u7528\u4e8eNTK\u611f\u77e5\u7f29\u653e\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u52a0\u8f7d\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_FREQ_BASE\uff09
\n–rope-freq-scale N RoPE\u9891\u7387\u7f29\u653e\u56e0\u5b50\uff0c\u5c06\u4e0a\u4e0b\u6587\u6269\u5c551\/N\u500d
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_FREQ_SCALE\uff09
\n–yarn-orig-ctx N YaRN\uff1a\u6a21\u578b\u7684\u539f\u59cb\u4e0a\u4e0b\u6587\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0 = \u6a21\u578b\u8bad\u7ec3
\n\u4e0a\u4e0b\u6587\u5927\u5c0f\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_ORIG_CTX\uff09
\n–yarn-ext-factor N YaRN\uff1a\u5916\u63a8\u6df7\u5408\u56e0\u5b50\uff08\u9ed8\u8ba4\uff1a-1.0\uff0c0.0 = \u5168\u90e8
\n\u63d2\u503c\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_EXT_FACTOR\uff09
\n–yarn-attn-factor N YaRN\uff1a\u7f29\u653esqrt(t)\u6216\u6ce8\u610f\u529b\u5e45\u5ea6\uff08\u9ed8\u8ba4\uff1a-1.0\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_ATTN_FACTOR\uff09
\n–yarn-beta-slow N YaRN\uff1a\u9ad8\u4fee\u6b63\u7ef4\u5ea6\u6216alpha\uff08\u9ed8\u8ba4\uff1a-1.0\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_BETA_SLOW\uff09
\n–yarn-beta-fast N YaRN\uff1a\u4f4e\u4fee\u6b63\u7ef4\u5ea6\u6216beta\uff08\u9ed8\u8ba4\uff1a-1.0\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_BETA_FAST\uff09
\n-nkvo, –no-kv-offload \u7981\u7528KV\u5378\u8f7d
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_KV_OFFLOAD\uff09
\n-nr, –no-repack \u7981\u7528\u6743\u91cd\u91cd\u65b0\u6253\u5305
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_REPACK\uff09
\n–no-host \u7ed5\u8fc7\u4e3b\u673a\u7f13\u51b2\u533a\u5141\u8bb8\u4f7f\u7528\u989d\u5916\u7f13\u51b2\u533a
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_HOST\uff09
\n-ctk, –cache-type-k TYPE K\u7684KV\u7f13\u5b58\u6570\u636e\u7c7b\u578b
\n\u5141\u8bb8\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
\n\uff08\u9ed8\u8ba4\uff1af16\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_K\uff09
\n-ctv, –cache-type-v TYPE V\u7684KV\u7f13\u5b58\u6570\u636e\u7c7b\u578b
\n\u5141\u8bb8\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
\n\uff08\u9ed8\u8ba4\uff1af16\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_V\uff09
\n-dt, –defrag-thold N KV\u7f13\u5b58\u788e\u7247\u6574\u7406\u9608\u503c\uff08\u5df2\u5f03\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DEFRAG_THOLD\uff09
\n-np, –parallel N \u89e3\u7801\u7684\u5e76\u884c\u5e8f\u5217\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a1\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_PARALLEL\uff09
\n–rpc SERVERS \u9017\u53f7\u5206\u9694\u7684RPC\u670d\u52a1\u5668\u5217\u8868
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_RPC\uff09
\n–mlock \u5f3a\u5236\u7cfb\u7edf\u5c06\u6a21\u578b\u4fdd\u7559\u5728RAM\u4e2d\u800c\u4e0d\u662f\u4ea4\u6362\u6216\u538b\u7f29
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MLOCK\uff09
\n–no-mmap \u4e0d\u4f7f\u7528\u5185\u5b58\u6620\u5c04\u6a21\u578b\uff08\u52a0\u8f7d\u8f83\u6162\u4f46\u53ef\u80fd\u51cf\u5c11\u9875\u9762\u4ea4\u6362\u5982\u679c\u672a\u4f7f\u7528mlock\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_MMAP\uff09
\n–numa TYPE \u5c1d\u8bd5\u5728\u67d0\u4e9bNUMA\u7cfb\u7edf\u4e0a\u4f18\u5316
\n– distribute: \u5747\u5300\u5206\u5e03\u5728\u6240\u6709\u8282\u70b9\u4e0a\u6267\u884c
\n– isolate: \u4ec5\u5728\u6267\u884c\u5f00\u59cb\u7684\u8282\u70b9\u7684CPU\u4e0a\u751f\u6210\u7ebf\u7a0b
\n– numactl: \u4f7f\u7528numactl\u63d0\u4f9b\u7684CPU\u6620\u5c04
\n\u5982\u679c\u4e4b\u524d\u672a\u4f7f\u7528\u6b64\u9009\u9879\uff0c\u5efa\u8bae\u5728\u4f7f\u7528\u524d\u6e05\u7a7a\u7cfb\u7edf\u9875\u9762\u7f13\u5b58
\n\u8bf7\u53c2\u9605 https:\/\/github.com\/ggml-org\/llama.cpp\/issues\/1437
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NUMA\uff09
\n-dev, –device <dev1,dev2,..> \u7528\u4e8e\u5378\u8f7d\u7684\u8bbe\u5907\u5217\u8868\uff08none = \u4e0d\u5378\u8f7d\uff09
\n\u4f7f\u7528–list-devices\u67e5\u770b\u53ef\u7528\u8bbe\u5907\u5217\u8868
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DEVICE\uff09
\n–list-devices \u6253\u5370\u53ef\u7528\u8bbe\u5907\u5217\u8868\u5e76\u9000\u51fa
\n–override-tensor, -ot <tensor name pattern>=<buffer type>,…
\n\u8986\u76d6\u5f20\u91cf\u7f13\u51b2\u533a\u7c7b\u578b
\n–cpu-moe, -cmoe \u5c06\u6240\u6709\u4e13\u5bb6\u6df7\u5408\uff08MoE\uff09\u6743\u91cd\u4fdd\u7559\u5728CPU\u4e0a
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CPU_MOE\uff09
\n–n-cpu-moe, -ncmoe N \u5c06\u524dN\u5c42\u7684\u4e13\u5bb6\u6df7\u5408\uff08MoE\uff09\u6743\u91cd\u4fdd\u7559\u5728CPU\u4e0a
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_CPU_MOE\uff09
\n-ngl, –gpu-layers, –n-gpu-layers N \u5b58\u50a8\u5728VRAM\u4e2d\u7684\u6700\u5927\u5c42\u6570\uff08\u9ed8\u8ba4\uff1a-1\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_GPU_LAYERS\uff09
\n-sm, –split-mode {none,layer,row} \u5982\u4f55\u5728\u591a\u4e2aGPU\u4e4b\u95f4\u62c6\u5206\u6a21\u578b\uff0c\u4ee5\u4e0b\u4e4b\u4e00\uff1a
\n– none: \u4ec5\u4f7f\u7528\u4e00\u4e2aGPU
\n– layer\uff08\u9ed8\u8ba4\uff09: \u5728GPU\u4e4b\u95f4\u62c6\u5206\u5c42\u548cKV
\n– row: \u5728GPU\u4e4b\u95f4\u62c6\u5206\u884c
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SPLIT_MODE\uff09
\n-ts, –tensor-split N0,N1,N2,… \u6bcf\u4e2aGPU\u5378\u8f7d\u6a21\u578b\u7684\u6bd4\u4f8b\uff0c\u9017\u53f7\u5206\u9694\u7684\u6bd4\u4f8b\u5217\u8868\uff0c\u4f8b\u59823,1
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_TENSOR_SPLIT\uff09
\n-mg, –main-gpu INDEX \u4f7f\u7528\u7684GPU\uff08split-mode = none\u65f6\uff09\uff0c\u6216\u7528\u4e8e
\n\u4e2d\u95f4\u7ed3\u679c\u548cKV\uff08split-mode = row\u65f6\uff09\uff08\u9ed8\u8ba4\uff1a0\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MAIN_GPU\uff09
\n–check-tensors \u68c0\u67e5\u6a21\u578b\u5f20\u91cf\u6570\u636e\u4e2d\u7684\u65e0\u6548\u503c\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n–override-kv KEY=TYPE:VALUE \u9ad8\u7ea7\u9009\u9879\uff0c\u901a\u8fc7\u952e\u8986\u76d6\u6a21\u578b\u5143\u6570\u636e\u3002\u53ef\u4ee5\u591a\u6b21\u6307\u5b9a\u3002
\n\u7c7b\u578b\uff1aint, float, bool, str\u3002\u793a\u4f8b\uff1a–override-kv
\ntokenizer.ggml.add_bos_token=bool:false
\n–no-op-offload \u7981\u7528\u5c06\u4e3b\u673a\u5f20\u91cf\u64cd\u4f5c\u5378\u8f7d\u5230\u8bbe\u5907\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n–lora FNAME LoRA\u9002\u914d\u5668\u8def\u5f84\uff08\u53ef\u91cd\u590d\u4f7f\u7528\u591a\u4e2a\u9002\u914d\u5668\uff09
\n–lora-scaled FNAME SCALE \u5e26\u7528\u6237\u5b9a\u4e49\u7f29\u653e\u7684LoRA\u9002\u914d\u5668\u8def\u5f84\uff08\u53ef\u91cd\u590d\u4f7f\u7528\u591a\u4e2a\u9002\u914d\u5668\uff09
\n–control-vector FNAME \u6dfb\u52a0\u63a7\u5236\u5411\u91cf
\n\u6ce8\u610f\uff1a\u6b64\u53c2\u6570\u53ef\u4ee5\u91cd\u590d\u6dfb\u52a0\u591a\u4e2a\u63a7\u5236\u5411\u91cf
\n–control-vector-scaled FNAME SCALE \u6dfb\u52a0\u5e26\u7528\u6237\u5b9a\u4e49\u7f29\u653eSCALE\u7684\u63a7\u5236\u5411\u91cf
\n\u6ce8\u610f\uff1a\u6b64\u53c2\u6570\u53ef\u4ee5\u91cd\u590d\u6dfb\u52a0\u591a\u4e2a\u7f29\u653e\u63a7\u5236\u5411\u91cf
\n–control-vector-layer-range START END
\n\u5e94\u7528\u63a7\u5236\u5411\u91cf\u7684\u5c42\u8303\u56f4\uff0c\u5305\u62ec\u8d77\u59cb\u548c\u7ed3\u675f
\n-m, –model FNAME \u52a0\u8f7d\u6a21\u578b\u7684\u8def\u5f84
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODEL\uff09
\n-mu, –model-url MODEL_URL \u6a21\u578b\u4e0b\u8f7dURL\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODEL_URL\uff09
\n-dr, –docker-repo [<repo>\/]<model>[:quant]
\nDocker Hub\u6a21\u578b\u4ed3\u5e93\u3002repo\u662f\u53ef\u9009\u7684\uff0c\u9ed8\u8ba4\u4e3aai\/. quant
\n\u662f\u53ef\u9009\u7684\uff0c\u9ed8\u8ba4\u4e3a:latest.
\n\u793a\u4f8b\uff1agemma3
\n\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DOCKER_REPO\uff09
\n-hf, -hfr, –hf-repo <user>\/<model>[:quant]
\nHugging Face\u6a21\u578b\u4ed3\u5e93\uff1bquant\u662f\u53ef\u9009\u7684\uff0c\u4e0d\u533a\u5206\u5927\u5c0f\u5199\uff0c
\n\u9ed8\u8ba4\u4e3aQ4_K_M\uff0c\u6216\u5982\u679cQ4_K_M\u4e0d\u5b58\u5728\uff0c\u5219\u56de\u9000\u5230\u4ed3\u5e93\u4e2d\u7684\u7b2c\u4e00\u4e2a\u6587\u4ef6\u3002
\n\u5982\u679c\u53ef\u7528\uff0cmmproj\u4e5f\u4f1a\u81ea\u52a8\u4e0b\u8f7d\u3002\u8981\u7981\u7528\uff0c\u8bf7\u6dfb\u52a0
\n–no-mmproj
\n\u793a\u4f8b\uff1aunsloth\/phi-4-GGUF:q4_k_m
\n\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_REPO\uff09
\n-hfd, -hfrd, –hf-repo-draft <user>\/<model>[:quant]
\n\u4e0e–hf-repo\u76f8\u540c\uff0c\u4f46\u7528\u4e8e\u8349\u7a3f\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HFD_REPO\uff09
\n-hff, –hf-file FILE Hugging Face\u6a21\u578b\u6587\u4ef6\u3002\u5982\u679c\u6307\u5b9a\uff0c\u5c06\u8986\u76d6–hf-repo\u4e2d\u7684quant\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_FILE\uff09
\n-hfv, -hfrv, –hf-repo-v <user>\/<model>[:quant]
\n\u8bed\u97f3\u7f16\u7801\u5668\u6a21\u578b\u7684Hugging Face\u6a21\u578b\u4ed3\u5e93\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_REPO_V\uff09
\n-hffv, –hf-file-v FILE \u8bed\u97f3\u7f16\u7801\u5668\u6a21\u578b\u7684Hugging Face\u6a21\u578b\u6587\u4ef6\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_FILE_V\uff09
\n-hft, –hf-token TOKEN Hugging Face\u8bbf\u95ee\u4ee4\u724c\uff08\u9ed8\u8ba4\uff1a\u6765\u81eaHF_TOKEN\u73af\u5883
\n\u53d8\u91cf\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aHF_TOKEN\uff09
\n–log-disable \u7981\u7528\u65e5\u5fd7
\n–log-file FNAME \u65e5\u5fd7\u5230\u6587\u4ef6
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_FILE\uff09
\n–log-colors [on|off|auto] \u8bbe\u7f6e\u5f69\u8272\u65e5\u5fd7\uff08’on’, ‘off’, \u6216 ‘auto’, \u9ed8\u8ba4\uff1a’auto’\uff09
\n‘auto’ \u5728\u8f93\u51fa\u5230\u7ec8\u7aef\u65f6\u542f\u7528\u989c\u8272
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_COLORS\uff09
\n-v, –verbose, –log-verbose \u8bbe\u7f6e\u8be6\u7ec6\u7ea7\u522b\u4e3a\u65e0\u9650\uff08\u5373\u8bb0\u5f55\u6240\u6709\u6d88\u606f\uff0c\u9002\u7528\u4e8e
\n\u8c03\u8bd5\uff09
\n–offline \u79bb\u7ebf\u6a21\u5f0f\uff1a\u5f3a\u5236\u4f7f\u7528\u7f13\u5b58\uff0c\u9632\u6b62\u7f51\u7edc\u8bbf\u95ee
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_OFFLINE\uff09
\n-lv, –verbosity, –log-verbosity N \u8bbe\u7f6e\u8be6\u7ec6\u7ea7\u522b\u9608\u503c\u3002\u8be6\u7ec6\u7ea7\u522b\u66f4\u9ad8\u7684\u6d88\u606f\u5c06\u88ab\u5ffd\u7565\u3002\u503c\uff1a
\n– 0: \u901a\u7528\u8f93\u51fa
\n– 1: \u9519\u8bef
\n– 2: \u8b66\u544a
\n– 3: \u4fe1\u606f
\n– 4: \u8c03\u8bd5
\n\uff08\u9ed8\u8ba4\uff1a3\uff09<\/p>\n

\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_VERBOSITY\uff09
\n–log-prefix \u5728\u65e5\u5fd7\u6d88\u606f\u4e2d\u542f\u7528\u524d\u7f00
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_PREFIX\uff09
\n–log-timestamps \u5728\u65e5\u5fd7\u6d88\u606f\u4e2d\u542f\u7528\u65f6\u95f4\u6233
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_TIMESTAMPS\uff09
\n-ctkd, –cache-type-k-draft TYPE \u8349\u7a3f\u6a21\u578bK\u7684KV\u7f13\u5b58\u6570\u636e\u7c7b\u578b
\n\u5141\u8bb8\u503c\uff1af32, f16, bf1\u7ebf, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
\n\uff08\u9ed8\u8ba4\uff1af16\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_K_DRAFT\uff09
\n-ctvd, –cache-type-v-draft TYPE \u8349\u7a3f\u6a21\u578bV\u7684KV\u7f13\u5b58\u6570\u636e\u7c7b\u578b
\n\u5141\u8bb8\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
\n\uff08\u9ed8\u8ba4\uff1af16\uff09
\n\uff08\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_V_DRAFT\uff09<\/p>\n

—– \u91c7\u6837 params —–<\/p>\n

–samplers SAMPLERS \u751f\u6210\u8fc7\u7a0b\u4e2d\u6309\u987a\u5e8f\u4f7f\u7528\u7684\u91c7\u6837\u5668\uff0c\u7528\u5206\u53f7\u5206\u9694
\n\uff08\u9ed8\u8ba4\uff1a
\npenalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature\uff09
\n-s, –seed SEED RNG\u79cd\u5b50\uff08\u9ed8\u8ba4\uff1a-1\uff0c-1\u65f6\u4f7f\u7528\u968f\u673a\u79cd\u5b50\uff09
\n–sampling-seq, –sampler-seq SEQUENCE
\n\u7528\u4e8e\u751f\u6210\u7684\u91c7\u6837\u5668\u7b80\u5316\u5e8f\u5217\uff08\u9ed8\u8ba4\uff1a
\nedskypmxt\uff09
\n–ignore-eos \u5ffd\u7565\u6d41\u7ed3\u675f\u6807\u8bb0\u5e76\u7ee7\u7eed\u751f\u6210\uff08\u610f\u5473\u7740
\n–logit-bias EOS-inf\uff09
\n–temp N \u6e29\u5ea6\uff08\u9ed8\u8ba4\uff1a0.8\uff09
\n–top-k N top-k\u91c7\u6837\uff08\u9ed8\u8ba4\uff1a40\uff0c0 = \u7981\u7528\uff09
\n–top-p N top-p\u91c7\u6837\uff08\u9ed8\u8ba4\uff1a0.9\uff0c1.0 = \u7981\u7528\uff09
\n–min-p N min-p\u91c7\u6837\uff08\u9ed8\u8ba4\uff1a0.1\uff0c0.0 = \u7981\u7528\uff09
\n–top-nsigma N top-n-sigma\u91c7\u6837\uff08\u9ed8\u8ba4\uff1a-1.0\uff0c-1.0 = \u7981\u7528\uff09
\n–xtc-probability N xtc\u6982\u7387\uff08\u9ed8\u8ba4\uff1a0.0\uff0c0.0 = \u7981\u7528\uff09
\n–xtc-threshold N xtc\u9608\u503c\uff08\u9ed8\u8ba4\uff1a0.1\uff0c1.0 = \u7981\u7528\uff09
\n–typical N \u672c\u5730\u5178\u578b\u91c7\u6837\uff0c\u53c2\u6570p\uff08\u9ed8\u8ba4\uff1a1.0\uff0c1.0 = \u7981\u7528\uff09
\n–repeat-last-n N \u7528\u4e8e\u60e9\u7f5a\u7684\u6700\u540en\u4e2atoken\uff08\u9ed8\u8ba4\uff1a64\uff0c0 = \u7981\u7528\uff0c-1
\n= ctx_size\uff09
\n–repeat-penalty N \u5bf9\u91cd\u590d\u7684token\u5e8f\u5217\u8fdb\u884c\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a1.0\uff0c1.0 = \u7981\u7528\uff09
\n–presence-penalty N \u91cd\u590dalpha\u5b58\u5728\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a0.0\uff0c0.0 = \u7981\u7528\uff09
\n–frequency-penalty N \u91cd\u590dalpha\u9891\u7387\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a0.0\uff0c0.0 = \u7981\u7528\uff09
\n–dry-multiplier N \u8bbe\u7f6eDRY\u91c7\u6837\u4e58\u6570\uff08\u9ed8\u8ba4\uff1a0.0\uff0c0.0 = \u7981\u7528\uff09
\n–dry-base N \u8bbe\u7f6eDRY\u91c7\u6837\u57fa\u7840\u503c\uff08\u9ed8\u8ba4\uff1a1.75\uff09
\n–dry-allowed-length N \u8bbe\u7f6eDRY\u91c7\u6837\u5141\u8bb8\u957f\u5ea6\uff08\u9ed8\u8ba4\uff1a2\uff09
\n–dry-penalty-last-n N \u8bbe\u7f6e\u6700\u540en\u4e2atoken\u7684DRY\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a-1\uff0c0 = \u7981\u7528\uff0c-1 =
\ncontext size\uff09
\n–dry-sequence-breaker STRING \u4e3aDRY\u91c7\u6837\u6dfb\u52a0\u5e8f\u5217\u7ec8\u6b62\u7b26\uff0c\u540c\u65f6\u6e05\u9664\u9ed8\u8ba4\u7ec8\u6b62\u7b26
\n\uff08’\\n’, ‘:’, ‘”‘, ‘*’\uff09\uff1b\u4f7f\u7528”none”\u8868\u793a\u4e0d\u4f7f\u7528\u4efb\u4f55
\n\u5e8f\u5217\u7ec8\u6b62\u7b26
\n–dynatemp-range N \u52a8\u6001\u6e29\u5ea6\u8303\u56f4\uff08\u9ed8\u8ba4\uff1a0.0\uff0c0.0 = \u7981\u7528\uff09
\n–dynatemp-exp N \u52a8\u6001\u6e29\u5ea6\u6307\u6570\uff08\u9ed8\u8ba4\uff1a1.0\uff09
\n–mirostat N \u4f7f\u7528Mirostat\u91c7\u6837\u3002
\n\u5982\u679c\u4f7f\u7528\uff0c\u5c06\u5ffd\u7565Top K\u3001Nucleus\u548c\u672c\u5730\u5178\u578b\u91c7\u6837\u5668\u3002
\n\uff08\u9ed8\u8ba4\uff1a0\uff0c0 = \u7981\u7528\uff0c1 = Mirostat\uff0c2 = Mirostat 2.0\uff09
\n–mirostat-lr N Mirostat\u5b66\u4e60\u7387\uff0c\u53c2\u6570eta\uff08\u9ed8\u8ba4\uff1a0.1\uff09
\n–mirostat-ent N Mirostat\u76ee\u6807\u71b5\uff0c\u53c2\u6570tau\uff08\u9ed8\u8ba4\uff1a5.0\uff09
\n-l, –logit-bias TOKEN_ID(+\/-)BIAS \u4fee\u6539token\u5728\u5b8c\u6210\u4e2d\u51fa\u73b0\u7684\u53ef\u80fd\u6027\uff0c
\n\u4f8b\u5982 `–logit-bias 15043+1` \u589e\u52a0token ‘ Hello’ \u51fa\u73b0\u7684\u53ef\u80fd\u6027\uff0c
\n\u6216 `–logit-bias 15043-1` \u51cf\u5c11token ‘ Hello’ \u51fa\u73b0\u7684\u53ef\u80fd\u6027
\n–grammar GRAMMAR \u7528\u4e8e\u7ea6\u675f\u751f\u6210\u7684\u7c7b\u4f3cBNF\u7684\u8bed\u6cd5\uff08\u53c2\u89c1grammars\/
\n\u76ee\u5f55\u4e2d\u7684\u793a\u4f8b\uff09\uff08\u9ed8\u8ba4\uff1a”\uff09
\n–grammar-file FNAME \u4ece\u6587\u4ef6\u8bfb\u53d6\u8bed\u6cd5
\n-j, –json-schema SCHEMA \u7528\u4e8e\u7ea6\u675f\u751f\u6210\u7684JSON\u6a21\u5f0f\uff08https:\/\/json-schema.org\/\uff09\uff0c\u4f8b\u5982
\n`{}` \u8868\u793a\u4efb\u4f55JSON\u5bf9\u8c61
\n\u5bf9\u4e8e\u5305\u542b\u5916\u90e8$ref\u7684\u6a21\u5f0f\uff0c\u8bf7\u4f7f\u7528 –grammar +
\nexample\/json_schema_to_grammar.py
\n-jf, –json-schema-file FILE \u5305\u542b\u7528\u4e8e\u7ea6\u675f\u751f\u6210\u7684JSON\u6a21\u5f0f\u7684\u6587\u4ef6
\n\uff08https:\/\/json-schema.org\/\uff09\uff0c\u4f8b\u5982 `{}` \u8868\u793a\u4efb\u4f55JSON\u5bf9\u8c61
\n\u5bf9\u4e8e\u5305\u542b\u5916\u90e8$ref\u7684\u6a21\u5f0f\uff0c\u8bf7\u4f7f\u7528 –grammar +
\nexample\/json_schema_to_grammar.py<\/p>\n

—– \u793a\u4f8b\u7279\u5b9a\u53c2\u6570 —–<\/p>\n

–ctx-checkpoints, –swa-checkpoints N
\n\u6bcf\u4e2a\u63d2\u69fd\u521b\u5efa\u7684\u4e0a\u4e0b\u6587\u68c0\u67e5\u70b9\u7684\u6700\u5927\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a8\uff09
\n[(\u66f4\u591a\u4fe1\u606f)](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/15293)
\n(env: LLAMA_ARG_CTX_CHECKPOINTS)
\n–cache-ram, -cram N \u8bbe\u7f6e\u6700\u5927\u7f13\u5b58\u5927\u5c0f\uff08\u5355\u4f4d\uff1aMiB\uff09\uff08\u9ed8\u8ba4\uff1a8192\uff0c-1 – \u65e0\u9650\u5236\uff0c0 – \u7981\u7528\uff09
\n[(\u66f4\u591a\u4fe1\u606f)](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/16391)
\n(env: LLAMA_ARG_CACHE_RAM)
\n–no-context-shift \u5728\u65e0\u9650\u6587\u672c\u751f\u6210\u4e2d\u7981\u7528\u4e0a\u4e0b\u6587\u5207\u6362\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_NO_CONTEXT_SHIFT)
\n–context-shift \u5728\u65e0\u9650\u6587\u672c\u751f\u6210\u4e2d\u542f\u7528\u4e0a\u4e0b\u6587\u5207\u6362\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_CONTEXT_SHIFT)
\n-r, –reverse-prompt PROMPT \u5728 PROMPT \u5904\u505c\u6b62\u751f\u6210\uff0c\u8fd4\u56de\u63a7\u5236\u6743\u81f3\u4ea4\u4e92\u6a21\u5f0f
\n-sp, –special \u542f\u7528\u7279\u6b8a\u6807\u8bb0\u8f93\u51fa\uff08\u9ed8\u8ba4\uff1afalse\uff09
\n–no-warmup \u8df3\u8fc7\u4f7f\u7528\u7a7a\u8fd0\u884c\u9884\u70ed\u6a21\u578b
\n–spm-infill \u4f7f\u7528\u540e\u7f00\/\u524d\u7f00\/\u4e2d\u95f4\u6a21\u5f0f\u8fdb\u884c\u586b\u5145\uff08\u800c\u975e\u524d\u7f00\/\u540e\u7f00\/\u4e2d\u95f4\u6a21\u5f0f\uff09\uff0c\u67d0\u4e9b\u6a21\u578b\u66f4\u504f\u597d\u6b64\u65b9\u5f0f\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n–pooling {none,mean,cls,last,rank} \u5d4c\u5165\u5411\u91cf\u7684\u6c60\u5316\u7c7b\u578b\uff0c\u672a\u6307\u5b9a\u65f6\u4f7f\u7528\u6a21\u578b\u9ed8\u8ba4\u503c
\n(env: LLAMA_ARG_POOLING)
\n-cb, –cont-batching \u542f\u7528\u8fde\u7eed\u6279\u5904\u7406\uff08\u53c8\u79f0\u52a8\u6001\u6279\u5904\u7406\uff09\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_CONT_BATCHING)
\n-nocb, –no-cont-batching \u7981\u7528\u8fde\u7eed\u6279\u5904\u7406
\n(env: LLAMA_ARG_NO_CONT_BATCHING)
\n–mmproj FILE \u591a\u6a21\u6001\u6295\u5f71\u5668\u6587\u4ef6\u8def\u5f84\u3002\u53c2\u89c1 tools\/mtmd\/README.md
\n\u6ce8\u610f\uff1a\u5982\u679c\u4f7f\u7528 -hf\uff0c\u53ef\u7701\u7565\u6b64\u53c2\u6570
\n(env: LLAMA_ARG_MMPROJ)
\n–mmproj-url URL \u591a\u6a21\u6001\u6295\u5f71\u5668\u6587\u4ef6\u7684 URL\u3002\u53c2\u89c1 tools\/mtmd\/README.md
\n(env: LLAMA_ARG_MMPROJ_URL)
\n–no-mmproj \u663e\u5f0f\u7981\u7528\u591a\u6a21\u6001\u6295\u5f71\u5668\uff0c\u9002\u7528\u4e8e\u4f7f\u7528 -hf \u7684\u60c5\u51b5
\n(env: LLAMA_ARG_NO_MMPROJ)
\n–no-mmproj-offload \u4e0d\u5c06\u591a\u6a21\u6001\u6295\u5f71\u5668\u5378\u8f7d\u5230 GPU
\n(env: LLAMA_ARG_NO_MMPROJ_OFFLOAD)
\n–image-min-tokens N \u6bcf\u5f20\u56fe\u50cf\u53ef\u5360\u7528\u7684\u6700\u5c0f\u6807\u8bb0\u6570\uff0c\u4ec5\u7528\u4e8e\u5177\u6709\u52a8\u6001\u5206\u8fa8\u7387\u7684\u89c6\u89c9\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u8bfb\u53d6\uff09
\n(env: LLAMA_ARG_IMAGE_MIN_TOKENS)
\n–image-max-tokens N \u6bcf\u5f20\u56fe\u50cf\u53ef\u5360\u7528\u7684\u6700\u5927\u6807\u8bb0\u6570\uff0c\u4ec5\u7528\u4e8e\u5177\u6709\u52a8\u6001\u5206\u8fa8\u7387\u7684\u89c6\u89c9\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u8bfb\u53d6\uff09
\n(env: LLAMA_ARG_IMAGE_MAX_TOKENS)
\n–override-tensor-draft, -otd <tensor name pattern>=<buffer type>,…
\n\u4e3a\u8349\u7a3f\u6a21\u578b\u8986\u76d6\u5f20\u91cf\u7f13\u51b2\u533a\u7c7b\u578b
\n–cpu-moe-draft, -cmoed \u5c06\u8349\u7a3f\u6a21\u578b\u7684\u6240\u6709\u4e13\u5bb6\u6df7\u5408\uff08MoE\uff09\u6743\u91cd\u4fdd\u7559\u5728 CPU \u4e2d
\n(env: LLAMA_ARG_CPU_MOE_DRAFT)
\n–n-cpu-moe-draft, -ncmoed N \u5c06\u8349\u7a3f\u6a21\u578b\u524d N \u5c42\u7684 MoE \u6743\u91cd\u4fdd\u7559\u5728 CPU \u4e2d
\n(env: LLAMA_ARG_N_CPU_MOE_DRAFT)
\n-a, –alias STRING \u4e3a\u6a21\u578b\u540d\u79f0\u8bbe\u7f6e\u522b\u540d\uff08\u4f9b REST API \u4f7f\u7528\uff09
\n(env: LLAMA_ARG_ALIAS)
\n–host HOST \u76d1\u542c\u7684 IP \u5730\u5740\uff0c\u6216\u82e5\u5730\u5740\u4ee5 .sock \u7ed3\u5c3e\u5219\u7ed1\u5b9a\u5230 UNIX \u5957\u63a5\u5b57\uff08\u9ed8\u8ba4\uff1a127.0.0.1\uff09
\n(env: LLAMA_ARG_HOST)
\n–port PORT \u76d1\u542c\u7aef\u53e3\uff08\u9ed8\u8ba4\uff1a8080\uff09
\n(env: LLAMA_ARG_PORT)
\n–path PATH \u63d0\u4f9b\u9759\u6001\u6587\u4ef6\u7684\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\uff09
\n(env: LLAMA_ARG_STATIC_PATH)
\n–api-prefix PREFIX \u670d\u52a1\u5668\u670d\u52a1\u7684\u524d\u7f00\u8def\u5f84\uff08\u4e0d\u5e26\u5c3e\u90e8\u659c\u6760\uff09
\n\uff08\u9ed8\u8ba4\uff1a\uff09
\n(env: LLAMA_ARG_API_PREFIX)
\n–no-webui \u7981\u7528 Web UI\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_NO_WEBUI)
\n–embedding, –embeddings \u4ec5\u652f\u6301\u5d4c\u5165\u7528\u4f8b\uff1b\u4ec5\u4e0e\u4e13\u7528\u5d4c\u5165\u6a21\u578b\u4e00\u8d77\u4f7f\u7528\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_EMBEDDINGS)
\n–reranking, –rerank \u542f\u7528\u670d\u52a1\u5668\u7684\u91cd\u6392\u5e8f\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_RERANKING)
\n–api-key KEY \u8ba4\u8bc1\u4f7f\u7528\u7684 API \u5bc6\u94a5\uff08\u9ed8\u8ba4\uff1a\u65e0\uff09
\n(env: LLAMA_API_KEY)
\n–api-key-file FNAME \u5305\u542b API \u5bc6\u94a5\u7684\u6587\u4ef6\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u65e0\uff09
\n–ssl-key-file FNAME PEM \u7f16\u7801 SSL \u79c1\u94a5\u6587\u4ef6\u8def\u5f84
\n(env: LLAMA_ARG_SSL_KEY_FILE)
\n–ssl-cert-file FNAME PEM \u7f16\u7801 SSL \u8bc1\u4e66\u6587\u4ef6\u8def\u5f84
\n(env: LLAMA_ARG_SSL_CERT_FILE)
\n–chat-template-kwargs STRING \u4e3a JSON \u6a21\u677f\u89e3\u6790\u5668\u8bbe\u7f6e\u9644\u52a0\u53c2\u6570
\n(env: LLAMA_CHAT_TEMPLATE_KWARGS)
\n-to, –timeout N \u670d\u52a1\u5668\u8bfb\/\u5199\u8d85\u65f6\u65f6\u95f4\uff08\u79d2\uff09\uff08\u9ed8\u8ba4\uff1a600\uff09
\n(env: LLAMA_ARG_TIMEOUT)
\n–threads-http N \u5904\u7406 HTTP \u8bf7\u6c42\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a-1\uff09
\n(env: LLAMA_ARG_THREADS_HTTP)
\n–cache-reuse N \u5c1d\u8bd5\u901a\u8fc7 KV \u5207\u6362\u4ece\u7f13\u5b58\u4e2d\u91cd\u7528\u7684\u6700\u5c0f\u5757\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0\uff09
\n[(\u5361\u7247)](https:\/\/ggml.ai\/f0.png)
\n(env: LLAMA_ARG_CACHE_REUSE)
\n–metrics \u542f\u7528 Prometheus \u517c\u5bb9\u7684\u6307\u6807\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_ENDPOINT_METRICS)
\n–props \u542f\u7528\u901a\u8fc7 POST \/props \u66f4\u6539\u5168\u5c40\u5c5e\u6027\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_ENDPOINT_PROPS)
\n–slots \u542f\u7528\u63d2\u69fd\u76d1\u63a7\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_ENDPOINT_SLOTS)
\n–no-slots \u7981\u7528\u63d2\u69fd\u76d1\u63a7\u7aef\u70b9
\n(env: LLAMA_ARG_NO_ENDPOINT_SLOTS)
\n–slot-save-path PATH \u4fdd\u5b58\u63d2\u69fd KV \u7f13\u5b58\u7684\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n–media-path PATH \u7528\u4e8e\u52a0\u8f7d\u672c\u5730\u5a92\u4f53\u6587\u4ef6\u7684\u76ee\u5f55\uff1b\u6587\u4ef6\u53ef\u901a\u8fc7 file:\/\/ URL \u4f7f\u7528\u76f8\u5bf9\u8def\u5f84\u8bbf\u95ee\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n–models-dir PATH \u8def\u7531\u5668\u670d\u52a1\u5668\u5305\u542b\u6a21\u578b\u7684\u76ee\u5f55\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_MODELS_DIR)
\n–models-preset PATH \u5305\u542b\u8def\u7531\u5668\u670d\u52a1\u5668\u6a21\u578b\u9884\u8bbe\u7684 INI \u6587\u4ef6\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_MODELS_PRESET)
\n–models-max N \u8def\u7531\u5668\u670d\u52a1\u5668\u540c\u65f6\u52a0\u8f7d\u6a21\u578b\u7684\u6700\u5927\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a4\uff0c0 = \u65e0\u9650\u5236\uff09
\n(env: LLAMA_ARG_MODELS_MAX)
\n–no-models-autoload \u7981\u7528\u6a21\u578b\u7684\u81ea\u52a8\u52a0\u8f7d\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_NO_MODELS_AUTOLOAD)
\n–jinja \u4f7f\u7528 Jinja \u6a21\u677f\u8fdb\u884c\u804a\u5929\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09
\n(env: LLAMA_ARG_JINJA)
\n–no-jinja \u7981\u7528 Jinja \u6a21\u677f\u8fdb\u884c\u804a\u5929\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n(env: LLAMA_ARG_NO_JINJA)
\n–reasoning-format FORMAT \u63a7\u5236\u662f\u5426\u5141\u8bb8\u63d0\u53d6\u54cd\u5e94\u4e2d\u7684\u601d\u8003\u6807\u8bb0\u4ee5\u53ca\u8fd4\u56de\u683c\u5f0f\uff1b\u53ef\u9009\u503c\uff1a
\n– none: \u4fdd\u7559 thoughts \u672a\u89e3\u6790\u5728 `message.content` \u4e2d
\n– deepseek: \u5c06 thoughts \u653e\u5165 `message.reasoning_content` \u4e2d
\n– deepseek-legacy: \u5728 `message.content` \u4e2d\u4fdd\u7559 `think` \u6807\u8bb0\uff0c\u540c\u65f6\u586b\u5145 `message.reasoning_content`
\n\uff08\u9ed8\u8ba4\uff1aauto\uff09
\n(env: LLAMA_ARG_THINK)
\n–reasoning-budget N \u63a7\u5236\u5141\u8bb8\u7684\u601d\u8003\u91cf\uff1b\u76ee\u524d\u4ec5\u652f\u6301\uff1a-1 \u8868\u793a\u65e0\u9650\u5236\uff0c0 \u8868\u793a\u7981\u7528\u601d\u8003\uff08\u9ed8\u8ba4\uff1a-1\uff09
\n(env: LLAMA_ARG_THINK_BUDGET)
\n–chat-template JINJA_TEMPLATE \u8bbe\u7f6e\u81ea\u5b9a\u4e49 Jinja \u804a\u5929\u6a21\u677f\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u5143\u6570\u636e\u4e2d\u83b7\u53d6\u6a21\u677f\uff09
\n\u5982\u679c\u6307\u5b9a\u4e86\u540e\u7f00\/\u524d\u7f00\uff0c\u6a21\u677f\u5c06\u88ab\u7981\u7528
\n\u4ec5\u63a5\u53d7\u5e38\u7528\u6a21\u677f\uff08\u9664\u975e\u5728\u8be5\u6807\u5fd7\u524d\u8bbe\u7f6e\u4e86 –jinja\uff09\uff1a
\n\u5185\u7f6e\u6a21\u677f\u5217\u8868\uff1a
\nbailing, bailing-think, bailing2, chatglm3, chatglm4, chatml,
\ncommand-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3,
\ngemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense,
\nhunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos,
\nllama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1,
\nmistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch,
\nopenchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss,
\nsmolvlm, vicuna, vicuna-orca, yandex, zephyr
\n(env: LLAMA_ARG_CHAT_TEMPLATE)
\n–chat-template-file JINJA_TEMPLATE_FILE
\n\u8bbe\u7f6e\u81ea\u5b9a\u4e49 Jinja \u804a\u5929\u6a21\u677f\u6587\u4ef6\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u5143\u6570\u636e\u4e2d\u83b7\u53d6\u6a21\u677f\uff09
\n\u5982\u679c\u6307\u5b9a\u4e86\u540e\u7f00\/\u524d\u7f00\uff0c\u6a21\u677f\u5c06\u88ab\u7981\u7528
\n\u4ec5\u63a5\u53d7\u5e38\u7528\u6a21\u677f\uff08\u9664\u975e\u5728\u8be5\u6807\u5fd7\u524d\u8bbe\u7f6e\u4e86 –jinja\uff09\uff1a
\n\u5185\u7f6e\u6a21\u677f\u5217\u8868\uff1a
\nbailing, bailing-think, bailing2, chatglm3, chatglm4, chatml,
\ncommand-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3,
\ngemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense,
\nhunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos,
\nllama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1,
\nmistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch,
\nopenchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss,
\nsmolvlm, vicuna, vicuna-orca, yandex, zephyr
\n(env: LLAMA_ARG_CHAT_TEMPLATE_FILE)
\n–no-prefill-assistant \u5982\u679c\u6700\u540e\u4e00\u6761\u6d88\u606f\u662f\u52a9\u7406\u6d88\u606f\uff0c\u662f\u5426\u9884\u586b\u5145\u52a9\u7406\u7684\u54cd\u5e94\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\u9884\u586b\u5145\uff09
\n\u5f53\u8bbe\u7f6e\u6b64\u6807\u5fd7\u65f6\uff0c\u5982\u679c\u6700\u540e\u4e00\u6761\u6d88\u606f\u662f\u52a9\u7406\u6d88\u606f\uff0c\u5c06\u89c6\u4e3a\u5b8c\u6574\u6d88\u606f\u5e76\u7981\u7528\u9884\u586b\u5145<\/p>\n

(env: LLAMA_ARG_NO_PREFILL_ASSISTANT)
\n-sps, –slot-prompt-similarity SIMILARITY
\n\u8bf7\u6c42\u7684\u63d0\u793a\u4e0e\u63d2\u69fd\u63d0\u793a\u5339\u914d\u7684\u76f8\u4f3c\u5ea6\u9608\u503c\uff0c\u4ee5\u4f7f\u7528\u8be5\u63d2\u69fd\uff08\u9ed8\u8ba4\uff1a0.10\uff0c0.0 = \u7981\u7528\uff09
\n–lora-init-without-apply \u52a0\u8f7d LoRA \u9002\u914d\u5668\u4f46\u4e0d\u7acb\u5373\u5e94\u7528\uff08\u7a0d\u540e\u901a\u8fc7 POST \/lora-adapters \u5e94\u7528\uff09\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09
\n-td, –threads-draft N \u751f\u6210\u671f\u95f4\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a\u4e0e –threads \u76f8\u540c\uff09
\n-tbd, –threads-batch-draft N \u6279\u5904\u7406\u548c\u63d0\u793a\u5904\u7406\u671f\u95f4\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a\u4e0e –threads-draft \u76f8\u540c\uff09
\n–draft-max, –draft, –draft-n N \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684\u8349\u7a3f\u6807\u8bb0\u6570\uff08\u9ed8\u8ba4\uff1a16\uff09
\n(env: LLAMA_ARG_DRAFT_MAX)
\n–draft-min, –draft-n-min N \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684\u6700\u5c0f\u8349\u7a3f\u6807\u8bb0\u6570\uff08\u9ed8\u8ba4\uff1a0\uff09
\n(env: LLAMA_ARG_DRAFT_MIN)
\n–draft-p-min P \u6700\u5c0f\u63a8\u6d4b\u89e3\u7801\u6982\u7387\uff08\u8d2a\u5a6a\uff09\uff08\u9ed8\u8ba4\uff1a0.8\uff09
\n(env: LLAMA_ARG_DRAFT_P_MIN)
\n-cd, –ctx-size-draft N \u8349\u7a3f\u6a21\u578b\u7684\u63d0\u793a\u4e0a\u4e0b\u6587\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0\uff0c0 = \u4ece\u6a21\u578b\u52a0\u8f7d\uff09
\n(env: LLAMA_ARG_CTX_SIZE_DRAFT)
\n-devd, –device-draft <dev1,dev2,..> \u7528\u4e8e\u5378\u8f7d\u8349\u7a3f\u6a21\u578b\u7684\u8bbe\u5907\u5217\u8868\uff08\u9017\u53f7\u5206\u9694\uff09\uff08none = \u4e0d\u5378\u8f7d\uff09
\n\u4f7f\u7528 –list-devices \u67e5\u770b\u53ef\u7528\u8bbe\u5907\u5217\u8868
\n-ngld, –gpu-layers-draft, –n-gpu-layers-draft N
\n\u4e3a\u8349\u7a3f\u6a21\u578b\u5b58\u50a8\u5728 VRAM \u4e2d\u7684\u5c42\u6570
\n(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT)
\n-md, –model-draft FNAME \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684\u8349\u7a3f\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n(env: LLAMA_ARG_MODEL_DRAFT)
\n–spec-replace TARGET DRAFT \u5982\u679c\u8349\u7a3f\u6a21\u578b\u548c\u4e3b\u6a21\u578b\u4e0d\u517c\u5bb9\uff0c\u5c06 TARGET \u4e2d\u7684\u5b57\u7b26\u4e32\u8f6c\u6362\u4e3a DRAFT
\n-mv, –model-vocoder FNAME \u7528\u4e8e\u97f3\u9891\u751f\u6210\u7684\u58f0\u7801\u5668\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09
\n–tts-use-guide-tokens \u4f7f\u7528\u5f15\u5bfc\u6807\u8bb0\u63d0\u9ad8 TTS \u8bcd\u6c47\u53ec\u56de\u7387
\n–embd-gemma-default \u4f7f\u7528\u9ed8\u8ba4 EmbeddingGemma \u6a21\u578b\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-1.5b-default \u4f7f\u7528\u9ed8\u8ba4 Qwen 2.5 Coder 1.5B\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-3b-default \u4f7f\u7528\u9ed8\u8ba4 Qwen 2.5 Coder 3B\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-7b-default \u4f7f\u7528\u9ed8\u8ba4 Qwen 2.5 Coder 7B\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-7b-spec \u4f7f\u7528 Qwen 2.5 Coder 7B + 0.5B \u8349\u7a3f\u8fdb\u884c\u63a8\u6d4b\u89e3\u7801\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-14b-spec \u4f7f\u7528 Qwen 2.5 Coder 14B + 0.5B \u8349\u7a3f\u8fdb\u884c\u63a8\u6d4b\u89e3\u7801\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–fim-qwen-30b-default \u4f7f\u7528\u9ed8\u8ba4 Qwen 3 Coder 30B A3B Instruct\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–gpt-oss-20b-default \u4f7f\u7528 gpt-oss-20b\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–gpt-oss-120b-default \u4f7f\u7528 gpt-oss-120b\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–vision-gemma-4b-default \u4f7f\u7528 Gemma 3 4B QAT\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09
\n–vision-gemma-12b-default \u4f7f\u7528 Gemma 3 12B QAT\uff08\u6ce8\u610f\uff1a\u53ef\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<\/p>\n","protected":false},"excerpt":{"rendered":"

—– \u5e38\u89c1\u53c2\u6570 —– -h, –help, &#8 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[21],"tags":[],"class_list":["post-1237","post","type-post","status-publish","format-standard","hentry","category-21"],"_links":{"self":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1237","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/comments?post=1237"}],"version-history":[{"count":1,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1237\/revisions"}],"predecessor-version":[{"id":1238,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1237\/revisions\/1238"}],"wp:attachment":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/media?parent=1237"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/categories?post=1237"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/tags?post=1237"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}