{"id":1973,"date":"2026-03-07T19:47:16","date_gmt":"2026-03-07T11:47:16","guid":{"rendered":"https:\/\/www.o6s.net\/?p=1973"},"modified":"2026-03-07T19:47:16","modified_gmt":"2026-03-07T11:47:16","slug":"llama-cpp-%e5%b7%a5%e5%85%b7%e4%b8%ad%e5%b8%b8%e8%a7%81%e5%8f%82%e6%95%b0%ef%bc%88llama-b8227-bin-win-cpu-x64%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.o6s.net\/index.php\/2026\/03\/07\/llama-cpp-%e5%b7%a5%e5%85%b7%e4%b8%ad%e5%b8%b8%e8%a7%81%e5%8f%82%e6%95%b0%ef%bc%88llama-b8227-bin-win-cpu-x64%ef%bc%89\/","title":{"rendered":"llama.cpp \u5de5\u5177\u4e2d\u5e38\u89c1\u53c2\u6570\uff08llama-b8227-bin-win-cpu-x64\uff09"},"content":{"rendered":"<p>\u4ee5\u4e0b\u662f `llama.cpp` \u5de5\u5177\u4e2d\u5e38\u89c1\u53c2\u6570\uff08Common Params\uff09\u3001\u91c7\u6837\u53c2\u6570\uff08Sampling Params\uff09\u53ca\u7279\u5b9a\u793a\u4f8b\u53c2\u6570\uff08Example-specific Params\uff09\u7684\u4e2d\u6587\u7ffb\u8bd1\u5bf9\u7167\u8868\u3002<\/p>\n<p>&#8212;<\/p>\n<p>### \u4e00\u3001\u901a\u7528\u53c2\u6570 (&#8212;&#8211; common params &#8212;&#8211;)<\/p>\n<p>* `-h, &#8211;help, &#8211;usage`: \u6253\u5370\u4f7f\u7528\u8bf4\u660e\u5e76\u9000\u51fa<br \/>\n* `&#8211;version`: \u663e\u793a\u7248\u672c\u53f7\u53ca\u6784\u5efa\u4fe1\u606f<br \/>\n* `&#8211;license`: \u663e\u793a\u6e90\u4ee3\u7801\u8bb8\u53ef\u8bc1\u53ca\u4f9d\u8d56\u9879<br \/>\n* `-cl, &#8211;cache-list`: \u663e\u793a\u7f13\u5b58\u4e2d\u7684\u6a21\u578b\u5217\u8868<br \/>\n* `&#8211;completion-bash`: \u6253\u5370\u53ef\u7528\u4e8e bash \u7684\u8865\u5168\u811a\u672c\u6e90\u6587\u4ef6<br \/>\n* `&#8211;verbose-prompt`: \u5728\u751f\u6210\u524d\u6253\u5370\u8be6\u7ec6\u7684\u63d0\u793a\uff08\u9ed8\u8ba4\uff1afalse\uff09<br \/>\n* `-t, &#8211;threads N`: \u751f\u6210\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684 CPU \u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a-1\uff0c\u8868\u793a\u81ea\u52a8\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_THREADS)*<br \/>\n* `&#8211;tb, &#8211;threads-batch N`: \u6279\u6b21\u5904\u7406\u548c\u63d0\u793a\u5904\u7406\u4e2d\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a\u4e0e `&#8211;threads` \u76f8\u540c\uff09<br \/>\n* `-C, &#8211;cpu-mask M`: CPU \u4eb2\u548c\u6027\u63a9\u7801\uff1a\u4efb\u610f\u957f\u5ea6\u7684\u5341\u516d\u8fdb\u5236\u6570\u3002\u4e0e `&#8211;cpu-range` \u4e92\u8865<br \/>\n* *(\u9ed8\u8ba4\uff1a\u7a7a\u5b57\u7b26\u4e32)*<br \/>\n* `-Cr, &#8211;cpu-range lo-hi`: CPU \u4eb2\u548c\u6027\u8303\u56f4\u3002\u4e0e `&#8211;cpu-mask` \u4e92\u8865<br \/>\n* `&#8211;cpu-strict &lt;0|1&gt;`: \u4f7f\u7528\u4e25\u683c\u7684 CPU \u653e\u7f6e\u7b56\u7565\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* `&#8211;prio N`: \u8bbe\u7f6e\u8fdb\u7a0b\/\u7ebf\u7a0b\u4f18\u5148\u7ea7\uff1alow(-1), normal(0), medium(1), high(2), realtime(3)\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* `&#8211;poll &lt;0&#8230;100&gt;`: \u4f7f\u7528\u8f6e\u8be2\u7ea7\u522b\u7b49\u5f85\u5de5\u4f5c\uff080 &#8211; \u65e0\u8f6e\u8be2\uff0c\u9ed8\u8ba4\uff1a50\uff09<br \/>\n* `-Cb, &#8211;cpu-mask-batch M`: CPU \u4eb2\u548c\u6027\u63a9\u7801\uff08\u6279\u6b21\uff09\uff1a\u4efb\u610f\u957f\u5ea6\u7684\u5341\u516d\u8fdb\u5236\u6570\u3002\u4e0e `&#8211;cpu-range-batch` \u4e92\u8865<br \/>\n* *(\u9ed8\u8ba4\uff1a\u4e0e `&#8211;cpu-mask` \u76f8\u540c)*<br \/>\n* `-Crb, &#8211;cpu-range-batch lo-hi`: CPU \u4eb2\u548c\u6027\u8303\u56f4\uff08\u6279\u6b21\uff09\u3002\u4e0e `&#8211;cpu-mask-batch` \u4e92\u8865<br \/>\n* `&#8211;cpu-strict-batch &lt;0|1&gt;`: \u4f7f\u7528\u4e25\u683c\u7684 CPU \u653e\u7f6e\u7b56\u7565\uff08\u6279\u6b21\uff09\uff08\u9ed8\u8ba4\uff1a\u4e0e `&#8211;cpu-strict` \u76f8\u540c\uff09<br \/>\n* `&#8211;prio-batch N`: \u8bbe\u7f6e\u8fdb\u7a0b\/\u7ebf\u7a0b\u4f18\u5148\u7ea7\uff08\u6279\u6b21\uff09\uff1a0-normal, 1-medium, 2-high, 3-realtime\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* `&#8211;poll-batch &lt;0|1&gt;`: \u4f7f\u7528\u8f6e\u8be2\u7b49\u5f85\u5de5\u4f5c\uff08\u6279\u6b21\uff09\uff08\u9ed8\u8ba4\uff1a\u4e0e `&#8211;poll` \u76f8\u540c\uff09<br \/>\n* `-c, &#8211;ctx-size N`: \u63d0\u793a\u4e0a\u4e0b\u6587\u7684\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0\uff0c0 \u8868\u793a\u4ece\u6a21\u578b\u52a0\u8f7d\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CTX_SIZE)*<br \/>\n* `-n, &#8211;predict, &#8211;n-predict N`: \u9884\u6d4b\u7684 token \u6570\u91cf\uff08\u9ed8\u8ba4\uff1a-1\uff0c-1 \u8868\u793a\u65e0\u9650\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_PREDICT)*<br \/>\n* `-b, &#8211;batch-size N`: \u903b\u8f91\u6700\u5927\u6279\u6b21\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a2048\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_BATCH)*<br \/>\n* `-ub, &#8211;ubatch-size N`: \u7269\u7406\u6700\u5927\u6279\u6b21\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a512\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_UBATCH)*<br \/>\n* `&#8211;keep N`: \u4ece\u521d\u59cb\u63d0\u793a\u4e2d\u4fdd\u7559\u7684 token \u6570\u91cf\uff08\u9ed8\u8ba4\uff1a0\uff0c-1 \u8868\u793a\u5168\u90e8\uff09<br \/>\n* `&#8211;swa-full`: \u4f7f\u7528\u5168\u5c3a\u5bf8 SWA \u7f13\u5b58\uff08\u9ed8\u8ba4\uff1afalse\uff09<br \/>\n* *(\u66f4\u591a\u4fe1\u606f\uff1a[PR #13194](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/13194#issuecomment-2868343055))*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SWA_FULL)*<br \/>\n* `-fa, &#8211;flash-attn [on|off|auto]`: \u8bbe\u7f6e Flash Attention \u7684\u4f7f\u7528\uff08&#8217;on&#8217;, &#8216;off&#8217;, \u6216 &#8216;auto&#8217;\uff0c\u9ed8\u8ba4\uff1a&#8217;auto&#8217;\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_FLASH_ATTN)*<br \/>\n* `&#8211;perf, &#8211;no-perf`: \u662f\u5426\u542f\u7528\u5185\u90e8 libllama \u6027\u80fd\u8ba1\u65f6\uff08\u9ed8\u8ba4\uff1afalse\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_PERF)*<br \/>\n* `-e, &#8211;escape, &#8211;no-escape`: \u662f\u5426\u5904\u7406\u8f6c\u4e49\u5e8f\u5217\uff08\\n, \\r, \\t, \\&#8217;, \\&#8221;, \\\\\uff09\uff08\u9ed8\u8ba4\uff1atrue\uff09<br \/>\n* `&#8211;rope-scaling {none,linear,yarn}`: RoPE \u9891\u7387\u7f29\u653e\u65b9\u6cd5\uff0c\u9664\u975e\u6a21\u578b\u6307\u5b9a\uff0c\u5426\u5219\u9ed8\u8ba4\u4e3a linear<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_SCALING_TYPE)*<br \/>\n* `&#8211;rope-scale N`: RoPE \u4e0a\u4e0b\u6587\u7f29\u653e\u56e0\u5b50\uff0c\u6309 N \u500d\u6269\u5c55\u4e0a\u4e0b\u6587<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_SCALE)*<br \/>\n* `&#8211;rope-freq-base N`: RoPE \u57fa\u7840\u9891\u7387\uff0c\u7528\u4e8e NTK \u611f\u77e5\u7f29\u653e\uff08\u9ed8\u8ba4\uff1a\u4ece\u6a21\u578b\u52a0\u8f7d\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_FREQ_BASE)*<br \/>\n* `&#8211;rope-freq-scale N`: RoPE \u9891\u7387\u7f29\u653e\u56e0\u5b50\uff0c\u6309 1\/N \u500d\u6269\u5c55\u4e0a\u4e0b\u6587<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ROPE_FREQ_SCALE)*<br \/>\n* `&#8211;yarn-orig-ctx N`: YaRN: \u6a21\u578b\u7684\u539f\u59cb\u4e0a\u4e0b\u6587\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0 = \u6a21\u578b\u8bad\u7ec3\u4e0a\u4e0b\u6587\u5927\u5c0f\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_ORIG_CTX)*<br \/>\n* `&#8211;yarn-ext-factor N`: YaRN: \u5916\u63a8\u6df7\u5408\u56e0\u5b50\uff08\u9ed8\u8ba4\uff1a-1.00\uff0c0.0 = \u5b8c\u5168\u63d2\u503c\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_EXT_FACTOR)*<br \/>\n* `&#8211;yarn-attn-factor N`: YaRN: \u7f29\u653e sqrt(t) \u6216\u6ce8\u610f\u529b\u5e45\u5ea6\uff08\u9ed8\u8ba4\uff1a-1.00\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_ATTN_FACTOR)*<br \/>\n* `&#8211;yarn-beta-slow N`: YaRN: \u9ad8\u4fee\u6b63\u7ef4\u5ea6\u6216 alpha\uff08\u9ed8\u8ba4\uff1a-1.00\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_BETA_SLOW)*<br \/>\n* `&#8211;yarn-beta-fast N`: YaRN: \u4f4e\u4fee\u6b63\u7ef4\u5ea6\u6216 beta\uff08\u9ed8\u8ba4\uff1a-1.00\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_YARN_BETA_FAST)*<br \/>\n* `-kvo, &#8211;kv-offload, -nkvo, &#8211;no-kv-offload`: \u662f\u5426\u542f\u7528 KV \u7f13\u5b58\u5378\u8f7d\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_KV_OFFLOAD)*<br \/>\n* `&#8211;repack, -nr, &#8211;no-repack`: \u662f\u5426\u542f\u7528\u6743\u91cd\u91cd\u65b0\u6253\u5305\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_REPACK)*<br \/>\n* `&#8211;no-host`: \u7ed5\u8fc7\u4e3b\u673a\u7f13\u51b2\u533a\uff0c\u5141\u8bb8\u4f7f\u7528\u989d\u5916\u7684\u7f13\u51b2\u533a<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NO_HOST)*<br \/>\n* `-ctk, &#8211;cache-type-k TYPE`: K \u7684 KV \u7f13\u5b58\u6570\u636e\u7c7b\u578b<br \/>\n* *\u5141\u8bb8\u7684\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1*<br \/>\n* *(\u9ed8\u8ba4\uff1af16)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_K)*<br \/>\n* `-ctv, &#8211;cache-type-v TYPE`: V \u7684 KV \u7f13\u5b58\u6570\u636e\u7c7b\u578b<br \/>\n* *\u5141\u8bb8\u7684\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1*<br \/>\n* *(\u9ed8\u8ba4\uff1af16)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_V)*<br \/>\n* `-dt, &#8211;defrag-thold N`: KV \u7f13\u5b58\u788e\u7247\u6574\u7406\u9608\u503c\uff08\u5df2\u5f03\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DEFRAG_THOLD)*<br \/>\n* `&#8211;rpc SERVERS`: \u9017\u53f7\u5206\u9694\u7684 RPC \u670d\u52a1\u5668\u5217\u8868 (host:port)<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_RPC)*<br \/>\n* `&#8211;mlock`: \u5f3a\u5236\u7cfb\u7edf\u4fdd\u7559\u6a21\u578b\u5728 RAM \u4e2d\uff0c\u800c\u4e0d\u662f\u4ea4\u6362\u6216\u538b\u7f29<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MLOCK)*<br \/>\n* `&#8211;mmap, &#8211;no-mmap`: \u662f\u5426\u5185\u5b58\u6620\u5c04\u6a21\u578b\u3002\uff08\u5982\u679c\u7981\u7528 mmap\uff0c\u52a0\u8f7d\u901f\u5ea6\u8f83\u6162\uff0c\u4f46\u5982\u679c\u4e0d\u4f7f\u7528 mlock \u53ef\u80fd\u4f1a\u51cf\u5c11\u9875\u9762\u6362\u51fa\uff09\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MMAP)*<br \/>\n* `-dio, &#8211;direct-io, -ndio, &#8211;no-direct-io`: \u5982\u679c\u53ef\u7528\u5219\u4f7f\u7528 DirectIO\u3002\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DIO)*<br \/>\n* `&#8211;numa TYPE`: \u5c1d\u8bd5\u4f18\u5316\u4ee5\u652f\u6301\u67d0\u4e9b NUMA \u7cfb\u7edf<br \/>\n* `distribute`: \u5728\u6240\u6709\u8282\u70b9\u4e0a\u5747\u5300\u5206\u5e03\u6267\u884c<br \/>\n* `isolate`: \u4ec5\u5728\u6267\u884c\u542f\u52a8\u6240\u5728\u7684\u8282\u70b9 CPU \u4e0a\u751f\u6210\u7ebf\u7a0b<br \/>\n* `numactl`: \u4f7f\u7528 numactl \u63d0\u4f9b\u7684 CPU \u6620\u5c04<br \/>\n* *\u5982\u679c\u4e4b\u524d\u672a\u4f7f\u7528\u6b64\u9009\u9879\u8fd0\u884c\uff0c\u5efa\u8bae\u5728\u8fd0\u884c\u6b64\u9009\u9879\u524d\u91ca\u653e\u7cfb\u7edf\u9875\u9762\u7f13\u5b58*<br \/>\n* *(\u53c2\u89c1\uff1ahttps:\/\/github.com\/ggml-org\/llama.cpp\/issues\/1437)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_NUMA)*<br \/>\n* `-dev, &#8211;device &lt;dev1,dev2,..&gt;`: \u7528\u4e8e\u5378\u8f7d\u7684\u9017\u53f7\u5206\u9694\u7684\u8bbe\u5907\u5217\u8868\uff08none = \u4e0d\u5378\u8f7d\uff09<br \/>\n* *\u4f7f\u7528 `&#8211;list-devices` \u67e5\u770b\u53ef\u7528\u8bbe\u5907\u5217\u8868*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DEVICE)*<br \/>\n* `&#8211;list-devices`: \u6253\u5370\u53ef\u7528\u8bbe\u5907\u5217\u8868\u5e76\u9000\u51fa<br \/>\n* `-ot, &#8211;override-tensor &lt;tensor name pattern&gt;=&lt;buffer type&gt;,&#8230;`: \u8986\u76d6\u5f20\u91cf\u7f13\u51b2\u533a\u7c7b\u578b<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_OVERRIDE_TENSOR)*<br \/>\n* `-cmoe, &#8211;cpu-moe`: \u5c06\u6240\u6709\u6df7\u5408\u4e13\u5bb6 (MoE) \u6743\u91cd\u4fdd\u7559\u5728 CPU<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CPU_MOE)*<br \/>\n* `-ncmoe, &#8211;n-cpu-moe N`: \u5c06\u524d N \u5c42\u7684\u6df7\u5408\u4e13\u5bb6 (MoE) \u6743\u91cd\u4fdd\u7559\u5728 CPU<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_CPU_MOE)*<br \/>\n* `-ngl, &#8211;gpu-layers, &#8211;n-gpu-layers N`: \u5b58\u50a8\u5728 VRAM \u4e2d\u7684\u6700\u5927\u5c42\u6570\uff0c\u53ef\u4ee5\u662f\u786e\u5207\u6570\u5b57\u3001&#8217;auto&#8217; \u6216 &#8216;all&#8217;\uff08\u9ed8\u8ba4\uff1aauto\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_GPU_LAYERS)*<br \/>\n* `-sm, &#8211;split-mode {none,layer,row}`: \u5982\u4f55\u5728\u591a\u4e2a GPU \u4e4b\u95f4\u5206\u5272\u6a21\u578b<br \/>\n* `none`: \u4ec5\u4f7f\u7528\u4e00\u4e2a GPU<br \/>\n* `layer` (\u9ed8\u8ba4): \u6309\u5c42\u548c KV \u5728 GPU \u95f4\u5206\u5272<br \/>\n* `row`: \u6309\u884c\u5728 GPU \u95f4\u5206\u5272<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SPLIT_MODE)*<br \/>\n* `-ts, &#8211;tensor-split N0,N1,N2,&#8230;`: \u5206\u914d\u7ed9\u6bcf\u4e2a GPU \u7684\u6a21\u578b\u6bd4\u4f8b\uff0c\u9017\u53f7\u5206\u9694\u7684\u6bd4\u4f8b\u5217\u8868\uff0c\u4f8b\u5982 3,1<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_TENSOR_SPLIT)*<br \/>\n* `-mg, &#8211;main-gpu INDEX`: \u7528\u4e8e\u6a21\u578b\u7684 GPU\uff08\u5f53 split-mode = none \u65f6\uff09\uff0c\u6216\u7528\u4e8e\u4e2d\u95f4\u7ed3\u679c\u548c KV\uff08\u5f53 split-mode = row \u65f6\uff09\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MAIN_GPU)*<br \/>\n* `-fit, &#8211;fit [on|off]`: \u662f\u5426\u8c03\u6574\u672a\u8bbe\u7f6e\u7684\u53c2\u6570\u4ee5\u9002\u5e94\u8bbe\u5907\u5185\u5b58\uff08&#8217;on&#8217; \u6216 &#8216;off&#8217;\uff0c\u9ed8\u8ba4\uff1a&#8217;on&#8217;\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_FIT)*<br \/>\n* `-fitt, &#8211;fit-target MiB0,MiB1,MiB2,&#8230;`: \u4e3a `&#8211;fit` \u8bbe\u7f6e\u7684\u6bcf\u4e2a\u8bbe\u5907\u7684\u76ee\u6807\u4f59\u91cf\uff0c\u9017\u53f7\u5206\u9694\u7684\u503c\u5217\u8868\uff0c\u5355\u4e2a\u503c\u5e7f\u64ad\u5230\u6240\u6709\u8bbe\u5907\uff0c\u9ed8\u8ba4\uff1a1024<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_FIT_TARGET)*<br \/>\n* `-fitc, &#8211;fit-ctx N`: `&#8211;fit` \u9009\u9879\u53ef\u4ee5\u8bbe\u7f6e\u7684\u6700\u4f4e ctx \u5927\u5c0f\uff0c\u9ed8\u8ba4\uff1a4096<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_FIT_CTX)*<br \/>\n* `&#8211;check-tensors`: \u68c0\u67e5\u6a21\u578b\u5f20\u91cf\u6570\u636e\u662f\u5426\u6709\u65e0\u6548\u503c\uff08\u9ed8\u8ba4\uff1afalse\uff09<br \/>\n* `&#8211;override-kv KEY=TYPE:VALUE,&#8230;`: \u9ad8\u7ea7\u9009\u9879\uff0c\u901a\u8fc7 key \u8986\u76d6\u6a21\u578b\u5143\u6570\u636e\u3002\u8981\u6307\u5b9a\u591a\u4e2a\u8986\u76d6\u9879\uff0c\u8bf7\u4f7f\u7528\u9017\u53f7\u5206\u9694\u7684\u503c\u3002<br \/>\n* *\u7c7b\u578b\uff1aint, float, bool, str\u3002\u793a\u4f8b\uff1a&#8211;override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false*<br \/>\n* `&#8211;op-offload, &#8211;no-op-offload`: \u662f\u5426\u5c06\u4e3b\u673a\u5f20\u91cf\u64cd\u4f5c\u5378\u8f7d\u5230\u8bbe\u5907\uff08\u9ed8\u8ba4\uff1atrue\uff09<br \/>\n* `-lora FNAME`: LoRA \u9002\u914d\u5668\u8def\u5f84\uff08\u4f7f\u7528\u9017\u53f7\u5206\u9694\u7684\u503c\u52a0\u8f7d\u591a\u4e2a\u9002\u914d\u5668\uff09<br \/>\n* `-lora-scaled FNAME:SCALE,&#8230;`: \u5e26\u7528\u6237\u5b9a\u4e49\u7f29\u653e\u7684 LoRA \u9002\u914d\u5668\u8def\u5f84\uff08\u683c\u5f0f\uff1aFNAME:SCALE,&#8230;\uff09<br \/>\n* *\u6ce8\u610f\uff1a\u4f7f\u7528\u9017\u53f7\u5206\u9694\u7684\u503c*<br \/>\n* `&#8211;control-vector FNAME`: \u6dfb\u52a0\u63a7\u5236\u5411\u91cf<br \/>\n* *\u6ce8\u610f\uff1a\u4f7f\u7528\u9017\u53f7\u5206\u9694\u7684\u503c\u6dfb\u52a0\u591a\u4e2a\u63a7\u5236\u5411\u91cf*<br \/>\n* `&#8211;control-vector-scaled FNAME:SCALE,&#8230;`: \u6dfb\u52a0\u5e26\u7528\u6237\u5b9a\u4e49\u7f29\u653e SCALE \u7684\u63a7\u5236\u5411\u91cf<br \/>\n* *\u6ce8\u610f\uff1a\u4f7f\u7528\u9017\u53f7\u5206\u9694\u7684\u503c\uff08\u683c\u5f0f\uff1aFNAME:SCALE,&#8230;\uff09*<br \/>\n* `&#8211;control-vector-layer-range START END`: \u5e94\u7528\u4e8e\u63a7\u5236\u5411\u91cf\u7684\u5c42\u8303\u56f4\uff0c\u8d77\u6b62\u5305\u542b<br \/>\n* `-m, &#8211;model FNAME`: \u8981\u52a0\u8f7d\u7684\u6a21\u578b\u8def\u5f84<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODEL)*<br \/>\n* `-mu, &#8211;model-url MODEL_URL`: \u6a21\u578b\u4e0b\u8f7d URL\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODEL_URL)*<br \/>\n* `-dr, &#8211;docker-repo [&lt;repo&gt;\/]&lt;model&gt;[:quant]`: Docker Hub \u6a21\u578b\u4ed3\u5e93\u3002repo \u53ef\u9009\uff0c\u9ed8\u8ba4\u4e3a ai\/\u3002quant \u53ef\u9009\uff0c\u9ed8\u8ba4\u4e3a :latest\u3002<br \/>\n* *\u793a\u4f8b\uff1agemma3*<br \/>\n* *(\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DOCKER_REPO)*<br \/>\n* `-hf, -hfr, &#8211;hf-repo &lt;user&gt;\/&lt;model&gt;[:quant]`: Hugging Face \u6a21\u578b\u4ed3\u5e93\uff1bquant \u53ef\u9009\uff0c\u4e0d\u533a\u5206\u5927\u5c0f\u5199\uff0c\u9ed8\u8ba4 Q4_K_M\uff0c\u5982\u679c Q4_K_M \u4e0d\u5b58\u5728\u5219\u56de\u9000\u5230\u4ed3\u5e93\u4e2d\u7684\u7b2c\u4e00\u4e2a\u6587\u4ef6\u3002<br \/>\n* *\u5982\u679c\u53ef\u7528\uff0c\u4e5f\u4f1a\u81ea\u52a8\u4e0b\u8f7d mmproj\u3002\u8981\u7981\u7528\uff0c\u6dfb\u52a0 &#8211;no-mmproj*<br \/>\n* *\u793a\u4f8b\uff1aunsloth\/phi-4-GGUF:q4_k_m*<br \/>\n* *(\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_REPO)*<br \/>\n* `-hfd, -hfrd, &#8211;hf-repo-draft &lt;user&gt;\/&lt;model&gt;[:quant]`: \u4e0e `&#8211;hf-repo` \u76f8\u540c\uff0c\u4f46\u7528\u4e8e draft \u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HFD_REPO)*<br \/>\n* `-hff, &#8211;hf-file FILE`: Hugging Face \u6a21\u578b\u6587\u4ef6\u3002\u5982\u679c\u6307\u5b9a\uff0c\u5b83\u5c06\u8986\u76d6 `&#8211;hf-repo` \u4e2d\u7684 quant\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_FILE)*<br \/>\n* `-hfv, -hfrv, &#8211;hf-repo-v &lt;user&gt;\/&lt;model&gt;[:quant]`: \u7528\u4e8e vocoder \u6a21\u578b\u7684 Hugging Face \u6a21\u578b\u4ed3\u5e93\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_REPO_V)*<br \/>\n* `-hffv, &#8211;hf-file-v FILE`: \u7528\u4e8e vocoder \u6a21\u578b\u7684 Hugging Face \u6a21\u578b\u6587\u4ef6\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HF_FILE_V)*<br \/>\n* `-hft, &#8211;hf-token TOKEN`: Hugging Face \u8bbf\u95ee\u4ee4\u724c\uff08\u9ed8\u8ba4\uff1aHF_TOKEN \u73af\u5883\u53d8\u91cf\u7684\u503c\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aHF_TOKEN)*<br \/>\n* `&#8211;log-disable`: \u7981\u7528\u65e5\u5fd7<br \/>\n* `&#8211;log-file FNAME`: \u65e5\u5fd7\u8f93\u51fa\u5230\u6587\u4ef6<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_FILE)*<br \/>\n* `&#8211;log-colors [on|off|auto]`: \u8bbe\u7f6e\u5f69\u8272\u65e5\u5fd7\uff08&#8217;on&#8217;, &#8216;off&#8217;, \u6216 &#8216;auto&#8217;\uff0c\u9ed8\u8ba4\uff1a&#8217;auto&#8217;\uff09<br \/>\n* *&#8217;auto&#8217; \u5728\u8f93\u51fa\u5230\u7ec8\u7aef\u65f6\u542f\u7528\u989c\u8272*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_COLORS)*<br \/>\n* `-v, &#8211;verbose, &#8211;log-verbose`: \u5c06\u8be6\u7ec6\u7a0b\u5ea6\u7ea7\u522b\u8bbe\u7f6e\u4e3a\u65e0\u7a77\u5927\uff08\u5373\u8bb0\u5f55\u6240\u6709\u6d88\u606f\uff0c\u9002\u7528\u4e8e\u8c03\u8bd5\uff09<br \/>\n* `&#8211;offline`: \u79bb\u7ebf\u6a21\u5f0f\uff1a\u5f3a\u5236\u4f7f\u7528\u7f13\u5b58\uff0c\u9632\u6b62\u7f51\u7edc\u8bbf\u95ee<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_OFFLINE)*<br \/>\n* `-lv, &#8211;verbosity, &#8211;log-verbosity N`: \u8bbe\u7f6e\u8be6\u7ec6\u7a0b\u5ea6\u9608\u503c\u3002\u8be6\u7ec6\u7a0b\u5ea6\u66f4\u9ad8\u7684\u6d88\u606f\u5c06\u88ab\u5ffd\u7565\u3002\u503c\uff1a<br \/>\n* `0`: \u901a\u7528\u8f93\u51fa<br \/>\n* `1`: \u9519\u8bef<br \/>\n* `2`: \u8b66\u544a<br \/>\n* `3`: \u4fe1\u606f<br \/>\n* `4`: \u8c03\u8bd5<br \/>\n* *(\u9ed8\u8ba4\uff1a3)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_VERBOSITY)*<br \/>\n* `&#8211;log-prefix`: \u5728\u65e5\u5fd7\u6d88\u606f\u4e2d\u542f\u7528\u524d\u7f00<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_PREFIX)*<br \/>\n* `&#8211;log-timestamps`: \u5728\u65e5\u5fd7\u6d88\u606f\u4e2d\u542f\u7528\u65f6\u95f4\u6233<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_LOG_TIMESTAMPS)*<br \/>\n* `-ctkd, &#8211;cache-type-k-draft TYPE`: draft \u6a21\u578b K \u7684 KV \u7f13\u5b58\u6570\u636e\u7c7b\u578b<br \/>\n* *\u5141\u8bb8\u7684\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1*<br \/>\n* *(\u9ed8\u8ba4\uff1af16)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_K_DRAFT)*<br \/>\n* `-ctvd, &#8211;cache-type-v-draft TYPE`: draft \u6a21\u578b V \u7684 KV \u7f13\u5b58\u6570\u636e\u7c7b\u578b<br \/>\n* *\u5141\u8bb8\u7684\u503c\uff1af32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1*<br \/>\n* *(\u9ed8\u8ba4\uff1af16)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_TYPE_V_DRAFT)*<\/p>\n<p>&#8212;<\/p>\n<p>### \u4e8c\u3001\u91c7\u6837\u53c2\u6570 (&#8212;&#8211; sampling params &#8212;&#8211;)<\/p>\n<p>* `&#8211;samplers SAMPLERS`: \u6309\u987a\u5e8f\u7528\u4e8e\u751f\u6210\u7684\u91c7\u6837\u5668\uff0c\u7531 &#8216;;&#8217; \u5206\u9694<br \/>\n* *(\u9ed8\u8ba4\uff1apenalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature)*<br \/>\n* `-s, &#8211;seed SEED`: RNG \u79cd\u5b50\uff08\u9ed8\u8ba4\uff1a-1\uff0c-1 \u8868\u793a\u4f7f\u7528\u968f\u673a\u79cd\u5b50\uff09<br \/>\n* `&#8211;sampler-seq, &#8211;sampling-seq SEQUENCE`: \u7b80\u5316\u7248\u7684\u91c7\u6837\u5668\u5e8f\u5217\uff08\u9ed8\u8ba4\uff1aedskypmxt\uff09<br \/>\n* `&#8211;ignore-eos`: \u5ffd\u7565\u7ed3\u675f\u6d41 token \u5e76\u7ee7\u7eed\u751f\u6210\uff08\u9690\u542b `&#8211;logit-bias EOS-inf`\uff09<br \/>\n* `&#8211;temp, &#8211;temperature N`: \u6e29\u5ea6\uff08\u9ed8\u8ba4\uff1a0.80\uff09<br \/>\n* `&#8211;top-k N`: top-k \u91c7\u6837\uff08\u9ed8\u8ba4\uff1a40\uff0c0 = \u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_TOP_K)*<br \/>\n* `&#8211;top-p N`: top-p \u91c7\u6837\uff08\u9ed8\u8ba4\uff1a0.95\uff0c1.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;min-p N`: min-p \u91c7\u6837\uff08\u9ed8\u8ba4\uff1a0.05\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;top-nsigma, &#8211;top-n-sigma N`: top-n-sigma \u91c7\u6837\uff08\u9ed8\u8ba4\uff1a-1.00\uff0c-1.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;xtc-probability N`: xtc \u6982\u7387\uff08\u9ed8\u8ba4\uff1a0.00\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;xtc-threshold N`: xtc \u9608\u503c\uff08\u9ed8\u8ba4\uff1a0.10\uff0c1.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;typical, &#8211;typical-p N`: \u5c40\u90e8\u5178\u578b\u91c7\u6837\uff0c\u53c2\u6570 p\uff08\u9ed8\u8ba4\uff1a1.00\uff0c1.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;repeat-last-n N`: \u8003\u8651\u7528\u4e8e\u60e9\u7f5a\u7684\u6700\u8fd1 n \u4e2a token\uff08\u9ed8\u8ba4\uff1a64\uff0c0 = \u7981\u7528\uff0c-1 = ctx_size\uff09<br \/>\n* `&#8211;repeat-penalty N`: \u60e9\u7f5a\u91cd\u590d\u7684 token \u5e8f\u5217\uff08\u9ed8\u8ba4\uff1a1.00\uff0c1.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;presence-penalty N`: \u91cd\u590d alpha \u5b58\u5728\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a0.00\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;frequency-penalty N`: \u91cd\u590d alpha \u9891\u7387\u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a0.00\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;dry-multiplier N`: \u8bbe\u7f6e DRY \u91c7\u6837\u4e58\u6570\uff08\u9ed8\u8ba4\uff1a0.00\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;dry-base N`: \u8bbe\u7f6e DRY \u91c7\u6837\u57fa\u7840\u503c\uff08\u9ed8\u8ba4\uff1a1.75\uff09<br \/>\n* `&#8211;dry-allowed-length N`: \u8bbe\u7f6e DRY \u91c7\u6837\u5141\u8bb8\u7684\u957f\u5ea6\uff08\u9ed8\u8ba4\uff1a2\uff09<br \/>\n* `&#8211;dry-penalty-last-n N`: \u8bbe\u7f6e\u6700\u8fd1 n \u4e2a token \u7684 DRY \u60e9\u7f5a\uff08\u9ed8\u8ba4\uff1a-1\uff0c0 = \u7981\u7528\uff0c-1 = \u4e0a\u4e0b\u6587\u5927\u5c0f\uff09<br \/>\n* `&#8211;dry-sequence-breaker STRING`: \u4e3a DRY \u91c7\u6837\u6dfb\u52a0\u5e8f\u5217\u5206\u9694\u7b26\uff0c\u6e05\u9664\u9ed8\u8ba4\u5206\u9694\u7b26\uff08&#8217;\\n&#8217;, &#8216;:&#8217;, &#8216;&#8221;&#8216;, &#8216;*&#8217;\uff09\uff1b\u4f7f\u7528 &#8220;none&#8221; \u4e0d\u4f7f\u7528\u4efb\u4f55\u5206\u9694\u7b26<br \/>\n* `&#8211;adaptive-target N`: adaptive-p: \u9009\u62e9\u63a5\u8fd1\u6b64\u6982\u7387\u7684 token\uff08\u6709\u6548\u8303\u56f4 0.0 \u5230 1.0\uff1b\u8d1f\u6570 = \u7981\u7528\uff09\uff08\u9ed8\u8ba4\uff1a-1.00\uff09<br \/>\n* *(\u66f4\u591a\u4fe1\u606f\uff1a[PR #17927](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/17927))*<br \/>\n* `&#8211;adaptive-decay N`: adaptive-p: \u76ee\u6807\u9002\u5e94\u7684\u65f6\u95f4\u8870\u51cf\u7387\u3002\u8f83\u4f4e\u7684\u503c\u66f4\u7075\u654f\uff0c\u8f83\u9ad8\u7684\u503c\u66f4\u7a33\u5b9a\u3002\uff08\u6709\u6548\u8303\u56f4 0.0 \u5230 0.99\uff09\uff08\u9ed8\u8ba4\uff1a0.90\uff09<br \/>\n* `&#8211;dynatemp-range N`: \u52a8\u6001\u6e29\u5ea6\u8303\u56f4\uff08\u9ed8\u8ba4\uff1a0.00\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;dynatemp-exp N`: \u52a8\u6001\u6e29\u5ea6\u6307\u6570\uff08\u9ed8\u8ba4\uff1a1.00\uff09<br \/>\n* `&#8211;mirostat N`: \u4f7f\u7528 Mirostat \u91c7\u6837\u3002\u5982\u679c\u540c\u65f6\u4f7f\u7528 Top K\u3001Nucleus \u548c Locally Typical \u91c7\u6837\u5668\uff0c\u5219\u4f1a\u88ab\u5ffd\u7565\u3002\uff08\u9ed8\u8ba4\uff1a0\uff0c0 = \u7981\u7528\uff0c1 = Mirostat, 2 = Mirostat 2.0\uff09<br \/>\n* `&#8211;mirostat-lr N`: Mirostat \u5b66\u4e60\u7387\uff0c\u53c2\u6570 eta\uff08\u9ed8\u8ba4\uff1a0.10\uff09<br \/>\n* `&#8211;mirostat-ent N`: Mirostat \u76ee\u6807\u71b5\uff0c\u53c2\u6570 tau\uff08\u9ed8\u8ba4\uff1a5.00\uff09<br \/>\n* `-l, &#8211;logit-bias TOKEN_ID(+\/-)BIAS`: \u4fee\u6539 token \u51fa\u73b0\u5728\u5b8c\u6210\u4e2d\u7684\u53ef\u80fd\u6027\uff0c\u4f8b\u5982 `&#8211;logit-bias 15043+1` \u589e\u52a0 token &#8216; Hello&#8217; \u7684\u53ef\u80fd\u6027\uff0c\u6216 `&#8211;logit-bias 15043-1` \u51cf\u5c11\u53ef\u80fd\u6027<br \/>\n* `&#8211;grammar GRAMMAR`: BNF \u98ce\u683c\u7684\u8bed\u6cd5\u4ee5\u9650\u5236\u751f\u6210\uff08\u53c2\u89c1 grammars\/ \u76ee\u5f55\u4e2d\u7684\u793a\u4f8b\uff09\uff08\u9ed8\u8ba4\uff1a&#8221;\uff09<br \/>\n* `&#8211;grammar-file FNAME`: \u8bfb\u53d6\u8bed\u6cd5\u7684\u6587\u4ef6<br \/>\n* `-j, &#8211;json-schema SCHEMA`: JSON \u6a21\u5f0f\u4ee5\u9650\u5236\u751f\u6210\uff08https:\/\/json-schema.org\/\uff09\uff0c\u4f8b\u5982 `{}` \u8868\u793a\u4efb\u4f55 JSON \u5bf9\u8c61<br \/>\n* *\u5bf9\u4e8e\u5e26\u6709\u5916\u90e8 $refs \u7684\u6a21\u5f0f\uff0c\u8bf7\u4f7f\u7528 `&#8211;grammar` + `example\/json_schema_to_grammar.py`*<br \/>\n* `-jf, &#8211;json-schema-file FILE`: \u5305\u542b\u7528\u4e8e\u9650\u5236\u751f\u6210\u7684 JSON \u6a21\u5f0f\u7684\u6587\u4ef6\uff08https:\/\/json-schema.org\/\uff09\uff0c\u4f8b\u5982 `{}` \u8868\u793a\u4efb\u4f55 JSON \u5bf9\u8c61<br \/>\n* *\u5bf9\u4e8e\u5e26\u6709\u5916\u90e8 $refs \u7684\u6a21\u5f0f\uff0c\u8bf7\u4f7f\u7528 `&#8211;grammar` + `example\/json_schema_to_grammar.py`*<br \/>\n* `-bs, &#8211;backend-sampling`: \u542f\u7528\u540e\u7aef\u91c7\u6837\uff08\u5b9e\u9a8c\u6027\uff09\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_BACKEND_SAMPLING)*<\/p>\n<p>&#8212;<\/p>\n<p>### \u4e09\u3001\u7279\u5b9a\u793a\u4f8b\u53c2\u6570 (&#8212;&#8211; example-specific params &#8212;&#8211;)<\/p>\n<p>* `-lcs, &#8211;lookup-cache-static FNAME`: \u7528\u4e8e\u67e5\u627e\u89e3\u7801\u7684\u9759\u6001\u67e5\u627e\u7f13\u5b58\u8def\u5f84\uff08\u4e0d\u7531\u751f\u6210\u66f4\u65b0\uff09<br \/>\n* `-lcd, &#8211;lookup-cache-dynamic FNAME`: \u7528\u4e8e\u67e5\u627e\u89e3\u7801\u7684\u52a8\u6001\u67e5\u627e\u7f13\u5b58\u8def\u5f84\uff08\u7531\u751f\u6210\u66f4\u65b0\uff09<br \/>\n* `-ctxcp, &#8211;ctx-checkpoints, &#8211;swa-checkpoints N`: \u6bcf\u4e2a\u69fd\u4f4d\u521b\u5efa\u7684\u6700\u5927\u4e0a\u4e0b\u6587\u68c0\u67e5\u70b9\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a32\uff09<br \/>\n* *(\u66f4\u591a\u4fe1\u606f\uff1a[PR #15293](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/15293))*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CTX_CHECKPOINTS)*<br \/>\n* `-cpent, &#8211;checkpoint-every-n-tokens N`: \u5728\u9884\u586b\u5145\uff08\u5904\u7406\uff09\u8fc7\u7a0b\u4e2d\u6bcf n \u4e2a token \u521b\u5efa\u4e00\u4e2a\u68c0\u67e5\u70b9\uff0c-1 \u7981\u7528\uff08\u9ed8\u8ba4\uff1a8192\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CHECKPOINT_EVERY_NT)*<br \/>\n* `-cram, &#8211;cache-ram N`: \u8bbe\u7f6e\u7f13\u5b58\u7684\u6700\u5927\u5927\u5c0f\uff08MiB\uff09\uff08\u9ed8\u8ba4\uff1a8192\uff0c-1 &#8211; \u65e0\u9650\u5236\uff0c0 &#8211; \u7981\u7528\uff09<br \/>\n* *(\u66f4\u591a\u4fe1\u606f\uff1a[PR #16391](https:\/\/github.com\/ggml-org\/llama.cpp\/pull\/16391))*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_RAM)*<br \/>\n* `-kvu, &#8211;kv-unified, -no-kvu, &#8211;no-kv-unified`: \u4f7f\u7528\u8de8\u6240\u6709\u5e8f\u5217\u5171\u4eab\u7684\u7edf\u4e00 KV \u7f13\u51b2\u533a\uff08\u9ed8\u8ba4\uff1a\u5982\u679c\u69fd\u4f4d\u6570\u91cf\u4e3a auto \u5219\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_KV_UNIFIED)*<br \/>\n* `&#8211;context-shift, &#8211;no-context-shift`: \u662f\u5426\u5728\u65e0\u9650\u6587\u672c\u751f\u6210\u4e2d\u4f7f\u7528\u4e0a\u4e0b\u6587\u504f\u79fb\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CONTEXT_SHIFT)*<br \/>\n* `-r, &#8211;reverse-prompt PROMPT`: \u5728 PROMPT \u5904\u6682\u505c\u751f\u6210\uff0c\u5728\u4ea4\u4e92\u6a21\u5f0f\u4e0b\u8fd4\u56de\u63a7\u5236\u6743<br \/>\n* `-sp, &#8211;special`: \u542f\u7528\u7279\u6b8a token \u8f93\u51fa\uff08\u9ed8\u8ba4\uff1afalse\uff09<br \/>\n* `&#8211;warmup, &#8211;no-warmup`: \u662f\u5426\u6267\u884c\u7a7a\u8fd0\u884c\u7684\u9884\u70ed\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* `&#8211;spm-infill`: \u4f7f\u7528 Suffix\/Prefix\/Middle \u6a21\u5f0f\u8fdb\u884c\u586b\u5145\uff08\u800c\u4e0d\u662f Prefix\/Suffix\/Middle\uff09\uff0c\u56e0\u4e3a\u67d0\u4e9b\u6a21\u578b\u66f4\u559c\u6b22\u8fd9\u79cd\u65b9\u5f0f\u3002\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* `&#8211;pooling {none,mean,cls,last,rank}`: \u5d4c\u5165\u7684\u6c60\u5316\u7c7b\u578b\uff0c\u5982\u679c\u672a\u6307\u5b9a\u5219\u4f7f\u7528\u6a21\u578b\u9ed8\u8ba4\u503c<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_POOLING)*<br \/>\n* `-np, &#8211;parallel N`: \u670d\u52a1\u5668\u69fd\u4f4d\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a-1\uff0c-1 = auto\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_PARALLEL)*<br \/>\n* `-cb, &#8211;cont-batching, -nocb, &#8211;no-cont-batching`: \u662f\u5426\u542f\u7528\u8fde\u7eed\u6279\u5904\u7406\uff08\u53c8\u79f0\u52a8\u6001\u6279\u5904\u7406\uff09\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CONT_BATCHING)*<br \/>\n* `-mm, &#8211;mmproj FILE`: \u591a\u6a21\u6001\u6295\u5f71\u6587\u4ef6\u7684\u8def\u5f84\u3002\u53c2\u89c1 tools\/mtmd\/README.md<br \/>\n* *\u6ce8\u610f\uff1a\u5982\u679c\u4f7f\u7528 -hf\uff0c\u6b64\u53c2\u6570\u53ef\u4ee5\u7701\u7565*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MMPROJ)*<br \/>\n* `-mmu, &#8211;mmproj-url URL`: \u591a\u6a21\u6001\u6295\u5f71\u6587\u4ef6\u7684 URL\u3002\u53c2\u89c1 tools\/mtmd\/README.md<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MMPROJ_URL)*<br \/>\n* `&#8211;mmproj-auto, &#8211;no-mmproj, &#8211;no-mmproj-auto`: \u662f\u5426\u4f7f\u7528\u591a\u6a21\u6001\u6295\u5f71\u6587\u4ef6\uff08\u5982\u679c\u53ef\u7528\uff09\uff0c\u5728\u4f7f\u7528 -hf \u65f6\u5f88\u6709\u7528\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MMPROJ_AUTO)*<br \/>\n* `&#8211;mmproj-offload, &#8211;no-mmproj-offload`: \u662f\u5426\u542f\u7528\u591a\u6a21\u6001\u6295\u5f71\u7684 GPU \u5378\u8f7d\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MMPROJ_OFFLOAD)*<br \/>\n* `&#8211;image-min-tokens N`: \u6bcf\u4e2a\u56fe\u50cf\u7684\u6700\u5c0f token \u6570\u91cf\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<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_IMAGE_MIN_TOKENS)*<br \/>\n* `&#8211;image-max-tokens N`: \u6bcf\u4e2a\u56fe\u50cf\u7684\u6700\u5927 token \u6570\u91cf\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<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_IMAGE_MAX_TOKENS)*<br \/>\n* `-otd, &#8211;override-tensor-draft &lt;tensor name pattern&gt;=&lt;buffer type&gt;,&#8230;`: \u8986\u76d6 draft \u6a21\u578b\u7684\u5f20\u91cf\u7f13\u51b2\u533a\u7c7b\u578b<br \/>\n* `-cmoed, &#8211;cpu-moe-draft`: \u5c06\u6240\u6709\u6df7\u5408\u4e13\u5bb6 (MoE) \u6743\u91cd\u4fdd\u7559\u5728 CPU\uff08\u7528\u4e8e draft \u6a21\u578b\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CPU_MOE_DRAFT)*<br \/>\n* `-ncmoed, &#8211;n-cpu-moe-draft N`: \u5c06\u524d N \u5c42\u7684\u6df7\u5408\u4e13\u5bb6 (MoE) \u6743\u91cd\u4fdd\u7559\u5728 CPU\uff08\u7528\u4e8e draft \u6a21\u578b\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_CPU_MOE_DRAFT)*<br \/>\n* `-a, &#8211;alias STRING`: \u8bbe\u7f6e\u6a21\u578b\u540d\u79f0\u522b\u540d\uff0c\u9017\u53f7\u5206\u9694\uff08\u4f9b API \u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ALIAS)*<br \/>\n* `&#8211;tags STRING`: \u8bbe\u7f6e\u6a21\u578b\u6807\u7b7e\uff0c\u9017\u53f7\u5206\u9694\uff08\u4fe1\u606f\u6027\uff0c\u4e0d\u7528\u4e8e\u8def\u7531\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_TAGS)*<br \/>\n* `&#8211;host HOST`: \u76d1\u542c\u7684 IP \u5730\u5740\uff0c\u6216\u4ee5 .sock \u7ed3\u5c3e\u7684\u5730\u5740\u7ed1\u5b9a\u5230 UNIX \u5957\u63a5\u5b57\uff08\u9ed8\u8ba4\uff1a127.0.0.1\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_HOST)*<br \/>\n* `&#8211;port PORT`: \u76d1\u542c\u7684\u7aef\u53e3\uff08\u9ed8\u8ba4\uff1a8080\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_PORT)*<br \/>\n* `&#8211;path PATH`: \u63d0\u4f9b\u9759\u6001\u6587\u4ef6\u7684\u670d\u52a1\u5668\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u7a7a\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_STATIC_PATH)*<br \/>\n* `&#8211;api-prefix PREFIX`: \u670d\u52a1\u5668\u63d0\u4f9b\u670d\u52a1\u7684\u6839\u8def\u5f84\uff08\u4e0d\u5e26\u5c3e\u968f\u659c\u6760\uff09<br \/>\n* *(\u9ed8\u8ba4\uff1a\u7a7a)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_API_PREFIX)*<br \/>\n* `&#8211;webui-config JSON`: \u63d0\u4f9b\u9ed8\u8ba4 WebUI \u8bbe\u7f6e\u7684 JSON\uff08\u8986\u76d6 WebUI \u9ed8\u8ba4\u503c\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_WEBUI_CONFIG)*<br \/>\n* `&#8211;webui-config-file PATH`: \u63d0\u4f9b\u9ed8\u8ba4 WebUI \u8bbe\u7f6e\u7684 JSON \u6587\u4ef6\uff08\u8986\u76d6 WebUI \u9ed8\u8ba4\u503c\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_WEBUI_CONFIG_FILE)*<br \/>\n* `&#8211;webui-mcp-proxy, &#8211;no-webui-mcp-proxy`: \u5b9e\u9a8c\u6027\uff1a\u662f\u5426\u542f\u7528 MCP CORS \u4ee3\u7406 &#8211; \u5728\u4e0d\u4fe1\u4efb\u7684\u73af\u5883\u4e2d\u4e0d\u8981\u542f\u7528\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_WEBUI_MCP_PROXY)*<br \/>\n* `&#8211;webui, &#8211;no-webui`: \u662f\u5426\u542f\u7528 Web UI\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_WEBUI)*<br \/>\n* `&#8211;embedding, &#8211;embeddings`: \u9650\u5236\u4ec5\u652f\u6301\u5d4c\u5165\u7528\u4f8b\uff1b\u4ec5\u7528\u4e8e\u4e13\u7528\u7684\u5d4c\u5165\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_EMBEDDINGS)*<br \/>\n* `&#8211;rerank, &#8211;reranking`: \u542f\u7528\u670d\u52a1\u5668\u4e0a\u7684\u91cd\u65b0\u6392\u5e8f\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_RERANKING)*<br \/>\n* `&#8211;api-key KEY`: \u7528\u4e8e\u8eab\u4efd\u9a8c\u8bc1\u7684 API \u5bc6\u94a5\uff0c\u53ef\u4ee5\u63d0\u4f9b\u591a\u4e2a\u5bc6\u94a5\uff0c\u7528\u9017\u53f7\u5206\u9694\uff08\u9ed8\u8ba4\uff1a\u65e0\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_API_KEY)*<br \/>\n* `&#8211;api-key-file FNAME`: \u5305\u542b API \u5bc6\u94a5\u7684\u6587\u4ef6\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u65e0\uff09<br \/>\n* `&#8211;ssl-key-file FNAME`: PEM \u7f16\u7801\u7684 SSL \u79c1\u94a5\u6587\u4ef6\u8def\u5f84<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SSL_KEY_FILE)*<br \/>\n* `&#8211;ssl-cert-file FNAME`: PEM \u7f16\u7801\u7684 SSL \u8bc1\u4e66\u6587\u4ef6\u8def\u5f84<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_SSL_CERT_FILE)*<br \/>\n* `&#8211;chat-template-kwargs STRING`: \u8bbe\u7f6e json \u6a21\u677f\u89e3\u6790\u5668\u7684\u9644\u52a0\u53c2\u6570\uff0c\u5fc5\u987b\u662f\u6709\u6548\u7684 JSON \u5bf9\u8c61\u5b57\u7b26\u4e32\uff0c\u4f8b\u5982 &#8216;{&#8220;key1&#8243;:&#8221;value1&#8243;,&#8221;key2&#8243;:&#8221;value2&#8221;}&#8217;<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_CHAT_TEMPLATE_KWARGS)*<br \/>\n* `-to, &#8211;timeout N`: \u670d\u52a1\u5668\u8bfb\u5199\u8d85\u65f6\uff08\u79d2\uff09\uff08\u9ed8\u8ba4\uff1a600\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_TIMEOUT)*<br \/>\n* `&#8211;threads-http N`: \u7528\u4e8e\u5904\u7406 HTTP \u8bf7\u6c42\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a-1\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_THREADS_HTTP)*<br \/>\n* `&#8211;cache-prompt, &#8211;no-cache-prompt`: \u662f\u5426\u542f\u7528\u63d0\u793a\u7f13\u5b58\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_PROMPT)*<br \/>\n* `&#8211;cache-reuse N`: \u5c1d\u8bd5\u901a\u8fc7 KV \u79fb\u4f4d\u4ece\u7f13\u5b58\u91cd\u7528\u5757\u7684\u6700\u5c0f\u5757\u5927\u5c0f\uff0c\u9700\u8981\u542f\u7528\u63d0\u793a\u7f13\u5b58\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* *(\u53c2\u89c1\uff1ahttps:\/\/ggml.ai\/f0.png)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CACHE_REUSE)*<br \/>\n* `&#8211;metrics`: \u542f\u7528\u517c\u5bb9 Prometheus \u7684\u6307\u6807\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ENDPOINT_METRICS)*<br \/>\n* `&#8211;props`: \u542f\u7528\u901a\u8fc7 POST \/props \u66f4\u6539\u5168\u5c40\u5c5e\u6027\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ENDPOINT_PROPS)*<br \/>\n* `&#8211;slots, &#8211;no-slots`: \u66b4\u9732\u69fd\u4f4d\u76d1\u63a7\u7aef\u70b9\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_ENDPOINT_SLOTS)*<br \/>\n* `&#8211;slot-save-path PATH`: \u4fdd\u5b58 slot kv \u7f13\u5b58\u7684\u8def\u5f84\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* `&#8211;media-path PATH`: \u52a0\u8f7d\u672c\u5730\u5a92\u4f53\u6587\u4ef6\u7684\u76ee\u5f55\uff1b\u53ef\u4ee5\u4f7f\u7528\u76f8\u5bf9\u8def\u5f84\u901a\u8fc7 file:\/\/ URL \u8bbf\u95ee\u6587\u4ef6\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* `&#8211;models-dir PATH`: \u5305\u542b\u8def\u7531\u5668\u670d\u52a1\u5668\u6a21\u578b\u7684\u76ee\u5f55\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODELS_DIR)*<br \/>\n* `&#8211;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<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODELS_PRESET)*<br \/>\n* `&#8211;models-max N`: \u5bf9\u4e8e\u8def\u7531\u5668\u670d\u52a1\u5668\uff0c\u540c\u65f6\u52a0\u8f7d\u7684\u6700\u5927\u6a21\u578b\u6570\u91cf\uff08\u9ed8\u8ba4\uff1a4\uff0c0 = \u65e0\u9650\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODELS_MAX)*<br \/>\n* `&#8211;models-autoload, &#8211;no-models-autoload`: \u5bf9\u4e8e\u8def\u7531\u5668\u670d\u52a1\u5668\uff0c\u662f\u5426\u81ea\u52a8\u52a0\u8f7d\u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODELS_AUTOLOAD)*<br \/>\n* `&#8211;jinja, &#8211;no-jinja`: \u662f\u5426\u4f7f\u7528 jinja \u6a21\u677f\u5f15\u64ce\u8fdb\u884c\u804a\u5929\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_JINJA)*<br \/>\n* `&#8211;reasoning-format FORMAT`: \u63a7\u5236\u662f\u5426\u5141\u8bb8\/\u4ece\u54cd\u5e94\u4e2d\u63d0\u53d6 thought \u6807\u7b7e\uff0c\u4ee5\u53ca\u8fd4\u56de\u683c\u5f0f\uff1b\u9009\u9879\u4e4b\u4e00\uff1a<br \/>\n* `none`: \u5728 `message.content` \u4e2d\u4fdd\u7559\u672a\u89e3\u6790\u7684\u601d\u60f3<br \/>\n* `deepseek`: \u5c06\u601d\u60f3\u653e\u5728 `message.reasoning_content` \u4e2d<br \/>\n* `deepseek-legacy`: \u5728 `message.content` \u4e2d\u4fdd\u7559 `&lt;think&gt;` \u6807\u7b7e\uff0c\u540c\u65f6\u586b\u5145 `message.reasoning_content`<br \/>\n* *(\u9ed8\u8ba4\uff1aauto)*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_THINK)*<br \/>\n* `&#8211;reasoning-budget N`: \u63a7\u5236\u5141\u8bb8\u7684\u201c\u601d\u8003\u201d\u6570\u91cf\uff1b\u76ee\u524d\u4ec5\u652f\u6301\uff1a-1 \u8868\u793a\u65e0\u9650\u5236\u601d\u8003\u9884\u7b97\uff0c\u6216 0 \u8868\u793a\u7981\u7528\u601d\u8003\uff08\u9ed8\u8ba4\uff1a-1\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_THINK_BUDGET)*<br \/>\n* `&#8211;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\u7684\u6a21\u677f\uff09<br \/>\n* *\u5982\u679c\u6307\u5b9a\u4e86 suffix\/prefix\uff0c\u6a21\u677f\u5c06\u88ab\u7981\u7528*<br \/>\n* *\u4ec5\u63a5\u53d7\u5e38\u7528\u6a21\u677f\uff08\u9664\u975e\u5728\u6b64\u6807\u5fd7\u4e4b\u524d\u8bbe\u7f6e\u4e86 &#8211;jinja\uff09\uff1a*<br \/>\n* *\u5185\u7f6e\u6a21\u677f\u5217\u8868\uff1abailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CHAT_TEMPLATE)*<br \/>\n* `&#8211;chat-template-file JINJA_TEMPLATE_FILE`: \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\u7684\u6a21\u677f\uff09<br \/>\n* *\u5982\u679c\u6307\u5b9a\u4e86 suffix\/prefix\uff0c\u6a21\u677f\u5c06\u88ab\u7981\u7528*<br \/>\n* *\u4ec5\u63a5\u53d7\u5e38\u7528\u6a21\u677f\uff08\u9664\u975e\u5728\u6b64\u6807\u5fd7\u4e4b\u524d\u8bbe\u7f6e\u4e86 &#8211;jinja\uff09\uff1a*<br \/>\n* *\u5185\u7f6e\u6a21\u677f\u5217\u8868\uff1abailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CHAT_TEMPLATE_FILE)*<br \/>\n* `&#8211;prefill-assistant, &#8211;no-prefill-assistant`: \u5982\u679c\u6700\u540e\u4e00\u6761\u6d88\u606f\u662f\u52a9\u624b\u6d88\u606f\uff0c\u662f\u5426\u9884\u586b\u5145\u52a9\u624b\u7684\u54cd\u5e94\uff08\u9ed8\u8ba4\uff1a\u542f\u7528\u9884\u586b\u5145\uff09<br \/>\n* *\u5f53\u8bbe\u7f6e\u6b64\u6807\u5fd7\u65f6\uff0c\u5982\u679c\u6700\u540e\u4e00\u6761\u6d88\u606f\u662f\u52a9\u624b\u6d88\u606f\uff0c\u5219\u5c06\u5176\u89c6\u4e3a\u5b8c\u6574\u6d88\u606f\u800c\u4e0d\u662f\u9884\u586b\u5145*<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_PREFILL_ASSISTANT)*<br \/>\n* `-sps, &#8211;slot-prompt-similarity SIMILARITY`: \u8bf7\u6c42\u7684\u63d0\u793a\u5fc5\u987b\u4e0e\u69fd\u4f4d\u7684\u63d0\u793a\u5339\u914d\u591a\u5c11\u624d\u80fd\u4f7f\u7528\u8be5\u69fd\u4f4d\uff08\u9ed8\u8ba4\uff1a0.10\uff0c0.0 = \u7981\u7528\uff09<br \/>\n* `&#8211;lora-init-without-apply`: \u52a0\u8f7d LoRA \u9002\u914d\u5668\u800c\u4e0d\u5e94\u7528\u5b83\u4eec\uff08\u7a0d\u540e\u901a\u8fc7 POST \/lora-adapters \u5e94\u7528\uff09\uff08\u9ed8\u8ba4\uff1a\u7981\u7528\uff09<br \/>\n* `&#8211;sleep-idle-seconds SECONDS`: \u7a7a\u95f2\u591a\u5c11\u79d2\u540e\u670d\u52a1\u5668\u5c06\u8fdb\u5165\u4f11\u7720\u72b6\u6001\uff08\u9ed8\u8ba4\uff1a-1; -1 = \u7981\u7528\uff09<br \/>\n* `-td, &#8211;threads-draft N`: \u751f\u6210\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a\u4e0e `&#8211;threads` \u76f8\u540c\uff09<br \/>\n* `-tbd, &#8211;threads-batch-draft N`: \u6279\u6b21\u5904\u7406\u548c\u63d0\u793a\u5904\u7406\u4e2d\u4f7f\u7528\u7684\u7ebf\u7a0b\u6570\uff08\u9ed8\u8ba4\uff1a\u4e0e `&#8211;threads-draft` \u76f8\u540c\uff09<br \/>\n* `&#8211;draft, &#8211;draft-n, &#8211;draft-max N`: \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684 draft token \u6570\u91cf\uff08\u9ed8\u8ba4\uff1a16\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DRAFT_MAX)*<br \/>\n* `&#8211;draft-min, &#8211;draft-n-min N`: \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684\u6700\u5c0f draft token \u6570\u91cf\uff08\u9ed8\u8ba4\uff1a0\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DRAFT_MIN)*<br \/>\n* `&#8211;draft-p-min P`: \u6700\u5c0f\u63a8\u6d4b\u89e3\u7801\u6982\u7387\uff08\u8d2a\u5a6a\uff09\uff08\u9ed8\u8ba4\uff1a0.75\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_DRAFT_P_MIN)*<br \/>\n* `-cd, &#8211;ctx-size-draft N`: draft \u6a21\u578b\u7684\u63d0\u793a\u4e0a\u4e0b\u6587\u5927\u5c0f\uff08\u9ed8\u8ba4\uff1a0\uff0c0 = \u4ece\u6a21\u578b\u52a0\u8f7d\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_CTX_SIZE_DRAFT)*<br \/>\n* `-devd, &#8211;device-draft &lt;dev1,dev2,..&gt;`: \u7528\u4e8e\u5378\u8f7d draft \u6a21\u578b\u7684\u9017\u53f7\u5206\u9694\u7684\u8bbe\u5907\u5217\u8868\uff08none = \u4e0d\u5378\u8f7d\uff09<br \/>\n* *\u4f7f\u7528 `&#8211;list-devices` \u67e5\u770b\u53ef\u7528\u8bbe\u5907\u5217\u8868*<br \/>\n* `-ngld, &#8211;gpu-layers-draft, &#8211;n-gpu-layers-draft N`: \u5b58\u50a8\u5728 VRAM \u4e2d\u7684\u6700\u5927 draft \u6a21\u578b\u5c42\u6570\uff0c\u53ef\u4ee5\u662f\u786e\u5207\u6570\u5b57\u3001&#8217;auto&#8217; \u6216 &#8216;all&#8217;\uff08\u9ed8\u8ba4\uff1aauto\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_N_GPU_LAYERS_DRAFT)*<br \/>\n* `-md, &#8211;model-draft FNAME`: \u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u7684 draft \u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* *(\u73af\u5883\u53d8\u91cf\uff1aLLAMA_ARG_MODEL_DRAFT)*<br \/>\n* `&#8211;spec-replace TARGET DRAFT`: \u5982\u679c draft \u6a21\u578b\u548c\u4e3b\u6a21\u578b\u4e0d\u517c\u5bb9\uff0c\u5219\u5c06 TARGET \u4e2d\u7684\u5b57\u7b26\u4e32\u7ffb\u8bd1\u4e3a DRAFT<br \/>\n* `&#8211;spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]`: \u5f53\u672a\u63d0\u4f9b draft \u6a21\u578b\u65f6\u8981\u4f7f\u7528\u7684\u63a8\u6d4b\u89e3\u7801\u7c7b\u578b<br \/>\n* *(\u9ed8\u8ba4\uff1anone)*<br \/>\n* `&#8211;spec-ngram-size-n N`: ngram-simple\/ngram-map \u63a8\u6d4b\u89e3\u7801\u7684 ngram \u5927\u5c0f N\uff0c\u67e5\u627e n-gram \u7684\u957f\u5ea6\uff08\u9ed8\u8ba4\uff1a12\uff09<br \/>\n* `&#8211;spec-ngram-size-m N`: ngram-simple\/ngram-map \u63a8\u6d4b\u89e3\u7801\u7684 ngram \u5927\u5c0f M\uff0cdraft m-gram \u7684\u957f\u5ea6\uff08\u9ed8\u8ba4\uff1a48\uff09<br \/>\n* `&#8211;spec-ngram-min-hits N`: ngram-map \u63a8\u6d4b\u89e3\u7801\u7684\u6700\u5c0f\u547d\u4e2d\u6b21\u6570\uff08\u9ed8\u8ba4\uff1a1\uff09<br \/>\n* `-mv, &#8211;model-vocoder FNAME`: \u7528\u4e8e\u97f3\u9891\u751f\u6210\u7684 vocoder \u6a21\u578b\uff08\u9ed8\u8ba4\uff1a\u672a\u4f7f\u7528\uff09<br \/>\n* `&#8211;tts-use-guide-tokens`: \u4f7f\u7528\u5f15\u5bfc token \u4ee5\u63d0\u9ad8 TTS \u5355\u8bcd\u53ec\u56de\u7387<br \/>\n* `&#8211;embd-gemma-default`: \u4f7f\u7528\u9ed8\u8ba4\u7684 EmbeddingGemma \u6a21\u578b\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-1.5b-default`: \u4f7f\u7528\u9ed8\u8ba4\u7684 Qwen 2.5 Coder 1.5B\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-3b-default`: \u4f7f\u7528\u9ed8\u8ba4\u7684 Qwen 2.5 Coder 3B\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-7b-default`: \u4f7f\u7528\u9ed8\u8ba4\u7684 Qwen 2.5 Coder 7B\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-7b-spec`: \u4f7f\u7528 Qwen 2.5 Coder 7B + 0.5B draft \u8fdb\u884c\u63a8\u6d4b\u89e3\u7801\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-14b-spec`: \u4f7f\u7528 Qwen 2.5 Coder 14B + 0.5B draft \u8fdb\u884c\u63a8\u6d4b\u89e3\u7801\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;fim-qwen-30b-default`: \u4f7f\u7528\u9ed8\u8ba4\u7684 Qwen 3 Coder 30B A3B Instruct\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;gpt-oss-20b-default`: \u4f7f\u7528 gpt-oss-20b\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;gpt-oss-120b-default`: \u4f7f\u7528 gpt-oss-120b\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;vision-gemma-4b-default`: \u4f7f\u7528 Gemma 3 4B QAT\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<br \/>\n* `&#8211;vision-gemma-12b-default`: \u4f7f\u7528 Gemma 3 12B QAT\uff08\u6ce8\u610f\uff1a\u53ef\u4ee5\u4ece\u4e92\u8054\u7f51\u4e0b\u8f7d\u6743\u91cd\uff09<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4ee5\u4e0b\u662f `llama.cpp` \u5de5\u5177\u4e2d\u5e38\u89c1\u53c2\u6570\uff08Common Params\uff09\u3001\u91c7\u6837\u53c2\u6570\uff08Sampling Pa [&hellip;]<\/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-1973","post","type-post","status-publish","format-standard","hentry","category-21"],"_links":{"self":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1973","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=1973"}],"version-history":[{"count":1,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1973\/revisions"}],"predecessor-version":[{"id":1974,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/posts\/1973\/revisions\/1974"}],"wp:attachment":[{"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/media?parent=1973"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/categories?post=1973"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.o6s.net\/index.php\/wp-json\/wp\/v2\/tags?post=1973"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}