[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M718NE3XGK23S8Z1CCTEXK19":3,"skill-same:01M718NE3XGK23S8Z1CCTEXK19":92},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":38,"linkOnly":39,"markdownContent":40,"markdownBody":41,"frontMatter":42,"refName":43,"commitSha":44,"skillDirectory":45,"forksCount":46,"repositoryHtmlUrl":47,"repositoryHomepageUrl":48,"ownerHtmlUrl":49,"ownerBlogUrl":48,"ownerDescription":50,"updatedAt":51,"install":52,"resources":65},"01M718NE3XGK23S8Z1CCTEXK19","llama-cpp-comparison",null,"Benchmark and profile Llaminar against llama.cpp on a matched real-model workload, authenticate prompt\u002Ftoken and runtime settings, and attribute prefill\u002Fdecode gaps to inference phases and kernels. Use for cross-engine performance comparisons, not ordinary parity-campaign execution.","Llaminar\u002Fllaminar","Llaminar","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F247404048?v=4","01M718MTZG2TQSA7DEB5DKQDN1","01M718MV06G9CV8S2JMN9XHX8Y",48,"AGPL-3.0","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002FSKILL.md",[],[18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37],"comparison","llama cpp comparison","llama cpp","Llama","cpp","benchmark","settings","decode","engine","kernels","gaps","prompt","use","prefill","execution","campaign","performance","runtime","for","model","2026-11-10 10:12:30",false,"---\nname: llama-cpp-comparison\ndescription: Benchmark and profile Llaminar against llama.cpp on a matched real-model workload, authenticate prompt\u002Ftoken and runtime settings, and attribute prefill\u002Fdecode gaps to inference phases and kernels. Use for cross-engine performance comparisons, not ordinary parity-campaign execution.\n---\n\n# Compare Llaminar with llama.cpp\n\nProduce a reproducible, numerical-contract-aware comparison of **Release**\ninference. Separate unprofiled end-to-end throughput from profiler attribution;\nneither a microbenchmark win nor Llaminar-only hotspot ranking establishes why\nLlaminar is slower than another engine.\n\n## Establish the comparison\n\nRead [matched benchmarks](references\u002Fbenchmarks.md) before running either engine.\nFreeze model bytes, device\u002Ftopology, exact prompt bytes, context capacity, cache\nprecision, sampling policy, MTP policy, warmup count and generated-token budget.\nRespect the user's constraints on formats, activation precision and VRAM.\nPin and record the llama.cpp commit\u002Fbuild; “current mainline” requires checking\nthe remote revision, not assuming an old checkout is still current.\n\nFor the short-prefill comparison, use a **512-token prompt aligned to the\n512-row Llaminar prefill bucket**. The supplied Qwen prompt fixture is a starting\npoint, not a tokenizer authority. Revalidate actual token counts, prompt-cache\nmisses, captured bucket selection and physical batches in both engines. Do not\nsilently replace a requested long-context workload with this short workload.\n\n## Attribute the gap\n\nRead [phase attribution](references\u002Fattribution.md). Capture both engines on the\nsame workload and classify physical work as projections\u002Freductions, attention,\nGDN\u002Fconvolution, normalization, activation preparation, KV\u002Fstate\u002Flayout,\nembedding, vocabulary head and sampling\u002Fcontrol. Keep fused operations whole.\n\nUse the applicable project backend workflow for attachment and kernel counters:\n\n- [CUDA](..\u002Fcuda-tuning\u002FSKILL.md): Nsight Systems\u002FCompute, captured-node evidence.\n- [ROCm](..\u002Frocm-tuning\u002FSKILL.md): HIP\u002FRCCL traces and instruction\u002Fresource proof.\n- [CPU](..\u002Fcpu-tuning\u002FSKILL.md): perf, ISA, affinity and physical-core scaling.\n\nThe optional [SQLite analyzer](scripts\u002Fattribute_nsight.py) accepts explicit\nphase windows and ordered classification rules. It reports category sums,\nunions, exclusive intervals, overlap and no-kernel intervals separately, plus\nper-kernel geometry. It never guesses token boundaries or vocabulary-head\nidentity. Keep those model\u002Frun-specific inputs alongside the raw trace.\n\n## Choose and validate a change\n\nRank **paired exposed-time deficits**, not only percent-of-total hot spots.\nCompare arithmetic, reduction order, operand formats, actual memory traffic,\noccupancy, spills, launch geometry and stream overlap for the selected kernels.\nFor GEMM dispatch work also use\n[NativeVNNI tuning](..\u002Fnativevnni-gemm-tuning\u002FSKILL.md); for MTP work use\n[MTP tuning](..\u002Fmtp-tuning\u002FSKILL.md). Never adopt a baseline's precision or\nnon-invariant reduction merely to match its timing.\n\nProfile one exact candidate\u002Fshape per launch, outside canonical timing. Then\nrepeat the fixed Release benchmark, preferably A\u002FB\u002FA or interleaved repetitions,\nand compare token identities and memory BOM as well as throughput. Retain a\nchange only when its model-level benefit exceeds the measured noise and the\napplicable correctness gates pass. Use\n[Llaminar testing](..\u002Fllaminar-testing\u002FSKILL.md) for the full Unit gate,\nProductionTestPreflight and affected canonical real-weight cells. Report\nmissing or still-running evidence explicitly; do not call a timing candidate\ncertified.\n\nSave dated results, exact commands, revisions, resource records and rejected\nexperiments under the project investigation, with large raw artifacts outside\nthe worktree. This skill owns the procedure, not a changing leaderboard or\nanother model\u002Ftopology matrix.\n","# Compare Llaminar with llama.cpp\n\nProduce a reproducible, numerical-contract-aware comparison of **Release**\ninference. Separate unprofiled end-to-end throughput from profiler attribution;\nneither a microbenchmark win nor Llaminar-only hotspot ranking establishes why\nLlaminar is slower than another engine.\n\n## Establish the comparison\n\nRead [matched benchmarks](references\u002Fbenchmarks.md) before running either engine.\nFreeze model bytes, device\u002Ftopology, exact prompt bytes, context capacity, cache\nprecision, sampling policy, MTP policy, warmup count and generated-token budget.\nRespect the user's constraints on formats, activation precision and VRAM.\nPin and record the llama.cpp commit\u002Fbuild; “current mainline” requires checking\nthe remote revision, not assuming an old checkout is still current.\n\nFor the short-prefill comparison, use a **512-token prompt aligned to the\n512-row Llaminar prefill bucket**. The supplied Qwen prompt fixture is a starting\npoint, not a tokenizer authority. Revalidate actual token counts, prompt-cache\nmisses, captured bucket selection and physical batches in both engines. Do not\nsilently replace a requested long-context workload with this short workload.\n\n## Attribute the gap\n\nRead [phase attribution](references\u002Fattribution.md). Capture both engines on the\nsame workload and classify physical work as projections\u002Freductions, attention,\nGDN\u002Fconvolution, normalization, activation preparation, KV\u002Fstate\u002Flayout,\nembedding, vocabulary head and sampling\u002Fcontrol. Keep fused operations whole.\n\nUse the applicable project backend workflow for attachment and kernel counters:\n\n- [CUDA](..\u002Fcuda-tuning\u002FSKILL.md): Nsight Systems\u002FCompute, captured-node evidence.\n- [ROCm](..\u002Frocm-tuning\u002FSKILL.md): HIP\u002FRCCL traces and instruction\u002Fresource proof.\n- [CPU](..\u002Fcpu-tuning\u002FSKILL.md): perf, ISA, affinity and physical-core scaling.\n\nThe optional [SQLite analyzer](scripts\u002Fattribute_nsight.py) accepts explicit\nphase windows and ordered classification rules. It reports category sums,\nunions, exclusive intervals, overlap and no-kernel intervals separately, plus\nper-kernel geometry. It never guesses token boundaries or vocabulary-head\nidentity. Keep those model\u002Frun-specific inputs alongside the raw trace.\n\n## Choose and validate a change\n\nRank **paired exposed-time deficits**, not only percent-of-total hot spots.\nCompare arithmetic, reduction order, operand formats, actual memory traffic,\noccupancy, spills, launch geometry and stream overlap for the selected kernels.\nFor GEMM dispatch work also use\n[NativeVNNI tuning](..\u002Fnativevnni-gemm-tuning\u002FSKILL.md); for MTP work use\n[MTP tuning](..\u002Fmtp-tuning\u002FSKILL.md). Never adopt a baseline's precision or\nnon-invariant reduction merely to match its timing.\n\nProfile one exact candidate\u002Fshape per launch, outside canonical timing. Then\nrepeat the fixed Release benchmark, preferably A\u002FB\u002FA or interleaved repetitions,\nand compare token identities and memory BOM as well as throughput. Retain a\nchange only when its model-level benefit exceeds the measured noise and the\napplicable correctness gates pass. Use\n[Llaminar testing](..\u002Fllaminar-testing\u002FSKILL.md) for the full Unit gate,\nProductionTestPreflight and affected canonical real-weight cells. Report\nmissing or still-running evidence explicitly; do not call a timing candidate\ncertified.\n\nSave dated results, exact commands, revisions, resource records and rejected\nexperiments under the project investigation, with large raw artifacts outside\nthe worktree. This skill owns the procedure, not a changing leaderboard or\nanother model\u002Ftopology matrix.\n",{"name":5,"description":7},"master","454741b1b2228bc6874e46ade75bbf523a8385d5",".agents\u002Fllama-cpp-comparison",4,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar","https:\u002F\u002Fllaminar.ai","https:\u002F\u002Fgithub.com\u002FLlaminar","Home of the Llaminar inferencing engine.","2026-11-15 04:04:16",{"source":8,"skillName":5,"commands":53,"prompt":63,"zipAvailable":64},[54,57,60],{"tool":55,"command":56},"npx","npx skills add Llaminar\u002Fllaminar --skill llama-cpp-comparison",{"tool":58,"command":59},"bunx","bunx skills add Llaminar\u002Fllaminar --skill llama-cpp-comparison",{"tool":61,"command":62},"pnpm","pnpm dlx skills add Llaminar\u002Fllaminar --skill llama-cpp-comparison","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002FSKILL.md",true,[66,72,78,82,88],{"path":67,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":70,"sourceHtmlUrl":71},"assets\u002Fprompt-512-qwen38.json","ASSET","资产文件",3038,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002Fassets\u002Fprompt-512-qwen38.json",{"path":73,"resourceType":74,"resourceTypeDesc":75,"sizeBytes":76,"sourceHtmlUrl":77},"references\u002Fattribution.md","REFERENCE","参考文档",7485,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002Freferences\u002Fattribution.md",{"path":79,"resourceType":74,"resourceTypeDesc":75,"sizeBytes":80,"sourceHtmlUrl":81},"references\u002Fbenchmarks.md",6967,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002Freferences\u002Fbenchmarks.md",{"path":83,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":86,"sourceHtmlUrl":87},"scripts\u002Fattribute_nsight.py","SCRIPT","脚本",10236,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002Fscripts\u002Fattribute_nsight.py",{"path":89,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":90,"sourceHtmlUrl":91},"scripts\u002Ftest_attribute_nsight.py",5909,"https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.agents\u002Fllama-cpp-comparison\u002Fscripts\u002Ftest_attribute_nsight.py",[93,102,110,120,128,133],{"skillNo":94,"name":95,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":38},"01M718P9ZZSG781K5PXM1V3KG2","rocm-tuning","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.codex\u002Fskills\u002Frocm-tuning\u002FSKILL.md",[],[99,100,101],"rocm tuning","tuning","rocm",{"skillNo":103,"name":104,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":38},"01M718P95WX88GYN9MTSC6AE40","mtp-tuning","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.codex\u002Fskills\u002Fmtp-tuning\u002FSKILL.md",[],[108,100,109],"mtp tuning","mtp",{"skillNo":111,"name":112,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":38},"01M718P7JK26A3N6VZYSE2J36T","cuda-kernel-profiling","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.codex\u002Fskills\u002Fcuda-kernel-profiling\u002FSKILL.md",[],[116,117,118,119],"profiling","cuda kernel profiling","cuda","kernel",{"skillNo":121,"name":122,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":38},"01M718P5YCB9J40NZZ0PRAKS5V","cpu-tuning","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.codex\u002Fskills\u002Fcpu-tuning\u002FSKILL.md",[],[126,127,100],"cpu tuning","CPU",{"skillNo":129,"name":95,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":130,"categories":131,"keywords":132,"publishedAt":38},"01M718P57PZ4G3X1B24ZVWR2PQ","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Frocm-tuning\u002FSKILL.md",[],[99,100,101],{"skillNo":134,"name":135,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":136,"categories":137,"keywords":138,"publishedAt":38},"01M718P4GF2X3EHBTY7ZB1Q2T2","nativevnni-gemm-tuning","https:\u002F\u002Fgithub.com\u002FLlaminar\u002Fllaminar\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fnativevnni-gemm-tuning\u002FSKILL.md",[],[100,139],"nativevnni gemm tuning"]