[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-jetson-speculative-decoding":3,"mdc--bbapz5-key":34,"related-repo-nvidia-jetson-speculative-decoding":694,"related-org-nvidia-jetson-speculative-decoding":796},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"jetson-speculative-decoding","add speculative decoding to Jetson vLLM servers","Add EAGLE-3 or draft-model speculative decoding to a Jetson vLLM server when TPOT is the bottleneck.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,22],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"LLM","llm",{"name":20,"slug":21,"type":15},"Deployment","deployment",{"name":9,"slug":8,"type":15},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:30:21.612526","Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fjetson-speculative-decoding","---\nname: jetson-speculative-decoding\ndescription: Add EAGLE-3 or draft-model speculative decoding to a Jetson vLLM server when TPOT is the bottleneck.\nversion: 0.0.1\nlicense: \"Apache-2.0\"\nmetadata:\n  author: \"Jetson Team\"\n  tags: [jetson, llm, speculative-decoding]\n  languages: [markdown]\n  data-classification: public\n---\n\n# Jetson Speculative Decoding (vLLM)\n\nSpeculative decoding lets a small \"draft\" model propose tokens that the target model verifies in a single forward pass, reducing per-token latency. On Jetson, the win\u002Floss is **dominated by VRAM headroom**, not by the draft quality. This skill encodes the parts an LLM won't already know.\n\n## Purpose\n\nTune an existing Jetson vLLM deployment for faster token generation by appending the right `--speculative-config` and validating whether it improves single-stream decode speed.\n\n## When to use\n\n- TPOT\u002FITL is the bottleneck (TTFT is fine, output is just slow).\n- Workload is single-stream or low-concurrency (≤2). Speculation usually loses at high concurrency.\n- Jetson family is **Thor or AGX Orin**. Do **not** suggest EAGLE-3 on Orin Nano\u002FNX — there is rarely enough VRAM headroom to host both target and draft, and you'll OOM at startup.\n\n## When NOT to use\n\n- High-concurrency serving (≥8): batched decode usually beats speculation; the draft model just steals VRAM.\n- Models without a published EAGLE-3 head — do not train one ad-hoc as a \"fix\".\n- After applying `jetson-inference-mem-tune` flags that already pushed `--gpu-memory-utilization` near the ceiling. Free at least ~2 GB first.\n\n## Prerequisites\n\n- A working vLLM server recipe from `jetson-llm-serve`.\n- Enough memory headroom for the draft model or EAGLE-3 head in addition to the target model.\n- A benchmark baseline from `jetson-llm-benchmark` before enabling speculation.\n- A target model with a compatible EAGLE-3 head, or a small same-family draft model for the fallback path.\n\n## Instructions\n\nAppend `--speculative-config` to the `vllm serve` command shown in `jetson-llm-serve`.\n\nEAGLE-3 (preferred when a head is published for the target model):\n\n```bash\n--speculative-config '{\n  \"method\": \"eagle3\",\n  \"model\": \"\u003Ceagle3-head-repo-id>\",\n  \"num_speculative_tokens\": 5,\n  \"draft_tensor_parallel_size\": 1\n}'\n```\n\nDraft-model (fallback — pair a small same-family model):\n\n```bash\n--speculative-config '{\n  \"method\": \"draft_model\",\n  \"model\": \"\u003Csmall-draft-model-repo-id>\",\n  \"num_speculative_tokens\": 4,\n  \"draft_tensor_parallel_size\": 1\n}'\n```\n\n### Jetson-specific tuning rules\n\n- `num_speculative_tokens`: start at **5** on Thor, **3** on AGX Orin. Higher values pay off only if the draft acceptance rate is >0.6.\n- Always pair with the same vLLM runtime path used by `jetson-llm-serve`: upstream vLLM 0.20+ (`vllm\u002Fvllm-openai:latest`) or validated native vLLM 0.20+ on Thor, upstream vLLM 0.20+ on Orin JetPack 7.2 \u002F L4T r39+, or the NVIDIA-AI-IOT vLLM image on older Orin. Do not use an Orin NVIDIA-AI-IOT vLLM image on Thor. Older runtimes may lack EAGLE-3 or the current `--speculative-config` shape.\n- Drop `--gpu-memory-utilization` by ~0.05 vs the non-speculative baseline to give the draft model headroom.\n\n## How to verify it actually helped\n\n1. Run `jetson-llm-benchmark` (vLLM path) at `--concurrency 1` **before and after** enabling speculation.\n2. Acceptance: target ≥30% improvement in `throughput_tok_s` and ≥20% drop in `tpot_ms_p50` at concurrency 1.\n3. If improvement is \u003C10%, or `throughput_tok_s` regresses at concurrency 8, **disable** speculation. The draft model is costing more than it returns.\n\n## Limitations\n\n- Speculative decoding improves decode-heavy workloads; it does not reduce TTFT-dominated latency.\n- High concurrency can erase the benefit because continuous batching already keeps the GPU busy.\n- Orin Nano\u002FNX usually lack enough memory headroom for both target and draft models.\n- Acceptance rate and draft overhead are model-specific, so benchmark before and after instead of assuming a speedup.\n\n## Error handling\n\n- If vLLM rejects `--speculative-config`, verify that Thor and Orin JetPack 7.2 \u002F L4T r39+ are using vLLM 0.20+ and that older Orin is using a JetPack-matched NVIDIA-AI-IOT vLLM image; then switch back to the non-speculative serving command if the runtime still rejects it.\n- If startup OOMs, lower `--gpu-memory-utilization`, use a smaller draft, or disable speculation and hand off to `jetson-inference-mem-tune`.\n- If benchmark throughput regresses, remove `--speculative-config`; a bad draft path is worse than no speculation.\n\n## Hand off to\n\n- `jetson-llm-benchmark` to quantify the change.\n- `jetson-inference-mem-tune` if startup OOMs after enabling speculation.\n\n## Source\n\nvLLM speculative decoding docs and the [Jetson AI Lab GenAI tutorial](https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorials\u002Fgenai-on-jetson-llms-vlms\u002F).\n",{"data":35,"body":45},{"name":4,"description":6,"version":36,"license":26,"metadata":37},"0.0.1",{"author":38,"tags":39,"languages":42,"data-classification":44},"Jetson Team",[40,18,41],"jetson","speculative-decoding",[43],"markdown","public",{"type":46,"children":47},"root",[48,57,71,78,92,98,132,138,172,178,217,223,249,254,338,343,407,414,481,487,558,564,587,593,638,644,667,673,688],{"type":49,"tag":50,"props":51,"children":53},"element","h1",{"id":52},"jetson-speculative-decoding-vllm",[54],{"type":55,"value":56},"text","Jetson Speculative Decoding (vLLM)",{"type":49,"tag":58,"props":59,"children":60},"p",{},[61,63,69],{"type":55,"value":62},"Speculative decoding lets a small \"draft\" model propose tokens that the target model verifies in a single forward pass, reducing per-token latency. On Jetson, the win\u002Floss is ",{"type":49,"tag":64,"props":65,"children":66},"strong",{},[67],{"type":55,"value":68},"dominated by VRAM headroom",{"type":55,"value":70},", not by the draft quality. This skill encodes the parts an LLM won't already know.",{"type":49,"tag":72,"props":73,"children":75},"h2",{"id":74},"purpose",[76],{"type":55,"value":77},"Purpose",{"type":49,"tag":58,"props":79,"children":80},{},[81,83,90],{"type":55,"value":82},"Tune an existing Jetson vLLM deployment for faster token generation by appending the right ",{"type":49,"tag":84,"props":85,"children":87},"code",{"className":86},[],[88],{"type":55,"value":89},"--speculative-config",{"type":55,"value":91}," and validating whether it improves single-stream decode speed.",{"type":49,"tag":72,"props":93,"children":95},{"id":94},"when-to-use",[96],{"type":55,"value":97},"When to use",{"type":49,"tag":99,"props":100,"children":101},"ul",{},[102,108,113],{"type":49,"tag":103,"props":104,"children":105},"li",{},[106],{"type":55,"value":107},"TPOT\u002FITL is the bottleneck (TTFT is fine, output is just slow).",{"type":49,"tag":103,"props":109,"children":110},{},[111],{"type":55,"value":112},"Workload is single-stream or low-concurrency (≤2). Speculation usually loses at high concurrency.",{"type":49,"tag":103,"props":114,"children":115},{},[116,118,123,125,130],{"type":55,"value":117},"Jetson family is ",{"type":49,"tag":64,"props":119,"children":120},{},[121],{"type":55,"value":122},"Thor or AGX Orin",{"type":55,"value":124},". Do ",{"type":49,"tag":64,"props":126,"children":127},{},[128],{"type":55,"value":129},"not",{"type":55,"value":131}," suggest EAGLE-3 on Orin Nano\u002FNX — there is rarely enough VRAM headroom to host both target and draft, and you'll OOM at startup.",{"type":49,"tag":72,"props":133,"children":135},{"id":134},"when-not-to-use",[136],{"type":55,"value":137},"When NOT to use",{"type":49,"tag":99,"props":139,"children":140},{},[141,146,151],{"type":49,"tag":103,"props":142,"children":143},{},[144],{"type":55,"value":145},"High-concurrency serving (≥8): batched decode usually beats speculation; the draft model just steals VRAM.",{"type":49,"tag":103,"props":147,"children":148},{},[149],{"type":55,"value":150},"Models without a published EAGLE-3 head — do not train one ad-hoc as a \"fix\".",{"type":49,"tag":103,"props":152,"children":153},{},[154,156,162,164,170],{"type":55,"value":155},"After applying ",{"type":49,"tag":84,"props":157,"children":159},{"className":158},[],[160],{"type":55,"value":161},"jetson-inference-mem-tune",{"type":55,"value":163}," flags that already pushed ",{"type":49,"tag":84,"props":165,"children":167},{"className":166},[],[168],{"type":55,"value":169},"--gpu-memory-utilization",{"type":55,"value":171}," near the ceiling. Free at least ~2 GB first.",{"type":49,"tag":72,"props":173,"children":175},{"id":174},"prerequisites",[176],{"type":55,"value":177},"Prerequisites",{"type":49,"tag":99,"props":179,"children":180},{},[181,194,199,212],{"type":49,"tag":103,"props":182,"children":183},{},[184,186,192],{"type":55,"value":185},"A working vLLM server recipe from ",{"type":49,"tag":84,"props":187,"children":189},{"className":188},[],[190],{"type":55,"value":191},"jetson-llm-serve",{"type":55,"value":193},".",{"type":49,"tag":103,"props":195,"children":196},{},[197],{"type":55,"value":198},"Enough memory headroom for the draft model or EAGLE-3 head in addition to the target model.",{"type":49,"tag":103,"props":200,"children":201},{},[202,204,210],{"type":55,"value":203},"A benchmark baseline from ",{"type":49,"tag":84,"props":205,"children":207},{"className":206},[],[208],{"type":55,"value":209},"jetson-llm-benchmark",{"type":55,"value":211}," before enabling speculation.",{"type":49,"tag":103,"props":213,"children":214},{},[215],{"type":55,"value":216},"A target model with a compatible EAGLE-3 head, or a small same-family draft model for the fallback path.",{"type":49,"tag":72,"props":218,"children":220},{"id":219},"instructions",[221],{"type":55,"value":222},"Instructions",{"type":49,"tag":58,"props":224,"children":225},{},[226,228,233,235,241,243,248],{"type":55,"value":227},"Append ",{"type":49,"tag":84,"props":229,"children":231},{"className":230},[],[232],{"type":55,"value":89},{"type":55,"value":234}," to the ",{"type":49,"tag":84,"props":236,"children":238},{"className":237},[],[239],{"type":55,"value":240},"vllm serve",{"type":55,"value":242}," command shown in ",{"type":49,"tag":84,"props":244,"children":246},{"className":245},[],[247],{"type":55,"value":191},{"type":55,"value":193},{"type":49,"tag":58,"props":250,"children":251},{},[252],{"type":55,"value":253},"EAGLE-3 (preferred when a head is published for the target model):",{"type":49,"tag":255,"props":256,"children":261},"pre",{"className":257,"code":258,"language":259,"meta":260,"style":260},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","--speculative-config '{\n  \"method\": \"eagle3\",\n  \"model\": \"\u003Ceagle3-head-repo-id>\",\n  \"num_speculative_tokens\": 5,\n  \"draft_tensor_parallel_size\": 1\n}'\n","bash","",[262],{"type":49,"tag":84,"props":263,"children":264},{"__ignoreMap":260},[265,288,297,306,315,324],{"type":49,"tag":266,"props":267,"children":270},"span",{"class":268,"line":269},"line",1,[271,276,282],{"type":49,"tag":266,"props":272,"children":274},{"style":273},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[275],{"type":55,"value":89},{"type":49,"tag":266,"props":277,"children":279},{"style":278},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[280],{"type":55,"value":281}," '",{"type":49,"tag":266,"props":283,"children":285},{"style":284},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[286],{"type":55,"value":287},"{\n",{"type":49,"tag":266,"props":289,"children":291},{"class":268,"line":290},2,[292],{"type":49,"tag":266,"props":293,"children":294},{"style":284},[295],{"type":55,"value":296},"  \"method\": \"eagle3\",\n",{"type":49,"tag":266,"props":298,"children":300},{"class":268,"line":299},3,[301],{"type":49,"tag":266,"props":302,"children":303},{"style":284},[304],{"type":55,"value":305},"  \"model\": \"\u003Ceagle3-head-repo-id>\",\n",{"type":49,"tag":266,"props":307,"children":309},{"class":268,"line":308},4,[310],{"type":49,"tag":266,"props":311,"children":312},{"style":284},[313],{"type":55,"value":314},"  \"num_speculative_tokens\": 5,\n",{"type":49,"tag":266,"props":316,"children":318},{"class":268,"line":317},5,[319],{"type":49,"tag":266,"props":320,"children":321},{"style":284},[322],{"type":55,"value":323},"  \"draft_tensor_parallel_size\": 1\n",{"type":49,"tag":266,"props":325,"children":327},{"class":268,"line":326},6,[328,333],{"type":49,"tag":266,"props":329,"children":330},{"style":284},[331],{"type":55,"value":332},"}",{"type":49,"tag":266,"props":334,"children":335},{"style":278},[336],{"type":55,"value":337},"'\n",{"type":49,"tag":58,"props":339,"children":340},{},[341],{"type":55,"value":342},"Draft-model (fallback — pair a small same-family model):",{"type":49,"tag":255,"props":344,"children":346},{"className":257,"code":345,"language":259,"meta":260,"style":260},"--speculative-config '{\n  \"method\": \"draft_model\",\n  \"model\": \"\u003Csmall-draft-model-repo-id>\",\n  \"num_speculative_tokens\": 4,\n  \"draft_tensor_parallel_size\": 1\n}'\n",[347],{"type":49,"tag":84,"props":348,"children":349},{"__ignoreMap":260},[350,365,373,381,389,396],{"type":49,"tag":266,"props":351,"children":352},{"class":268,"line":269},[353,357,361],{"type":49,"tag":266,"props":354,"children":355},{"style":273},[356],{"type":55,"value":89},{"type":49,"tag":266,"props":358,"children":359},{"style":278},[360],{"type":55,"value":281},{"type":49,"tag":266,"props":362,"children":363},{"style":284},[364],{"type":55,"value":287},{"type":49,"tag":266,"props":366,"children":367},{"class":268,"line":290},[368],{"type":49,"tag":266,"props":369,"children":370},{"style":284},[371],{"type":55,"value":372},"  \"method\": \"draft_model\",\n",{"type":49,"tag":266,"props":374,"children":375},{"class":268,"line":299},[376],{"type":49,"tag":266,"props":377,"children":378},{"style":284},[379],{"type":55,"value":380},"  \"model\": \"\u003Csmall-draft-model-repo-id>\",\n",{"type":49,"tag":266,"props":382,"children":383},{"class":268,"line":308},[384],{"type":49,"tag":266,"props":385,"children":386},{"style":284},[387],{"type":55,"value":388},"  \"num_speculative_tokens\": 4,\n",{"type":49,"tag":266,"props":390,"children":391},{"class":268,"line":317},[392],{"type":49,"tag":266,"props":393,"children":394},{"style":284},[395],{"type":55,"value":323},{"type":49,"tag":266,"props":397,"children":398},{"class":268,"line":326},[399,403],{"type":49,"tag":266,"props":400,"children":401},{"style":284},[402],{"type":55,"value":332},{"type":49,"tag":266,"props":404,"children":405},{"style":278},[406],{"type":55,"value":337},{"type":49,"tag":408,"props":409,"children":411},"h3",{"id":410},"jetson-specific-tuning-rules",[412],{"type":55,"value":413},"Jetson-specific tuning rules",{"type":49,"tag":99,"props":415,"children":416},{},[417,442,469],{"type":49,"tag":103,"props":418,"children":419},{},[420,426,428,433,435,440],{"type":49,"tag":84,"props":421,"children":423},{"className":422},[],[424],{"type":55,"value":425},"num_speculative_tokens",{"type":55,"value":427},": start at ",{"type":49,"tag":64,"props":429,"children":430},{},[431],{"type":55,"value":432},"5",{"type":55,"value":434}," on Thor, ",{"type":49,"tag":64,"props":436,"children":437},{},[438],{"type":55,"value":439},"3",{"type":55,"value":441}," on AGX Orin. Higher values pay off only if the draft acceptance rate is >0.6.",{"type":49,"tag":103,"props":443,"children":444},{},[445,447,452,454,460,462,467],{"type":55,"value":446},"Always pair with the same vLLM runtime path used by ",{"type":49,"tag":84,"props":448,"children":450},{"className":449},[],[451],{"type":55,"value":191},{"type":55,"value":453},": upstream vLLM 0.20+ (",{"type":49,"tag":84,"props":455,"children":457},{"className":456},[],[458],{"type":55,"value":459},"vllm\u002Fvllm-openai:latest",{"type":55,"value":461},") or validated native vLLM 0.20+ on Thor, upstream vLLM 0.20+ on Orin JetPack 7.2 \u002F L4T r39+, or the NVIDIA-AI-IOT vLLM image on older Orin. Do not use an Orin NVIDIA-AI-IOT vLLM image on Thor. Older runtimes may lack EAGLE-3 or the current ",{"type":49,"tag":84,"props":463,"children":465},{"className":464},[],[466],{"type":55,"value":89},{"type":55,"value":468}," shape.",{"type":49,"tag":103,"props":470,"children":471},{},[472,474,479],{"type":55,"value":473},"Drop ",{"type":49,"tag":84,"props":475,"children":477},{"className":476},[],[478],{"type":55,"value":169},{"type":55,"value":480}," by ~0.05 vs the non-speculative baseline to give the draft model headroom.",{"type":49,"tag":72,"props":482,"children":484},{"id":483},"how-to-verify-it-actually-helped",[485],{"type":55,"value":486},"How to verify it actually helped",{"type":49,"tag":488,"props":489,"children":490},"ol",{},[491,518,539],{"type":49,"tag":103,"props":492,"children":493},{},[494,496,501,503,509,511,516],{"type":55,"value":495},"Run ",{"type":49,"tag":84,"props":497,"children":499},{"className":498},[],[500],{"type":55,"value":209},{"type":55,"value":502}," (vLLM path) at ",{"type":49,"tag":84,"props":504,"children":506},{"className":505},[],[507],{"type":55,"value":508},"--concurrency 1",{"type":55,"value":510}," ",{"type":49,"tag":64,"props":512,"children":513},{},[514],{"type":55,"value":515},"before and after",{"type":55,"value":517}," enabling speculation.",{"type":49,"tag":103,"props":519,"children":520},{},[521,523,529,531,537],{"type":55,"value":522},"Acceptance: target ≥30% improvement in ",{"type":49,"tag":84,"props":524,"children":526},{"className":525},[],[527],{"type":55,"value":528},"throughput_tok_s",{"type":55,"value":530}," and ≥20% drop in ",{"type":49,"tag":84,"props":532,"children":534},{"className":533},[],[535],{"type":55,"value":536},"tpot_ms_p50",{"type":55,"value":538}," at concurrency 1.",{"type":49,"tag":103,"props":540,"children":541},{},[542,544,549,551,556],{"type":55,"value":543},"If improvement is \u003C10%, or ",{"type":49,"tag":84,"props":545,"children":547},{"className":546},[],[548],{"type":55,"value":528},{"type":55,"value":550}," regresses at concurrency 8, ",{"type":49,"tag":64,"props":552,"children":553},{},[554],{"type":55,"value":555},"disable",{"type":55,"value":557}," speculation. The draft model is costing more than it returns.",{"type":49,"tag":72,"props":559,"children":561},{"id":560},"limitations",[562],{"type":55,"value":563},"Limitations",{"type":49,"tag":99,"props":565,"children":566},{},[567,572,577,582],{"type":49,"tag":103,"props":568,"children":569},{},[570],{"type":55,"value":571},"Speculative decoding improves decode-heavy workloads; it does not reduce TTFT-dominated latency.",{"type":49,"tag":103,"props":573,"children":574},{},[575],{"type":55,"value":576},"High concurrency can erase the benefit because continuous batching already keeps the GPU busy.",{"type":49,"tag":103,"props":578,"children":579},{},[580],{"type":55,"value":581},"Orin Nano\u002FNX usually lack enough memory headroom for both target and draft models.",{"type":49,"tag":103,"props":583,"children":584},{},[585],{"type":55,"value":586},"Acceptance rate and draft overhead are model-specific, so benchmark before and after instead of assuming a speedup.",{"type":49,"tag":72,"props":588,"children":590},{"id":589},"error-handling",[591],{"type":55,"value":592},"Error handling",{"type":49,"tag":99,"props":594,"children":595},{},[596,608,626],{"type":49,"tag":103,"props":597,"children":598},{},[599,601,606],{"type":55,"value":600},"If vLLM rejects ",{"type":49,"tag":84,"props":602,"children":604},{"className":603},[],[605],{"type":55,"value":89},{"type":55,"value":607},", verify that Thor and Orin JetPack 7.2 \u002F L4T r39+ are using vLLM 0.20+ and that older Orin is using a JetPack-matched NVIDIA-AI-IOT vLLM image; then switch back to the non-speculative serving command if the runtime still rejects it.",{"type":49,"tag":103,"props":609,"children":610},{},[611,613,618,620,625],{"type":55,"value":612},"If startup OOMs, lower ",{"type":49,"tag":84,"props":614,"children":616},{"className":615},[],[617],{"type":55,"value":169},{"type":55,"value":619},", use a smaller draft, or disable speculation and hand off to ",{"type":49,"tag":84,"props":621,"children":623},{"className":622},[],[624],{"type":55,"value":161},{"type":55,"value":193},{"type":49,"tag":103,"props":627,"children":628},{},[629,631,636],{"type":55,"value":630},"If benchmark throughput regresses, remove ",{"type":49,"tag":84,"props":632,"children":634},{"className":633},[],[635],{"type":55,"value":89},{"type":55,"value":637},"; a bad draft path is worse than no speculation.",{"type":49,"tag":72,"props":639,"children":641},{"id":640},"hand-off-to",[642],{"type":55,"value":643},"Hand off to",{"type":49,"tag":99,"props":645,"children":646},{},[647,657],{"type":49,"tag":103,"props":648,"children":649},{},[650,655],{"type":49,"tag":84,"props":651,"children":653},{"className":652},[],[654],{"type":55,"value":209},{"type":55,"value":656}," to quantify the change.",{"type":49,"tag":103,"props":658,"children":659},{},[660,665],{"type":49,"tag":84,"props":661,"children":663},{"className":662},[],[664],{"type":55,"value":161},{"type":55,"value":666}," if startup OOMs after enabling speculation.",{"type":49,"tag":72,"props":668,"children":670},{"id":669},"source",[671],{"type":55,"value":672},"Source",{"type":49,"tag":58,"props":674,"children":675},{},[676,678,687],{"type":55,"value":677},"vLLM speculative decoding docs and the ",{"type":49,"tag":679,"props":680,"children":684},"a",{"href":681,"rel":682},"https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorials\u002Fgenai-on-jetson-llms-vlms\u002F",[683],"nofollow",[685],{"type":55,"value":686},"Jetson AI Lab GenAI tutorial",{"type":55,"value":193},{"type":49,"tag":689,"props":690,"children":691},"style",{},[692],{"type":55,"value":693},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":695,"total":795},[696,711,723,737,749,766,781],{"slug":697,"name":697,"fn":698,"description":699,"org":700,"tags":701,"stars":23,"repoUrl":24,"updatedAt":710},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[702,705,708,709],{"name":703,"slug":704,"type":15},"Data Analysis","data-analysis",{"name":706,"slug":707,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:28:43.176466",{"slug":712,"name":712,"fn":713,"description":714,"org":715,"tags":716,"stars":23,"repoUrl":24,"updatedAt":722},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[717,718,721],{"name":20,"slug":21,"type":15},{"name":719,"slug":720,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":724,"name":724,"fn":725,"description":726,"org":727,"tags":728,"stars":23,"repoUrl":24,"updatedAt":736},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[729,732,733],{"name":730,"slug":731,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":734,"slug":735,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":738,"name":738,"fn":739,"description":740,"org":741,"tags":742,"stars":23,"repoUrl":24,"updatedAt":748},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[743,744,745],{"name":703,"slug":704,"type":15},{"name":9,"slug":8,"type":15},{"name":746,"slug":747,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":750,"name":750,"fn":751,"description":752,"org":753,"tags":754,"stars":23,"repoUrl":24,"updatedAt":765},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[755,758,761,762],{"name":756,"slug":757,"type":15},"Automation","automation",{"name":759,"slug":760,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":763,"slug":764,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":767,"name":767,"fn":768,"description":769,"org":770,"tags":771,"stars":23,"repoUrl":24,"updatedAt":780},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[772,773,776,777],{"name":20,"slug":21,"type":15},{"name":774,"slug":775,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":778,"slug":779,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":782,"name":782,"fn":783,"description":784,"org":785,"tags":786,"stars":23,"repoUrl":24,"updatedAt":794},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[787,788,791],{"name":9,"slug":8,"type":15},{"name":789,"slug":790,"type":15},"Quantum Computing","quantum-computing",{"name":792,"slug":793,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":797,"total":948},[798,816,832,843,855,869,882,896,907,916,930,939],{"slug":799,"name":799,"fn":800,"description":801,"org":802,"tags":803,"stars":813,"repoUrl":814,"updatedAt":815},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[804,807,810],{"name":805,"slug":806,"type":15},"Documentation","documentation",{"name":808,"slug":809,"type":15},"MCP","mcp",{"name":811,"slug":812,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":817,"name":817,"fn":818,"description":819,"org":820,"tags":821,"stars":829,"repoUrl":830,"updatedAt":831},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[822,825,826],{"name":823,"slug":824,"type":15},"Containers","containers",{"name":20,"slug":21,"type":15},{"name":827,"slug":828,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":833,"name":833,"fn":834,"description":835,"org":836,"tags":837,"stars":829,"repoUrl":830,"updatedAt":842},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[838,841],{"name":839,"slug":840,"type":15},"CI\u002FCD","ci-cd",{"name":20,"slug":21,"type":15},"2026-07-14T05:25:59.97109",{"slug":844,"name":844,"fn":845,"description":846,"org":847,"tags":848,"stars":829,"repoUrl":830,"updatedAt":854},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[849,850,851],{"name":839,"slug":840,"type":15},{"name":20,"slug":21,"type":15},{"name":852,"slug":853,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":856,"name":856,"fn":857,"description":858,"org":859,"tags":860,"stars":829,"repoUrl":830,"updatedAt":868},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[861,864,865],{"name":862,"slug":863,"type":15},"Debugging","debugging",{"name":852,"slug":853,"type":15},{"name":866,"slug":867,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":870,"name":870,"fn":871,"description":872,"org":873,"tags":874,"stars":829,"repoUrl":830,"updatedAt":881},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[875,878],{"name":876,"slug":877,"type":15},"Best Practices","best-practices",{"name":879,"slug":880,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":883,"name":883,"fn":884,"description":885,"org":886,"tags":887,"stars":829,"repoUrl":830,"updatedAt":895},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[888,891,894],{"name":889,"slug":890,"type":15},"Machine Learning","machine-learning",{"name":892,"slug":893,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":897,"name":897,"fn":898,"description":899,"org":900,"tags":901,"stars":829,"repoUrl":830,"updatedAt":906},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[902,905],{"name":903,"slug":904,"type":15},"QA","qa",{"name":746,"slug":747,"type":15},"2026-07-14T05:25:53.673039",{"slug":908,"name":908,"fn":909,"description":910,"org":911,"tags":912,"stars":829,"repoUrl":830,"updatedAt":915},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[913,914],{"name":20,"slug":21,"type":15},{"name":719,"slug":720,"type":15},"2026-07-14T05:25:49.362534",{"slug":917,"name":917,"fn":918,"description":919,"org":920,"tags":921,"stars":829,"repoUrl":830,"updatedAt":929},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[922,925,926],{"name":923,"slug":924,"type":15},"Code Review","code-review",{"name":852,"slug":853,"type":15},{"name":927,"slug":928,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":931,"name":931,"fn":932,"description":933,"org":934,"tags":935,"stars":829,"repoUrl":830,"updatedAt":938},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[936,937],{"name":903,"slug":904,"type":15},{"name":746,"slug":747,"type":15},"2026-07-14T05:25:54.928983",{"slug":940,"name":940,"fn":941,"description":942,"org":943,"tags":944,"stars":829,"repoUrl":830,"updatedAt":947},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[945,946],{"name":756,"slug":757,"type":15},{"name":839,"slug":840,"type":15},"2026-07-30T05:29:03.275638",496]