[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-compileiq-booster-pack":3,"mdc--6r8wwh-key":34,"related-org-nvidia-compileiq-booster-pack":1546,"related-repo-nvidia-compileiq-booster-pack":1705},{"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},"compileiq-booster-pack","apply CompileIQ booster packs to compilers","Use BEFORE running a full CompileIQ search. Walks through downloading a Booster Pack from NVIDIA\u002FCompileIQ GitHub Releases, applying ACF candidates one at a time to the user's compiler (raw PTXAS, NVCC, Triton, Helion, FlashInfer), and keeping only candidates that compile, pass correctness, and beat the no-ACF baseline. Includes the mandatory Debug-pack O0\u002FO3 ACF-injection canary that proves the ACF is reaching PTXAS. Triggers on \"booster pack\", \"ACF\", \"apply-controls\", \"speed up without searching\", \"helion fp8\", \"flashinfer batch decode\", \"debug pack\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Optimization","optimization",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Engineering","engineering",107,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FCompileIQ","2026-07-14T05:32:12.791444","Apache-2.0",8,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"An Optimizer for Nvidia Compilers.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FCompileIQ\u002Ftree\u002FHEAD\u002Fagent-skills\u002Fcompileiq-booster-pack","---\nname: compileiq-booster-pack\ndescription: >\n  Use BEFORE running a full CompileIQ search. Walks through downloading a\n  Booster Pack from NVIDIA\u002FCompileIQ GitHub Releases, applying ACF candidates\n  one at a time to the user's compiler (raw PTXAS, NVCC, Triton, Helion,\n  FlashInfer), and keeping only candidates that compile, pass correctness,\n  and beat the no-ACF baseline. Includes the mandatory Debug-pack O0\u002FO3\n  ACF-injection canary that proves the ACF is reaching PTXAS. Triggers on\n  \"booster pack\", \"ACF\", \"apply-controls\", \"speed up without searching\",\n  \"helion fp8\", \"flashinfer batch decode\", \"debug pack\".\nwhen_to_use: |\n  - Workload is close to a published pack (Helion FP8 quant \u002F causal\n    depthwise conv \u002F Gated DeltaNet fwd; FlashInfer BatchDecode is a known\n    related workload).\n  - User wants a fast shortcut before paying for a full CompileIQ search.\n  - User asks \"is there a known-good config for X?\"\n  Don't use when:\n  - User lacks a stable baseline, correctness check, or benchmark setup.\n  - Workload has nothing in common with available packs; skip to\n    compileiq-run-search.\nlicense: Apache-2.0\nmetadata:\n  version: \"1.0.0\"\n  author: NVIDIA CompileIQ\n  domain: compiler-optimization\nallowed-tools: Bash Read\npaths: [\"**\u002F*.cu\", \"**\u002F*.cuh\", \"**\u002F*.acf\", \"**\u002F*.py\", \"**\u002F*.sh\"]\n---\n\n# compileiq-booster-pack\n\nTry curated `.acf` candidates *before* running a full CompileIQ search.\nA Booster Pack is a zip of ACFs that NVIDIA validated against a specific\nworkload family. They are not guaranteed speedups; treat every candidate\nas workload-specific and validate it on your own benchmark.\n\nAuthoritative narrative: `docs\u002Fbooster_packs.md`, `docs\u002Fflashinfer_booster.md`.\n\n## When\n\n| If this is true | Use this path |\n|---|---|\n| Workload is close to a Booster Pack's intended workload, compiler, GPU, and validation context. | Try the Booster Pack first. |\n| Workload differs materially or no pack candidate helps. | Run a full CompileIQ search (`compileiq-run-search`). |\n| Baseline, correctness check, compiler path, or benchmark setup are not in place. | Wait. Fix those before applying any ACF. |\n\n## Available packs (today)\n\n| Pack | Workloads it was validated against | Notes |\n|---|---|---|\n| `booster-pack-helion.zip` | Helion FP8 Quantization, Causal Depthwise Convolution, Gated DeltaNet Forward | Has shown benefit on FlashInfer `BatchDecodeWithPagedKVCacheWrapper`; related attention workloads worth testing. |\n| `booster-pack-debug.zip` | Diagnostic ACFs (`O0`, `O3`, others that disable or alter selected optimizations) | Not for speed; for **debugging**. Use the O0\u002FO3 canary below before trusting any other pack. |\n\nThe public release shape is documented in `docs\u002Fbooster_packs.md`. There is no\nruntime download API today. Don't invent one.\n\n## Steps\n\n### 0. Pre-flight: the O0\u002FO3 ACF-injection canary (mandatory first step)\n\nThe most common silent failure when applying ACFs is a framework cache (Triton,\nHelion, FlashInfer's `flashinfer_cubin`\u002F`flashinfer_jit_cache`, NVCC build\ncache) serving a stale binary that ignored the ACF. The Debug pack has two\nACFs with predictable, opposite-direction signatures:\n\n- **`O0` ACF**: forces unoptimized compilation. Applied → expect a **measurable\n  regression** (often 2-10x slower) vs. baseline.\n- **`O3` ACF**: forces the default optimization level. Applied → expect to\n  **match baseline** (the no-ACF default is already `-O3`).\n\n```bash\n# Baseline\nT_BASE_MS=$(.\u002Frun-benchmark.sh)\n\n# O0 must regress\nPTXAS_OPTIONS=\"--apply-controls=debug-pack\u002FO0.acf\" T_O0_MS=$(.\u002Frun-benchmark.sh)\n\n# O3 must match baseline\nPTXAS_OPTIONS=\"--apply-controls=debug-pack\u002FO3.acf\" T_O3_MS=$(.\u002Frun-benchmark.sh)\n\npython -c \"\nimport sys\nbase, o0, o3 = $T_BASE_MS, $T_O0_MS, $T_O3_MS\nif o0 \u003C base * 1.05:\n    print('FAIL: O0 did not regress; ACF is NOT reaching PTXAS. Fix the cache-bust.')\n    sys.exit(1)\nif abs(o3 - base) \u002F base > 0.05:\n    print(f'WARN: O3 differs from baseline by >5%; baseline may not be -O3 or framework caching differs.')\nprint('PASS: ACF injection is wired up correctly.')\n\"\n```\n\nIf this fails, **stop**. Fix the cache-bust before trying any real pack candidate:\n\n- Triton: `export TRITON_ALWAYS_COMPILE=1`, unique `TRITON_CACHE_DIR` per eval.\n- Helion: `export HELION_SKIP_CACHE=1`.\n- FlashInfer: confirm `flashinfer_cubin` and `flashinfer_jit_cache` packages are absent (`docs\u002Fflashinfer_booster.md:56-64`).\n- Raw nvcc: clean the build dir between candidates.\n\n### 1. Download\n\nBrowse `https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FCompileIQ\u002Freleases`, find the latest tag\nmatching `booster-packs-*`, and download the relevant pack zip plus the\ntop-level `booster-pack-catalog.json`.\n\n```bash\nBOOSTER_TAG=\"$(gh release list -R NVIDIA\u002FCompileIQ --limit 100 --json tagName,isDraft \\\n  --jq '.[] | select(.isDraft == false) | select(.tagName | startswith(\"booster-packs-\")) | .tagName' \\\n  | head -n 1)\"\necho \"Using $BOOSTER_TAG\"\ngh release download \"$BOOSTER_TAG\" -R NVIDIA\u002FCompileIQ -p 'booster-pack-helion.zip' -p 'booster-pack-catalog.json' -D .\u002Fpacks\nunzip .\u002Fpacks\u002Fbooster-pack-helion.zip -d .\u002Fpacks\ncat .\u002Fpacks\u002Fbooster-pack-helion\u002Fbooster-pack-manifest.json\n```\n\nAlways read the per-pack manifest before applying: it lists the intended\nworkload, compiler version, GPU target, validation context, and known caveats.\nFor a reproducible rerun, set `BOOSTER_TAG` to the exact tag printed above.\n\n### 2. Apply one ACF at a time\n\n| Target | Injection |\n|---|---|\n| Raw PTXAS | `ptxas -v -arch=sm_100 --apply-controls candidate.acf kernel.ptx` |\n| NVCC (CUDA source) | `nvcc -Xptxas --apply-controls=candidate.acf -arch=sm_100 kernel.cu -o exe` |\n| Triton | `PTXAS_OPTIONS=\"--apply-controls=candidate.acf\" TRITON_ALWAYS_COMPILE=1 python bench.py` |\n| Helion | Helion's official ACF API + `HELION_SKIP_CACHE=1` (see `helionlang.com\u002Fexamples\u002Facfs\u002Fsoftmax_acf.html`). |\n| FlashInfer | `FLASHINFER_EXTRA_CUDAFLAGS=\"--ptxas-options=--apply-controls=$ACF_FILE\" python bench.py` (see `docs\u002Fflashinfer_booster.md:107`). |\n\nApply exactly one ACF per run. If it fails to compile, hangs, crashes, returns\nwrong answers, or regresses, **reject** that candidate and move to the next.\n\n### 3. Validate every candidate\n\n- Compare against a known-good reference (correctness, not just speed).\n- Test multiple input shapes when shape matters.\n- Use compile and runtime timeouts to bound runaway candidates.\n- Run multiple performance trials if the benchmark is noisy.\n- Record the reproducibility checklist below (one row per candidate).\n\n### 4. Reproducibility log\n\nFor every candidate you accept or reject, append a row to\n`booster-pack-log.csv` with:\n\n- ACF filename (and sha256)\n- Manifest \u002F release version\n- Benchmark command\n- GPU model + driver version\n- CTK version\n- `nvcc` and `ptxas` paths + versions\n- Framework version or commit\n- Input shape\n- Baseline result (mean ± std)\n- Candidate result (mean ± std)\n- Correctness status\n- Decision: `KEPT` or `REJECTED:\u003Creason>`\n\nThis is the same checklist `docs\u002Fflashinfer_booster.md:135-148` recommends.\nThe `scripts\u002Fapply_one_acf.sh` helper does most of this automatically.\n\n## Self-test\n\n```bash\nbash scripts\u002Fapply_one_acf.sh --self-test\n```\n\nDry-runs a \"baseline vs baseline\" comparison (no ACF applied to either side)\nand confirms the helper correctly reports \"NOT a real improvement\". Catches\nmisconfigured script invocations before they pollute the reproducibility log.\n\n## Gotchas\n\n- **Pack name is not a hard boundary.** Helion Pack helps some FlashInfer\n  cases (`docs\u002Fbooster_packs.md:34`); test before assuming.\n- **Booster Packs are not search-space inputs.** Don't try to feed an ACF\n  through `PtxasSearchSpace(...)`; packs are already-generated `.acf`\n  candidate bundles, not inputs to `PtxasSearchSpace` or `NvccSearchSpace`.\n- **Force recompilation.** If you can't *prove* a recompile happened between\n  candidates, don't trust the measurement. See the cache-bust hints under the\n  pre-flight canary.\n\n## Next\n\n- If no pack candidate helps your workload, go to `compileiq-run-search` for a\n  full CompileIQ search over `PtxasSearchSpace()`.\n- For attention workloads specifically, also see\n  `compileiq-search-space` (variant=\"att\").\n",{"data":35,"body":48},{"name":4,"description":6,"when_to_use":36,"license":26,"metadata":37,"allowed-tools":41,"paths":42},"- Workload is close to a published pack (Helion FP8 quant \u002F causal\n  depthwise conv \u002F Gated DeltaNet fwd; FlashInfer BatchDecode is a known\n  related workload).\n- User wants a fast shortcut before paying for a full CompileIQ search.\n- User asks \"is there a known-good config for X?\"\nDon't use when:\n- User lacks a stable baseline, correctness check, or benchmark setup.\n- Workload has nothing in common with available packs; skip to\n  compileiq-run-search.\n",{"version":38,"author":39,"domain":40},"1.0.0","NVIDIA CompileIQ","compiler-optimization","Bash Read",[43,44,45,46,47],"**\u002F*.cu","**\u002F*.cuh","**\u002F*.acf","**\u002F*.py","**\u002F*.sh",{"type":49,"children":50},"root",[51,58,81,102,109,184,190,292,304,310,317,338,393,697,709,776,782,810,1061,1073,1079,1205,1217,1223,1251,1257,1270,1360,1381,1387,1411,1416,1422,1499,1505,1540],{"type":52,"tag":53,"props":54,"children":55},"element","h1",{"id":4},[56],{"type":57,"value":4},"text",{"type":52,"tag":59,"props":60,"children":61},"p",{},[62,64,71,73,79],{"type":57,"value":63},"Try curated ",{"type":52,"tag":65,"props":66,"children":68},"code",{"className":67},[],[69],{"type":57,"value":70},".acf",{"type":57,"value":72}," candidates ",{"type":52,"tag":74,"props":75,"children":76},"em",{},[77],{"type":57,"value":78},"before",{"type":57,"value":80}," running a full CompileIQ search.\nA Booster Pack is a zip of ACFs that NVIDIA validated against a specific\nworkload family. They are not guaranteed speedups; treat every candidate\nas workload-specific and validate it on your own benchmark.",{"type":52,"tag":59,"props":82,"children":83},{},[84,86,92,94,100],{"type":57,"value":85},"Authoritative narrative: ",{"type":52,"tag":65,"props":87,"children":89},{"className":88},[],[90],{"type":57,"value":91},"docs\u002Fbooster_packs.md",{"type":57,"value":93},", ",{"type":52,"tag":65,"props":95,"children":97},{"className":96},[],[98],{"type":57,"value":99},"docs\u002Fflashinfer_booster.md",{"type":57,"value":101},".",{"type":52,"tag":103,"props":104,"children":106},"h2",{"id":105},"when",[107],{"type":57,"value":108},"When",{"type":52,"tag":110,"props":111,"children":112},"table",{},[113,132],{"type":52,"tag":114,"props":115,"children":116},"thead",{},[117],{"type":52,"tag":118,"props":119,"children":120},"tr",{},[121,127],{"type":52,"tag":122,"props":123,"children":124},"th",{},[125],{"type":57,"value":126},"If this is true",{"type":52,"tag":122,"props":128,"children":129},{},[130],{"type":57,"value":131},"Use this path",{"type":52,"tag":133,"props":134,"children":135},"tbody",{},[136,150,171],{"type":52,"tag":118,"props":137,"children":138},{},[139,145],{"type":52,"tag":140,"props":141,"children":142},"td",{},[143],{"type":57,"value":144},"Workload is close to a Booster Pack's intended workload, compiler, GPU, and validation context.",{"type":52,"tag":140,"props":146,"children":147},{},[148],{"type":57,"value":149},"Try the Booster Pack first.",{"type":52,"tag":118,"props":151,"children":152},{},[153,158],{"type":52,"tag":140,"props":154,"children":155},{},[156],{"type":57,"value":157},"Workload differs materially or no pack candidate helps.",{"type":52,"tag":140,"props":159,"children":160},{},[161,163,169],{"type":57,"value":162},"Run a full CompileIQ search (",{"type":52,"tag":65,"props":164,"children":166},{"className":165},[],[167],{"type":57,"value":168},"compileiq-run-search",{"type":57,"value":170},").",{"type":52,"tag":118,"props":172,"children":173},{},[174,179],{"type":52,"tag":140,"props":175,"children":176},{},[177],{"type":57,"value":178},"Baseline, correctness check, compiler path, or benchmark setup are not in place.",{"type":52,"tag":140,"props":180,"children":181},{},[182],{"type":57,"value":183},"Wait. Fix those before applying any ACF.",{"type":52,"tag":103,"props":185,"children":187},{"id":186},"available-packs-today",[188],{"type":57,"value":189},"Available packs (today)",{"type":52,"tag":110,"props":191,"children":192},{},[193,214],{"type":52,"tag":114,"props":194,"children":195},{},[196],{"type":52,"tag":118,"props":197,"children":198},{},[199,204,209],{"type":52,"tag":122,"props":200,"children":201},{},[202],{"type":57,"value":203},"Pack",{"type":52,"tag":122,"props":205,"children":206},{},[207],{"type":57,"value":208},"Workloads it was validated against",{"type":52,"tag":122,"props":210,"children":211},{},[212],{"type":57,"value":213},"Notes",{"type":52,"tag":133,"props":215,"children":216},{},[217,247],{"type":52,"tag":118,"props":218,"children":219},{},[220,229,234],{"type":52,"tag":140,"props":221,"children":222},{},[223],{"type":52,"tag":65,"props":224,"children":226},{"className":225},[],[227],{"type":57,"value":228},"booster-pack-helion.zip",{"type":52,"tag":140,"props":230,"children":231},{},[232],{"type":57,"value":233},"Helion FP8 Quantization, Causal Depthwise Convolution, Gated DeltaNet Forward",{"type":52,"tag":140,"props":235,"children":236},{},[237,239,245],{"type":57,"value":238},"Has shown benefit on FlashInfer ",{"type":52,"tag":65,"props":240,"children":242},{"className":241},[],[243],{"type":57,"value":244},"BatchDecodeWithPagedKVCacheWrapper",{"type":57,"value":246},"; related attention workloads worth testing.",{"type":52,"tag":118,"props":248,"children":249},{},[250,259,279],{"type":52,"tag":140,"props":251,"children":252},{},[253],{"type":52,"tag":65,"props":254,"children":256},{"className":255},[],[257],{"type":57,"value":258},"booster-pack-debug.zip",{"type":52,"tag":140,"props":260,"children":261},{},[262,264,270,271,277],{"type":57,"value":263},"Diagnostic ACFs (",{"type":52,"tag":65,"props":265,"children":267},{"className":266},[],[268],{"type":57,"value":269},"O0",{"type":57,"value":93},{"type":52,"tag":65,"props":272,"children":274},{"className":273},[],[275],{"type":57,"value":276},"O3",{"type":57,"value":278},", others that disable or alter selected optimizations)",{"type":52,"tag":140,"props":280,"children":281},{},[282,284,290],{"type":57,"value":283},"Not for speed; for ",{"type":52,"tag":285,"props":286,"children":287},"strong",{},[288],{"type":57,"value":289},"debugging",{"type":57,"value":291},". Use the O0\u002FO3 canary below before trusting any other pack.",{"type":52,"tag":59,"props":293,"children":294},{},[295,297,302],{"type":57,"value":296},"The public release shape is documented in ",{"type":52,"tag":65,"props":298,"children":300},{"className":299},[],[301],{"type":57,"value":91},{"type":57,"value":303},". There is no\nruntime download API today. Don't invent one.",{"type":52,"tag":103,"props":305,"children":307},{"id":306},"steps",[308],{"type":57,"value":309},"Steps",{"type":52,"tag":311,"props":312,"children":314},"h3",{"id":313},"_0-pre-flight-the-o0o3-acf-injection-canary-mandatory-first-step",[315],{"type":57,"value":316},"0. Pre-flight: the O0\u002FO3 ACF-injection canary (mandatory first step)",{"type":52,"tag":59,"props":318,"children":319},{},[320,322,328,330,336],{"type":57,"value":321},"The most common silent failure when applying ACFs is a framework cache (Triton,\nHelion, FlashInfer's ",{"type":52,"tag":65,"props":323,"children":325},{"className":324},[],[326],{"type":57,"value":327},"flashinfer_cubin",{"type":57,"value":329},"\u002F",{"type":52,"tag":65,"props":331,"children":333},{"className":332},[],[334],{"type":57,"value":335},"flashinfer_jit_cache",{"type":57,"value":337},", NVCC build\ncache) serving a stale binary that ignored the ACF. The Debug pack has two\nACFs with predictable, opposite-direction signatures:",{"type":52,"tag":339,"props":340,"children":341},"ul",{},[342,365],{"type":52,"tag":343,"props":344,"children":345},"li",{},[346,356,358,363],{"type":52,"tag":285,"props":347,"children":348},{},[349,354],{"type":52,"tag":65,"props":350,"children":352},{"className":351},[],[353],{"type":57,"value":269},{"type":57,"value":355}," ACF",{"type":57,"value":357},": forces unoptimized compilation. Applied → expect a ",{"type":52,"tag":285,"props":359,"children":360},{},[361],{"type":57,"value":362},"measurable\nregression",{"type":57,"value":364}," (often 2-10x slower) vs. baseline.",{"type":52,"tag":343,"props":366,"children":367},{},[368,377,379,384,386,392],{"type":52,"tag":285,"props":369,"children":370},{},[371,376],{"type":52,"tag":65,"props":372,"children":374},{"className":373},[],[375],{"type":57,"value":276},{"type":57,"value":355},{"type":57,"value":378},": forces the default optimization level. Applied → expect to\n",{"type":52,"tag":285,"props":380,"children":381},{},[382],{"type":57,"value":383},"match baseline",{"type":57,"value":385}," (the no-ACF default is already ",{"type":52,"tag":65,"props":387,"children":389},{"className":388},[],[390],{"type":57,"value":391},"-O3",{"type":57,"value":170},{"type":52,"tag":394,"props":395,"children":400},"pre",{"className":396,"code":397,"language":398,"meta":399,"style":399},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","# Baseline\nT_BASE_MS=$(.\u002Frun-benchmark.sh)\n\n# O0 must regress\nPTXAS_OPTIONS=\"--apply-controls=debug-pack\u002FO0.acf\" T_O0_MS=$(.\u002Frun-benchmark.sh)\n\n# O3 must match baseline\nPTXAS_OPTIONS=\"--apply-controls=debug-pack\u002FO3.acf\" T_O3_MS=$(.\u002Frun-benchmark.sh)\n\npython -c \"\nimport sys\nbase, o0, o3 = $T_BASE_MS, $T_O0_MS, $T_O3_MS\nif o0 \u003C base * 1.05:\n    print('FAIL: O0 did not regress; ACF is NOT reaching PTXAS. Fix the cache-bust.')\n    sys.exit(1)\nif abs(o3 - base) \u002F base > 0.05:\n    print(f'WARN: O3 differs from baseline by >5%; baseline may not be -O3 or framework caching differs.')\nprint('PASS: ACF injection is wired up correctly.')\n\"\n","bash","",[401],{"type":52,"tag":65,"props":402,"children":403},{"__ignoreMap":399},[404,416,443,453,462,508,516,525,566,574,593,602,634,643,652,661,670,679,688],{"type":52,"tag":405,"props":406,"children":409},"span",{"class":407,"line":408},"line",1,[410],{"type":52,"tag":405,"props":411,"children":413},{"style":412},"--shiki-light:#90A4AE;--shiki-light-font-style:italic;--shiki-default:#546E7A;--shiki-default-font-style:italic;--shiki-dark:#676E95;--shiki-dark-font-style:italic",[414],{"type":57,"value":415},"# Baseline\n",{"type":52,"tag":405,"props":417,"children":419},{"class":407,"line":418},2,[420,426,432,438],{"type":52,"tag":405,"props":421,"children":423},{"style":422},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[424],{"type":57,"value":425},"T_BASE_MS",{"type":52,"tag":405,"props":427,"children":429},{"style":428},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[430],{"type":57,"value":431},"=$(",{"type":52,"tag":405,"props":433,"children":435},{"style":434},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[436],{"type":57,"value":437},".\u002Frun-benchmark.sh",{"type":52,"tag":405,"props":439,"children":440},{"style":428},[441],{"type":57,"value":442},")\n",{"type":52,"tag":405,"props":444,"children":446},{"class":407,"line":445},3,[447],{"type":52,"tag":405,"props":448,"children":450},{"emptyLinePlaceholder":449},true,[451],{"type":57,"value":452},"\n",{"type":52,"tag":405,"props":454,"children":456},{"class":407,"line":455},4,[457],{"type":52,"tag":405,"props":458,"children":459},{"style":412},[460],{"type":57,"value":461},"# O0 must regress\n",{"type":52,"tag":405,"props":463,"children":465},{"class":407,"line":464},5,[466,471,476,481,487,491,496,500,504],{"type":52,"tag":405,"props":467,"children":468},{"style":422},[469],{"type":57,"value":470},"PTXAS_OPTIONS",{"type":52,"tag":405,"props":472,"children":473},{"style":428},[474],{"type":57,"value":475},"=",{"type":52,"tag":405,"props":477,"children":478},{"style":428},[479],{"type":57,"value":480},"\"",{"type":52,"tag":405,"props":482,"children":484},{"style":483},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[485],{"type":57,"value":486},"--apply-controls=debug-pack\u002FO0.acf",{"type":52,"tag":405,"props":488,"children":489},{"style":428},[490],{"type":57,"value":480},{"type":52,"tag":405,"props":492,"children":493},{"style":422},[494],{"type":57,"value":495}," T_O0_MS",{"type":52,"tag":405,"props":497,"children":498},{"style":428},[499],{"type":57,"value":431},{"type":52,"tag":405,"props":501,"children":502},{"style":434},[503],{"type":57,"value":437},{"type":52,"tag":405,"props":505,"children":506},{"style":428},[507],{"type":57,"value":442},{"type":52,"tag":405,"props":509,"children":511},{"class":407,"line":510},6,[512],{"type":52,"tag":405,"props":513,"children":514},{"emptyLinePlaceholder":449},[515],{"type":57,"value":452},{"type":52,"tag":405,"props":517,"children":519},{"class":407,"line":518},7,[520],{"type":52,"tag":405,"props":521,"children":522},{"style":412},[523],{"type":57,"value":524},"# O3 must match baseline\n",{"type":52,"tag":405,"props":526,"children":527},{"class":407,"line":27},[528,532,536,540,545,549,554,558,562],{"type":52,"tag":405,"props":529,"children":530},{"style":422},[531],{"type":57,"value":470},{"type":52,"tag":405,"props":533,"children":534},{"style":428},[535],{"type":57,"value":475},{"type":52,"tag":405,"props":537,"children":538},{"style":428},[539],{"type":57,"value":480},{"type":52,"tag":405,"props":541,"children":542},{"style":483},[543],{"type":57,"value":544},"--apply-controls=debug-pack\u002FO3.acf",{"type":52,"tag":405,"props":546,"children":547},{"style":428},[548],{"type":57,"value":480},{"type":52,"tag":405,"props":550,"children":551},{"style":422},[552],{"type":57,"value":553}," T_O3_MS",{"type":52,"tag":405,"props":555,"children":556},{"style":428},[557],{"type":57,"value":431},{"type":52,"tag":405,"props":559,"children":560},{"style":434},[561],{"type":57,"value":437},{"type":52,"tag":405,"props":563,"children":564},{"style":428},[565],{"type":57,"value":442},{"type":52,"tag":405,"props":567,"children":569},{"class":407,"line":568},9,[570],{"type":52,"tag":405,"props":571,"children":572},{"emptyLinePlaceholder":449},[573],{"type":57,"value":452},{"type":52,"tag":405,"props":575,"children":577},{"class":407,"line":576},10,[578,583,588],{"type":52,"tag":405,"props":579,"children":580},{"style":434},[581],{"type":57,"value":582},"python",{"type":52,"tag":405,"props":584,"children":585},{"style":483},[586],{"type":57,"value":587}," -c",{"type":52,"tag":405,"props":589,"children":590},{"style":428},[591],{"type":57,"value":592}," \"\n",{"type":52,"tag":405,"props":594,"children":596},{"class":407,"line":595},11,[597],{"type":52,"tag":405,"props":598,"children":599},{"style":483},[600],{"type":57,"value":601},"import sys\n",{"type":52,"tag":405,"props":603,"children":605},{"class":407,"line":604},12,[606,611,616,620,625,629],{"type":52,"tag":405,"props":607,"children":608},{"style":483},[609],{"type":57,"value":610},"base, o0, o3 = ",{"type":52,"tag":405,"props":612,"children":613},{"style":422},[614],{"type":57,"value":615},"$T_BASE_MS",{"type":52,"tag":405,"props":617,"children":618},{"style":483},[619],{"type":57,"value":93},{"type":52,"tag":405,"props":621,"children":622},{"style":422},[623],{"type":57,"value":624},"$T_O0_MS",{"type":52,"tag":405,"props":626,"children":627},{"style":483},[628],{"type":57,"value":93},{"type":52,"tag":405,"props":630,"children":631},{"style":422},[632],{"type":57,"value":633},"$T_O3_MS\n",{"type":52,"tag":405,"props":635,"children":637},{"class":407,"line":636},13,[638],{"type":52,"tag":405,"props":639,"children":640},{"style":483},[641],{"type":57,"value":642},"if o0 \u003C base * 1.05:\n",{"type":52,"tag":405,"props":644,"children":646},{"class":407,"line":645},14,[647],{"type":52,"tag":405,"props":648,"children":649},{"style":483},[650],{"type":57,"value":651},"    print('FAIL: O0 did not regress; ACF is NOT reaching PTXAS. Fix the cache-bust.')\n",{"type":52,"tag":405,"props":653,"children":655},{"class":407,"line":654},15,[656],{"type":52,"tag":405,"props":657,"children":658},{"style":483},[659],{"type":57,"value":660},"    sys.exit(1)\n",{"type":52,"tag":405,"props":662,"children":664},{"class":407,"line":663},16,[665],{"type":52,"tag":405,"props":666,"children":667},{"style":483},[668],{"type":57,"value":669},"if abs(o3 - base) \u002F base > 0.05:\n",{"type":52,"tag":405,"props":671,"children":673},{"class":407,"line":672},17,[674],{"type":52,"tag":405,"props":675,"children":676},{"style":483},[677],{"type":57,"value":678},"    print(f'WARN: O3 differs from baseline by >5%; baseline may not be -O3 or framework caching differs.')\n",{"type":52,"tag":405,"props":680,"children":682},{"class":407,"line":681},18,[683],{"type":52,"tag":405,"props":684,"children":685},{"style":483},[686],{"type":57,"value":687},"print('PASS: ACF injection is wired up correctly.')\n",{"type":52,"tag":405,"props":689,"children":691},{"class":407,"line":690},19,[692],{"type":52,"tag":405,"props":693,"children":694},{"style":428},[695],{"type":57,"value":696},"\"\n",{"type":52,"tag":59,"props":698,"children":699},{},[700,702,707],{"type":57,"value":701},"If this fails, ",{"type":52,"tag":285,"props":703,"children":704},{},[705],{"type":57,"value":706},"stop",{"type":57,"value":708},". Fix the cache-bust before trying any real pack candidate:",{"type":52,"tag":339,"props":710,"children":711},{},[712,733,745,771],{"type":52,"tag":343,"props":713,"children":714},{},[715,717,723,725,731],{"type":57,"value":716},"Triton: ",{"type":52,"tag":65,"props":718,"children":720},{"className":719},[],[721],{"type":57,"value":722},"export TRITON_ALWAYS_COMPILE=1",{"type":57,"value":724},", unique ",{"type":52,"tag":65,"props":726,"children":728},{"className":727},[],[729],{"type":57,"value":730},"TRITON_CACHE_DIR",{"type":57,"value":732}," per eval.",{"type":52,"tag":343,"props":734,"children":735},{},[736,738,744],{"type":57,"value":737},"Helion: ",{"type":52,"tag":65,"props":739,"children":741},{"className":740},[],[742],{"type":57,"value":743},"export HELION_SKIP_CACHE=1",{"type":57,"value":101},{"type":52,"tag":343,"props":746,"children":747},{},[748,750,755,757,762,764,770],{"type":57,"value":749},"FlashInfer: confirm ",{"type":52,"tag":65,"props":751,"children":753},{"className":752},[],[754],{"type":57,"value":327},{"type":57,"value":756}," and ",{"type":52,"tag":65,"props":758,"children":760},{"className":759},[],[761],{"type":57,"value":335},{"type":57,"value":763}," packages are absent (",{"type":52,"tag":65,"props":765,"children":767},{"className":766},[],[768],{"type":57,"value":769},"docs\u002Fflashinfer_booster.md:56-64",{"type":57,"value":170},{"type":52,"tag":343,"props":772,"children":773},{},[774],{"type":57,"value":775},"Raw nvcc: clean the build dir between candidates.",{"type":52,"tag":311,"props":777,"children":779},{"id":778},"_1-download",[780],{"type":57,"value":781},"1. Download",{"type":52,"tag":59,"props":783,"children":784},{},[785,787,793,795,801,803,809],{"type":57,"value":786},"Browse ",{"type":52,"tag":65,"props":788,"children":790},{"className":789},[],[791],{"type":57,"value":792},"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FCompileIQ\u002Freleases",{"type":57,"value":794},", find the latest tag\nmatching ",{"type":52,"tag":65,"props":796,"children":798},{"className":797},[],[799],{"type":57,"value":800},"booster-packs-*",{"type":57,"value":802},", and download the relevant pack zip plus the\ntop-level ",{"type":52,"tag":65,"props":804,"children":806},{"className":805},[],[807],{"type":57,"value":808},"booster-pack-catalog.json",{"type":57,"value":101},{"type":52,"tag":394,"props":811,"children":813},{"className":396,"code":812,"language":398,"meta":399,"style":399},"BOOSTER_TAG=\"$(gh release list -R NVIDIA\u002FCompileIQ --limit 100 --json tagName,isDraft \\\n  --jq '.[] | select(.isDraft == false) | select(.tagName | startswith(\"booster-packs-\")) | .tagName' \\\n  | head -n 1)\"\necho \"Using $BOOSTER_TAG\"\ngh release download \"$BOOSTER_TAG\" -R NVIDIA\u002FCompileIQ -p 'booster-pack-helion.zip' -p 'booster-pack-catalog.json' -D .\u002Fpacks\nunzip .\u002Fpacks\u002Fbooster-pack-helion.zip -d .\u002Fpacks\ncat .\u002Fpacks\u002Fbooster-pack-helion\u002Fbooster-pack-manifest.json\n",[814],{"type":52,"tag":65,"props":815,"children":816},{"__ignoreMap":399},[817,860,887,915,943,1026,1048],{"type":52,"tag":405,"props":818,"children":819},{"class":407,"line":408},[820,825,829,834,839,844,850,855],{"type":52,"tag":405,"props":821,"children":822},{"style":422},[823],{"type":57,"value":824},"BOOSTER_TAG",{"type":52,"tag":405,"props":826,"children":827},{"style":428},[828],{"type":57,"value":475},{"type":52,"tag":405,"props":830,"children":831},{"style":428},[832],{"type":57,"value":833},"\"$(",{"type":52,"tag":405,"props":835,"children":836},{"style":434},[837],{"type":57,"value":838},"gh",{"type":52,"tag":405,"props":840,"children":841},{"style":483},[842],{"type":57,"value":843}," release list -R NVIDIA\u002FCompileIQ --limit ",{"type":52,"tag":405,"props":845,"children":847},{"style":846},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[848],{"type":57,"value":849},"100",{"type":52,"tag":405,"props":851,"children":852},{"style":483},[853],{"type":57,"value":854}," --json tagName,isDraft ",{"type":52,"tag":405,"props":856,"children":857},{"style":422},[858],{"type":57,"value":859},"\\\n",{"type":52,"tag":405,"props":861,"children":862},{"class":407,"line":418},[863,868,873,878,882],{"type":52,"tag":405,"props":864,"children":865},{"style":483},[866],{"type":57,"value":867},"  --jq ",{"type":52,"tag":405,"props":869,"children":870},{"style":428},[871],{"type":57,"value":872},"'",{"type":52,"tag":405,"props":874,"children":875},{"style":483},[876],{"type":57,"value":877},".[] | select(.isDraft == false) | select(.tagName | startswith(\"booster-packs-\")) | .tagName",{"type":52,"tag":405,"props":879,"children":880},{"style":428},[881],{"type":57,"value":872},{"type":52,"tag":405,"props":883,"children":884},{"style":422},[885],{"type":57,"value":886}," \\\n",{"type":52,"tag":405,"props":888,"children":889},{"class":407,"line":445},[890,895,900,905,910],{"type":52,"tag":405,"props":891,"children":892},{"style":428},[893],{"type":57,"value":894},"  |",{"type":52,"tag":405,"props":896,"children":897},{"style":434},[898],{"type":57,"value":899}," head",{"type":52,"tag":405,"props":901,"children":902},{"style":483},[903],{"type":57,"value":904}," -n ",{"type":52,"tag":405,"props":906,"children":907},{"style":846},[908],{"type":57,"value":909},"1",{"type":52,"tag":405,"props":911,"children":912},{"style":428},[913],{"type":57,"value":914},")\"\n",{"type":52,"tag":405,"props":916,"children":917},{"class":407,"line":455},[918,924,929,934,939],{"type":52,"tag":405,"props":919,"children":921},{"style":920},"--shiki-light:#6182B8;--shiki-default:#82AAFF;--shiki-dark:#82AAFF",[922],{"type":57,"value":923},"echo",{"type":52,"tag":405,"props":925,"children":926},{"style":428},[927],{"type":57,"value":928}," \"",{"type":52,"tag":405,"props":930,"children":931},{"style":483},[932],{"type":57,"value":933},"Using ",{"type":52,"tag":405,"props":935,"children":936},{"style":422},[937],{"type":57,"value":938},"$BOOSTER_TAG",{"type":52,"tag":405,"props":940,"children":941},{"style":428},[942],{"type":57,"value":696},{"type":52,"tag":405,"props":944,"children":945},{"class":407,"line":464},[946,950,955,960,964,968,972,977,982,987,992,996,1000,1004,1008,1012,1016,1021],{"type":52,"tag":405,"props":947,"children":948},{"style":434},[949],{"type":57,"value":838},{"type":52,"tag":405,"props":951,"children":952},{"style":483},[953],{"type":57,"value":954}," release",{"type":52,"tag":405,"props":956,"children":957},{"style":483},[958],{"type":57,"value":959}," download",{"type":52,"tag":405,"props":961,"children":962},{"style":428},[963],{"type":57,"value":928},{"type":52,"tag":405,"props":965,"children":966},{"style":422},[967],{"type":57,"value":938},{"type":52,"tag":405,"props":969,"children":970},{"style":428},[971],{"type":57,"value":480},{"type":52,"tag":405,"props":973,"children":974},{"style":483},[975],{"type":57,"value":976}," -R",{"type":52,"tag":405,"props":978,"children":979},{"style":483},[980],{"type":57,"value":981}," NVIDIA\u002FCompileIQ",{"type":52,"tag":405,"props":983,"children":984},{"style":483},[985],{"type":57,"value":986}," -p",{"type":52,"tag":405,"props":988,"children":989},{"style":428},[990],{"type":57,"value":991}," '",{"type":52,"tag":405,"props":993,"children":994},{"style":483},[995],{"type":57,"value":228},{"type":52,"tag":405,"props":997,"children":998},{"style":428},[999],{"type":57,"value":872},{"type":52,"tag":405,"props":1001,"children":1002},{"style":483},[1003],{"type":57,"value":986},{"type":52,"tag":405,"props":1005,"children":1006},{"style":428},[1007],{"type":57,"value":991},{"type":52,"tag":405,"props":1009,"children":1010},{"style":483},[1011],{"type":57,"value":808},{"type":52,"tag":405,"props":1013,"children":1014},{"style":428},[1015],{"type":57,"value":872},{"type":52,"tag":405,"props":1017,"children":1018},{"style":483},[1019],{"type":57,"value":1020}," -D",{"type":52,"tag":405,"props":1022,"children":1023},{"style":483},[1024],{"type":57,"value":1025}," .\u002Fpacks\n",{"type":52,"tag":405,"props":1027,"children":1028},{"class":407,"line":510},[1029,1034,1039,1044],{"type":52,"tag":405,"props":1030,"children":1031},{"style":434},[1032],{"type":57,"value":1033},"unzip",{"type":52,"tag":405,"props":1035,"children":1036},{"style":483},[1037],{"type":57,"value":1038}," .\u002Fpacks\u002Fbooster-pack-helion.zip",{"type":52,"tag":405,"props":1040,"children":1041},{"style":483},[1042],{"type":57,"value":1043}," -d",{"type":52,"tag":405,"props":1045,"children":1046},{"style":483},[1047],{"type":57,"value":1025},{"type":52,"tag":405,"props":1049,"children":1050},{"class":407,"line":518},[1051,1056],{"type":52,"tag":405,"props":1052,"children":1053},{"style":434},[1054],{"type":57,"value":1055},"cat",{"type":52,"tag":405,"props":1057,"children":1058},{"style":483},[1059],{"type":57,"value":1060}," .\u002Fpacks\u002Fbooster-pack-helion\u002Fbooster-pack-manifest.json\n",{"type":52,"tag":59,"props":1062,"children":1063},{},[1064,1066,1071],{"type":57,"value":1065},"Always read the per-pack manifest before applying: it lists the intended\nworkload, compiler version, GPU target, validation context, and known caveats.\nFor a reproducible rerun, set ",{"type":52,"tag":65,"props":1067,"children":1069},{"className":1068},[],[1070],{"type":57,"value":824},{"type":57,"value":1072}," to the exact tag printed above.",{"type":52,"tag":311,"props":1074,"children":1076},{"id":1075},"_2-apply-one-acf-at-a-time",[1077],{"type":57,"value":1078},"2. Apply one ACF at a time",{"type":52,"tag":110,"props":1080,"children":1081},{},[1082,1098],{"type":52,"tag":114,"props":1083,"children":1084},{},[1085],{"type":52,"tag":118,"props":1086,"children":1087},{},[1088,1093],{"type":52,"tag":122,"props":1089,"children":1090},{},[1091],{"type":57,"value":1092},"Target",{"type":52,"tag":122,"props":1094,"children":1095},{},[1096],{"type":57,"value":1097},"Injection",{"type":52,"tag":133,"props":1099,"children":1100},{},[1101,1118,1135,1152,1180],{"type":52,"tag":118,"props":1102,"children":1103},{},[1104,1109],{"type":52,"tag":140,"props":1105,"children":1106},{},[1107],{"type":57,"value":1108},"Raw PTXAS",{"type":52,"tag":140,"props":1110,"children":1111},{},[1112],{"type":52,"tag":65,"props":1113,"children":1115},{"className":1114},[],[1116],{"type":57,"value":1117},"ptxas -v -arch=sm_100 --apply-controls candidate.acf kernel.ptx",{"type":52,"tag":118,"props":1119,"children":1120},{},[1121,1126],{"type":52,"tag":140,"props":1122,"children":1123},{},[1124],{"type":57,"value":1125},"NVCC (CUDA source)",{"type":52,"tag":140,"props":1127,"children":1128},{},[1129],{"type":52,"tag":65,"props":1130,"children":1132},{"className":1131},[],[1133],{"type":57,"value":1134},"nvcc -Xptxas --apply-controls=candidate.acf -arch=sm_100 kernel.cu -o exe",{"type":52,"tag":118,"props":1136,"children":1137},{},[1138,1143],{"type":52,"tag":140,"props":1139,"children":1140},{},[1141],{"type":57,"value":1142},"Triton",{"type":52,"tag":140,"props":1144,"children":1145},{},[1146],{"type":52,"tag":65,"props":1147,"children":1149},{"className":1148},[],[1150],{"type":57,"value":1151},"PTXAS_OPTIONS=\"--apply-controls=candidate.acf\" TRITON_ALWAYS_COMPILE=1 python bench.py",{"type":52,"tag":118,"props":1153,"children":1154},{},[1155,1160],{"type":52,"tag":140,"props":1156,"children":1157},{},[1158],{"type":57,"value":1159},"Helion",{"type":52,"tag":140,"props":1161,"children":1162},{},[1163,1165,1171,1173,1179],{"type":57,"value":1164},"Helion's official ACF API + ",{"type":52,"tag":65,"props":1166,"children":1168},{"className":1167},[],[1169],{"type":57,"value":1170},"HELION_SKIP_CACHE=1",{"type":57,"value":1172}," (see ",{"type":52,"tag":65,"props":1174,"children":1176},{"className":1175},[],[1177],{"type":57,"value":1178},"helionlang.com\u002Fexamples\u002Facfs\u002Fsoftmax_acf.html",{"type":57,"value":170},{"type":52,"tag":118,"props":1181,"children":1182},{},[1183,1188],{"type":52,"tag":140,"props":1184,"children":1185},{},[1186],{"type":57,"value":1187},"FlashInfer",{"type":52,"tag":140,"props":1189,"children":1190},{},[1191,1197,1198,1204],{"type":52,"tag":65,"props":1192,"children":1194},{"className":1193},[],[1195],{"type":57,"value":1196},"FLASHINFER_EXTRA_CUDAFLAGS=\"--ptxas-options=--apply-controls=$ACF_FILE\" python bench.py",{"type":57,"value":1172},{"type":52,"tag":65,"props":1199,"children":1201},{"className":1200},[],[1202],{"type":57,"value":1203},"docs\u002Fflashinfer_booster.md:107",{"type":57,"value":170},{"type":52,"tag":59,"props":1206,"children":1207},{},[1208,1210,1215],{"type":57,"value":1209},"Apply exactly one ACF per run. If it fails to compile, hangs, crashes, returns\nwrong answers, or regresses, ",{"type":52,"tag":285,"props":1211,"children":1212},{},[1213],{"type":57,"value":1214},"reject",{"type":57,"value":1216}," that candidate and move to the next.",{"type":52,"tag":311,"props":1218,"children":1220},{"id":1219},"_3-validate-every-candidate",[1221],{"type":57,"value":1222},"3. Validate every candidate",{"type":52,"tag":339,"props":1224,"children":1225},{},[1226,1231,1236,1241,1246],{"type":52,"tag":343,"props":1227,"children":1228},{},[1229],{"type":57,"value":1230},"Compare against a known-good reference (correctness, not just speed).",{"type":52,"tag":343,"props":1232,"children":1233},{},[1234],{"type":57,"value":1235},"Test multiple input shapes when shape matters.",{"type":52,"tag":343,"props":1237,"children":1238},{},[1239],{"type":57,"value":1240},"Use compile and runtime timeouts to bound runaway candidates.",{"type":52,"tag":343,"props":1242,"children":1243},{},[1244],{"type":57,"value":1245},"Run multiple performance trials if the benchmark is noisy.",{"type":52,"tag":343,"props":1247,"children":1248},{},[1249],{"type":57,"value":1250},"Record the reproducibility checklist below (one row per candidate).",{"type":52,"tag":311,"props":1252,"children":1254},{"id":1253},"_4-reproducibility-log",[1255],{"type":57,"value":1256},"4. Reproducibility log",{"type":52,"tag":59,"props":1258,"children":1259},{},[1260,1262,1268],{"type":57,"value":1261},"For every candidate you accept or reject, append a row to\n",{"type":52,"tag":65,"props":1263,"children":1265},{"className":1264},[],[1266],{"type":57,"value":1267},"booster-pack-log.csv",{"type":57,"value":1269}," with:",{"type":52,"tag":339,"props":1271,"children":1272},{},[1273,1278,1283,1288,1293,1298,1316,1321,1326,1331,1336,1341],{"type":52,"tag":343,"props":1274,"children":1275},{},[1276],{"type":57,"value":1277},"ACF filename (and sha256)",{"type":52,"tag":343,"props":1279,"children":1280},{},[1281],{"type":57,"value":1282},"Manifest \u002F release version",{"type":52,"tag":343,"props":1284,"children":1285},{},[1286],{"type":57,"value":1287},"Benchmark command",{"type":52,"tag":343,"props":1289,"children":1290},{},[1291],{"type":57,"value":1292},"GPU model + driver version",{"type":52,"tag":343,"props":1294,"children":1295},{},[1296],{"type":57,"value":1297},"CTK version",{"type":52,"tag":343,"props":1299,"children":1300},{},[1301,1307,1308,1314],{"type":52,"tag":65,"props":1302,"children":1304},{"className":1303},[],[1305],{"type":57,"value":1306},"nvcc",{"type":57,"value":756},{"type":52,"tag":65,"props":1309,"children":1311},{"className":1310},[],[1312],{"type":57,"value":1313},"ptxas",{"type":57,"value":1315}," paths + versions",{"type":52,"tag":343,"props":1317,"children":1318},{},[1319],{"type":57,"value":1320},"Framework version or commit",{"type":52,"tag":343,"props":1322,"children":1323},{},[1324],{"type":57,"value":1325},"Input shape",{"type":52,"tag":343,"props":1327,"children":1328},{},[1329],{"type":57,"value":1330},"Baseline result (mean ± std)",{"type":52,"tag":343,"props":1332,"children":1333},{},[1334],{"type":57,"value":1335},"Candidate result (mean ± std)",{"type":52,"tag":343,"props":1337,"children":1338},{},[1339],{"type":57,"value":1340},"Correctness status",{"type":52,"tag":343,"props":1342,"children":1343},{},[1344,1346,1352,1354],{"type":57,"value":1345},"Decision: ",{"type":52,"tag":65,"props":1347,"children":1349},{"className":1348},[],[1350],{"type":57,"value":1351},"KEPT",{"type":57,"value":1353}," or ",{"type":52,"tag":65,"props":1355,"children":1357},{"className":1356},[],[1358],{"type":57,"value":1359},"REJECTED:\u003Creason>",{"type":52,"tag":59,"props":1361,"children":1362},{},[1363,1365,1371,1373,1379],{"type":57,"value":1364},"This is the same checklist ",{"type":52,"tag":65,"props":1366,"children":1368},{"className":1367},[],[1369],{"type":57,"value":1370},"docs\u002Fflashinfer_booster.md:135-148",{"type":57,"value":1372}," recommends.\nThe ",{"type":52,"tag":65,"props":1374,"children":1376},{"className":1375},[],[1377],{"type":57,"value":1378},"scripts\u002Fapply_one_acf.sh",{"type":57,"value":1380}," helper does most of this automatically.",{"type":52,"tag":103,"props":1382,"children":1384},{"id":1383},"self-test",[1385],{"type":57,"value":1386},"Self-test",{"type":52,"tag":394,"props":1388,"children":1390},{"className":396,"code":1389,"language":398,"meta":399,"style":399},"bash scripts\u002Fapply_one_acf.sh --self-test\n",[1391],{"type":52,"tag":65,"props":1392,"children":1393},{"__ignoreMap":399},[1394],{"type":52,"tag":405,"props":1395,"children":1396},{"class":407,"line":408},[1397,1401,1406],{"type":52,"tag":405,"props":1398,"children":1399},{"style":434},[1400],{"type":57,"value":398},{"type":52,"tag":405,"props":1402,"children":1403},{"style":483},[1404],{"type":57,"value":1405}," scripts\u002Fapply_one_acf.sh",{"type":52,"tag":405,"props":1407,"children":1408},{"style":483},[1409],{"type":57,"value":1410}," --self-test\n",{"type":52,"tag":59,"props":1412,"children":1413},{},[1414],{"type":57,"value":1415},"Dry-runs a \"baseline vs baseline\" comparison (no ACF applied to either side)\nand confirms the helper correctly reports \"NOT a real improvement\". Catches\nmisconfigured script invocations before they pollute the reproducibility log.",{"type":52,"tag":103,"props":1417,"children":1419},{"id":1418},"gotchas",[1420],{"type":57,"value":1421},"Gotchas",{"type":52,"tag":339,"props":1423,"children":1424},{},[1425,1443,1482],{"type":52,"tag":343,"props":1426,"children":1427},{},[1428,1433,1435,1441],{"type":52,"tag":285,"props":1429,"children":1430},{},[1431],{"type":57,"value":1432},"Pack name is not a hard boundary.",{"type":57,"value":1434}," Helion Pack helps some FlashInfer\ncases (",{"type":52,"tag":65,"props":1436,"children":1438},{"className":1437},[],[1439],{"type":57,"value":1440},"docs\u002Fbooster_packs.md:34",{"type":57,"value":1442},"); test before assuming.",{"type":52,"tag":343,"props":1444,"children":1445},{},[1446,1451,1453,1459,1461,1466,1468,1474,1475,1481],{"type":52,"tag":285,"props":1447,"children":1448},{},[1449],{"type":57,"value":1450},"Booster Packs are not search-space inputs.",{"type":57,"value":1452}," Don't try to feed an ACF\nthrough ",{"type":52,"tag":65,"props":1454,"children":1456},{"className":1455},[],[1457],{"type":57,"value":1458},"PtxasSearchSpace(...)",{"type":57,"value":1460},"; packs are already-generated ",{"type":52,"tag":65,"props":1462,"children":1464},{"className":1463},[],[1465],{"type":57,"value":70},{"type":57,"value":1467},"\ncandidate bundles, not inputs to ",{"type":52,"tag":65,"props":1469,"children":1471},{"className":1470},[],[1472],{"type":57,"value":1473},"PtxasSearchSpace",{"type":57,"value":1353},{"type":52,"tag":65,"props":1476,"children":1478},{"className":1477},[],[1479],{"type":57,"value":1480},"NvccSearchSpace",{"type":57,"value":101},{"type":52,"tag":343,"props":1483,"children":1484},{},[1485,1490,1492,1497],{"type":52,"tag":285,"props":1486,"children":1487},{},[1488],{"type":57,"value":1489},"Force recompilation.",{"type":57,"value":1491}," If you can't ",{"type":52,"tag":74,"props":1493,"children":1494},{},[1495],{"type":57,"value":1496},"prove",{"type":57,"value":1498}," a recompile happened between\ncandidates, don't trust the measurement. See the cache-bust hints under the\npre-flight canary.",{"type":52,"tag":103,"props":1500,"children":1502},{"id":1501},"next",[1503],{"type":57,"value":1504},"Next",{"type":52,"tag":339,"props":1506,"children":1507},{},[1508,1527],{"type":52,"tag":343,"props":1509,"children":1510},{},[1511,1513,1518,1520,1526],{"type":57,"value":1512},"If no pack candidate helps your workload, go to ",{"type":52,"tag":65,"props":1514,"children":1516},{"className":1515},[],[1517],{"type":57,"value":168},{"type":57,"value":1519}," for a\nfull CompileIQ search over ",{"type":52,"tag":65,"props":1521,"children":1523},{"className":1522},[],[1524],{"type":57,"value":1525},"PtxasSearchSpace()",{"type":57,"value":101},{"type":52,"tag":343,"props":1528,"children":1529},{},[1530,1532,1538],{"type":57,"value":1531},"For attention workloads specifically, also see\n",{"type":52,"tag":65,"props":1533,"children":1535},{"className":1534},[],[1536],{"type":57,"value":1537},"compileiq-search-space",{"type":57,"value":1539}," (variant=\"att\").",{"type":52,"tag":1541,"props":1542,"children":1543},"style",{},[1544],{"type":57,"value":1545},"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":1547,"total":1704},[1548,1566,1583,1594,1606,1619,1632,1646,1659,1670,1684,1693],{"slug":1549,"name":1549,"fn":1550,"description":1551,"org":1552,"tags":1553,"stars":1563,"repoUrl":1564,"updatedAt":1565},"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},[1554,1557,1560],{"name":1555,"slug":1556,"type":15},"Documentation","documentation",{"name":1558,"slug":1559,"type":15},"MCP","mcp",{"name":1561,"slug":1562,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1567,"name":1567,"fn":1568,"description":1569,"org":1570,"tags":1571,"stars":1580,"repoUrl":1581,"updatedAt":1582},"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},[1572,1575,1578],{"name":1573,"slug":1574,"type":15},"Containers","containers",{"name":1576,"slug":1577,"type":15},"Deployment","deployment",{"name":1579,"slug":582,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1584,"name":1584,"fn":1585,"description":1586,"org":1587,"tags":1588,"stars":1580,"repoUrl":1581,"updatedAt":1593},"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},[1589,1592],{"name":1590,"slug":1591,"type":15},"CI\u002FCD","ci-cd",{"name":1576,"slug":1577,"type":15},"2026-07-14T05:25:59.97109",{"slug":1595,"name":1595,"fn":1596,"description":1597,"org":1598,"tags":1599,"stars":1580,"repoUrl":1581,"updatedAt":1605},"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},[1600,1601,1602],{"name":1590,"slug":1591,"type":15},{"name":1576,"slug":1577,"type":15},{"name":1603,"slug":1604,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1607,"name":1607,"fn":1608,"description":1609,"org":1610,"tags":1611,"stars":1580,"repoUrl":1581,"updatedAt":1618},"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},[1612,1614,1615],{"name":1613,"slug":289,"type":15},"Debugging",{"name":1603,"slug":1604,"type":15},{"name":1616,"slug":1617,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1620,"name":1620,"fn":1621,"description":1622,"org":1623,"tags":1624,"stars":1580,"repoUrl":1581,"updatedAt":1631},"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},[1625,1628],{"name":1626,"slug":1627,"type":15},"Best Practices","best-practices",{"name":1629,"slug":1630,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1633,"name":1633,"fn":1634,"description":1635,"org":1636,"tags":1637,"stars":1580,"repoUrl":1581,"updatedAt":1645},"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},[1638,1641,1644],{"name":1639,"slug":1640,"type":15},"Machine Learning","machine-learning",{"name":1642,"slug":1643,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1647,"name":1647,"fn":1648,"description":1649,"org":1650,"tags":1651,"stars":1580,"repoUrl":1581,"updatedAt":1658},"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},[1652,1655],{"name":1653,"slug":1654,"type":15},"QA","qa",{"name":1656,"slug":1657,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1660,"name":1660,"fn":1661,"description":1662,"org":1663,"tags":1664,"stars":1580,"repoUrl":1581,"updatedAt":1669},"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},[1665,1666],{"name":1576,"slug":1577,"type":15},{"name":1667,"slug":1668,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1671,"name":1671,"fn":1672,"description":1673,"org":1674,"tags":1675,"stars":1580,"repoUrl":1581,"updatedAt":1683},"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},[1676,1679,1680],{"name":1677,"slug":1678,"type":15},"Code Review","code-review",{"name":1603,"slug":1604,"type":15},{"name":1681,"slug":1682,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1685,"name":1685,"fn":1686,"description":1687,"org":1688,"tags":1689,"stars":1580,"repoUrl":1581,"updatedAt":1692},"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},[1690,1691],{"name":1653,"slug":1654,"type":15},{"name":1656,"slug":1657,"type":15},"2026-07-14T05:25:54.928983",{"slug":1694,"name":1694,"fn":1695,"description":1696,"org":1697,"tags":1698,"stars":1580,"repoUrl":1581,"updatedAt":1703},"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},[1699,1702],{"name":1700,"slug":1701,"type":15},"Automation","automation",{"name":1590,"slug":1591,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1706,"total":518},[1707,1717,1724,1736,1749,1758,1769],{"slug":1708,"name":1708,"fn":1709,"description":1710,"org":1711,"tags":1712,"stars":23,"repoUrl":24,"updatedAt":1716},"compileiq-author-objective","author CompileIQ objective functions","Use when writing the objective_function= passed to Search(). Covers the two legal signatures (compiler-only str vs mixed list), the baseline-knockout branch, per-eval cache busting, framework-specific --apply-controls injection (raw PTXAS, NVCC, Triton, Helion, cuTeDSL\u002FFA4, FlashInfer), correctness-before-timing, INVALID_SCORE handling, and the Debug-pack O0\u002FO3 ACF-injection canary that must pass before launching a search. Triggers on \"objective function\", \"apply-controls\", \"INVALID_SCORE\", \"save_compiler_config\", \"baseline knockout\", \"BASELINE_CONFIG\", \"every config returns the same score\", \"TypeError fromhex\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1713,1714,1715],{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:32:10.176779",{"slug":4,"name":4,"fn":5,"description":6,"org":1718,"tags":1719,"stars":23,"repoUrl":24,"updatedAt":25},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1720,1721,1722,1723],{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":17,"slug":18,"type":15},{"name":13,"slug":14,"type":15},{"slug":1725,"name":1725,"fn":1726,"description":1727,"org":1728,"tags":1729,"stars":23,"repoUrl":24,"updatedAt":1735},"compileiq-bootstrap","bootstrap CompileIQ project environments","Use when starting a fresh CompileIQ project, hitting a socket timeout, or before running any other compileiq-* skill. Verifies CUDA 13.3+, ptxas, GPU access, that `from compileiq.ciq import Search` and friends resolve, and that `PtxasSearchSpace().retrieve()` returns a real path. Documents the env vars that control timeouts, caching, and search-space mirroring. Triggers on \"set up compileiq\", \"compileiq doesn't work\", \"socket timeout\", \"where do search spaces come from\", \"air-gapped compileiq\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1730,1731,1732],{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":1733,"slug":1734,"type":15},"Onboarding","onboarding","2026-07-14T05:32:11.472149",{"slug":1737,"name":1737,"fn":1738,"description":1739,"org":1740,"tags":1741,"stars":23,"repoUrl":24,"updatedAt":1748},"compileiq-debug","debug CompileIQ search and evaluation failures","Use when something is wrong: Search() hangs, all evaluations return INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors fill the log, CV% is too high, or a winning ACF candidate needs NCU profiling to explain. Symptom-indexed table on top. Triggers on \"compileiq hang\", \"socket timeout\", \"INVALID_SCORE\", \"not converging\", \"every score is the same\", \"TypeError fromhex\", \"ncu profile\", \"register spill\", \"ptxas error\", \"not in expected format\", \"high cv\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1742,1743,1746,1747],{"name":1613,"slug":289,"type":15},{"name":1744,"slug":1745,"type":15},"Evals","evals",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:32:05.08078",{"slug":168,"name":168,"fn":1750,"description":1751,"org":1752,"tags":1753,"stars":23,"repoUrl":24,"updatedAt":1757},"execute CompileIQ search workflows","Use when composing the Search(...) call and calling .start(). Covers the four worker classes (MultiProcessWorker \u002F IsoMultiProcessWorker \u002F RayWorker \u002F AsyncWorker) and when to pick each, SearchConfiguration sizing rules, dump_results checkpointing, tracker_config choice (Disabled \u002F Loguru \u002F MLflow), num_workers\u002Ftask_timeout semantics, and GPU clock locking for stable measurements. Triggers on \"Search()\", \"tuner.start()\", \"pool_size\", \"num_workers\", \"task_timeout\", \"IsoMultiProcessWorker\", \"RayWorker\", \"dump_results\", \"MLflow\", \"GPU clocks\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1754,1755,1756],{"name":1700,"slug":1701,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:32:08.913442",{"slug":1537,"name":1537,"fn":1759,"description":1760,"org":1761,"tags":1762,"stars":23,"repoUrl":24,"updatedAt":1768},"configure CompileIQ search spaces","Use when picking the search_space= argument for Search(). Covers the three provider classes (PtxasSearchSpace, NvccSearchSpace, LocalSearchSpaceBin), how to pin a version\u002Fvariant\u002Ftag, the attention-focused 'att' variant for attention kernels (FlashAttention, GQA, MHA, MLA, FlashInfer Batch Decode), air-gapped mirroring via CIQ_SEARCH_SPACES_DIR, and custom user-defined search spaces built from compileiq.search_spaces.base primitives. Triggers on \"search space\", \"PtxasSearchSpace\", \"NvccSearchSpace\", \"air-gapped compileiq\", \"offline compileiq\", \"CIQ_SEARCH_SPACES_DIR\", \"attention variant\", \"ptxas att\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1763,1766,1767],{"name":1764,"slug":1765,"type":15},"Configuration","configuration",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:32:07.605639",{"slug":1770,"name":1770,"fn":1771,"description":1772,"org":1773,"tags":1774,"stars":23,"repoUrl":24,"updatedAt":1781},"compileiq-validate-result","validate CompileIQ search results","Use AFTER a Search has completed and BEFORE claiming any speedup or shipping an ACF. Loads the dump_results CSV, extracts top-K candidates (single-objective) or the Pareto front (multi-objective), re-measures each against the no-ACF baseline with 100+ trials on fresh caches, runs Welch's t-test plus Cohen's d, rejects three classic false-positive patterns (lucky-min \u002F higher-variance \u002F multiple-comparisons-of-N), and saves the validated winner as best.acf. Triggers on \"validate result\", \"extract best config\", \"Welch's t-test\", \"is my speedup real\", \"save best ACF\", \"pareto front\", \"claim speedup\", \"ship config\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1775,1778,1779,1780],{"name":1776,"slug":1777,"type":15},"Data Analysis","data-analysis",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},{"name":1653,"slug":1654,"type":15},"2026-07-14T05:32:06.343444"]