[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-warp-compile-time-optimizer":3,"mdc-3dktzm-key":37,"related-org-nvidia-warp-compile-time-optimizer":1697,"related-repo-nvidia-warp-compile-time-optimizer":1857},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":32,"sourceUrl":35,"mdContent":36},"warp-compile-time-optimizer","optimize Warp compile and startup times","Use when compile time or startup time is the problem in code that uses Warp: a request to improve, optimize, or cut compile times; an app that is slow to start or stalls at the first wp.launch; seconds of compiling before real work begins; JIT modules recompiling on every run or every CI job. Only applies when the code being optimized uses Warp kernels. Not for steady-state kernel runtime, memory, correctness, building Warp itself from source, or nvcc\u002FC++ build times.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,17],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Engineering","engineering",6864,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fwarp","2026-08-05T05:58:13.178762","Apache-2.0",555,[26,27,28,29,8,30,31],"cuda","differentiable-programming","gpu","gpu-acceleration","nvidia-warp","python",{"repoUrl":21,"stars":20,"forks":24,"topics":33,"description":34},[26,27,28,29,8,30,31],"A Python framework for GPU-accelerated simulation, robotics, and machine learning.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fwarp\u002Ftree\u002FHEAD\u002Fskills\u002Fwarp-compile-time-optimizer","---\nname: warp-compile-time-optimizer\ndescription: >-\n  Use when compile time or startup time is the problem in code that uses Warp:\n  a request to improve, optimize, or cut compile times; an app that is slow to\n  start or stalls at the first wp.launch; seconds of compiling before real work\n  begins; JIT modules recompiling on every run or every CI job. Only applies\n  when the code being optimized uses Warp kernels. Not for steady-state kernel\n  runtime, memory, correctness, building Warp itself from source, or nvcc\u002FC++\n  build times.\nlicense: Apache-2.0\ncompatibility: Requires Python 3.10+ and an installed warp-lang package. A CUDA device is needed to diagnose CUDA-specific mechanisms.\nallowed-tools: Bash, Read, Edit, Write, Glob, Grep, env\nmetadata:\n  version: \"0.1.0\"\n  author: \"Warp Team \u003Cwarp-python@nvidia.com>\"\n  tags:\n    - warp\n    - compilation\n    - cold-start\n    - startup-latency\n    - kernel-cache\n    - gpu\n---\n\n# Warp cold-start compile time\n\n## Start with a runnable command\n\nThe probe runs the target in a subprocess, writes only to temporary\ndirectories, and needs no network, external tool servers, or Warp checkout.\n\n| Script | Purpose | Arguments |\n| --- | --- | --- |\n| `scripts\u002Fwarp_compile_probe.py` | Measure isolated cold\u002Fwarm compilation and launches. | `measure [OPTIONS] -- COMMAND...`; use `--help`. |\n\nUse `run_script(\"scripts\u002Fwarp_compile_probe.py\", args=[...])` when supported;\notherwise use the Python command below. The target must run to completion.\n\n## Compilation model\n\nWarp compiles modules, not individual kernels. A module's identity is:\n\n```text\n(live kernel & function set) x (module options) x (CUDA block_dim) x (generic instances)\n```\n\nEach identity requires code generation and native compilation for the full\nmodule.\n\nCold-start cost is roughly:\n\n```text\nnumber of distinct module identities you touch  x  size of each module\n```\n\nReduce it in two ways:\n\n1. Stop identity churn. A module that changes after loading compiles again.\n2. Stop source duplication. A module used at three block dimensions compiles\n   every kernel three times.\n\nDeleting one kernel from a module that still builds saves only part of one\nbuild. Removing an unnecessary module identity saves the full build.\n\nWhen neither applies, overlap independent CUDA builds (CS-13). This changes\nwhen work happens, not how much is compiled, so judge it on elapsed time.\n\n## When a module's options are fixed\n\nOptions have two deadlines:\n\n1. **Module creation (Python import).** The module copies `enable_backward`,\n   `max_unroll`, `lineinfo`, `deterministic`, `deterministic_max_records`, and\n   `compile_time_trace` from `warp.config`. Setting a global later is silently\n   ignored by that module. `default_grid_stride` is the exception.\n2. **First load.** Before compilation, change the existing module with\n   `wp.set_module_options()` or `wp.get_module(name).options`. Changing an\n   option after load creates a new identity and rebuilds the module (CS-3).\n\n| Missed deadline | Symptom | Cost |\n| --- | --- | --- |\n| `wp.config.*` set after import | hash unchanged, option silently absent | the entire benefit, invisibly |\n| module options set after load | a second hash, module builds twice | one extra build, visible in the trace |\n\nAfter changing an option, confirm the hash moved for every target module. An\nunchanged hash means the option never arrived.\n\n## Preserve behavior\n\nChange how Warp compiles the code, not the workload.\n\nDo not delete or merge kernels to claim a gain. Apparently redundant stages\nmay preserve ownership, aliasing, retained outputs, numerical boundaries, or\nAPI behavior. Fix duplication at the module level.\n\nPreserve every launch and its order, dimensions, dtypes, devices, block\ndimensions, gradients, numerical modes, dynamic\u002Fplugin behavior, and public\nAPI signatures. Keep kernel names when moving definitions to module scope\nbecause logs, cache artifacts, and external tools expose them.\n\n## Instructions\n\n### 1. Find the real cost, and confirm it is compilation\n\nAsk what command the user actually waits on, then measure it cold:\n\n```bash\npython scripts\u002Fwarp_compile_probe.py measure --samples 3 \\\n    --json baseline.json -- \u003Cthe user's command>\n```\n\nThe probe gives each sample private `WARP_CACHE_PATH`, `WARP_CACHE_ROOT`, and\n`CUDA_CACHE_PATH` directories, enables module timers, and records launches.\nNever clear a live cache with `wp.clear_kernel_cache()` or\n`wp.clear_lto_cache()`; clearing is not isolated and can disrupt other\nprocesses.\n\nRead the probe output before source. If compilation is a small part of wall\ntime, report the real bottleneck and stop. For libraries and tests, use the\nsmallest command that compiles the workload's modules.\n\nModules that each compiled once, with no repeated hashes, block-dimension\nvariants, or LTO, have no structural churn. This rules out redundant builds,\nnot oversized builds; still check cache reuse (CS-2), backward codegen\n(CS-10), unrolling (CS-11), the precompiled header (CS-12), and overlap when\nseveral CUDA modules remain (CS-13).\n\nEvery sample also re-runs the command against the cache it just populated. If\nwarm module work is not near zero, diagnose cache reuse (CS-2) before changing\nmodule structure.\n\n### 2. Ask about runtime tradeoffs when needed\n\nApply ordering fixes, lifecycle grouping, and option hoists without asking.\nAsk before changing `fast_math`, `max_unroll`, or a MathDx\u002Ftile implementation:\n\n> Some of these knobs cut compile time but can make the compiled kernels\n> slower or change numerics. Are you optimizing a fast edit-run loop (where\n> slower kernels are usually fine), or production startup (where they usually\n> are not)?\n\nIf the user is unavailable:\n\n- Leave numerics-changing options and implementation swaps alone.\n- Before removing a capability such as backward codegen, test whether it is\n  used; report what is removed, the measured benefit, and how to revert.\n- Set global `wp.config.*` options at application entry points, not in library\n  code.\n\nRecord declined options and their measured benefits in the step 6 ledger.\n\n#### Match the scope of the change to the scope of the evidence\n\nA profile supports a change to the measured application, not every consumer of\na shared library. Repository searches also miss out-of-tree and future callers.\nFor example, a forward-only application does not justify disabling gradients\ninside a solver library that another application differentiates through.\n\nScope the option to the measured process, before importing the library:\n\n```python\nimport warp as wp\n\nwp.config.enable_backward = False   # must precede the library import\n\nimport the_library\n```\n\nThis also reaches every module the application loads. CS-10 covers the silent\nimport-order trap. If only a library change works, send its maintainers the\nmeasurement and let them decide the contract.\n\n### 3. Diagnose from the measurement, not from reading the source\n\nThe probe prints every compiled module identity with its name, hash, device,\nand block dimension, then names which modules built more than once. Match what\nyou see:\n\n| What the probe shows | What it means | Where to look |\n| --- | --- | --- |\n| One module name, several **hashes** | Identity churn: its kernel set, options, or generic instances changed after it first loaded | CS-1, CS-3, CS-6 |\n| One module name, several **block_dim** values | The whole module is recompiled per block dimension (CUDA) | CS-5 |\n| Many one-kernel modules in one feature | Fixed per-module cost repeated | CS-4 |\n| A hash-named module per kernel | `module=\"unique\"` used on stable kernels | CS-9 |\n| Big gap between module time and native compile time, plus `.lto` artifacts | MathDx\u002FLTO setup | CS-7 |\n| `(compiled)` on a run that should have been warm | Cache is not being reused | CS-2 |\n| Modules load, then \"Failed to find module\" | Concurrent CPU JIT first-use race | CS-8 |\n| Large generated source, no rebuild problem | Unroll budget | CS-11 |\n| Adjoint code in a module nothing differentiates | Backward codegen | CS-10 |\n| Compiles slow across the board, or a few small modules on CUDA below toolkit 13 | The precompiled header is turned off, or is not paying for itself | CS-12 |\n| Several independent modules, each built once, `overlap_factor` near 1.0 | Builds are running one at a time; parallel loading is off by default | CS-13 |\n| An option you set changed nothing, and that module's hash is unchanged | It was assigned after the module was created, so it never arrived | \"When a module's options are fixed\" |\n| No row above fires | Nothing is being built redundantly; the cost is the size of the builds themselves | Step 6 |\n\n`references\u002Fmechanisms.md` has one section per mechanism: how to confirm it,\nthe fix, its limits, and its failure mode. Read only the sections selected by\nthe measurement.\n\n### 4. Choose module boundaries deliberately\n\nGroup kernels in one module only when they share:\n\n- lifecycle: they are defined, loaded, and invalidated together;\n- option set: they need the same `fast_math`, `enable_backward`, `max_unroll`,\n  and MathDx settings;\n- stable block dimension on CUDA.\n\nKernels with the same lifecycle but different stable block dimensions should\nnot share a module because each would compile twice. Separate kernels with\nindependent lifecycles too.\n\nKernels whose block dimension varies at runtime (chosen from input size, say)\nhave no stable mapping, so keep them in their own module rather than dragging\na whole shared module into an extra variant.\n\nPrefer the least invasive change that removes a build. Ordering fixes and\noption hoists are cheaper and safer than re-architecting module layout;\nregroup only when fixed per-module cost or block-dimension duplication\ndominates.\n\n`wp.set_module_options()` targets its calling Python module, not kernels with\nan explicit `module=\"pkg.name\"`. Either use a real Python module or update the\nnamed module before it loads:\n\n```python\nwp.set_module_options({\"enable_backward\": False})  # at module scope\n\nwp.get_module(\"pkg.name\").options.update({\"enable_backward\": False})\n```\n\nDo not pass `wp.get_module()` to `wp.set_module_options(module=...)`, or use\n`@wp.kernel(module_options={...})` without `module=\"unique\"`. Per-kernel\n`enable_backward=False` has a tile-module exception covered by CS-10. Confirm\nthe module hash after every option change.\n\n### 5. Verify\n\n```bash\npython scripts\u002Fwarp_compile_probe.py measure --samples 3 \\\n    --json candidate.json -- \u003Cthe same command>\npython scripts\u002Fwarp_compile_probe.py compare baseline.json candidate.json\n```\n\n`compare` rejects changed launch topology and treats a result inside\n`max(1% of baseline, 2 x baseline MAD)` as inconclusive.\n\nFor `BUILDS OVERLAPPED`, judge scheduling changes on compile elapsed rather\nthan summed module timers. The required warm pass supplies that clock. See\n`references\u002Fmeasurement.md`.\n\nThen check what the probe cannot see:\n\n- Diff numeric output and run the project's tests or entry points.\n- After changing `enable_backward` or boundaries, verify a gradient path.\n- After changing `fast_math`, `max_unroll`, MathDx, or an implementation,\n  benchmark steady-state runtime.\n- Exercise uncovered dynamic kernels, dtypes, profiles, and modes.\n\n### 6. Report results\n\nRead \"Reporting results\" in `references\u002Fmeasurement.md`. Report:\n\nDescribe every optimization in plain language: name the behavior, the evidence,\nand the effect. For example, write \"moved module options before the first load\nto avoid a redundant rebuild,\" not \"applied CS-3.\" Treat `CS-*` labels as\ninternal navigation aids, not user-facing explanations.\n\n| Option | Measured | Why not taken | To take it |\n| --- | ---: | --- | --- |\n| `enable_backward=False` on `pkg.solver` | −38% cold | a live tape traverses these kernels | set at the entry point, then re-check adjoints |\n| `max_unroll=4` | −2%, inside noise | changes generated code for no measured gain | — |\n\n- before\u002Fafter medians and sample counts;\n- the reduction and residual cost, sized against the original complaint;\n- mechanisms fixed, ruled out, measured and declined, or not reached;\n- tradeoffs and behavior not verified;\n- the ledger above for declined or incompletely verified options;\n- the measured value of relaxing any constraint that blocked a fix.\n\nState only what the evidence supports. \"No structural churn\" does not mean\n\"optimal\" or \"irreducible.\" Measure declined levers when practical; label any\nestimate untested. Before reporting no available fix, check CS-13. For a\npossible module split, first measure a one-kernel module with the same options\nto establish the repeated fixed cost.\n\n## Troubleshooting\n\nRun the target command directly before debugging the probe. See\n`references\u002Fmeasurement.md` for cache\u002Fnoise issues and\n`references\u002Fmechanisms.md` for mechanism-specific failures.\n\n## Limitations\n\nTwo rules override any gain:\n\n- Isolate both Warp and CUDA caches for every cold sample.\n- Keep `max_workers \u003C= 1` when a load can target CPU, including `device=None`\n  and mixed device lists. Concurrent CPU first loads can lose kernels; retries\n  do not make them safe. CUDA-only loading is unaffected.\n\nMeasurements are environment-specific: cold times move with CPU, GPU, driver,\ntoolchain, and Warp version. Mechanisms transfer; numbers do not. Read the\nknown unknowns in `references\u002Fmechanisms.md` before making broad claims.\n\n## Reference files\n\n- `references\u002Fmechanisms.md`: the thirteen compile-time mechanisms, each with\n  its confirming signal, fix, applicability limits, and failure mode. Read the\n  sections your measurement points to.\n- `references\u002Fmeasurement.md`: measurement protocol, what each metric does\n  and does not mean, reporting guidance, log examples, and manual measurement.\n",{"data":38,"body":50},{"name":4,"description":6,"license":23,"compatibility":39,"allowed-tools":40,"metadata":41},"Requires Python 3.10+ and an installed warp-lang package. A CUDA device is needed to diagnose CUDA-specific mechanisms.","Bash, Read, Edit, Write, Glob, Grep, env",{"version":42,"author":43,"tags":44},"0.1.0","Warp Team \u003Cwarp-python@nvidia.com>",[45,46,47,48,49,28],"warp","compilation","cold-start","startup-latency","kernel-cache",{"type":51,"children":52},"root",[53,62,69,75,145,158,164,169,181,186,191,200,205,220,225,230,236,241,343,412,417,423,428,433,438,444,451,456,549,592,597,602,607,613,632,641,646,672,677,684,689,694,745,750,756,761,1061,1072,1078,1083,1120,1125,1130,1135,1153,1183,1227,1233,1336,1355,1375,1380,1422,1428,1440,1453,1547,1580,1585,1591,1610,1616,1621,1650,1662,1668,1691],{"type":54,"tag":55,"props":56,"children":58},"element","h1",{"id":57},"warp-cold-start-compile-time",[59],{"type":60,"value":61},"text","Warp cold-start compile time",{"type":54,"tag":63,"props":64,"children":66},"h2",{"id":65},"start-with-a-runnable-command",[67],{"type":60,"value":68},"Start with a runnable command",{"type":54,"tag":70,"props":71,"children":72},"p",{},[73],{"type":60,"value":74},"The probe runs the target in a subprocess, writes only to temporary\ndirectories, and needs no network, external tool servers, or Warp checkout.",{"type":54,"tag":76,"props":77,"children":78},"table",{},[79,103],{"type":54,"tag":80,"props":81,"children":82},"thead",{},[83],{"type":54,"tag":84,"props":85,"children":86},"tr",{},[87,93,98],{"type":54,"tag":88,"props":89,"children":90},"th",{},[91],{"type":60,"value":92},"Script",{"type":54,"tag":88,"props":94,"children":95},{},[96],{"type":60,"value":97},"Purpose",{"type":54,"tag":88,"props":99,"children":100},{},[101],{"type":60,"value":102},"Arguments",{"type":54,"tag":104,"props":105,"children":106},"tbody",{},[107],{"type":54,"tag":84,"props":108,"children":109},{},[110,121,126],{"type":54,"tag":111,"props":112,"children":113},"td",{},[114],{"type":54,"tag":115,"props":116,"children":118},"code",{"className":117},[],[119],{"type":60,"value":120},"scripts\u002Fwarp_compile_probe.py",{"type":54,"tag":111,"props":122,"children":123},{},[124],{"type":60,"value":125},"Measure isolated cold\u002Fwarm compilation and launches.",{"type":54,"tag":111,"props":127,"children":128},{},[129,135,137,143],{"type":54,"tag":115,"props":130,"children":132},{"className":131},[],[133],{"type":60,"value":134},"measure [OPTIONS] -- COMMAND...",{"type":60,"value":136},"; use ",{"type":54,"tag":115,"props":138,"children":140},{"className":139},[],[141],{"type":60,"value":142},"--help",{"type":60,"value":144},".",{"type":54,"tag":70,"props":146,"children":147},{},[148,150,156],{"type":60,"value":149},"Use ",{"type":54,"tag":115,"props":151,"children":153},{"className":152},[],[154],{"type":60,"value":155},"run_script(\"scripts\u002Fwarp_compile_probe.py\", args=[...])",{"type":60,"value":157}," when supported;\notherwise use the Python command below. The target must run to completion.",{"type":54,"tag":63,"props":159,"children":161},{"id":160},"compilation-model",[162],{"type":60,"value":163},"Compilation model",{"type":54,"tag":70,"props":165,"children":166},{},[167],{"type":60,"value":168},"Warp compiles modules, not individual kernels. A module's identity is:",{"type":54,"tag":170,"props":171,"children":176},"pre",{"className":172,"code":174,"language":60,"meta":175},[173],"language-text","(live kernel & function set) x (module options) x (CUDA block_dim) x (generic instances)\n","",[177],{"type":54,"tag":115,"props":178,"children":179},{"__ignoreMap":175},[180],{"type":60,"value":174},{"type":54,"tag":70,"props":182,"children":183},{},[184],{"type":60,"value":185},"Each identity requires code generation and native compilation for the full\nmodule.",{"type":54,"tag":70,"props":187,"children":188},{},[189],{"type":60,"value":190},"Cold-start cost is roughly:",{"type":54,"tag":170,"props":192,"children":195},{"className":193,"code":194,"language":60,"meta":175},[173],"number of distinct module identities you touch  x  size of each module\n",[196],{"type":54,"tag":115,"props":197,"children":198},{"__ignoreMap":175},[199],{"type":60,"value":194},{"type":54,"tag":70,"props":201,"children":202},{},[203],{"type":60,"value":204},"Reduce it in two ways:",{"type":54,"tag":206,"props":207,"children":208},"ol",{},[209,215],{"type":54,"tag":210,"props":211,"children":212},"li",{},[213],{"type":60,"value":214},"Stop identity churn. A module that changes after loading compiles again.",{"type":54,"tag":210,"props":216,"children":217},{},[218],{"type":60,"value":219},"Stop source duplication. A module used at three block dimensions compiles\nevery kernel three times.",{"type":54,"tag":70,"props":221,"children":222},{},[223],{"type":60,"value":224},"Deleting one kernel from a module that still builds saves only part of one\nbuild. Removing an unnecessary module identity saves the full build.",{"type":54,"tag":70,"props":226,"children":227},{},[228],{"type":60,"value":229},"When neither applies, overlap independent CUDA builds (CS-13). This changes\nwhen work happens, not how much is compiled, so judge it on elapsed time.",{"type":54,"tag":63,"props":231,"children":233},{"id":232},"when-a-modules-options-are-fixed",[234],{"type":60,"value":235},"When a module's options are fixed",{"type":54,"tag":70,"props":237,"children":238},{},[239],{"type":60,"value":240},"Options have two deadlines:",{"type":54,"tag":206,"props":242,"children":243},{},[244,317],{"type":54,"tag":210,"props":245,"children":246},{},[247,253,255,261,263,269,271,277,278,284,285,291,293,299,301,307,309,315],{"type":54,"tag":248,"props":249,"children":250},"strong",{},[251],{"type":60,"value":252},"Module creation (Python import).",{"type":60,"value":254}," The module copies ",{"type":54,"tag":115,"props":256,"children":258},{"className":257},[],[259],{"type":60,"value":260},"enable_backward",{"type":60,"value":262},",\n",{"type":54,"tag":115,"props":264,"children":266},{"className":265},[],[267],{"type":60,"value":268},"max_unroll",{"type":60,"value":270},", ",{"type":54,"tag":115,"props":272,"children":274},{"className":273},[],[275],{"type":60,"value":276},"lineinfo",{"type":60,"value":270},{"type":54,"tag":115,"props":279,"children":281},{"className":280},[],[282],{"type":60,"value":283},"deterministic",{"type":60,"value":270},{"type":54,"tag":115,"props":286,"children":288},{"className":287},[],[289],{"type":60,"value":290},"deterministic_max_records",{"type":60,"value":292},", and\n",{"type":54,"tag":115,"props":294,"children":296},{"className":295},[],[297],{"type":60,"value":298},"compile_time_trace",{"type":60,"value":300}," from ",{"type":54,"tag":115,"props":302,"children":304},{"className":303},[],[305],{"type":60,"value":306},"warp.config",{"type":60,"value":308},". Setting a global later is silently\nignored by that module. ",{"type":54,"tag":115,"props":310,"children":312},{"className":311},[],[313],{"type":60,"value":314},"default_grid_stride",{"type":60,"value":316}," is the exception.",{"type":54,"tag":210,"props":318,"children":319},{},[320,325,327,333,335,341],{"type":54,"tag":248,"props":321,"children":322},{},[323],{"type":60,"value":324},"First load.",{"type":60,"value":326}," Before compilation, change the existing module with\n",{"type":54,"tag":115,"props":328,"children":330},{"className":329},[],[331],{"type":60,"value":332},"wp.set_module_options()",{"type":60,"value":334}," or ",{"type":54,"tag":115,"props":336,"children":338},{"className":337},[],[339],{"type":60,"value":340},"wp.get_module(name).options",{"type":60,"value":342},". Changing an\noption after load creates a new identity and rebuilds the module (CS-3).",{"type":54,"tag":76,"props":344,"children":345},{},[346,367],{"type":54,"tag":80,"props":347,"children":348},{},[349],{"type":54,"tag":84,"props":350,"children":351},{},[352,357,362],{"type":54,"tag":88,"props":353,"children":354},{},[355],{"type":60,"value":356},"Missed deadline",{"type":54,"tag":88,"props":358,"children":359},{},[360],{"type":60,"value":361},"Symptom",{"type":54,"tag":88,"props":363,"children":364},{},[365],{"type":60,"value":366},"Cost",{"type":54,"tag":104,"props":368,"children":369},{},[370,394],{"type":54,"tag":84,"props":371,"children":372},{},[373,384,389],{"type":54,"tag":111,"props":374,"children":375},{},[376,382],{"type":54,"tag":115,"props":377,"children":379},{"className":378},[],[380],{"type":60,"value":381},"wp.config.*",{"type":60,"value":383}," set after import",{"type":54,"tag":111,"props":385,"children":386},{},[387],{"type":60,"value":388},"hash unchanged, option silently absent",{"type":54,"tag":111,"props":390,"children":391},{},[392],{"type":60,"value":393},"the entire benefit, invisibly",{"type":54,"tag":84,"props":395,"children":396},{},[397,402,407],{"type":54,"tag":111,"props":398,"children":399},{},[400],{"type":60,"value":401},"module options set after load",{"type":54,"tag":111,"props":403,"children":404},{},[405],{"type":60,"value":406},"a second hash, module builds twice",{"type":54,"tag":111,"props":408,"children":409},{},[410],{"type":60,"value":411},"one extra build, visible in the trace",{"type":54,"tag":70,"props":413,"children":414},{},[415],{"type":60,"value":416},"After changing an option, confirm the hash moved for every target module. An\nunchanged hash means the option never arrived.",{"type":54,"tag":63,"props":418,"children":420},{"id":419},"preserve-behavior",[421],{"type":60,"value":422},"Preserve behavior",{"type":54,"tag":70,"props":424,"children":425},{},[426],{"type":60,"value":427},"Change how Warp compiles the code, not the workload.",{"type":54,"tag":70,"props":429,"children":430},{},[431],{"type":60,"value":432},"Do not delete or merge kernels to claim a gain. Apparently redundant stages\nmay preserve ownership, aliasing, retained outputs, numerical boundaries, or\nAPI behavior. Fix duplication at the module level.",{"type":54,"tag":70,"props":434,"children":435},{},[436],{"type":60,"value":437},"Preserve every launch and its order, dimensions, dtypes, devices, block\ndimensions, gradients, numerical modes, dynamic\u002Fplugin behavior, and public\nAPI signatures. Keep kernel names when moving definitions to module scope\nbecause logs, cache artifacts, and external tools expose them.",{"type":54,"tag":63,"props":439,"children":441},{"id":440},"instructions",[442],{"type":60,"value":443},"Instructions",{"type":54,"tag":445,"props":446,"children":448},"h3",{"id":447},"_1-find-the-real-cost-and-confirm-it-is-compilation",[449],{"type":60,"value":450},"1. Find the real cost, and confirm it is compilation",{"type":54,"tag":70,"props":452,"children":453},{},[454],{"type":60,"value":455},"Ask what command the user actually waits on, then measure it cold:",{"type":54,"tag":170,"props":457,"children":461},{"className":458,"code":459,"language":460,"meta":175,"style":175},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","python scripts\u002Fwarp_compile_probe.py measure --samples 3 \\\n    --json baseline.json -- \u003Cthe user's command>\n","bash",[462],{"type":54,"tag":115,"props":463,"children":464},{"__ignoreMap":175},[465,504],{"type":54,"tag":466,"props":467,"children":470},"span",{"class":468,"line":469},"line",1,[471,476,482,487,492,498],{"type":54,"tag":466,"props":472,"children":474},{"style":473},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[475],{"type":60,"value":31},{"type":54,"tag":466,"props":477,"children":479},{"style":478},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[480],{"type":60,"value":481}," scripts\u002Fwarp_compile_probe.py",{"type":54,"tag":466,"props":483,"children":484},{"style":478},[485],{"type":60,"value":486}," measure",{"type":54,"tag":466,"props":488,"children":489},{"style":478},[490],{"type":60,"value":491}," --samples",{"type":54,"tag":466,"props":493,"children":495},{"style":494},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[496],{"type":60,"value":497}," 3",{"type":54,"tag":466,"props":499,"children":501},{"style":500},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[502],{"type":60,"value":503}," \\\n",{"type":54,"tag":466,"props":505,"children":507},{"class":468,"line":506},2,[508,513,518,523,529,534,539,544],{"type":54,"tag":466,"props":509,"children":510},{"style":478},[511],{"type":60,"value":512},"    --json",{"type":54,"tag":466,"props":514,"children":515},{"style":478},[516],{"type":60,"value":517}," baseline.json",{"type":54,"tag":466,"props":519,"children":520},{"style":478},[521],{"type":60,"value":522}," --",{"type":54,"tag":466,"props":524,"children":526},{"style":525},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[527],{"type":60,"value":528}," \u003C",{"type":54,"tag":466,"props":530,"children":531},{"style":478},[532],{"type":60,"value":533},"the",{"type":54,"tag":466,"props":535,"children":536},{"style":478},[537],{"type":60,"value":538}," user",{"type":54,"tag":466,"props":540,"children":541},{"style":525},[542],{"type":60,"value":543},"'",{"type":54,"tag":466,"props":545,"children":546},{"style":478},[547],{"type":60,"value":548},"s command>\n",{"type":54,"tag":70,"props":550,"children":551},{},[552,554,560,561,567,568,574,576,582,584,590],{"type":60,"value":553},"The probe gives each sample private ",{"type":54,"tag":115,"props":555,"children":557},{"className":556},[],[558],{"type":60,"value":559},"WARP_CACHE_PATH",{"type":60,"value":270},{"type":54,"tag":115,"props":562,"children":564},{"className":563},[],[565],{"type":60,"value":566},"WARP_CACHE_ROOT",{"type":60,"value":292},{"type":54,"tag":115,"props":569,"children":571},{"className":570},[],[572],{"type":60,"value":573},"CUDA_CACHE_PATH",{"type":60,"value":575}," directories, enables module timers, and records launches.\nNever clear a live cache with ",{"type":54,"tag":115,"props":577,"children":579},{"className":578},[],[580],{"type":60,"value":581},"wp.clear_kernel_cache()",{"type":60,"value":583}," or\n",{"type":54,"tag":115,"props":585,"children":587},{"className":586},[],[588],{"type":60,"value":589},"wp.clear_lto_cache()",{"type":60,"value":591},"; clearing is not isolated and can disrupt other\nprocesses.",{"type":54,"tag":70,"props":593,"children":594},{},[595],{"type":60,"value":596},"Read the probe output before source. If compilation is a small part of wall\ntime, report the real bottleneck and stop. For libraries and tests, use the\nsmallest command that compiles the workload's modules.",{"type":54,"tag":70,"props":598,"children":599},{},[600],{"type":60,"value":601},"Modules that each compiled once, with no repeated hashes, block-dimension\nvariants, or LTO, have no structural churn. This rules out redundant builds,\nnot oversized builds; still check cache reuse (CS-2), backward codegen\n(CS-10), unrolling (CS-11), the precompiled header (CS-12), and overlap when\nseveral CUDA modules remain (CS-13).",{"type":54,"tag":70,"props":603,"children":604},{},[605],{"type":60,"value":606},"Every sample also re-runs the command against the cache it just populated. If\nwarm module work is not near zero, diagnose cache reuse (CS-2) before changing\nmodule structure.",{"type":54,"tag":445,"props":608,"children":610},{"id":609},"_2-ask-about-runtime-tradeoffs-when-needed",[611],{"type":60,"value":612},"2. Ask about runtime tradeoffs when needed",{"type":54,"tag":70,"props":614,"children":615},{},[616,618,624,625,630],{"type":60,"value":617},"Apply ordering fixes, lifecycle grouping, and option hoists without asking.\nAsk before changing ",{"type":54,"tag":115,"props":619,"children":621},{"className":620},[],[622],{"type":60,"value":623},"fast_math",{"type":60,"value":270},{"type":54,"tag":115,"props":626,"children":628},{"className":627},[],[629],{"type":60,"value":268},{"type":60,"value":631},", or a MathDx\u002Ftile implementation:",{"type":54,"tag":633,"props":634,"children":635},"blockquote",{},[636],{"type":54,"tag":70,"props":637,"children":638},{},[639],{"type":60,"value":640},"Some of these knobs cut compile time but can make the compiled kernels\nslower or change numerics. Are you optimizing a fast edit-run loop (where\nslower kernels are usually fine), or production startup (where they usually\nare not)?",{"type":54,"tag":70,"props":642,"children":643},{},[644],{"type":60,"value":645},"If the user is unavailable:",{"type":54,"tag":647,"props":648,"children":649},"ul",{},[650,655,660],{"type":54,"tag":210,"props":651,"children":652},{},[653],{"type":60,"value":654},"Leave numerics-changing options and implementation swaps alone.",{"type":54,"tag":210,"props":656,"children":657},{},[658],{"type":60,"value":659},"Before removing a capability such as backward codegen, test whether it is\nused; report what is removed, the measured benefit, and how to revert.",{"type":54,"tag":210,"props":661,"children":662},{},[663,665,670],{"type":60,"value":664},"Set global ",{"type":54,"tag":115,"props":666,"children":668},{"className":667},[],[669],{"type":60,"value":381},{"type":60,"value":671}," options at application entry points, not in library\ncode.",{"type":54,"tag":70,"props":673,"children":674},{},[675],{"type":60,"value":676},"Record declined options and their measured benefits in the step 6 ledger.",{"type":54,"tag":678,"props":679,"children":681},"h4",{"id":680},"match-the-scope-of-the-change-to-the-scope-of-the-evidence",[682],{"type":60,"value":683},"Match the scope of the change to the scope of the evidence",{"type":54,"tag":70,"props":685,"children":686},{},[687],{"type":60,"value":688},"A profile supports a change to the measured application, not every consumer of\na shared library. Repository searches also miss out-of-tree and future callers.\nFor example, a forward-only application does not justify disabling gradients\ninside a solver library that another application differentiates through.",{"type":54,"tag":70,"props":690,"children":691},{},[692],{"type":60,"value":693},"Scope the option to the measured process, before importing the library:",{"type":54,"tag":170,"props":695,"children":698},{"className":696,"code":697,"language":31,"meta":175,"style":175},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","import warp as wp\n\nwp.config.enable_backward = False   # must precede the library import\n\nimport the_library\n",[699],{"type":54,"tag":115,"props":700,"children":701},{"__ignoreMap":175},[702,710,719,728,736],{"type":54,"tag":466,"props":703,"children":704},{"class":468,"line":469},[705],{"type":54,"tag":466,"props":706,"children":707},{},[708],{"type":60,"value":709},"import warp as wp\n",{"type":54,"tag":466,"props":711,"children":712},{"class":468,"line":506},[713],{"type":54,"tag":466,"props":714,"children":716},{"emptyLinePlaceholder":715},true,[717],{"type":60,"value":718},"\n",{"type":54,"tag":466,"props":720,"children":722},{"class":468,"line":721},3,[723],{"type":54,"tag":466,"props":724,"children":725},{},[726],{"type":60,"value":727},"wp.config.enable_backward = False   # must precede the library import\n",{"type":54,"tag":466,"props":729,"children":731},{"class":468,"line":730},4,[732],{"type":54,"tag":466,"props":733,"children":734},{"emptyLinePlaceholder":715},[735],{"type":60,"value":718},{"type":54,"tag":466,"props":737,"children":739},{"class":468,"line":738},5,[740],{"type":54,"tag":466,"props":741,"children":742},{},[743],{"type":60,"value":744},"import the_library\n",{"type":54,"tag":70,"props":746,"children":747},{},[748],{"type":60,"value":749},"This also reaches every module the application loads. CS-10 covers the silent\nimport-order trap. If only a library change works, send its maintainers the\nmeasurement and let them decide the contract.",{"type":54,"tag":445,"props":751,"children":753},{"id":752},"_3-diagnose-from-the-measurement-not-from-reading-the-source",[754],{"type":60,"value":755},"3. Diagnose from the measurement, not from reading the source",{"type":54,"tag":70,"props":757,"children":758},{},[759],{"type":60,"value":760},"The probe prints every compiled module identity with its name, hash, device,\nand block dimension, then names which modules built more than once. Match what\nyou see:",{"type":54,"tag":76,"props":762,"children":763},{},[764,785],{"type":54,"tag":80,"props":765,"children":766},{},[767],{"type":54,"tag":84,"props":768,"children":769},{},[770,775,780],{"type":54,"tag":88,"props":771,"children":772},{},[773],{"type":60,"value":774},"What the probe shows",{"type":54,"tag":88,"props":776,"children":777},{},[778],{"type":60,"value":779},"What it means",{"type":54,"tag":88,"props":781,"children":782},{},[783],{"type":60,"value":784},"Where to look",{"type":54,"tag":104,"props":786,"children":787},{},[788,811,835,853,877,903,927,945,963,981,999,1025,1043],{"type":54,"tag":84,"props":789,"children":790},{},[791,801,806],{"type":54,"tag":111,"props":792,"children":793},{},[794,796],{"type":60,"value":795},"One module name, several ",{"type":54,"tag":248,"props":797,"children":798},{},[799],{"type":60,"value":800},"hashes",{"type":54,"tag":111,"props":802,"children":803},{},[804],{"type":60,"value":805},"Identity churn: its kernel set, options, or generic instances changed after it first loaded",{"type":54,"tag":111,"props":807,"children":808},{},[809],{"type":60,"value":810},"CS-1, CS-3, CS-6",{"type":54,"tag":84,"props":812,"children":813},{},[814,825,830],{"type":54,"tag":111,"props":815,"children":816},{},[817,818,823],{"type":60,"value":795},{"type":54,"tag":248,"props":819,"children":820},{},[821],{"type":60,"value":822},"block_dim",{"type":60,"value":824}," values",{"type":54,"tag":111,"props":826,"children":827},{},[828],{"type":60,"value":829},"The whole module is recompiled per block dimension (CUDA)",{"type":54,"tag":111,"props":831,"children":832},{},[833],{"type":60,"value":834},"CS-5",{"type":54,"tag":84,"props":836,"children":837},{},[838,843,848],{"type":54,"tag":111,"props":839,"children":840},{},[841],{"type":60,"value":842},"Many one-kernel modules in one feature",{"type":54,"tag":111,"props":844,"children":845},{},[846],{"type":60,"value":847},"Fixed per-module cost repeated",{"type":54,"tag":111,"props":849,"children":850},{},[851],{"type":60,"value":852},"CS-4",{"type":54,"tag":84,"props":854,"children":855},{},[856,861,872],{"type":54,"tag":111,"props":857,"children":858},{},[859],{"type":60,"value":860},"A hash-named module per kernel",{"type":54,"tag":111,"props":862,"children":863},{},[864,870],{"type":54,"tag":115,"props":865,"children":867},{"className":866},[],[868],{"type":60,"value":869},"module=\"unique\"",{"type":60,"value":871}," used on stable kernels",{"type":54,"tag":111,"props":873,"children":874},{},[875],{"type":60,"value":876},"CS-9",{"type":54,"tag":84,"props":878,"children":879},{},[880,893,898],{"type":54,"tag":111,"props":881,"children":882},{},[883,885,891],{"type":60,"value":884},"Big gap between module time and native compile time, plus ",{"type":54,"tag":115,"props":886,"children":888},{"className":887},[],[889],{"type":60,"value":890},".lto",{"type":60,"value":892}," artifacts",{"type":54,"tag":111,"props":894,"children":895},{},[896],{"type":60,"value":897},"MathDx\u002FLTO setup",{"type":54,"tag":111,"props":899,"children":900},{},[901],{"type":60,"value":902},"CS-7",{"type":54,"tag":84,"props":904,"children":905},{},[906,917,922],{"type":54,"tag":111,"props":907,"children":908},{},[909,915],{"type":54,"tag":115,"props":910,"children":912},{"className":911},[],[913],{"type":60,"value":914},"(compiled)",{"type":60,"value":916}," on a run that should have been warm",{"type":54,"tag":111,"props":918,"children":919},{},[920],{"type":60,"value":921},"Cache is not being reused",{"type":54,"tag":111,"props":923,"children":924},{},[925],{"type":60,"value":926},"CS-2",{"type":54,"tag":84,"props":928,"children":929},{},[930,935,940],{"type":54,"tag":111,"props":931,"children":932},{},[933],{"type":60,"value":934},"Modules load, then \"Failed to find module\"",{"type":54,"tag":111,"props":936,"children":937},{},[938],{"type":60,"value":939},"Concurrent CPU JIT first-use race",{"type":54,"tag":111,"props":941,"children":942},{},[943],{"type":60,"value":944},"CS-8",{"type":54,"tag":84,"props":946,"children":947},{},[948,953,958],{"type":54,"tag":111,"props":949,"children":950},{},[951],{"type":60,"value":952},"Large generated source, no rebuild problem",{"type":54,"tag":111,"props":954,"children":955},{},[956],{"type":60,"value":957},"Unroll budget",{"type":54,"tag":111,"props":959,"children":960},{},[961],{"type":60,"value":962},"CS-11",{"type":54,"tag":84,"props":964,"children":965},{},[966,971,976],{"type":54,"tag":111,"props":967,"children":968},{},[969],{"type":60,"value":970},"Adjoint code in a module nothing differentiates",{"type":54,"tag":111,"props":972,"children":973},{},[974],{"type":60,"value":975},"Backward codegen",{"type":54,"tag":111,"props":977,"children":978},{},[979],{"type":60,"value":980},"CS-10",{"type":54,"tag":84,"props":982,"children":983},{},[984,989,994],{"type":54,"tag":111,"props":985,"children":986},{},[987],{"type":60,"value":988},"Compiles slow across the board, or a few small modules on CUDA below toolkit 13",{"type":54,"tag":111,"props":990,"children":991},{},[992],{"type":60,"value":993},"The precompiled header is turned off, or is not paying for itself",{"type":54,"tag":111,"props":995,"children":996},{},[997],{"type":60,"value":998},"CS-12",{"type":54,"tag":84,"props":1000,"children":1001},{},[1002,1015,1020],{"type":54,"tag":111,"props":1003,"children":1004},{},[1005,1007,1013],{"type":60,"value":1006},"Several independent modules, each built once, ",{"type":54,"tag":115,"props":1008,"children":1010},{"className":1009},[],[1011],{"type":60,"value":1012},"overlap_factor",{"type":60,"value":1014}," near 1.0",{"type":54,"tag":111,"props":1016,"children":1017},{},[1018],{"type":60,"value":1019},"Builds are running one at a time; parallel loading is off by default",{"type":54,"tag":111,"props":1021,"children":1022},{},[1023],{"type":60,"value":1024},"CS-13",{"type":54,"tag":84,"props":1026,"children":1027},{},[1028,1033,1038],{"type":54,"tag":111,"props":1029,"children":1030},{},[1031],{"type":60,"value":1032},"An option you set changed nothing, and that module's hash is unchanged",{"type":54,"tag":111,"props":1034,"children":1035},{},[1036],{"type":60,"value":1037},"It was assigned after the module was created, so it never arrived",{"type":54,"tag":111,"props":1039,"children":1040},{},[1041],{"type":60,"value":1042},"\"When a module's options are fixed\"",{"type":54,"tag":84,"props":1044,"children":1045},{},[1046,1051,1056],{"type":54,"tag":111,"props":1047,"children":1048},{},[1049],{"type":60,"value":1050},"No row above fires",{"type":54,"tag":111,"props":1052,"children":1053},{},[1054],{"type":60,"value":1055},"Nothing is being built redundantly; the cost is the size of the builds themselves",{"type":54,"tag":111,"props":1057,"children":1058},{},[1059],{"type":60,"value":1060},"Step 6",{"type":54,"tag":70,"props":1062,"children":1063},{},[1064,1070],{"type":54,"tag":115,"props":1065,"children":1067},{"className":1066},[],[1068],{"type":60,"value":1069},"references\u002Fmechanisms.md",{"type":60,"value":1071}," has one section per mechanism: how to confirm it,\nthe fix, its limits, and its failure mode. Read only the sections selected by\nthe measurement.",{"type":54,"tag":445,"props":1073,"children":1075},{"id":1074},"_4-choose-module-boundaries-deliberately",[1076],{"type":60,"value":1077},"4. Choose module boundaries deliberately",{"type":54,"tag":70,"props":1079,"children":1080},{},[1081],{"type":60,"value":1082},"Group kernels in one module only when they share:",{"type":54,"tag":647,"props":1084,"children":1085},{},[1086,1091,1115],{"type":54,"tag":210,"props":1087,"children":1088},{},[1089],{"type":60,"value":1090},"lifecycle: they are defined, loaded, and invalidated together;",{"type":54,"tag":210,"props":1092,"children":1093},{},[1094,1096,1101,1102,1107,1108,1113],{"type":60,"value":1095},"option set: they need the same ",{"type":54,"tag":115,"props":1097,"children":1099},{"className":1098},[],[1100],{"type":60,"value":623},{"type":60,"value":270},{"type":54,"tag":115,"props":1103,"children":1105},{"className":1104},[],[1106],{"type":60,"value":260},{"type":60,"value":270},{"type":54,"tag":115,"props":1109,"children":1111},{"className":1110},[],[1112],{"type":60,"value":268},{"type":60,"value":1114},",\nand MathDx settings;",{"type":54,"tag":210,"props":1116,"children":1117},{},[1118],{"type":60,"value":1119},"stable block dimension on CUDA.",{"type":54,"tag":70,"props":1121,"children":1122},{},[1123],{"type":60,"value":1124},"Kernels with the same lifecycle but different stable block dimensions should\nnot share a module because each would compile twice. Separate kernels with\nindependent lifecycles too.",{"type":54,"tag":70,"props":1126,"children":1127},{},[1128],{"type":60,"value":1129},"Kernels whose block dimension varies at runtime (chosen from input size, say)\nhave no stable mapping, so keep them in their own module rather than dragging\na whole shared module into an extra variant.",{"type":54,"tag":70,"props":1131,"children":1132},{},[1133],{"type":60,"value":1134},"Prefer the least invasive change that removes a build. Ordering fixes and\noption hoists are cheaper and safer than re-architecting module layout;\nregroup only when fixed per-module cost or block-dimension duplication\ndominates.",{"type":54,"tag":70,"props":1136,"children":1137},{},[1138,1143,1145,1151],{"type":54,"tag":115,"props":1139,"children":1141},{"className":1140},[],[1142],{"type":60,"value":332},{"type":60,"value":1144}," targets its calling Python module, not kernels with\nan explicit ",{"type":54,"tag":115,"props":1146,"children":1148},{"className":1147},[],[1149],{"type":60,"value":1150},"module=\"pkg.name\"",{"type":60,"value":1152},". Either use a real Python module or update the\nnamed module before it loads:",{"type":54,"tag":170,"props":1154,"children":1156},{"className":696,"code":1155,"language":31,"meta":175,"style":175},"wp.set_module_options({\"enable_backward\": False})  # at module scope\n\nwp.get_module(\"pkg.name\").options.update({\"enable_backward\": False})\n",[1157],{"type":54,"tag":115,"props":1158,"children":1159},{"__ignoreMap":175},[1160,1168,1175],{"type":54,"tag":466,"props":1161,"children":1162},{"class":468,"line":469},[1163],{"type":54,"tag":466,"props":1164,"children":1165},{},[1166],{"type":60,"value":1167},"wp.set_module_options({\"enable_backward\": False})  # at module scope\n",{"type":54,"tag":466,"props":1169,"children":1170},{"class":468,"line":506},[1171],{"type":54,"tag":466,"props":1172,"children":1173},{"emptyLinePlaceholder":715},[1174],{"type":60,"value":718},{"type":54,"tag":466,"props":1176,"children":1177},{"class":468,"line":721},[1178],{"type":54,"tag":466,"props":1179,"children":1180},{},[1181],{"type":60,"value":1182},"wp.get_module(\"pkg.name\").options.update({\"enable_backward\": False})\n",{"type":54,"tag":70,"props":1184,"children":1185},{},[1186,1188,1194,1196,1202,1204,1210,1212,1217,1219,1225],{"type":60,"value":1187},"Do not pass ",{"type":54,"tag":115,"props":1189,"children":1191},{"className":1190},[],[1192],{"type":60,"value":1193},"wp.get_module()",{"type":60,"value":1195}," to ",{"type":54,"tag":115,"props":1197,"children":1199},{"className":1198},[],[1200],{"type":60,"value":1201},"wp.set_module_options(module=...)",{"type":60,"value":1203},", or use\n",{"type":54,"tag":115,"props":1205,"children":1207},{"className":1206},[],[1208],{"type":60,"value":1209},"@wp.kernel(module_options={...})",{"type":60,"value":1211}," without ",{"type":54,"tag":115,"props":1213,"children":1215},{"className":1214},[],[1216],{"type":60,"value":869},{"type":60,"value":1218},". Per-kernel\n",{"type":54,"tag":115,"props":1220,"children":1222},{"className":1221},[],[1223],{"type":60,"value":1224},"enable_backward=False",{"type":60,"value":1226}," has a tile-module exception covered by CS-10. Confirm\nthe module hash after every option change.",{"type":54,"tag":445,"props":1228,"children":1230},{"id":1229},"_5-verify",[1231],{"type":60,"value":1232},"5. Verify",{"type":54,"tag":170,"props":1234,"children":1236},{"className":458,"code":1235,"language":460,"meta":175,"style":175},"python scripts\u002Fwarp_compile_probe.py measure --samples 3 \\\n    --json candidate.json -- \u003Cthe same command>\npython scripts\u002Fwarp_compile_probe.py compare baseline.json candidate.json\n",[1237],{"type":54,"tag":115,"props":1238,"children":1239},{"__ignoreMap":175},[1240,1267,1311],{"type":54,"tag":466,"props":1241,"children":1242},{"class":468,"line":469},[1243,1247,1251,1255,1259,1263],{"type":54,"tag":466,"props":1244,"children":1245},{"style":473},[1246],{"type":60,"value":31},{"type":54,"tag":466,"props":1248,"children":1249},{"style":478},[1250],{"type":60,"value":481},{"type":54,"tag":466,"props":1252,"children":1253},{"style":478},[1254],{"type":60,"value":486},{"type":54,"tag":466,"props":1256,"children":1257},{"style":478},[1258],{"type":60,"value":491},{"type":54,"tag":466,"props":1260,"children":1261},{"style":494},[1262],{"type":60,"value":497},{"type":54,"tag":466,"props":1264,"children":1265},{"style":500},[1266],{"type":60,"value":503},{"type":54,"tag":466,"props":1268,"children":1269},{"class":468,"line":506},[1270,1274,1279,1283,1287,1291,1296,1301,1306],{"type":54,"tag":466,"props":1271,"children":1272},{"style":478},[1273],{"type":60,"value":512},{"type":54,"tag":466,"props":1275,"children":1276},{"style":478},[1277],{"type":60,"value":1278}," candidate.json",{"type":54,"tag":466,"props":1280,"children":1281},{"style":478},[1282],{"type":60,"value":522},{"type":54,"tag":466,"props":1284,"children":1285},{"style":525},[1286],{"type":60,"value":528},{"type":54,"tag":466,"props":1288,"children":1289},{"style":478},[1290],{"type":60,"value":533},{"type":54,"tag":466,"props":1292,"children":1293},{"style":478},[1294],{"type":60,"value":1295}," same",{"type":54,"tag":466,"props":1297,"children":1298},{"style":478},[1299],{"type":60,"value":1300}," comman",{"type":54,"tag":466,"props":1302,"children":1303},{"style":500},[1304],{"type":60,"value":1305},"d",{"type":54,"tag":466,"props":1307,"children":1308},{"style":525},[1309],{"type":60,"value":1310},">\n",{"type":54,"tag":466,"props":1312,"children":1313},{"class":468,"line":721},[1314,1318,1322,1327,1331],{"type":54,"tag":466,"props":1315,"children":1316},{"style":473},[1317],{"type":60,"value":31},{"type":54,"tag":466,"props":1319,"children":1320},{"style":478},[1321],{"type":60,"value":481},{"type":54,"tag":466,"props":1323,"children":1324},{"style":478},[1325],{"type":60,"value":1326}," compare",{"type":54,"tag":466,"props":1328,"children":1329},{"style":478},[1330],{"type":60,"value":517},{"type":54,"tag":466,"props":1332,"children":1333},{"style":478},[1334],{"type":60,"value":1335}," candidate.json\n",{"type":54,"tag":70,"props":1337,"children":1338},{},[1339,1345,1347,1353],{"type":54,"tag":115,"props":1340,"children":1342},{"className":1341},[],[1343],{"type":60,"value":1344},"compare",{"type":60,"value":1346}," rejects changed launch topology and treats a result inside\n",{"type":54,"tag":115,"props":1348,"children":1350},{"className":1349},[],[1351],{"type":60,"value":1352},"max(1% of baseline, 2 x baseline MAD)",{"type":60,"value":1354}," as inconclusive.",{"type":54,"tag":70,"props":1356,"children":1357},{},[1358,1360,1366,1368,1374],{"type":60,"value":1359},"For ",{"type":54,"tag":115,"props":1361,"children":1363},{"className":1362},[],[1364],{"type":60,"value":1365},"BUILDS OVERLAPPED",{"type":60,"value":1367},", judge scheduling changes on compile elapsed rather\nthan summed module timers. The required warm pass supplies that clock. See\n",{"type":54,"tag":115,"props":1369,"children":1371},{"className":1370},[],[1372],{"type":60,"value":1373},"references\u002Fmeasurement.md",{"type":60,"value":144},{"type":54,"tag":70,"props":1376,"children":1377},{},[1378],{"type":60,"value":1379},"Then check what the probe cannot see:",{"type":54,"tag":647,"props":1381,"children":1382},{},[1383,1388,1400,1417],{"type":54,"tag":210,"props":1384,"children":1385},{},[1386],{"type":60,"value":1387},"Diff numeric output and run the project's tests or entry points.",{"type":54,"tag":210,"props":1389,"children":1390},{},[1391,1393,1398],{"type":60,"value":1392},"After changing ",{"type":54,"tag":115,"props":1394,"children":1396},{"className":1395},[],[1397],{"type":60,"value":260},{"type":60,"value":1399}," or boundaries, verify a gradient path.",{"type":54,"tag":210,"props":1401,"children":1402},{},[1403,1404,1409,1410,1415],{"type":60,"value":1392},{"type":54,"tag":115,"props":1405,"children":1407},{"className":1406},[],[1408],{"type":60,"value":623},{"type":60,"value":270},{"type":54,"tag":115,"props":1411,"children":1413},{"className":1412},[],[1414],{"type":60,"value":268},{"type":60,"value":1416},", MathDx, or an implementation,\nbenchmark steady-state runtime.",{"type":54,"tag":210,"props":1418,"children":1419},{},[1420],{"type":60,"value":1421},"Exercise uncovered dynamic kernels, dtypes, profiles, and modes.",{"type":54,"tag":445,"props":1423,"children":1425},{"id":1424},"_6-report-results",[1426],{"type":60,"value":1427},"6. Report results",{"type":54,"tag":70,"props":1429,"children":1430},{},[1431,1433,1438],{"type":60,"value":1432},"Read \"Reporting results\" in ",{"type":54,"tag":115,"props":1434,"children":1436},{"className":1435},[],[1437],{"type":60,"value":1373},{"type":60,"value":1439},". Report:",{"type":54,"tag":70,"props":1441,"children":1442},{},[1443,1445,1451],{"type":60,"value":1444},"Describe every optimization in plain language: name the behavior, the evidence,\nand the effect. For example, write \"moved module options before the first load\nto avoid a redundant rebuild,\" not \"applied CS-3.\" Treat ",{"type":54,"tag":115,"props":1446,"children":1448},{"className":1447},[],[1449],{"type":60,"value":1450},"CS-*",{"type":60,"value":1452}," labels as\ninternal navigation aids, not user-facing explanations.",{"type":54,"tag":76,"props":1454,"children":1455},{},[1456,1483],{"type":54,"tag":80,"props":1457,"children":1458},{},[1459],{"type":54,"tag":84,"props":1460,"children":1461},{},[1462,1467,1473,1478],{"type":54,"tag":88,"props":1463,"children":1464},{},[1465],{"type":60,"value":1466},"Option",{"type":54,"tag":88,"props":1468,"children":1470},{"align":1469},"right",[1471],{"type":60,"value":1472},"Measured",{"type":54,"tag":88,"props":1474,"children":1475},{},[1476],{"type":60,"value":1477},"Why not taken",{"type":54,"tag":88,"props":1479,"children":1480},{},[1481],{"type":60,"value":1482},"To take it",{"type":54,"tag":104,"props":1484,"children":1485},{},[1486,1520],{"type":54,"tag":84,"props":1487,"children":1488},{},[1489,1505,1510,1515],{"type":54,"tag":111,"props":1490,"children":1491},{},[1492,1497,1499],{"type":54,"tag":115,"props":1493,"children":1495},{"className":1494},[],[1496],{"type":60,"value":1224},{"type":60,"value":1498}," on ",{"type":54,"tag":115,"props":1500,"children":1502},{"className":1501},[],[1503],{"type":60,"value":1504},"pkg.solver",{"type":54,"tag":111,"props":1506,"children":1507},{"align":1469},[1508],{"type":60,"value":1509},"−38% cold",{"type":54,"tag":111,"props":1511,"children":1512},{},[1513],{"type":60,"value":1514},"a live tape traverses these kernels",{"type":54,"tag":111,"props":1516,"children":1517},{},[1518],{"type":60,"value":1519},"set at the entry point, then re-check adjoints",{"type":54,"tag":84,"props":1521,"children":1522},{},[1523,1532,1537,1542],{"type":54,"tag":111,"props":1524,"children":1525},{},[1526],{"type":54,"tag":115,"props":1527,"children":1529},{"className":1528},[],[1530],{"type":60,"value":1531},"max_unroll=4",{"type":54,"tag":111,"props":1533,"children":1534},{"align":1469},[1535],{"type":60,"value":1536},"−2%, inside noise",{"type":54,"tag":111,"props":1538,"children":1539},{},[1540],{"type":60,"value":1541},"changes generated code for no measured gain",{"type":54,"tag":111,"props":1543,"children":1544},{},[1545],{"type":60,"value":1546},"—",{"type":54,"tag":647,"props":1548,"children":1549},{},[1550,1555,1560,1565,1570,1575],{"type":54,"tag":210,"props":1551,"children":1552},{},[1553],{"type":60,"value":1554},"before\u002Fafter medians and sample counts;",{"type":54,"tag":210,"props":1556,"children":1557},{},[1558],{"type":60,"value":1559},"the reduction and residual cost, sized against the original complaint;",{"type":54,"tag":210,"props":1561,"children":1562},{},[1563],{"type":60,"value":1564},"mechanisms fixed, ruled out, measured and declined, or not reached;",{"type":54,"tag":210,"props":1566,"children":1567},{},[1568],{"type":60,"value":1569},"tradeoffs and behavior not verified;",{"type":54,"tag":210,"props":1571,"children":1572},{},[1573],{"type":60,"value":1574},"the ledger above for declined or incompletely verified options;",{"type":54,"tag":210,"props":1576,"children":1577},{},[1578],{"type":60,"value":1579},"the measured value of relaxing any constraint that blocked a fix.",{"type":54,"tag":70,"props":1581,"children":1582},{},[1583],{"type":60,"value":1584},"State only what the evidence supports. \"No structural churn\" does not mean\n\"optimal\" or \"irreducible.\" Measure declined levers when practical; label any\nestimate untested. Before reporting no available fix, check CS-13. For a\npossible module split, first measure a one-kernel module with the same options\nto establish the repeated fixed cost.",{"type":54,"tag":63,"props":1586,"children":1588},{"id":1587},"troubleshooting",[1589],{"type":60,"value":1590},"Troubleshooting",{"type":54,"tag":70,"props":1592,"children":1593},{},[1594,1596,1601,1603,1608],{"type":60,"value":1595},"Run the target command directly before debugging the probe. See\n",{"type":54,"tag":115,"props":1597,"children":1599},{"className":1598},[],[1600],{"type":60,"value":1373},{"type":60,"value":1602}," for cache\u002Fnoise issues and\n",{"type":54,"tag":115,"props":1604,"children":1606},{"className":1605},[],[1607],{"type":60,"value":1069},{"type":60,"value":1609}," for mechanism-specific failures.",{"type":54,"tag":63,"props":1611,"children":1613},{"id":1612},"limitations",[1614],{"type":60,"value":1615},"Limitations",{"type":54,"tag":70,"props":1617,"children":1618},{},[1619],{"type":60,"value":1620},"Two rules override any gain:",{"type":54,"tag":647,"props":1622,"children":1623},{},[1624,1629],{"type":54,"tag":210,"props":1625,"children":1626},{},[1627],{"type":60,"value":1628},"Isolate both Warp and CUDA caches for every cold sample.",{"type":54,"tag":210,"props":1630,"children":1631},{},[1632,1634,1640,1642,1648],{"type":60,"value":1633},"Keep ",{"type":54,"tag":115,"props":1635,"children":1637},{"className":1636},[],[1638],{"type":60,"value":1639},"max_workers \u003C= 1",{"type":60,"value":1641}," when a load can target CPU, including ",{"type":54,"tag":115,"props":1643,"children":1645},{"className":1644},[],[1646],{"type":60,"value":1647},"device=None",{"type":60,"value":1649},"\nand mixed device lists. Concurrent CPU first loads can lose kernels; retries\ndo not make them safe. CUDA-only loading is unaffected.",{"type":54,"tag":70,"props":1651,"children":1652},{},[1653,1655,1660],{"type":60,"value":1654},"Measurements are environment-specific: cold times move with CPU, GPU, driver,\ntoolchain, and Warp version. Mechanisms transfer; numbers do not. Read the\nknown unknowns in ",{"type":54,"tag":115,"props":1656,"children":1658},{"className":1657},[],[1659],{"type":60,"value":1069},{"type":60,"value":1661}," before making broad claims.",{"type":54,"tag":63,"props":1663,"children":1665},{"id":1664},"reference-files",[1666],{"type":60,"value":1667},"Reference files",{"type":54,"tag":647,"props":1669,"children":1670},{},[1671,1681],{"type":54,"tag":210,"props":1672,"children":1673},{},[1674,1679],{"type":54,"tag":115,"props":1675,"children":1677},{"className":1676},[],[1678],{"type":60,"value":1069},{"type":60,"value":1680},": the thirteen compile-time mechanisms, each with\nits confirming signal, fix, applicability limits, and failure mode. Read the\nsections your measurement points to.",{"type":54,"tag":210,"props":1682,"children":1683},{},[1684,1689],{"type":54,"tag":115,"props":1685,"children":1687},{"className":1686},[],[1688],{"type":60,"value":1373},{"type":60,"value":1690},": measurement protocol, what each metric does\nand does not mean, reporting guidance, log examples, and manual measurement.",{"type":54,"tag":1692,"props":1693,"children":1694},"style",{},[1695],{"type":60,"value":1696},"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":1698,"total":1856},[1699,1717,1734,1745,1757,1771,1784,1798,1811,1822,1836,1845],{"slug":1700,"name":1700,"fn":1701,"description":1702,"org":1703,"tags":1704,"stars":1714,"repoUrl":1715,"updatedAt":1716},"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},[1705,1708,1711],{"name":1706,"slug":1707,"type":15},"Documentation","documentation",{"name":1709,"slug":1710,"type":15},"MCP","mcp",{"name":1712,"slug":1713,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1718,"name":1718,"fn":1719,"description":1720,"org":1721,"tags":1722,"stars":1731,"repoUrl":1732,"updatedAt":1733},"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},[1723,1726,1729],{"name":1724,"slug":1725,"type":15},"Containers","containers",{"name":1727,"slug":1728,"type":15},"Deployment","deployment",{"name":1730,"slug":31,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1735,"name":1735,"fn":1736,"description":1737,"org":1738,"tags":1739,"stars":1731,"repoUrl":1732,"updatedAt":1744},"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},[1740,1743],{"name":1741,"slug":1742,"type":15},"CI\u002FCD","ci-cd",{"name":1727,"slug":1728,"type":15},"2026-07-14T05:25:59.97109",{"slug":1746,"name":1746,"fn":1747,"description":1748,"org":1749,"tags":1750,"stars":1731,"repoUrl":1732,"updatedAt":1756},"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},[1751,1752,1753],{"name":1741,"slug":1742,"type":15},{"name":1727,"slug":1728,"type":15},{"name":1754,"slug":1755,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1758,"name":1758,"fn":1759,"description":1760,"org":1761,"tags":1762,"stars":1731,"repoUrl":1732,"updatedAt":1770},"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},[1763,1766,1767],{"name":1764,"slug":1765,"type":15},"Debugging","debugging",{"name":1754,"slug":1755,"type":15},{"name":1768,"slug":1769,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1772,"name":1772,"fn":1773,"description":1774,"org":1775,"tags":1776,"stars":1731,"repoUrl":1732,"updatedAt":1783},"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},[1777,1780],{"name":1778,"slug":1779,"type":15},"Best Practices","best-practices",{"name":1781,"slug":1782,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1785,"name":1785,"fn":1786,"description":1787,"org":1788,"tags":1789,"stars":1731,"repoUrl":1732,"updatedAt":1797},"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},[1790,1793,1796],{"name":1791,"slug":1792,"type":15},"Machine Learning","machine-learning",{"name":1794,"slug":1795,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1799,"name":1799,"fn":1800,"description":1801,"org":1802,"tags":1803,"stars":1731,"repoUrl":1732,"updatedAt":1810},"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},[1804,1807],{"name":1805,"slug":1806,"type":15},"QA","qa",{"name":1808,"slug":1809,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1812,"name":1812,"fn":1813,"description":1814,"org":1815,"tags":1816,"stars":1731,"repoUrl":1732,"updatedAt":1821},"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},[1817,1818],{"name":1727,"slug":1728,"type":15},{"name":1819,"slug":1820,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1823,"name":1823,"fn":1824,"description":1825,"org":1826,"tags":1827,"stars":1731,"repoUrl":1732,"updatedAt":1835},"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},[1828,1831,1832],{"name":1829,"slug":1830,"type":15},"Code Review","code-review",{"name":1754,"slug":1755,"type":15},{"name":1833,"slug":1834,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1837,"name":1837,"fn":1838,"description":1839,"org":1840,"tags":1841,"stars":1731,"repoUrl":1732,"updatedAt":1844},"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},[1842,1843],{"name":1805,"slug":1806,"type":15},{"name":1808,"slug":1809,"type":15},"2026-07-14T05:25:54.928983",{"slug":1846,"name":1846,"fn":1847,"description":1848,"org":1849,"tags":1850,"stars":1731,"repoUrl":1732,"updatedAt":1855},"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},[1851,1854],{"name":1852,"slug":1853,"type":15},"Automation","automation",{"name":1741,"slug":1742,"type":15},"2026-07-30T05:29:03.275638",525,{"items":1858,"total":721},[1859,1865,1876],{"slug":4,"name":4,"fn":5,"description":6,"org":1860,"tags":1861,"stars":20,"repoUrl":21,"updatedAt":22},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1862,1863,1864],{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},{"slug":1866,"name":1866,"fn":1867,"description":1868,"org":1869,"tags":1870,"stars":20,"repoUrl":21,"updatedAt":1875},"warp-debug-gradients","debug gradients in Warp programs","Use to diagnose and fix incorrect gradients in differentiable Warp programs. Anything trained, optimized, calibrated, or fit through Warp kernels depends on wp.Tape gradients, so treat any misbehavior of such a workflow as a gradient problem until proven otherwise — use this when training diverges or NaNs, won't train at all, stalls or plateaus above the expected loss, converges to a wrong or biased answer, is worse than a reference implementation, works at small scale but fails at production scale, or fails a QA\u002Fvalidation recheck. Also for explicit symptoms — exploding, NaN\u002Finf, zero, or subtly wrong gradients, suspected wp.Tape\u002Fbackward issues, gradcheck failures — but users usually describe only the surface symptom (\"the sim explodes\", \"the fit gets dragged toward outliers\") without mentioning gradients: make that leap. Not for forward-only Warp work, build\u002Finstall problems, or autograd issues in other frameworks without Warp.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1871,1872,1873,1874],{"name":1764,"slug":1765,"type":15},{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-08-05T05:58:27.639091",{"slug":1877,"name":1877,"fn":1878,"description":1879,"org":1880,"tags":1881,"stars":20,"repoUrl":21,"updatedAt":1888},"warp-eval","evaluate NVIDIA Warp simulation candidates","Evaluate whether an existing hot path is a credible NVIDIA Warp candidate. Use for irregular or spatial queries, particle or geometry simulation, branch-heavy loops, many small launches, host fallbacks, or large intermediates. CPU-only code and absent GPU dependencies are normal unless NVIDIA is prohibited. Exclude required cross-vendor or CPU-only deployment, vendor-lowered dense or NN layers, general Warp API questions, and already-selected Warp kernels. Contribution policy alone is not exclusion.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1882,1883,1884,1885],{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},{"name":1886,"slug":1887,"type":15},"Simulation","simulation","2026-08-05T05:58:27.275636"]