[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-warp-eval":3,"mdc-1vtr8q-key":40,"related-org-nvidia-warp-eval":1637,"related-repo-nvidia-warp-eval":1797},{"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":35,"sourceUrl":38,"mdContent":39},"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},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,17,20],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Simulation","simulation",{"name":21,"slug":22,"type":15},"Engineering","engineering",6864,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fwarp","2026-08-05T05:58:27.275636","Apache-2.0",555,[29,30,31,32,8,33,34],"cuda","differentiable-programming","gpu","gpu-acceleration","nvidia-warp","python",{"repoUrl":24,"stars":23,"forks":27,"topics":36,"description":37},[29,30,31,32,8,33,34],"A Python framework for GPU-accelerated simulation, robotics, and machine learning.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fwarp\u002Ftree\u002FHEAD\u002Fskills\u002Fwarp-eval","---\nname: warp-eval\ndescription: >\n  Evaluate whether an existing hot path is a credible NVIDIA Warp candidate.\n  Use for irregular or spatial queries, particle or geometry simulation,\n  branch-heavy loops, many small launches, host fallbacks, or large\n  intermediates. CPU-only code and absent GPU dependencies are normal unless\n  NVIDIA is prohibited. Exclude required cross-vendor or CPU-only deployment,\n  vendor-lowered dense or NN layers, general Warp API questions, and\n  already-selected Warp kernels. Contribution policy alone is not exclusion.\nlicense: Apache-2.0\nmetadata:\n  author: NVIDIA Corporation \u003Cwarp-python@nvidia.com>\n  tags:\n    - warp\n    - gpu-acceleration\n    - performance\n    - simulation\n    - evaluation\ncompatibility: >\n  Screening, static evaluation and reporting need no GPU. Measuring Warp\n  requires an NVIDIA CUDA GPU, the target project's dependencies and a\n  representative workload; without them, abort before profiling.\n---\n\n# Warp evaluation\n\n## Purpose\n\nCollect reproducible evidence about how a **narrow seam** in an existing\ncodebase would behave in NVIDIA Warp. Report facts; the user decides.\n\nName Warp as the option under evaluation in the first line, state that no\nadoption recommendation will follow, and do not treat the triggering performance\nrequest as authorization to experiment.\n\nMeasured evaluations produce `warp-evaluation-report\u002F`: the report, one\nindependently applicable diff per solution, the drivers, and raw results. Never\nmodify production code. Exits before measured work create no directory.\n\n## Hard rules\n\nThese override any local reasoning.\n\n1. **Default objectives are latency, throughput, and peak or retained memory.**\n   Count maintainability, ergonomics, packaging, extensibility, autodiff or new\n   functionality only when the user names it; otherwise report them as\n   constraints or costs, not benefits.\n2. **Correctness is a gate.** If the incumbent is buggy or its contract unclear,\n   abort that comparison until an independent oracle or clarified contract\n   exists. Report the defect without prescribing a response.\n3. **Never predict savings from source shape or another workload.** Materialized\n   memory comes from profiler or allocator evidence, not source a compiler may\n   fuse.\n4. **\"It can be written in Warp\" is a hypothesis, never a proven opportunity.**\n5. **Gates are exact, never adjacent or analogous.** Fire a gate only when every\n   condition in its definition is established. Gate D requires a maintained\n   implementation confirmed to execute with CUDA on an NVIDIA GPU; a fast native\n   CPU library is a baseline, not Gate D.\n6. **Label every claim** as observed fact, measurement, hypothesis or unknown.\n7. **Time the user-visible stage, not the kernel.** Include Warp's cold\n   import\u002Finit\u002FJIT in the real process regime, transfers, launches, Python\n   launch loops, structure build\u002Frefit, allocation, conversion, validation,\n   compaction and synchronization. Report each cost and the end-to-end\n   difference.\n8. **No measured work without explicit authorization.** After stage 1, stop and\n   ask before profiling, environment changes, prototyping, benchmarking or GPU\n   use.\n9. **Authorized measured work includes Warp.** If profiling exposes no gate,\n   continue through the strongest in-project baseline, minimum Warp prototype\n   and end-to-end comparison. If Warp is out of scope, stop before profiling.\n10. **Never infer environment intent.** NVIDIA deployment and ownership of an\n    optional compiled dependency are product decisions. Abort only on a stated\n    constraint; otherwise ask once and stop (`AWAITING INTENT`).\n11. **Warp measurements are CUDA-only and synchronized.** Resolve an explicit\n    NVIDIA CUDA device; discard CPU resolution and dispatch-only timing.\n12. **Never recommend or rank adoption options.** Report facts, measurements,\n    hypotheses and unknowns per seam and regime.\n13. **Stop at the report.** No rollout or production edits.\n\n## Requirements\n\nStatic screening requires the target repository and its stated product\nconstraints. Measured work additionally requires explicit authorization, an\nNVIDIA CUDA GPU, target-project dependencies and a representative workload.\n\n## Limitations\n\n- **Warp CPU kernels are serial and outside this skill's evaluation scope.**\n  \"Run it on the CPU instead\" is not a performance fallback.\n- **Mesh and geometry queries compute in float32\u002Fint32** even behind a float64\n  public API. Absolute error grows with coordinate magnitude.\n- **Warp does not follow semver.** Feature releases can break APIs, only the\n  newest feature line is maintained, deprecations run roughly four monthly\n  releases.\n- **`latest` docs track development, not the shipped release.** Pin to the\n  target project's own Warp version; if it has none, use the current stable and\n  say which.\n- **Kernel-only builtins are not resolvable from Python** —\n  `hasattr(wp, \"mesh_query_point\")` is `False` on a release that has it. Probe\n  the stub file or that version's docs before concluding a builtin is absent.\n- **`enable_backward=False`** (kernel, module or global) removes adjoint\n  codegen. If nothing differentiates through the seam, set it before measuring\n  compile cost.\n\n## Output Format\n\n| State | Reached when | Report directory |\n|---|---|---|\n| `ABORT` | Any gate establishes that the evaluated seam cannot satisfy the stated scope | No if nothing was measured; otherwise preserve the evidence already collected |\n| `AWAITING INTENT` | The environment gates turn on a fact only the user has | No — one question, both branches concrete |\n| `AWAITING AUTHORIZATION` | A candidate pattern survives stage 1 | No — early findings plus scope\u002Fresource preview |\n| `INCOMPLETE` | Authorized work cannot obtain representative evidence required for the scoped evaluation | Yes — preserve collected evidence and name the one missing artifact |\n| Report delivered | Authorized work produced measurements, or stopped after the report directory existed | Yes — facts per seam and regime, with missing evidence explicit |\n\nUse this exact shape for an early exit:\n`ABORT — Gate \u003Cletter>: \u003Ccited fact>; \u003Cwhy the scoped Warp evaluation cannot proceed>.`\nDo not name a preferred alternative.\nReporting rules:\n[references\u002Fevidence-and-reporting.md](references\u002Fevidence-and-reporting.md).\nA delivered directory follows the template's fixed order: schema and provenance,\nauthorization and evaluation state, stage census, one `B\u003Cn>` evidence section\nper seam\u002Fregime, caveats, then environment\u002Freproduction. `solutions\u002F`,\n`benchmarks\u002F` and `results\u002F` contain every linked artifact.\n\n## Examples\n\n**Gate exit:** `ABORT — Gate A: deployment.md requires one implementation with\nAMD, Apple and NVIDIA parity; a Warp-specific path cannot satisfy this scope.`\n\n**Surviving candidate:** Name the seam and pattern, label inferred facts as\nassumptions, state that no gate has fired, preview the profile\u002Fbaseline\u002FWarp\nprototype\u002Fbenchmark scope and its cost, then ask the separate intent and\nauthorization questions from\n[references\u002Fauthorization-checkpoint.md](references\u002Fauthorization-checkpoint.md).\n\n## Inputs\n\n**Required:** the target repository and a performance, memory or scale problem\nwith a candidate seam. **Optional:** explicit deployment\u002Fpackaging constraints,\nexisting profiles or logs, representative datasets and acceptance criteria.\nPrompt constraints take precedence over repository policy\u002Fconfiguration, then\nexisting logs. User corrections override inference. Never substitute an\nassumption for a stated fact or measurement.\n\n## Available scripts\n\n| Script | Purpose | Arguments |\n|---|---|---|\n| `scripts\u002Fdriver-template.py` | Copy once per bottleneck; define workloads and variants | Edit placeholders, then run the copied driver |\n| `scripts\u002Fmeasure.py` | Import from drivers for synchronized timing, memory and isolated cases | Python API; do not execute directly |\n| `scripts\u002Fvalidate_report_schema.py` | Validate the delivered report and evidence links | `\u003Creport-directory>` |\n\nUse `run_script(\"scripts\u002Fvalidate_report_schema.py\", args=[\"warp-evaluation-report\"])`\nwhen supported; otherwise invoke the script with Python and the report directory.\n\n## Troubleshooting\n\n- No representative workload: mark the scope `INCOMPLETE` and name the missing\n  artifact; do not invent data or fire Gate F.\n- Warp resolves to CPU or no CUDA device: discard the run and stop before\n  correctness or timing claims.\n- Report validation fails: fix the report or referenced artifact; never waive\n  the schema error.\n\n## Instructions\n\nEvery stage before the last can end the evaluation. Stop as soon as a gate\nfires; do not gather evidence that cannot change the scoped facts.\n\n### 1. Read the code, derive the contract, check the gates\n\n- Identify a candidate and its metric with\n  [references\u002Ftarget-patterns.md](references\u002Ftarget-patterns.md).\n- Derive devices\u002Fresidency, dtypes\u002Fshapes, sizes, frequency, process lifetime,\n  gradients and packaging from the repository. Infer before asking.\n- State the inferred contract in one line and invite correction. Unknown\n  hardware, counts, sizes and tolerances are assumptions, never measurements.\n  Every inference remains open to correction and cannot satisfy a gate that\n  requires a stated fact or measurement.\n- Check Gates A–E before profiling. Check Gate F now only if representative\n  evidence already exists; otherwise carry it into stage 2. Every gate uses only\n  the exact boundaries in\n  [references\u002Frejection-gates.md](references\u002Frejection-gates.md).\n\n| Gate | Fires when |\n|---|---|\n| **A** | Production is *stated* CPU-only or to need non-NVIDIA portability, with no acceptable optional CUDA path |\n| **B** | Data must cross the host\u002Fdevice boundary per small or infrequent call and the boundary cannot be widened |\n| **C** | The region is dense tensor algebra already mapped to a tuned framework or vendor library |\n| **D** | A mature CUDA implementation already meets the contract, and no non-performance objective was requested |\n| **E** | A stated policy blocks Warp's dependency, compilation, cache or fallback obligations |\n| **F** | Representative evidence proves the region too small a share of its requested metric for any backend to move it |\n\n- Gate F can fire in stage 1 only from representative evidence that already\n  exists — a supplied profile, structural bound, or arithmetic on figures the\n  user quoted. If that evidence does not exist, Gate F remains open until stage\n  2 profiling; inferred values never fire it.\n- Gates A and E need a stated constraint. A CPU implementation, another\n  accelerator, no Warp dependency, or a small dependency list proves nothing.\n- When A\u002FE are unresolved and a pattern survives, ask whether an optional NVIDIA\n  path is acceptable: named extra, soft import, existing fallback, default\n  install unchanged. Every affirmative answer must say explicitly that Warp will\n  be prototyped and benchmarked; conditions constrain only that Warp scope. A\n  negative or undecided answer means `ABORT`.\n- Do not ask when another gate fired, the repository answers, or no pattern\n  matched. No pattern means no profiling.\n- If a candidate survives, combine any intent question with the\n  [authorization checkpoint](references\u002Fauthorization-checkpoint.md) — early\n  findings, exact scope, stages, resource cost — then stop. Stage 2 requires\n  both settled intent and explicit authorization.\n\n### 2. Profile the real application\n\nRequires explicit authorization and a settled intent question.\n\n- Profile with the project's own profiler and representative entry points.\n- Measure synchronized end-to-end stage time and peak memory before choosing a\n  backend. Report which entry points were profiled and which a gate screened.\n- Prioritize further measurement by observed cost, not source appearance.\n- Name each measurement by the public method and variant actually invoked. A\n  fallback is an execution regime of that public seam, not a different\n  operation, and a cheaper sibling method cannot screen out the named method.\n- Confirm the timed branch ran on the intended device. Unchanged cost and\n  near-zero device allocation between host and device inputs exposes a host\n  fallback.\n- Measure the stage's free-stage ceiling and stubbed floor through the public\n  boundary. Do not subtract per-op timings.\n- If the measured ceiling proves the candidate cannot move its own metric,\n  `ABORT` the affected scope under Gate F, preserve the evidence already\n  collected, and stop. The existing authorization already covered this\n  materiality check; do not ask for authorization again.\n- If representative coverage is unavailable — no representative dataset,\n  runnable entry point or production distribution — record the single missing\n  artifact, mark the affected scope `INCOMPLETE`, and stop. This is missing\n  evidence, not Gate F and not `ABORT`. An invented workload cannot prove\n  materiality.\n\nProtocol: [references\u002Fbenchmark-protocol.md](references\u002Fbenchmark-protocol.md).\n\n### 3. Form falsifiable hypotheses\n\nRecord per candidate: source, bottleneck evidence, objective, narrow seam,\nmechanism Warp could change, strongest incumbent, risks, acceptance threshold,\ncheapest falsifying experiment. Screen against\n[references\u002Ftarget-patterns.md](references\u002Ftarget-patterns.md); if none\nsurvives, write the report and stop.\n\n### 4. Write the contract before the prototype\n\nDefine values, dtypes, shapes, devices, errors, mutation, ordering, ties,\ncapacity\u002Foverflow, topology\u002Fdegeneracy, tolerances, required gradients, streams,\nownership, aliasing, invalidation, concurrency, capture, teardown and fallback.\n\n- Fix tolerances before seeing Warp output. Never weaken a contract after a\n  mismatch.\n- `ABORT` before prototyping if the proposed seam cannot satisfy a required\n  contract.\n- Pre-register, before timing: workload provenance, the state variable and\n  production range controlling cost, tuning knobs, incumbent run-to-run spread,\n  and the oracle applied to every implementation.\n\nHazards and adversarial checks:\n[references\u002Fsemantic-contract.md](references\u002Fsemantic-contract.md).\n\n### 5. Improve the baseline first\n\nAlgorithm before backend: (1) a better or output-sensitive algorithm;\n(2) chunking, tiling, sparse output, layout, rematerialization; (3) the\nincumbent framework's compiler and native primitives; (4) **what the project\nalready depends on** — its own accelerator backend, a parallel idiom it ships\nbut leaves off, or a capability an existing dependency exposes and nobody wired\nup; (5) only then narrow Warp.\n\n- Compare only in-scope options: the improved incumbent, capabilities reachable\n  through current dependencies, and Warp. Do not add unrelated libraries.\n- Search declared dependencies for dormant backends, flags and bindings before\n  calling a route absent.\n- Keep the improved incumbent as the oracle. A result against an untuned,\n  incorrect or asymptotically inferior baseline does not establish a backend\n  comparison.\n\nClose every in-project route before prototyping Warp:\n\n| State | What it takes to claim it |\n|---|---|\n| **measured** | timed through the same boundary as the baseline |\n| **absent** | a cited declaration, symbol table or missing flag proves it is unavailable |\n| **waived** | you asked the user and they chose to skip it; record their words |\n\nA capability present but unbound is reachable, not absent. If exposing it costs\nno more than the planned Warp seam, measure it first. Ladder details:\n[references\u002Fbaselines.md](references\u002Fbaselines.md).\nWaived routes do not block stage 6; every route not explicitly waived must be\nmeasured or evidenced absent before the Warp prototype begins.\n\n### 6. Prototype the minimum unit\n\n- Work in a separate copy, production unchanged, integration seam off by\n  default. Prototype only enough to test the hypothesis.\n- Select an explicit CUDA device for every Warp prototype and verify that Warp\n  resolves it as CUDA before correctness or performance work. Never exercise or\n  report Warp's CPU backend.\n- Size the unit by shared data and structure lifetime, not function boundaries:\n  build\u002Frefit\u002Fquery costs that amortize together are one unit.\n- Name variants before timing. Preserve each solution as an **independent**\n  patch against the pinned baseline, and prove it applies cleanly and reproduces\n  the measured result.\n- Surface unsupported cases and overflow before timing.\n- Run the adversarial contract checks, audit correctness at production scale,\n  and deliberately break a branch to prove the tests fail. Any failed required\n  check returns `ABORT` for the affected scope.\n- Before treating a disappointing compute-bound or reuse-heavy result as\n  representative, consider the stronger formulations in\n  [references\u002Ftarget-patterns.md](references\u002Ftarget-patterns.md). One naive\n  kernel does not bound Warp's potential.\n\n### 7. Benchmark the whole boundary\n\n- Copy [scripts\u002Fdriver-template.py](scripts\u002Fdriver-template.py) once per\n  bottleneck and measure through [scripts\u002Fmeasure.py](scripts\u002Fmeasure.py). Do\n  not hand-roll timing or memory.\n- Warp launches are asynchronous. The measurement helper must synchronize the\n  selected device immediately before starting and after enqueueing every timed\n  region, before stopping its wall timer.\n- Include every rule-7 cost, the public call, the immediate downstream stage,\n  realistic sizes, and cold\u002Fwarm process regimes.\n- Transcribe every report cell from emitted JSON; absent records read\n  `not measured`.\n- Report `null_test`, and one-time costs both separately and amortized.\n  Serialize GPU measurements under an exclusive device lock. Below 1.5× is no\n  measured difference.\n- `ABORT` for a seam and regime whose predeclared end-to-end performance or\n  memory requirement fails, after preserving the measurements.\n\nFull protocol:\n[references\u002Fbenchmark-protocol.md](references\u002Fbenchmark-protocol.md).\n\n### 8. Summarize the evidence\n\n- Per seam and execution regime, report semantic results, strongest-baseline\n  measurements, workload provenance, time, memory, lifecycle, portability and\n  ownership facts.\n- Use `pass`, `fail`, `not measured`, `not available`, `no representative data`,\n  `unknown` or `n\u002Fa` only where a stated criterion makes that status objective.\n  Missing evidence remains missing.\n- Mark the report `complete` only when every evaluated seam and regime includes\n  an end-to-end Warp measurement. Use `aborted — \u003Cgate and scope>` when a gate\n  fired, or `incomplete — \u003Cmissing evidence and scope>` when representative\n  evidence was unavailable. Preserve everything collected. Never deliver an\n  incumbent-only report as a completed `warp-eval`.\n- Give every stage worth ≥10 % of the measured total a census row with a status\n  and a one-line reason, including the screened-out stages.\n- Check each gate against the metric *that candidate's* pattern exhausts, not\n  the study's headline objective.\n\n### 9. Hand back\n\n- Use\n  [assets\u002Fwarp-evaluation-report-template.md](assets\u002Fwarp-evaluation-report-template.md)\n  unchanged in schema. Record authorization and scope.\n- One table per bottleneck: baseline, current-dependency solutions, then Warp;\n  absolute time and peak memory, ratios, contract status, evidence gaps.\n- Include prose only where it explains the measurements or their bounds.\n  Working artifacts belong in `results\u002F`.\n\n```bash\nuv run python scripts\u002Fvalidate_report_schema.py \u003Creport-directory>\n```\n\n- Run the validator once the first measurement establishes a census\u002Ftable, and\n  again before delivery. Fix every error.\n- Verify builds and imports from artifacts, not exit status. Ship the exact\n  drivers run. Then stop.\n",{"data":41,"body":48},{"name":4,"description":6,"license":26,"metadata":42,"compatibility":47},{"author":43,"tags":44},"NVIDIA Corporation \u003Cwarp-python@nvidia.com>",[45,32,14,19,46],"warp","evaluation","Screening, static evaluation and reporting need no GPU. Measuring Warp requires an NVIDIA CUDA GPU, the target project's dependencies and a representative workload; without them, abort before profiling.\n",{"type":49,"children":50},"root",[51,60,67,81,86,100,106,111,252,258,263,269,359,365,503,556,562,578,595,601,618,624,720,733,739,764,770,775,782,817,943,984,990,995,1059,1070,1076,1087,1093,1098,1121,1132,1138,1150,1168,1173,1242,1254,1260,1318,1324,1389,1399,1405,1520,1526,1557,1618,1631],{"type":52,"tag":53,"props":54,"children":56},"element","h1",{"id":55},"warp-evaluation",[57],{"type":58,"value":59},"text","Warp evaluation",{"type":52,"tag":61,"props":62,"children":64},"h2",{"id":63},"purpose",[65],{"type":58,"value":66},"Purpose",{"type":52,"tag":68,"props":69,"children":70},"p",{},[71,73,79],{"type":58,"value":72},"Collect reproducible evidence about how a ",{"type":52,"tag":74,"props":75,"children":76},"strong",{},[77],{"type":58,"value":78},"narrow seam",{"type":58,"value":80}," in an existing\ncodebase would behave in NVIDIA Warp. Report facts; the user decides.",{"type":52,"tag":68,"props":82,"children":83},{},[84],{"type":58,"value":85},"Name Warp as the option under evaluation in the first line, state that no\nadoption recommendation will follow, and do not treat the triggering performance\nrequest as authorization to experiment.",{"type":52,"tag":68,"props":87,"children":88},{},[89,91,98],{"type":58,"value":90},"Measured evaluations produce ",{"type":52,"tag":92,"props":93,"children":95},"code",{"className":94},[],[96],{"type":58,"value":97},"warp-evaluation-report\u002F",{"type":58,"value":99},": the report, one\nindependently applicable diff per solution, the drivers, and raw results. Never\nmodify production code. Exits before measured work create no directory.",{"type":52,"tag":61,"props":101,"children":103},{"id":102},"hard-rules",[104],{"type":58,"value":105},"Hard rules",{"type":52,"tag":68,"props":107,"children":108},{},[109],{"type":58,"value":110},"These override any local reasoning.",{"type":52,"tag":112,"props":113,"children":114},"ol",{},[115,126,136,146,154,164,174,184,194,204,222,232,242],{"type":52,"tag":116,"props":117,"children":118},"li",{},[119,124],{"type":52,"tag":74,"props":120,"children":121},{},[122],{"type":58,"value":123},"Default objectives are latency, throughput, and peak or retained memory.",{"type":58,"value":125},"\nCount maintainability, ergonomics, packaging, extensibility, autodiff or new\nfunctionality only when the user names it; otherwise report them as\nconstraints or costs, not benefits.",{"type":52,"tag":116,"props":127,"children":128},{},[129,134],{"type":52,"tag":74,"props":130,"children":131},{},[132],{"type":58,"value":133},"Correctness is a gate.",{"type":58,"value":135}," If the incumbent is buggy or its contract unclear,\nabort that comparison until an independent oracle or clarified contract\nexists. Report the defect without prescribing a response.",{"type":52,"tag":116,"props":137,"children":138},{},[139,144],{"type":52,"tag":74,"props":140,"children":141},{},[142],{"type":58,"value":143},"Never predict savings from source shape or another workload.",{"type":58,"value":145}," Materialized\nmemory comes from profiler or allocator evidence, not source a compiler may\nfuse.",{"type":52,"tag":116,"props":147,"children":148},{},[149],{"type":52,"tag":74,"props":150,"children":151},{},[152],{"type":58,"value":153},"\"It can be written in Warp\" is a hypothesis, never a proven opportunity.",{"type":52,"tag":116,"props":155,"children":156},{},[157,162],{"type":52,"tag":74,"props":158,"children":159},{},[160],{"type":58,"value":161},"Gates are exact, never adjacent or analogous.",{"type":58,"value":163}," Fire a gate only when every\ncondition in its definition is established. Gate D requires a maintained\nimplementation confirmed to execute with CUDA on an NVIDIA GPU; a fast native\nCPU library is a baseline, not Gate D.",{"type":52,"tag":116,"props":165,"children":166},{},[167,172],{"type":52,"tag":74,"props":168,"children":169},{},[170],{"type":58,"value":171},"Label every claim",{"type":58,"value":173}," as observed fact, measurement, hypothesis or unknown.",{"type":52,"tag":116,"props":175,"children":176},{},[177,182],{"type":52,"tag":74,"props":178,"children":179},{},[180],{"type":58,"value":181},"Time the user-visible stage, not the kernel.",{"type":58,"value":183}," Include Warp's cold\nimport\u002Finit\u002FJIT in the real process regime, transfers, launches, Python\nlaunch loops, structure build\u002Frefit, allocation, conversion, validation,\ncompaction and synchronization. Report each cost and the end-to-end\ndifference.",{"type":52,"tag":116,"props":185,"children":186},{},[187,192],{"type":52,"tag":74,"props":188,"children":189},{},[190],{"type":58,"value":191},"No measured work without explicit authorization.",{"type":58,"value":193}," After stage 1, stop and\nask before profiling, environment changes, prototyping, benchmarking or GPU\nuse.",{"type":52,"tag":116,"props":195,"children":196},{},[197,202],{"type":52,"tag":74,"props":198,"children":199},{},[200],{"type":58,"value":201},"Authorized measured work includes Warp.",{"type":58,"value":203}," If profiling exposes no gate,\ncontinue through the strongest in-project baseline, minimum Warp prototype\nand end-to-end comparison. If Warp is out of scope, stop before profiling.",{"type":52,"tag":116,"props":205,"children":206},{},[207,212,214,220],{"type":52,"tag":74,"props":208,"children":209},{},[210],{"type":58,"value":211},"Never infer environment intent.",{"type":58,"value":213}," NVIDIA deployment and ownership of an\noptional compiled dependency are product decisions. Abort only on a stated\nconstraint; otherwise ask once and stop (",{"type":52,"tag":92,"props":215,"children":217},{"className":216},[],[218],{"type":58,"value":219},"AWAITING INTENT",{"type":58,"value":221},").",{"type":52,"tag":116,"props":223,"children":224},{},[225,230],{"type":52,"tag":74,"props":226,"children":227},{},[228],{"type":58,"value":229},"Warp measurements are CUDA-only and synchronized.",{"type":58,"value":231}," Resolve an explicit\nNVIDIA CUDA device; discard CPU resolution and dispatch-only timing.",{"type":52,"tag":116,"props":233,"children":234},{},[235,240],{"type":52,"tag":74,"props":236,"children":237},{},[238],{"type":58,"value":239},"Never recommend or rank adoption options.",{"type":58,"value":241}," Report facts, measurements,\nhypotheses and unknowns per seam and regime.",{"type":52,"tag":116,"props":243,"children":244},{},[245,250],{"type":52,"tag":74,"props":246,"children":247},{},[248],{"type":58,"value":249},"Stop at the report.",{"type":58,"value":251}," No rollout or production edits.",{"type":52,"tag":61,"props":253,"children":255},{"id":254},"requirements",[256],{"type":58,"value":257},"Requirements",{"type":52,"tag":68,"props":259,"children":260},{},[261],{"type":58,"value":262},"Static screening requires the target repository and its stated product\nconstraints. Measured work additionally requires explicit authorization, an\nNVIDIA CUDA GPU, target-project dependencies and a representative workload.",{"type":52,"tag":61,"props":264,"children":266},{"id":265},"limitations",[267],{"type":58,"value":268},"Limitations",{"type":52,"tag":270,"props":271,"children":272},"ul",{},[273,283,293,303,319,345],{"type":52,"tag":116,"props":274,"children":275},{},[276,281],{"type":52,"tag":74,"props":277,"children":278},{},[279],{"type":58,"value":280},"Warp CPU kernels are serial and outside this skill's evaluation scope.",{"type":58,"value":282},"\n\"Run it on the CPU instead\" is not a performance fallback.",{"type":52,"tag":116,"props":284,"children":285},{},[286,291],{"type":52,"tag":74,"props":287,"children":288},{},[289],{"type":58,"value":290},"Mesh and geometry queries compute in float32\u002Fint32",{"type":58,"value":292}," even behind a float64\npublic API. Absolute error grows with coordinate magnitude.",{"type":52,"tag":116,"props":294,"children":295},{},[296,301],{"type":52,"tag":74,"props":297,"children":298},{},[299],{"type":58,"value":300},"Warp does not follow semver.",{"type":58,"value":302}," Feature releases can break APIs, only the\nnewest feature line is maintained, deprecations run roughly four monthly\nreleases.",{"type":52,"tag":116,"props":304,"children":305},{},[306,317],{"type":52,"tag":74,"props":307,"children":308},{},[309,315],{"type":52,"tag":92,"props":310,"children":312},{"className":311},[],[313],{"type":58,"value":314},"latest",{"type":58,"value":316}," docs track development, not the shipped release.",{"type":58,"value":318}," Pin to the\ntarget project's own Warp version; if it has none, use the current stable and\nsay which.",{"type":52,"tag":116,"props":320,"children":321},{},[322,327,329,335,337,343],{"type":52,"tag":74,"props":323,"children":324},{},[325],{"type":58,"value":326},"Kernel-only builtins are not resolvable from Python",{"type":58,"value":328}," —\n",{"type":52,"tag":92,"props":330,"children":332},{"className":331},[],[333],{"type":58,"value":334},"hasattr(wp, \"mesh_query_point\")",{"type":58,"value":336}," is ",{"type":52,"tag":92,"props":338,"children":340},{"className":339},[],[341],{"type":58,"value":342},"False",{"type":58,"value":344}," on a release that has it. Probe\nthe stub file or that version's docs before concluding a builtin is absent.",{"type":52,"tag":116,"props":346,"children":347},{},[348,357],{"type":52,"tag":74,"props":349,"children":350},{},[351],{"type":52,"tag":92,"props":352,"children":354},{"className":353},[],[355],{"type":58,"value":356},"enable_backward=False",{"type":58,"value":358}," (kernel, module or global) removes adjoint\ncodegen. If nothing differentiates through the seam, set it before measuring\ncompile cost.",{"type":52,"tag":61,"props":360,"children":362},{"id":361},"output-format",[363],{"type":58,"value":364},"Output Format",{"type":52,"tag":366,"props":367,"children":368},"table",{},[369,393],{"type":52,"tag":370,"props":371,"children":372},"thead",{},[373],{"type":52,"tag":374,"props":375,"children":376},"tr",{},[377,383,388],{"type":52,"tag":378,"props":379,"children":380},"th",{},[381],{"type":58,"value":382},"State",{"type":52,"tag":378,"props":384,"children":385},{},[386],{"type":58,"value":387},"Reached when",{"type":52,"tag":378,"props":389,"children":390},{},[391],{"type":58,"value":392},"Report directory",{"type":52,"tag":394,"props":395,"children":396},"tbody",{},[397,420,441,463,485],{"type":52,"tag":374,"props":398,"children":399},{},[400,410,415],{"type":52,"tag":401,"props":402,"children":403},"td",{},[404],{"type":52,"tag":92,"props":405,"children":407},{"className":406},[],[408],{"type":58,"value":409},"ABORT",{"type":52,"tag":401,"props":411,"children":412},{},[413],{"type":58,"value":414},"Any gate establishes that the evaluated seam cannot satisfy the stated scope",{"type":52,"tag":401,"props":416,"children":417},{},[418],{"type":58,"value":419},"No if nothing was measured; otherwise preserve the evidence already collected",{"type":52,"tag":374,"props":421,"children":422},{},[423,431,436],{"type":52,"tag":401,"props":424,"children":425},{},[426],{"type":52,"tag":92,"props":427,"children":429},{"className":428},[],[430],{"type":58,"value":219},{"type":52,"tag":401,"props":432,"children":433},{},[434],{"type":58,"value":435},"The environment gates turn on a fact only the user has",{"type":52,"tag":401,"props":437,"children":438},{},[439],{"type":58,"value":440},"No — one question, both branches concrete",{"type":52,"tag":374,"props":442,"children":443},{},[444,453,458],{"type":52,"tag":401,"props":445,"children":446},{},[447],{"type":52,"tag":92,"props":448,"children":450},{"className":449},[],[451],{"type":58,"value":452},"AWAITING AUTHORIZATION",{"type":52,"tag":401,"props":454,"children":455},{},[456],{"type":58,"value":457},"A candidate pattern survives stage 1",{"type":52,"tag":401,"props":459,"children":460},{},[461],{"type":58,"value":462},"No — early findings plus scope\u002Fresource preview",{"type":52,"tag":374,"props":464,"children":465},{},[466,475,480],{"type":52,"tag":401,"props":467,"children":468},{},[469],{"type":52,"tag":92,"props":470,"children":472},{"className":471},[],[473],{"type":58,"value":474},"INCOMPLETE",{"type":52,"tag":401,"props":476,"children":477},{},[478],{"type":58,"value":479},"Authorized work cannot obtain representative evidence required for the scoped evaluation",{"type":52,"tag":401,"props":481,"children":482},{},[483],{"type":58,"value":484},"Yes — preserve collected evidence and name the one missing artifact",{"type":52,"tag":374,"props":486,"children":487},{},[488,493,498],{"type":52,"tag":401,"props":489,"children":490},{},[491],{"type":58,"value":492},"Report delivered",{"type":52,"tag":401,"props":494,"children":495},{},[496],{"type":58,"value":497},"Authorized work produced measurements, or stopped after the report directory existed",{"type":52,"tag":401,"props":499,"children":500},{},[501],{"type":58,"value":502},"Yes — facts per seam and regime, with missing evidence explicit",{"type":52,"tag":68,"props":504,"children":505},{},[506,508,514,516,522,524,530,532,538,540,546,548,554],{"type":58,"value":507},"Use this exact shape for an early exit:\n",{"type":52,"tag":92,"props":509,"children":511},{"className":510},[],[512],{"type":58,"value":513},"ABORT — Gate \u003Cletter>: \u003Ccited fact>; \u003Cwhy the scoped Warp evaluation cannot proceed>.",{"type":58,"value":515},"\nDo not name a preferred alternative.\nReporting rules:\n",{"type":52,"tag":517,"props":518,"children":520},"a",{"href":519},"references\u002Fevidence-and-reporting.md",[521],{"type":58,"value":519},{"type":58,"value":523},".\nA delivered directory follows the template's fixed order: schema and provenance,\nauthorization and evaluation state, stage census, one ",{"type":52,"tag":92,"props":525,"children":527},{"className":526},[],[528],{"type":58,"value":529},"B\u003Cn>",{"type":58,"value":531}," evidence section\nper seam\u002Fregime, caveats, then environment\u002Freproduction. ",{"type":52,"tag":92,"props":533,"children":535},{"className":534},[],[536],{"type":58,"value":537},"solutions\u002F",{"type":58,"value":539},",\n",{"type":52,"tag":92,"props":541,"children":543},{"className":542},[],[544],{"type":58,"value":545},"benchmarks\u002F",{"type":58,"value":547}," and ",{"type":52,"tag":92,"props":549,"children":551},{"className":550},[],[552],{"type":58,"value":553},"results\u002F",{"type":58,"value":555}," contain every linked artifact.",{"type":52,"tag":61,"props":557,"children":559},{"id":558},"examples",[560],{"type":58,"value":561},"Examples",{"type":52,"tag":68,"props":563,"children":564},{},[565,570,572],{"type":52,"tag":74,"props":566,"children":567},{},[568],{"type":58,"value":569},"Gate exit:",{"type":58,"value":571}," ",{"type":52,"tag":92,"props":573,"children":575},{"className":574},[],[576],{"type":58,"value":577},"ABORT — Gate A: deployment.md requires one implementation with AMD, Apple and NVIDIA parity; a Warp-specific path cannot satisfy this scope.",{"type":52,"tag":68,"props":579,"children":580},{},[581,586,588,593],{"type":52,"tag":74,"props":582,"children":583},{},[584],{"type":58,"value":585},"Surviving candidate:",{"type":58,"value":587}," Name the seam and pattern, label inferred facts as\nassumptions, state that no gate has fired, preview the profile\u002Fbaseline\u002FWarp\nprototype\u002Fbenchmark scope and its cost, then ask the separate intent and\nauthorization questions from\n",{"type":52,"tag":517,"props":589,"children":591},{"href":590},"references\u002Fauthorization-checkpoint.md",[592],{"type":58,"value":590},{"type":58,"value":594},".",{"type":52,"tag":61,"props":596,"children":598},{"id":597},"inputs",[599],{"type":58,"value":600},"Inputs",{"type":52,"tag":68,"props":602,"children":603},{},[604,609,611,616],{"type":52,"tag":74,"props":605,"children":606},{},[607],{"type":58,"value":608},"Required:",{"type":58,"value":610}," the target repository and a performance, memory or scale problem\nwith a candidate seam. ",{"type":52,"tag":74,"props":612,"children":613},{},[614],{"type":58,"value":615},"Optional:",{"type":58,"value":617}," explicit deployment\u002Fpackaging constraints,\nexisting profiles or logs, representative datasets and acceptance criteria.\nPrompt constraints take precedence over repository policy\u002Fconfiguration, then\nexisting logs. User corrections override inference. Never substitute an\nassumption for a stated fact or measurement.",{"type":52,"tag":61,"props":619,"children":621},{"id":620},"available-scripts",[622],{"type":58,"value":623},"Available scripts",{"type":52,"tag":366,"props":625,"children":626},{},[627,647],{"type":52,"tag":370,"props":628,"children":629},{},[630],{"type":52,"tag":374,"props":631,"children":632},{},[633,638,642],{"type":52,"tag":378,"props":634,"children":635},{},[636],{"type":58,"value":637},"Script",{"type":52,"tag":378,"props":639,"children":640},{},[641],{"type":58,"value":66},{"type":52,"tag":378,"props":643,"children":644},{},[645],{"type":58,"value":646},"Arguments",{"type":52,"tag":394,"props":648,"children":649},{},[650,672,694],{"type":52,"tag":374,"props":651,"children":652},{},[653,662,667],{"type":52,"tag":401,"props":654,"children":655},{},[656],{"type":52,"tag":92,"props":657,"children":659},{"className":658},[],[660],{"type":58,"value":661},"scripts\u002Fdriver-template.py",{"type":52,"tag":401,"props":663,"children":664},{},[665],{"type":58,"value":666},"Copy once per bottleneck; define workloads and variants",{"type":52,"tag":401,"props":668,"children":669},{},[670],{"type":58,"value":671},"Edit placeholders, then run the copied driver",{"type":52,"tag":374,"props":673,"children":674},{},[675,684,689],{"type":52,"tag":401,"props":676,"children":677},{},[678],{"type":52,"tag":92,"props":679,"children":681},{"className":680},[],[682],{"type":58,"value":683},"scripts\u002Fmeasure.py",{"type":52,"tag":401,"props":685,"children":686},{},[687],{"type":58,"value":688},"Import from drivers for synchronized timing, memory and isolated cases",{"type":52,"tag":401,"props":690,"children":691},{},[692],{"type":58,"value":693},"Python API; do not execute directly",{"type":52,"tag":374,"props":695,"children":696},{},[697,706,711],{"type":52,"tag":401,"props":698,"children":699},{},[700],{"type":52,"tag":92,"props":701,"children":703},{"className":702},[],[704],{"type":58,"value":705},"scripts\u002Fvalidate_report_schema.py",{"type":52,"tag":401,"props":707,"children":708},{},[709],{"type":58,"value":710},"Validate the delivered report and evidence links",{"type":52,"tag":401,"props":712,"children":713},{},[714],{"type":52,"tag":92,"props":715,"children":717},{"className":716},[],[718],{"type":58,"value":719},"\u003Creport-directory>",{"type":52,"tag":68,"props":721,"children":722},{},[723,725,731],{"type":58,"value":724},"Use ",{"type":52,"tag":92,"props":726,"children":728},{"className":727},[],[729],{"type":58,"value":730},"run_script(\"scripts\u002Fvalidate_report_schema.py\", args=[\"warp-evaluation-report\"])",{"type":58,"value":732},"\nwhen supported; otherwise invoke the script with Python and the report directory.",{"type":52,"tag":61,"props":734,"children":736},{"id":735},"troubleshooting",[737],{"type":58,"value":738},"Troubleshooting",{"type":52,"tag":270,"props":740,"children":741},{},[742,754,759],{"type":52,"tag":116,"props":743,"children":744},{},[745,747,752],{"type":58,"value":746},"No representative workload: mark the scope ",{"type":52,"tag":92,"props":748,"children":750},{"className":749},[],[751],{"type":58,"value":474},{"type":58,"value":753}," and name the missing\nartifact; do not invent data or fire Gate F.",{"type":52,"tag":116,"props":755,"children":756},{},[757],{"type":58,"value":758},"Warp resolves to CPU or no CUDA device: discard the run and stop before\ncorrectness or timing claims.",{"type":52,"tag":116,"props":760,"children":761},{},[762],{"type":58,"value":763},"Report validation fails: fix the report or referenced artifact; never waive\nthe schema error.",{"type":52,"tag":61,"props":765,"children":767},{"id":766},"instructions",[768],{"type":58,"value":769},"Instructions",{"type":52,"tag":68,"props":771,"children":772},{},[773],{"type":58,"value":774},"Every stage before the last can end the evaluation. Stop as soon as a gate\nfires; do not gather evidence that cannot change the scoped facts.",{"type":52,"tag":776,"props":777,"children":779},"h3",{"id":778},"_1-read-the-code-derive-the-contract-check-the-gates",[780],{"type":58,"value":781},"1. Read the code, derive the contract, check the gates",{"type":52,"tag":270,"props":783,"children":784},{},[785,796,801,806],{"type":52,"tag":116,"props":786,"children":787},{},[788,790,795],{"type":58,"value":789},"Identify a candidate and its metric with\n",{"type":52,"tag":517,"props":791,"children":793},{"href":792},"references\u002Ftarget-patterns.md",[794],{"type":58,"value":792},{"type":58,"value":594},{"type":52,"tag":116,"props":797,"children":798},{},[799],{"type":58,"value":800},"Derive devices\u002Fresidency, dtypes\u002Fshapes, sizes, frequency, process lifetime,\ngradients and packaging from the repository. Infer before asking.",{"type":52,"tag":116,"props":802,"children":803},{},[804],{"type":58,"value":805},"State the inferred contract in one line and invite correction. Unknown\nhardware, counts, sizes and tolerances are assumptions, never measurements.\nEvery inference remains open to correction and cannot satisfy a gate that\nrequires a stated fact or measurement.",{"type":52,"tag":116,"props":807,"children":808},{},[809,811,816],{"type":58,"value":810},"Check Gates A–E before profiling. Check Gate F now only if representative\nevidence already exists; otherwise carry it into stage 2. Every gate uses only\nthe exact boundaries in\n",{"type":52,"tag":517,"props":812,"children":814},{"href":813},"references\u002Frejection-gates.md",[815],{"type":58,"value":813},{"type":58,"value":594},{"type":52,"tag":366,"props":818,"children":819},{},[820,836],{"type":52,"tag":370,"props":821,"children":822},{},[823],{"type":52,"tag":374,"props":824,"children":825},{},[826,831],{"type":52,"tag":378,"props":827,"children":828},{},[829],{"type":58,"value":830},"Gate",{"type":52,"tag":378,"props":832,"children":833},{},[834],{"type":58,"value":835},"Fires when",{"type":52,"tag":394,"props":837,"children":838},{},[839,863,879,895,911,927],{"type":52,"tag":374,"props":840,"children":841},{},[842,850],{"type":52,"tag":401,"props":843,"children":844},{},[845],{"type":52,"tag":74,"props":846,"children":847},{},[848],{"type":58,"value":849},"A",{"type":52,"tag":401,"props":851,"children":852},{},[853,855,861],{"type":58,"value":854},"Production is ",{"type":52,"tag":856,"props":857,"children":858},"em",{},[859],{"type":58,"value":860},"stated",{"type":58,"value":862}," CPU-only or to need non-NVIDIA portability, with no acceptable optional CUDA path",{"type":52,"tag":374,"props":864,"children":865},{},[866,874],{"type":52,"tag":401,"props":867,"children":868},{},[869],{"type":52,"tag":74,"props":870,"children":871},{},[872],{"type":58,"value":873},"B",{"type":52,"tag":401,"props":875,"children":876},{},[877],{"type":58,"value":878},"Data must cross the host\u002Fdevice boundary per small or infrequent call and the boundary cannot be widened",{"type":52,"tag":374,"props":880,"children":881},{},[882,890],{"type":52,"tag":401,"props":883,"children":884},{},[885],{"type":52,"tag":74,"props":886,"children":887},{},[888],{"type":58,"value":889},"C",{"type":52,"tag":401,"props":891,"children":892},{},[893],{"type":58,"value":894},"The region is dense tensor algebra already mapped to a tuned framework or vendor library",{"type":52,"tag":374,"props":896,"children":897},{},[898,906],{"type":52,"tag":401,"props":899,"children":900},{},[901],{"type":52,"tag":74,"props":902,"children":903},{},[904],{"type":58,"value":905},"D",{"type":52,"tag":401,"props":907,"children":908},{},[909],{"type":58,"value":910},"A mature CUDA implementation already meets the contract, and no non-performance objective was requested",{"type":52,"tag":374,"props":912,"children":913},{},[914,922],{"type":52,"tag":401,"props":915,"children":916},{},[917],{"type":52,"tag":74,"props":918,"children":919},{},[920],{"type":58,"value":921},"E",{"type":52,"tag":401,"props":923,"children":924},{},[925],{"type":58,"value":926},"A stated policy blocks Warp's dependency, compilation, cache or fallback obligations",{"type":52,"tag":374,"props":928,"children":929},{},[930,938],{"type":52,"tag":401,"props":931,"children":932},{},[933],{"type":52,"tag":74,"props":934,"children":935},{},[936],{"type":58,"value":937},"F",{"type":52,"tag":401,"props":939,"children":940},{},[941],{"type":58,"value":942},"Representative evidence proves the region too small a share of its requested metric for any backend to move it",{"type":52,"tag":270,"props":944,"children":945},{},[946,951,956,967,972],{"type":52,"tag":116,"props":947,"children":948},{},[949],{"type":58,"value":950},"Gate F can fire in stage 1 only from representative evidence that already\nexists — a supplied profile, structural bound, or arithmetic on figures the\nuser quoted. If that evidence does not exist, Gate F remains open until stage\n2 profiling; inferred values never fire it.",{"type":52,"tag":116,"props":952,"children":953},{},[954],{"type":58,"value":955},"Gates A and E need a stated constraint. A CPU implementation, another\naccelerator, no Warp dependency, or a small dependency list proves nothing.",{"type":52,"tag":116,"props":957,"children":958},{},[959,961,966],{"type":58,"value":960},"When A\u002FE are unresolved and a pattern survives, ask whether an optional NVIDIA\npath is acceptable: named extra, soft import, existing fallback, default\ninstall unchanged. Every affirmative answer must say explicitly that Warp will\nbe prototyped and benchmarked; conditions constrain only that Warp scope. A\nnegative or undecided answer means ",{"type":52,"tag":92,"props":962,"children":964},{"className":963},[],[965],{"type":58,"value":409},{"type":58,"value":594},{"type":52,"tag":116,"props":968,"children":969},{},[970],{"type":58,"value":971},"Do not ask when another gate fired, the repository answers, or no pattern\nmatched. No pattern means no profiling.",{"type":52,"tag":116,"props":973,"children":974},{},[975,977,982],{"type":58,"value":976},"If a candidate survives, combine any intent question with the\n",{"type":52,"tag":517,"props":978,"children":979},{"href":590},[980],{"type":58,"value":981},"authorization checkpoint",{"type":58,"value":983}," — early\nfindings, exact scope, stages, resource cost — then stop. Stage 2 requires\nboth settled intent and explicit authorization.",{"type":52,"tag":776,"props":985,"children":987},{"id":986},"_2-profile-the-real-application",[988],{"type":58,"value":989},"2. Profile the real application",{"type":52,"tag":68,"props":991,"children":992},{},[993],{"type":58,"value":994},"Requires explicit authorization and a settled intent question.",{"type":52,"tag":270,"props":996,"children":997},{},[998,1003,1008,1013,1018,1023,1028,1040],{"type":52,"tag":116,"props":999,"children":1000},{},[1001],{"type":58,"value":1002},"Profile with the project's own profiler and representative entry points.",{"type":52,"tag":116,"props":1004,"children":1005},{},[1006],{"type":58,"value":1007},"Measure synchronized end-to-end stage time and peak memory before choosing a\nbackend. Report which entry points were profiled and which a gate screened.",{"type":52,"tag":116,"props":1009,"children":1010},{},[1011],{"type":58,"value":1012},"Prioritize further measurement by observed cost, not source appearance.",{"type":52,"tag":116,"props":1014,"children":1015},{},[1016],{"type":58,"value":1017},"Name each measurement by the public method and variant actually invoked. A\nfallback is an execution regime of that public seam, not a different\noperation, and a cheaper sibling method cannot screen out the named method.",{"type":52,"tag":116,"props":1019,"children":1020},{},[1021],{"type":58,"value":1022},"Confirm the timed branch ran on the intended device. Unchanged cost and\nnear-zero device allocation between host and device inputs exposes a host\nfallback.",{"type":52,"tag":116,"props":1024,"children":1025},{},[1026],{"type":58,"value":1027},"Measure the stage's free-stage ceiling and stubbed floor through the public\nboundary. Do not subtract per-op timings.",{"type":52,"tag":116,"props":1029,"children":1030},{},[1031,1033,1038],{"type":58,"value":1032},"If the measured ceiling proves the candidate cannot move its own metric,\n",{"type":52,"tag":92,"props":1034,"children":1036},{"className":1035},[],[1037],{"type":58,"value":409},{"type":58,"value":1039}," the affected scope under Gate F, preserve the evidence already\ncollected, and stop. The existing authorization already covered this\nmateriality check; do not ask for authorization again.",{"type":52,"tag":116,"props":1041,"children":1042},{},[1043,1045,1050,1052,1057],{"type":58,"value":1044},"If representative coverage is unavailable — no representative dataset,\nrunnable entry point or production distribution — record the single missing\nartifact, mark the affected scope ",{"type":52,"tag":92,"props":1046,"children":1048},{"className":1047},[],[1049],{"type":58,"value":474},{"type":58,"value":1051},", and stop. This is missing\nevidence, not Gate F and not ",{"type":52,"tag":92,"props":1053,"children":1055},{"className":1054},[],[1056],{"type":58,"value":409},{"type":58,"value":1058},". An invented workload cannot prove\nmateriality.",{"type":52,"tag":68,"props":1060,"children":1061},{},[1062,1064,1069],{"type":58,"value":1063},"Protocol: ",{"type":52,"tag":517,"props":1065,"children":1067},{"href":1066},"references\u002Fbenchmark-protocol.md",[1068],{"type":58,"value":1066},{"type":58,"value":594},{"type":52,"tag":776,"props":1071,"children":1073},{"id":1072},"_3-form-falsifiable-hypotheses",[1074],{"type":58,"value":1075},"3. Form falsifiable hypotheses",{"type":52,"tag":68,"props":1077,"children":1078},{},[1079,1081,1085],{"type":58,"value":1080},"Record per candidate: source, bottleneck evidence, objective, narrow seam,\nmechanism Warp could change, strongest incumbent, risks, acceptance threshold,\ncheapest falsifying experiment. Screen against\n",{"type":52,"tag":517,"props":1082,"children":1083},{"href":792},[1084],{"type":58,"value":792},{"type":58,"value":1086},"; if none\nsurvives, write the report and stop.",{"type":52,"tag":776,"props":1088,"children":1090},{"id":1089},"_4-write-the-contract-before-the-prototype",[1091],{"type":58,"value":1092},"4. Write the contract before the prototype",{"type":52,"tag":68,"props":1094,"children":1095},{},[1096],{"type":58,"value":1097},"Define values, dtypes, shapes, devices, errors, mutation, ordering, ties,\ncapacity\u002Foverflow, topology\u002Fdegeneracy, tolerances, required gradients, streams,\nownership, aliasing, invalidation, concurrency, capture, teardown and fallback.",{"type":52,"tag":270,"props":1099,"children":1100},{},[1101,1106,1116],{"type":52,"tag":116,"props":1102,"children":1103},{},[1104],{"type":58,"value":1105},"Fix tolerances before seeing Warp output. Never weaken a contract after a\nmismatch.",{"type":52,"tag":116,"props":1107,"children":1108},{},[1109,1114],{"type":52,"tag":92,"props":1110,"children":1112},{"className":1111},[],[1113],{"type":58,"value":409},{"type":58,"value":1115}," before prototyping if the proposed seam cannot satisfy a required\ncontract.",{"type":52,"tag":116,"props":1117,"children":1118},{},[1119],{"type":58,"value":1120},"Pre-register, before timing: workload provenance, the state variable and\nproduction range controlling cost, tuning knobs, incumbent run-to-run spread,\nand the oracle applied to every implementation.",{"type":52,"tag":68,"props":1122,"children":1123},{},[1124,1126,1131],{"type":58,"value":1125},"Hazards and adversarial checks:\n",{"type":52,"tag":517,"props":1127,"children":1129},{"href":1128},"references\u002Fsemantic-contract.md",[1130],{"type":58,"value":1128},{"type":58,"value":594},{"type":52,"tag":776,"props":1133,"children":1135},{"id":1134},"_5-improve-the-baseline-first",[1136],{"type":58,"value":1137},"5. Improve the baseline first",{"type":52,"tag":68,"props":1139,"children":1140},{},[1141,1143,1148],{"type":58,"value":1142},"Algorithm before backend: (1) a better or output-sensitive algorithm;\n(2) chunking, tiling, sparse output, layout, rematerialization; (3) the\nincumbent framework's compiler and native primitives; (4) ",{"type":52,"tag":74,"props":1144,"children":1145},{},[1146],{"type":58,"value":1147},"what the project\nalready depends on",{"type":58,"value":1149}," — its own accelerator backend, a parallel idiom it ships\nbut leaves off, or a capability an existing dependency exposes and nobody wired\nup; (5) only then narrow Warp.",{"type":52,"tag":270,"props":1151,"children":1152},{},[1153,1158,1163],{"type":52,"tag":116,"props":1154,"children":1155},{},[1156],{"type":58,"value":1157},"Compare only in-scope options: the improved incumbent, capabilities reachable\nthrough current dependencies, and Warp. Do not add unrelated libraries.",{"type":52,"tag":116,"props":1159,"children":1160},{},[1161],{"type":58,"value":1162},"Search declared dependencies for dormant backends, flags and bindings before\ncalling a route absent.",{"type":52,"tag":116,"props":1164,"children":1165},{},[1166],{"type":58,"value":1167},"Keep the improved incumbent as the oracle. A result against an untuned,\nincorrect or asymptotically inferior baseline does not establish a backend\ncomparison.",{"type":52,"tag":68,"props":1169,"children":1170},{},[1171],{"type":58,"value":1172},"Close every in-project route before prototyping Warp:",{"type":52,"tag":366,"props":1174,"children":1175},{},[1176,1191],{"type":52,"tag":370,"props":1177,"children":1178},{},[1179],{"type":52,"tag":374,"props":1180,"children":1181},{},[1182,1186],{"type":52,"tag":378,"props":1183,"children":1184},{},[1185],{"type":58,"value":382},{"type":52,"tag":378,"props":1187,"children":1188},{},[1189],{"type":58,"value":1190},"What it takes to claim it",{"type":52,"tag":394,"props":1192,"children":1193},{},[1194,1210,1226],{"type":52,"tag":374,"props":1195,"children":1196},{},[1197,1205],{"type":52,"tag":401,"props":1198,"children":1199},{},[1200],{"type":52,"tag":74,"props":1201,"children":1202},{},[1203],{"type":58,"value":1204},"measured",{"type":52,"tag":401,"props":1206,"children":1207},{},[1208],{"type":58,"value":1209},"timed through the same boundary as the baseline",{"type":52,"tag":374,"props":1211,"children":1212},{},[1213,1221],{"type":52,"tag":401,"props":1214,"children":1215},{},[1216],{"type":52,"tag":74,"props":1217,"children":1218},{},[1219],{"type":58,"value":1220},"absent",{"type":52,"tag":401,"props":1222,"children":1223},{},[1224],{"type":58,"value":1225},"a cited declaration, symbol table or missing flag proves it is unavailable",{"type":52,"tag":374,"props":1227,"children":1228},{},[1229,1237],{"type":52,"tag":401,"props":1230,"children":1231},{},[1232],{"type":52,"tag":74,"props":1233,"children":1234},{},[1235],{"type":58,"value":1236},"waived",{"type":52,"tag":401,"props":1238,"children":1239},{},[1240],{"type":58,"value":1241},"you asked the user and they chose to skip it; record their words",{"type":52,"tag":68,"props":1243,"children":1244},{},[1245,1247,1252],{"type":58,"value":1246},"A capability present but unbound is reachable, not absent. If exposing it costs\nno more than the planned Warp seam, measure it first. Ladder details:\n",{"type":52,"tag":517,"props":1248,"children":1250},{"href":1249},"references\u002Fbaselines.md",[1251],{"type":58,"value":1249},{"type":58,"value":1253},".\nWaived routes do not block stage 6; every route not explicitly waived must be\nmeasured or evidenced absent before the Warp prototype begins.",{"type":52,"tag":776,"props":1255,"children":1257},{"id":1256},"_6-prototype-the-minimum-unit",[1258],{"type":58,"value":1259},"6. Prototype the minimum unit",{"type":52,"tag":270,"props":1261,"children":1262},{},[1263,1268,1273,1278,1290,1295,1307],{"type":52,"tag":116,"props":1264,"children":1265},{},[1266],{"type":58,"value":1267},"Work in a separate copy, production unchanged, integration seam off by\ndefault. Prototype only enough to test the hypothesis.",{"type":52,"tag":116,"props":1269,"children":1270},{},[1271],{"type":58,"value":1272},"Select an explicit CUDA device for every Warp prototype and verify that Warp\nresolves it as CUDA before correctness or performance work. Never exercise or\nreport Warp's CPU backend.",{"type":52,"tag":116,"props":1274,"children":1275},{},[1276],{"type":58,"value":1277},"Size the unit by shared data and structure lifetime, not function boundaries:\nbuild\u002Frefit\u002Fquery costs that amortize together are one unit.",{"type":52,"tag":116,"props":1279,"children":1280},{},[1281,1283,1288],{"type":58,"value":1282},"Name variants before timing. Preserve each solution as an ",{"type":52,"tag":74,"props":1284,"children":1285},{},[1286],{"type":58,"value":1287},"independent",{"type":58,"value":1289},"\npatch against the pinned baseline, and prove it applies cleanly and reproduces\nthe measured result.",{"type":52,"tag":116,"props":1291,"children":1292},{},[1293],{"type":58,"value":1294},"Surface unsupported cases and overflow before timing.",{"type":52,"tag":116,"props":1296,"children":1297},{},[1298,1300,1305],{"type":58,"value":1299},"Run the adversarial contract checks, audit correctness at production scale,\nand deliberately break a branch to prove the tests fail. Any failed required\ncheck returns ",{"type":52,"tag":92,"props":1301,"children":1303},{"className":1302},[],[1304],{"type":58,"value":409},{"type":58,"value":1306}," for the affected scope.",{"type":52,"tag":116,"props":1308,"children":1309},{},[1310,1312,1316],{"type":58,"value":1311},"Before treating a disappointing compute-bound or reuse-heavy result as\nrepresentative, consider the stronger formulations in\n",{"type":52,"tag":517,"props":1313,"children":1314},{"href":792},[1315],{"type":58,"value":792},{"type":58,"value":1317},". One naive\nkernel does not bound Warp's potential.",{"type":52,"tag":776,"props":1319,"children":1321},{"id":1320},"_7-benchmark-the-whole-boundary",[1322],{"type":58,"value":1323},"7. Benchmark the whole boundary",{"type":52,"tag":270,"props":1325,"children":1326},{},[1327,1344,1349,1354,1366,1379],{"type":52,"tag":116,"props":1328,"children":1329},{},[1330,1332,1336,1338,1342],{"type":58,"value":1331},"Copy ",{"type":52,"tag":517,"props":1333,"children":1334},{"href":661},[1335],{"type":58,"value":661},{"type":58,"value":1337}," once per\nbottleneck and measure through ",{"type":52,"tag":517,"props":1339,"children":1340},{"href":683},[1341],{"type":58,"value":683},{"type":58,"value":1343},". Do\nnot hand-roll timing or memory.",{"type":52,"tag":116,"props":1345,"children":1346},{},[1347],{"type":58,"value":1348},"Warp launches are asynchronous. The measurement helper must synchronize the\nselected device immediately before starting and after enqueueing every timed\nregion, before stopping its wall timer.",{"type":52,"tag":116,"props":1350,"children":1351},{},[1352],{"type":58,"value":1353},"Include every rule-7 cost, the public call, the immediate downstream stage,\nrealistic sizes, and cold\u002Fwarm process regimes.",{"type":52,"tag":116,"props":1355,"children":1356},{},[1357,1359,1365],{"type":58,"value":1358},"Transcribe every report cell from emitted JSON; absent records read\n",{"type":52,"tag":92,"props":1360,"children":1362},{"className":1361},[],[1363],{"type":58,"value":1364},"not measured",{"type":58,"value":594},{"type":52,"tag":116,"props":1367,"children":1368},{},[1369,1371,1377],{"type":58,"value":1370},"Report ",{"type":52,"tag":92,"props":1372,"children":1374},{"className":1373},[],[1375],{"type":58,"value":1376},"null_test",{"type":58,"value":1378},", and one-time costs both separately and amortized.\nSerialize GPU measurements under an exclusive device lock. Below 1.5× is no\nmeasured difference.",{"type":52,"tag":116,"props":1380,"children":1381},{},[1382,1387],{"type":52,"tag":92,"props":1383,"children":1385},{"className":1384},[],[1386],{"type":58,"value":409},{"type":58,"value":1388}," for a seam and regime whose predeclared end-to-end performance or\nmemory requirement fails, after preserving the measurements.",{"type":52,"tag":68,"props":1390,"children":1391},{},[1392,1394,1398],{"type":58,"value":1393},"Full protocol:\n",{"type":52,"tag":517,"props":1395,"children":1396},{"href":1066},[1397],{"type":58,"value":1066},{"type":58,"value":594},{"type":52,"tag":776,"props":1400,"children":1402},{"id":1401},"_8-summarize-the-evidence",[1403],{"type":58,"value":1404},"8. Summarize the evidence",{"type":52,"tag":270,"props":1406,"children":1407},{},[1408,1413,1468,1503,1508],{"type":52,"tag":116,"props":1409,"children":1410},{},[1411],{"type":58,"value":1412},"Per seam and execution regime, report semantic results, strongest-baseline\nmeasurements, workload provenance, time, memory, lifecycle, portability and\nownership facts.",{"type":52,"tag":116,"props":1414,"children":1415},{},[1416,1417,1423,1425,1431,1432,1437,1438,1444,1445,1451,1452,1458,1460,1466],{"type":58,"value":724},{"type":52,"tag":92,"props":1418,"children":1420},{"className":1419},[],[1421],{"type":58,"value":1422},"pass",{"type":58,"value":1424},", ",{"type":52,"tag":92,"props":1426,"children":1428},{"className":1427},[],[1429],{"type":58,"value":1430},"fail",{"type":58,"value":1424},{"type":52,"tag":92,"props":1433,"children":1435},{"className":1434},[],[1436],{"type":58,"value":1364},{"type":58,"value":1424},{"type":52,"tag":92,"props":1439,"children":1441},{"className":1440},[],[1442],{"type":58,"value":1443},"not available",{"type":58,"value":1424},{"type":52,"tag":92,"props":1446,"children":1448},{"className":1447},[],[1449],{"type":58,"value":1450},"no representative data",{"type":58,"value":539},{"type":52,"tag":92,"props":1453,"children":1455},{"className":1454},[],[1456],{"type":58,"value":1457},"unknown",{"type":58,"value":1459}," or ",{"type":52,"tag":92,"props":1461,"children":1463},{"className":1462},[],[1464],{"type":58,"value":1465},"n\u002Fa",{"type":58,"value":1467}," only where a stated criterion makes that status objective.\nMissing evidence remains missing.",{"type":52,"tag":116,"props":1469,"children":1470},{},[1471,1473,1479,1481,1487,1489,1495,1497,1502],{"type":58,"value":1472},"Mark the report ",{"type":52,"tag":92,"props":1474,"children":1476},{"className":1475},[],[1477],{"type":58,"value":1478},"complete",{"type":58,"value":1480}," only when every evaluated seam and regime includes\nan end-to-end Warp measurement. Use ",{"type":52,"tag":92,"props":1482,"children":1484},{"className":1483},[],[1485],{"type":58,"value":1486},"aborted — \u003Cgate and scope>",{"type":58,"value":1488}," when a gate\nfired, or ",{"type":52,"tag":92,"props":1490,"children":1492},{"className":1491},[],[1493],{"type":58,"value":1494},"incomplete — \u003Cmissing evidence and scope>",{"type":58,"value":1496}," when representative\nevidence was unavailable. Preserve everything collected. Never deliver an\nincumbent-only report as a completed ",{"type":52,"tag":92,"props":1498,"children":1500},{"className":1499},[],[1501],{"type":58,"value":4},{"type":58,"value":594},{"type":52,"tag":116,"props":1504,"children":1505},{},[1506],{"type":58,"value":1507},"Give every stage worth ≥10 % of the measured total a census row with a status\nand a one-line reason, including the screened-out stages.",{"type":52,"tag":116,"props":1509,"children":1510},{},[1511,1513,1518],{"type":58,"value":1512},"Check each gate against the metric ",{"type":52,"tag":856,"props":1514,"children":1515},{},[1516],{"type":58,"value":1517},"that candidate's",{"type":58,"value":1519}," pattern exhausts, not\nthe study's headline objective.",{"type":52,"tag":776,"props":1521,"children":1523},{"id":1522},"_9-hand-back",[1524],{"type":58,"value":1525},"9. Hand back",{"type":52,"tag":270,"props":1527,"children":1528},{},[1529,1541,1546],{"type":52,"tag":116,"props":1530,"children":1531},{},[1532,1534,1539],{"type":58,"value":1533},"Use\n",{"type":52,"tag":517,"props":1535,"children":1537},{"href":1536},"assets\u002Fwarp-evaluation-report-template.md",[1538],{"type":58,"value":1536},{"type":58,"value":1540},"\nunchanged in schema. Record authorization and scope.",{"type":52,"tag":116,"props":1542,"children":1543},{},[1544],{"type":58,"value":1545},"One table per bottleneck: baseline, current-dependency solutions, then Warp;\nabsolute time and peak memory, ratios, contract status, evidence gaps.",{"type":52,"tag":116,"props":1547,"children":1548},{},[1549,1551,1556],{"type":58,"value":1550},"Include prose only where it explains the measurements or their bounds.\nWorking artifacts belong in ",{"type":52,"tag":92,"props":1552,"children":1554},{"className":1553},[],[1555],{"type":58,"value":553},{"type":58,"value":594},{"type":52,"tag":1558,"props":1559,"children":1564},"pre",{"className":1560,"code":1561,"language":1562,"meta":1563,"style":1563},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","uv run python scripts\u002Fvalidate_report_schema.py \u003Creport-directory>\n","bash","",[1565],{"type":52,"tag":92,"props":1566,"children":1567},{"__ignoreMap":1563},[1568],{"type":52,"tag":1569,"props":1570,"children":1573},"span",{"class":1571,"line":1572},"line",1,[1574,1580,1586,1591,1596,1602,1607,1613],{"type":52,"tag":1569,"props":1575,"children":1577},{"style":1576},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[1578],{"type":58,"value":1579},"uv",{"type":52,"tag":1569,"props":1581,"children":1583},{"style":1582},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[1584],{"type":58,"value":1585}," run",{"type":52,"tag":1569,"props":1587,"children":1588},{"style":1582},[1589],{"type":58,"value":1590}," python",{"type":52,"tag":1569,"props":1592,"children":1593},{"style":1582},[1594],{"type":58,"value":1595}," scripts\u002Fvalidate_report_schema.py",{"type":52,"tag":1569,"props":1597,"children":1599},{"style":1598},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[1600],{"type":58,"value":1601}," \u003C",{"type":52,"tag":1569,"props":1603,"children":1604},{"style":1582},[1605],{"type":58,"value":1606},"report-director",{"type":52,"tag":1569,"props":1608,"children":1610},{"style":1609},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[1611],{"type":58,"value":1612},"y",{"type":52,"tag":1569,"props":1614,"children":1615},{"style":1598},[1616],{"type":58,"value":1617},">\n",{"type":52,"tag":270,"props":1619,"children":1620},{},[1621,1626],{"type":52,"tag":116,"props":1622,"children":1623},{},[1624],{"type":58,"value":1625},"Run the validator once the first measurement establishes a census\u002Ftable, and\nagain before delivery. Fix every error.",{"type":52,"tag":116,"props":1627,"children":1628},{},[1629],{"type":58,"value":1630},"Verify builds and imports from artifacts, not exit status. Ship the exact\ndrivers run. Then stop.",{"type":52,"tag":1632,"props":1633,"children":1634},"style",{},[1635],{"type":58,"value":1636},"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":1638,"total":1796},[1639,1657,1674,1685,1697,1711,1724,1738,1751,1762,1776,1785],{"slug":1640,"name":1640,"fn":1641,"description":1642,"org":1643,"tags":1644,"stars":1654,"repoUrl":1655,"updatedAt":1656},"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},[1645,1648,1651],{"name":1646,"slug":1647,"type":15},"Documentation","documentation",{"name":1649,"slug":1650,"type":15},"MCP","mcp",{"name":1652,"slug":1653,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1658,"name":1658,"fn":1659,"description":1660,"org":1661,"tags":1662,"stars":1671,"repoUrl":1672,"updatedAt":1673},"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},[1663,1666,1669],{"name":1664,"slug":1665,"type":15},"Containers","containers",{"name":1667,"slug":1668,"type":15},"Deployment","deployment",{"name":1670,"slug":34,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1675,"name":1675,"fn":1676,"description":1677,"org":1678,"tags":1679,"stars":1671,"repoUrl":1672,"updatedAt":1684},"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},[1680,1683],{"name":1681,"slug":1682,"type":15},"CI\u002FCD","ci-cd",{"name":1667,"slug":1668,"type":15},"2026-07-14T05:25:59.97109",{"slug":1686,"name":1686,"fn":1687,"description":1688,"org":1689,"tags":1690,"stars":1671,"repoUrl":1672,"updatedAt":1696},"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},[1691,1692,1693],{"name":1681,"slug":1682,"type":15},{"name":1667,"slug":1668,"type":15},{"name":1694,"slug":1695,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1698,"name":1698,"fn":1699,"description":1700,"org":1701,"tags":1702,"stars":1671,"repoUrl":1672,"updatedAt":1710},"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},[1703,1706,1707],{"name":1704,"slug":1705,"type":15},"Debugging","debugging",{"name":1694,"slug":1695,"type":15},{"name":1708,"slug":1709,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1712,"name":1712,"fn":1713,"description":1714,"org":1715,"tags":1716,"stars":1671,"repoUrl":1672,"updatedAt":1723},"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},[1717,1720],{"name":1718,"slug":1719,"type":15},"Best Practices","best-practices",{"name":1721,"slug":1722,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1725,"name":1725,"fn":1726,"description":1727,"org":1728,"tags":1729,"stars":1671,"repoUrl":1672,"updatedAt":1737},"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},[1730,1733,1736],{"name":1731,"slug":1732,"type":15},"Machine Learning","machine-learning",{"name":1734,"slug":1735,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1739,"name":1739,"fn":1740,"description":1741,"org":1742,"tags":1743,"stars":1671,"repoUrl":1672,"updatedAt":1750},"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},[1744,1747],{"name":1745,"slug":1746,"type":15},"QA","qa",{"name":1748,"slug":1749,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1752,"name":1752,"fn":1753,"description":1754,"org":1755,"tags":1756,"stars":1671,"repoUrl":1672,"updatedAt":1761},"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},[1757,1758],{"name":1667,"slug":1668,"type":15},{"name":1759,"slug":1760,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1763,"name":1763,"fn":1764,"description":1765,"org":1766,"tags":1767,"stars":1671,"repoUrl":1672,"updatedAt":1775},"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},[1768,1771,1772],{"name":1769,"slug":1770,"type":15},"Code Review","code-review",{"name":1694,"slug":1695,"type":15},{"name":1773,"slug":1774,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1777,"name":1777,"fn":1778,"description":1779,"org":1780,"tags":1781,"stars":1671,"repoUrl":1672,"updatedAt":1784},"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},[1782,1783],{"name":1745,"slug":1746,"type":15},{"name":1748,"slug":1749,"type":15},"2026-07-14T05:25:54.928983",{"slug":1786,"name":1786,"fn":1787,"description":1788,"org":1789,"tags":1790,"stars":1671,"repoUrl":1672,"updatedAt":1795},"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},[1791,1794],{"name":1792,"slug":1793,"type":15},"Automation","automation",{"name":1681,"slug":1682,"type":15},"2026-07-30T05:29:03.275638",525,{"items":1798,"total":1827},[1799,1809,1820],{"slug":1800,"name":1800,"fn":1801,"description":1802,"org":1803,"tags":1804,"stars":23,"repoUrl":24,"updatedAt":1808},"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},[1805,1806,1807],{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-08-05T05:58:13.178762",{"slug":1810,"name":1810,"fn":1811,"description":1812,"org":1813,"tags":1814,"stars":23,"repoUrl":24,"updatedAt":1819},"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},[1815,1816,1817,1818],{"name":1704,"slug":1705,"type":15},{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-08-05T05:58:27.639091",{"slug":4,"name":4,"fn":5,"description":6,"org":1821,"tags":1822,"stars":23,"repoUrl":24,"updatedAt":25},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1823,1824,1825,1826],{"name":21,"slug":22,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},{"name":18,"slug":19,"type":15},3]