[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-cupynumeric-migration-readiness":3,"mdc--4aa4fz-key":34,"related-repo-nvidia-cupynumeric-migration-readiness":2593,"related-org-nvidia-cupynumeric-migration-readiness":2695},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"cupynumeric-migration-readiness","assess NumPy to cuPyNumeric migration readiness","Pre-migration readiness assessor for porting NumPy to cuPyNumeric. Use BEFORE substantial porting work begins when the user asks whether code will scale on GPU, whether they should migrate to cuPyNumeric, which NumPy patterns transfer cleanly, what must be refactored before porting, or mentions pre-port assessment, scaling analysis, or refactor planning. Inspect the user's source code, look up NumPy usage, cross-reference the cuPyNumeric API support manifest, and distinguish distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, Python\u002Fobject-heavy control flow, shape\u002Fdata-dependent branching, and in-place mutation hazards. Produce a verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED, with concrete refactor pointers.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,22],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Data Analysis","data-analysis",{"name":20,"slug":21,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:28:13.154911","CC-BY-4.0 OR Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fcupynumeric-migration-readiness","---\nname: cupynumeric-migration-readiness\ndescription: Pre-migration readiness assessor for porting NumPy to cuPyNumeric. Use BEFORE substantial porting work begins when the user asks whether code will scale on GPU, whether they should migrate to cuPyNumeric, which NumPy patterns transfer cleanly, what must be refactored before porting, or mentions pre-port assessment, scaling analysis, or refactor planning. Inspect the user's source code, look up NumPy usage, cross-reference the cuPyNumeric API support manifest, and distinguish distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, Python\u002Fobject-heavy control flow, shape\u002Fdata-dependent branching, and in-place mutation hazards. Produce a verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED, with concrete refactor pointers.\nlicense: CC-BY-4.0 OR Apache-2.0\ncompatibility: Knowledge-driven assessment; no cuPyNumeric install required. Runtime claims target Linux x86_64\u002Faarch64 with NVIDIA compute capability >= 7.0 and CUDA 12.x\u002F13.x. Runtime validation is delegated to cuPyNumeric Doctor.\nmetadata:\n  author: \"NVIDIA Corporation \u003Clegate@nvidia.com>\"\n  version: \"2.0.0\"\n  tags:\n  - cupynumeric\n  - legate\n  - numpy\n  - gpu\n  - distributed-computing\n  upstream: https:\u002F\u002Fgithub.com\u002Fnv-legate\u002Fcupynumeric\n  docs: https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002F\n---\n\n# cuPyNumeric Migration Readiness\n\n## Purpose\n\n**Use this skill BEFORE the migration, not during.** Answer one question: *which of the user's existing NumPy APIs will scale on cuPyNumeric, and which need refactoring, before they commit engineer-weeks to porting?* To answer it: read the source, classify each NumPy idiom by its expected multi-GPU scaling on the Legate\u002FNVIDIA GPU stack, cross-reference the bundled API-support manifest, and produce a structured verdict with per-finding reasoning and recipe pointers.\n\n**This is a static, read-only assessment.** Inspect the user's source with `Read`, `Grep`, and `Glob`. Do **not** execute the user's code, modify or write files, or print environment variables or secrets. The `legate`, and cuPyNumeric Doctor commands shown below are suggestions for the *user* to run — not actions this skill performs.\n\nIf this skill has never been seen before, head to [`references\u002Fgetting-started.md`](references\u002Fgetting-started.md) first.\n\n## When to use this skill\n\nUse when the user is **about to** migrate NumPy code to GPU and asks whether it will scale on cuPyNumeric \u002F GPU, whether they should migrate, which parts will benefit, what must change before porting, or whether the port is worth it — or mentions pre-port assessment, scaling analysis, idiom analysis, GPU refactor planning, or identifying NumPy anti-patterns for GPU.\n\n**Decline and redirect** when the request is *not* a pre-migration assessment:\n\n- **Post-migration performance \u002F profiling** (\"already ported, why is it slow?\") → point to `legate --profile` and the upstream [profiling and debugging](https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002Fuser\u002Fprofiling_debugging.html) walkthrough.\n- **Custom CUDA \u002F kernel authoring** (\"write\u002Foptimize a CUDA kernel\")\n\nA graph \u002F sparse \u002F ML \u002F NLP  workload that the user *is* asking to migrate is still **in scope**: assess it and return **NOT RECOMMENDED** via Gate 4. That is a verdict, not a decline.\n\n## Instructions\n\nRun all five steps below, in order. Read the user's code and reason about it semantically; do not emit a one-shot prose verdict.\n\n### Step 1 — Gather context\n\nElicit before scanning code. Each item below has a default tuned to the typical workload — use the default when the user does not volunteer specifics; do not block on questions.\n\n- **Source location.** Default to the current working directory when no path is given.\n- **Approximate hot-path array sizes at runtime.** Default to 30–50 million elements. Map the user's numbers (or this default) to the [Gate 2 tiers](references\u002Fdecision-framework.md#gate-2-problem-size) (65K per-GPU floor; 10M+ for real single-GPU speedup; 100M+ for multi-GPU).\n- **Target hardware.** Default to 1–4 GPUs, single-node. Confirm before assuming multi-node. For CPU-only runs, ask about RAM per node instead of FBMEM.\n- **Dominant compute pattern.** Stencil \u002F GEMM \u002F Monte Carlo \u002F reductions \u002F mixed-with-SciPy. Ask the user to name it; otherwise infer it from the code in Step 3.\n\nState the defaults you applied at the top of the assessment so the user can correct them. If a value is indeterminable, say so plainly and proceed with the qualitative-only assessment — do not fabricate numbers beyond the defaults above.\n\n### Step 2 — Load the API support manifest\n\nRead [`assets\u002Fapi-support.md`](assets\u002Fapi-support.md), the committed snapshot of the upstream NumPy-vs-cuPyNumeric comparison table. For each NumPy API the code calls, find its line and read the leading glyph:\n\n- `✓✓ numpy.X` — implemented and works on multi-GPU (the best path).\n- `✓ numpy.X` — implemented but single-GPU\u002FCPU only (caveats multi-node).\n- `🟡 numpy.X — \u003Cnote>` — partial support; read the note.\n- `✗ numpy.X` — not implemented on the cuPyNumeric distributed path. Behavior on call is version-specific (some unsupported APIs route through host NumPy, others raise an exception) — either way, hot-path use is a migration blocker. Do not promise users a silent fallback to host-NumPy.\n\nIf the `Fetched:` line is more than ~90 days old, refresh the snapshot — see the **Available Scripts** section.\n\n### Step 3 — Read the code semantically\n\nWalk the user's files with `Read` and `Grep` and classify each region of array math against [`references\u002Fidioms-that-scale.md`](references\u002Fidioms-that-scale.md) and [`references\u002Fidioms-that-block.md`](references\u002Fidioms-that-block.md) (full rationale and R-codes live there). Read semantically, not by regex: before flagging, confirm `arr` traces back to a `cupynumeric` array (or `np.*` aliased to it) and check whether the access sits inside a hot loop. Apply these rules:\n\n- **Flag element loops** (`for i in range(n): arr[i] = ...`) as blockers; treat an epoch\u002Fstep\u002Ffile loop with a vectorized body as fine — distinguish the two.\n- **Flag scalar sync** — `.item()` \u002F `float()` \u002F `int()` \u002F `bool()` \u002F `complex()` on a cuPyNumeric array inside a hot loop (per-iteration host sync); allow it at the boundary.\n- **Flag reducing conditions** — `if`\u002F`while` over an array reduction (`while np.max(err) > tol:`) syncs every iteration.\n- **Flag hoistable allocation in a loop** as a fixable inefficiency.\n- **Flag `mpi4py`** in runtime code that partitions\u002Fcommunicates array data alongside `cupynumeric` ([R108](references\u002Fidioms-that-block.md#r108)) — but first confirm it issues MPI calls on a hot path; ignore a grep hit in a README, build script, or alt-launcher.\n- **Flag `order=`** on `reshape` \u002F `asarray` \u002F `flatten` as [R109](references\u002Fidioms-that-block.md#r109) — always, regardless of whether the version warns or silently no-ops.\n- **Always cite [R304](references\u002Fidioms-that-scale.md#r304)** in INFO for `np.random.*` under multi-GPU: cross-GPU bit-identical reproducibility is impossible by default (`--gpus N` \u002F `LEGATE_GPUS` is the [Legate launcher arg](https:\u002F\u002Fdocs.nvidia.com\u002Flegate\u002Flatest\u002Fmanual\u002Fusage\u002Frunning.html)).\n- **Flag Python builtins on arrays** (`sum`\u002F`max`\u002F`min`\u002F`any`\u002F`iter(arr)`) — host-iteration fallback ([R110](references\u002Fidioms-that-block.md#r110); [upstream best practices](https:\u002F\u002Fnv-legate.github.io\u002Fcupynumeric\u002Fuser\u002Fpractices.html#use-numpy-s-functions-avoid-using-python-s-built-in-functions)). Allow `len(arr)` (shape lookup; prefer `arr.shape[0]` \u002F `arr.size` for 0-d safety).\n- **Flag `cupy` mixed with `cupynumeric`** in a hot loop ([R111](references\u002Fidioms-that-block.md#r111)); the runtimes don't share GPU memory, so every hop goes through host NumPy.\n- **Look up every NumPy API the code calls** in `assets\u002Fapi-support.md` (glyph legend in Step 2).\n\nFor the deep \"why,\" read [`references\u002Fgpu-stack.md`](references\u002Fgpu-stack.md) (memory, SM, communication, dispatch) and [`references\u002Fexecution-model.md`](references\u002Fexecution-model.md) (lazy execution, sync points, mapper).\n\n### Step 4 — Produce a structured assessment\n\nDeliver the report in this order. Cite `file:line` for every finding so the user can navigate.\n\n1. **Verdict** in one sentence — see \"Verdict framework\" below.\n1. **What works (SCALES findings)** — quote representative lines so the user sees what will speed up after the import swap.\n1. **What blocks (BLOCKS findings)** — each tied to [`idioms-that-block.md`](references\u002Fidioms-that-block.md) and a recipe in [`refactor-recipes.md`](references\u002Frefactor-recipes.md).\n1. **What's fixable (REFACTOR findings)** — group by recipe; one recipe often fixes many sites.\n1. **Compatibility \u002F cost notes (INFO findings)** — SciPy boundaries, single-GPU-only linalg \u002F FFT, RNG layout vs `--gpus N`.\n1. **API support gaps** — APIs the code calls that are unimplemented or single-GPU only per the manifest.\n1. **Decision-framework summary** — Gates 1–6 from [`references\u002Fdecision-framework.md`](references\u002Fdecision-framework.md), marked pass \u002F fail \u002F uncertain.\n1. **Recommended next steps** — which recipes to apply first, whether to port one module first, and when to involve cuPyNumeric Doctor.\n\n**All 8 sections must appear**, even when the verdict is READY or NOT RECOMMENDED. Under an empty section write **\"None for this code\"** or **\"n\u002Fa — see verdict\"** in one line — do NOT omit the heading; the headings are the structural contract the report is graded on. See [`assets\u002Fsample_report.md`](assets\u002Fsample_report.md) for worked reports.\n\n### Step 5 — Hand off to cuPyNumeric Doctor for runtime validation\n\nDirect the user to run [cuPyNumeric Doctor](https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002Fuser\u002Fdoctor.html) once they have applied the recipes and the code runs:\n\n```bash\nCUPYNUMERIC_DOCTOR=1 CUPYNUMERIC_DOCTOR_FORMAT=json CUPYNUMERIC_DOCTOR_FILENAME=doctor-report.json legate --gpus 1 main.py\n```\n\ncuPyNumeric Doctor catches at runtime what source review can miss (scalar item access, ndarray iteration, advanced indexing, `nonzero` misuse, `mpi4py` import, in-place ops on views). End the assessment at: \"now run with cuPyNumeric Doctor enabled; here is what to look for in its output.\"\n\n## Verdict framework\n\nAssign the verdict **qualitatively**, from the *kinds* of findings, not a score:\n\n| Verdict | When | Action |\n|---|---|---|\n| **READY** | No BLOCKS; few\u002Fno REFACTOR | Swap the import; benchmark |\n| **LIGHT REFACTOR** | A few recipe-fixable patterns ([R201](references\u002Fidioms-that-block.md#r201)–[R206](references\u002Fidioms-that-block.md#r206)), or one or two simple BLOCKS | Apply 1–3 recipes from [`refactor-recipes.md`](references\u002Frefactor-recipes.md); re-walk to READY |\n| **SIGNIFICANT REFACTOR** | Multiple BLOCKS in hot paths, or any [R108](references\u002Fidioms-that-block.md#r108) (`mpi4py`) — rewrites, not disqualifications | Real project; budget 1–3 engineer-weeks per module |\n| **NOT RECOMMENDED** | Only two failures: Gate 2 (arrays below the 65,536 floor) or Gate 4 (wrong compute pattern). A pile of BLOCKS does *not* land here | Restructure first or use a different runtime |\n\nApply these in order; the first match wins:\n\n1. **Gate 4 fails** (sparse \u002F graph \u002F ML \u002F sequential \u002F string) → **NOT RECOMMENDED**.\n1. **Gate 2 fails** (hot-path arrays \u003C 65,536 elements\u002FGPU, no realistic batching path) → **NOT RECOMMENDED**.\n1. **Any [R108](references\u002Fidioms-that-block.md#r108) (`mpi4py`)** → **SIGNIFICANT REFACTOR** (the parallelism-layer rewrite is the cost, not a disqualification).\n1. **Multiple BLOCKS** ([R101](references\u002Fidioms-that-block.md#r101)–[R111](references\u002Fidioms-that-block.md#r111)) across hot paths → **SIGNIFICANT REFACTOR** (count does not escalate past this — each BLOCKS has a documented recipe).\n1. **One or two recipe-fixable BLOCKS** (e.g., R101–R104 element-loop \u002F sync) → **LIGHT REFACTOR**.\n1. **Only REFACTOR patterns** (R201–R206) → **LIGHT REFACTOR**; recipes are mechanical.\n1. **No BLOCKS, no REFACTOR** → **READY**.\n1. **APIs missing from the manifest on the hot path** → demote one tier (SIGNIFICANT stays SIGNIFICANT, never NOT RECOMMENDED). Single-GPU-only APIs matter only for multi-node.\n\n**Weigh the *kinds* of findings, not their count.** One R101 in a hot loop outranks ten R001s — it destroys the scaling the R001s would have delivered. Conversely a pile of BLOCKS + R108 is *still* SIGNIFICANT, not NOT RECOMMENDED — the tiers measure engineering cost, not despair. NOT RECOMMENDED requires a *size* or *compute-pattern* failure. Full framework: [`references\u002Fdecision-framework.md`](references\u002Fdecision-framework.md).\n\n## What scales vs what blocks (at-a-glance)\n\n- **SCALES** (keep as-is) — vectorized elementwise, reductions, matmul \u002F einsum, `np.where`, large-per-GPU stencil slicing `arr[1:-1, 1:-1]`, `out=`, boolean-mask indexing.\n- **BLOCKS** (remove before migration) — element loops, `np.vectorize`, `for row in arr`, `.item()\u002F.tolist()\u002Fbool(arr)` in a hot loop, reducing `if`\u002F`while` in a loop, `arr[::2]`, `dtype=object`, `mpi4py`, `order=`, `min\u002Fmax\u002Fsum(arr)`.\n- **REFACTOR** (apply a [recipe](references\u002Frefactor-recipes.md)) — alloc in a loop, `x = x + y` rebind in a loop, `vstack\u002Fhstack\u002Fconcatenate` in a loop, `np.nonzero()` + indexing, view-mutation of `diag\u002Fflip\u002Fflatten`, `reshape` in a hot loop.\n- **INFO** (cost note, not a blocker) — SciPy imports, single-device `linalg.qr\u002Fsvd`, single-transform `fft.*`, size-thresholded `linalg.solve\u002Fcholesky`.\n\nFull taxonomy in [`idioms-that-scale.md`](references\u002Fidioms-that-scale.md) and [`idioms-that-block.md`](references\u002Fidioms-that-block.md). Pass over silently any API the manifest doesn't list (out of scope of the upstream table — flagging it would be noise).\n\n## Reading order\n\nThe canonical, read-in-order guide lives in [`references\u002Fgetting-started.md`](references\u002Fgetting-started.md#must-read-references-in-order) — read it once for orientation.\n\nFor a non-trivial assessment the must-reads are [`idioms-that-block.md`](references\u002Fidioms-that-block.md), [`refactor-recipes.md`](references\u002Frefactor-recipes.md), and [`decision-framework.md`](references\u002Fdecision-framework.md); the rest ([`idioms-that-scale.md`](references\u002Fidioms-that-scale.md), [`gpu-stack.md`](references\u002Fgpu-stack.md), [`execution-model.md`](references\u002Fexecution-model.md), [`partitioning-and-balance.md`](references\u002Fpartitioning-and-balance.md), [`case-studies.md`](references\u002Fcase-studies.md)) are read on demand.\n\n## Limitations\n\n- **Does not run cuPyNumeric.** No runtime required; this is the pre-port check. Actual speedup measurement happens after migration.\n- **Does not auto-generate refactored code.** It identifies what to change and points to recipes; the user (or a follow-up agent) applies them.\n- **Does not profile the workload.** For runtime measurement use `legate.timing.time()` and the upstream [profiling and debugging](https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002Fuser\u002Fprofiling_debugging.html) guide.\n- **Does not replace judgment.** Pattern matching misses implicit syncs inside logging, decorators that hide `.tolist()`, runtime-data-dependent partition mismatches. Read the source too, especially in borderline cases.\n\n## Examples\n\nA worked assessment of the bundled `assets\u002Fexamples\u002F` fixtures (an example, not a template):\n\n> **Verdict: LIGHT REFACTOR.** `scales_well.py` translates cleanly; `needs_refactor.py` needs one allocation hoisted; `blocks_scaling.py` syncs every iteration via `.item()`.\n>\n> **What works:** `scales_well.py:23-31` (stencil R005), `:40-44` (reduction R002), `:18-22` (elementwise R001).\n> **What blocks:** `blocks_scaling.py:51-58` ([R104](references\u002Fidioms-that-block.md#r104) — `.item()` in hot loop) → [RR-sync](references\u002Frefactor-recipes.md#rr-sync).\n> **What's fixable:** `needs_refactor.py:21-28` ([R201](references\u002Fidioms-that-block.md#r201) — alloc in loop) → [RR-alloc](references\u002Frefactor-recipes.md#rr-alloc).\n> **Next:** apply the recipes; re-walk to READY; enable `CUPYNUMERIC_DOCTOR=1` on the first real run.\n\nThe full worked report is in [`assets\u002Fsample_report.md`](assets\u002Fsample_report.md).\n\n## Authoritative upstream references\n\n- **Comparison table** (source for `assets\u002Fapi-support.md`): https:\u002F\u002Fnv-legate.github.io\u002Fcupynumeric\u002Fapi\u002Fcomparison.html (mirror, most current) \u002F `...\u002Flatest\u002Fapi\u002Fcomparison.html` on docs.nvidia.com (canonical)\n- **Best practices**, **Doctor**, **profiling**, **differences with NumPy**, **Legate launcher** — under https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002F (`user\u002Fpractices.html`, `user\u002Fdoctor.html`, `user\u002Fprofiling_debugging.html`, `user\u002Fdifferences.html`) and https:\u002F\u002Fdocs.nvidia.com\u002Flegate\u002Flatest\u002Fmanual\u002Fusage\u002Frunning.html\n- **Source**: https:\u002F\u002Fgithub.com\u002Fnv-legate\u002Fcupynumeric\n\n## Available Scripts\n\n| Script | Purpose | Arguments |\n|---|---|---|\n| `scripts\u002Ffetch_api_support.py` | Scrape the upstream comparison table into `assets\u002Fapi-support.md`. Python stdlib only; standalone. | `--default-path` (write the committed `assets\u002Fapi-support.md`); `--docs-nvidia-url` (use canonical `docs.nvidia.com` instead of the default GitHub Pages mirror) |\n\nThe user runs this to refresh the manifest (`python scripts\u002Ffetch_api_support.py --default-path`).\n\n## Bundled references and assets\n\nThe `references\u002F` files are enumerated under **Required reading order** above (R-code ranges: idioms-that-scale.md = R001–R007 \u002F R301–R305; idioms-that-block.md = R101–R111 \u002F R201–R206). Assets: `assets\u002Fapi-support.md` (committed API snapshot, load in Step 2), `assets\u002Fsample_report.md` and `assets\u002Fexamples\u002F*.py` (worked report and fixtures).\n\n## Troubleshooting\n\n| Symptom | Cause | Fix |\n|---|---|---|\n| `Fetched:` line in the manifest > ~90 days old | Stale snapshot | Run `fetch_api_support.py --default-path` (user-run) |\n| Manifest missing or scraper fails | Upstream HTML changed | `WebFetch` the [comparison table](https:\u002F\u002Fnv-legate.github.io\u002Fcupynumeric\u002Fapi\u002Fcomparison.html) for that assessment |\n| NOT RECOMMENDED for many fixable BLOCKS | Heuristics applied out of order | Re-apply order: Gate 4 → Gate 2 → R108 → BLOCKS → REFACTOR; weigh *kinds*, not count |\n| Kernel authoring or post-migration profiling | Out of scope | Decline and redirect (see \"When to use\") — no verdict |\n",{"data":35,"body":48},{"name":4,"description":6,"license":26,"compatibility":36,"metadata":37},"Knowledge-driven assessment; no cuPyNumeric install required. Runtime claims target Linux x86_64\u002Faarch64 with NVIDIA compute capability >= 7.0 and CUDA 12.x\u002F13.x. Runtime validation is delegated to cuPyNumeric Doctor.",{"author":38,"version":39,"tags":40,"upstream":46,"docs":47},"NVIDIA Corporation \u003Clegate@nvidia.com>","2.0.0",[41,42,43,44,45],"cupynumeric","legate","numpy","gpu","distributed-computing","https:\u002F\u002Fgithub.com\u002Fnv-legate\u002Fcupynumeric","https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002F",{"type":49,"children":50},"root",[51,59,66,86,142,159,165,177,193,236,262,268,273,280,285,336,341,347,363,410,430,436,499,865,892,898,911,1035,1070,1076,1090,1175,1195,1201,1220,1379,1384,1528,1573,1579,1780,1805,1811,1827,1913,1919,1984,1990,2003,2156,2170,2176,2301,2306,2390,2402,2408,2449,2455,2587],{"type":52,"tag":53,"props":54,"children":55},"element","h1",{"id":4},[56],{"type":57,"value":58},"text","cuPyNumeric Migration Readiness",{"type":52,"tag":60,"props":61,"children":63},"h2",{"id":62},"purpose",[64],{"type":57,"value":65},"Purpose",{"type":52,"tag":67,"props":68,"children":69},"p",{},[70,76,78,84],{"type":52,"tag":71,"props":72,"children":73},"strong",{},[74],{"type":57,"value":75},"Use this skill BEFORE the migration, not during.",{"type":57,"value":77}," Answer one question: ",{"type":52,"tag":79,"props":80,"children":81},"em",{},[82],{"type":57,"value":83},"which of the user's existing NumPy APIs will scale on cuPyNumeric, and which need refactoring, before they commit engineer-weeks to porting?",{"type":57,"value":85}," To answer it: read the source, classify each NumPy idiom by its expected multi-GPU scaling on the Legate\u002FNVIDIA GPU stack, cross-reference the bundled API-support manifest, and produce a structured verdict with per-finding reasoning and recipe pointers.",{"type":52,"tag":67,"props":87,"children":88},{},[89,94,96,103,105,111,113,119,121,126,128,133,135,140],{"type":52,"tag":71,"props":90,"children":91},{},[92],{"type":57,"value":93},"This is a static, read-only assessment.",{"type":57,"value":95}," Inspect the user's source with ",{"type":52,"tag":97,"props":98,"children":100},"code",{"className":99},[],[101],{"type":57,"value":102},"Read",{"type":57,"value":104},", ",{"type":52,"tag":97,"props":106,"children":108},{"className":107},[],[109],{"type":57,"value":110},"Grep",{"type":57,"value":112},", and ",{"type":52,"tag":97,"props":114,"children":116},{"className":115},[],[117],{"type":57,"value":118},"Glob",{"type":57,"value":120},". Do ",{"type":52,"tag":71,"props":122,"children":123},{},[124],{"type":57,"value":125},"not",{"type":57,"value":127}," execute the user's code, modify or write files, or print environment variables or secrets. The ",{"type":52,"tag":97,"props":129,"children":131},{"className":130},[],[132],{"type":57,"value":42},{"type":57,"value":134},", and cuPyNumeric Doctor commands shown below are suggestions for the ",{"type":52,"tag":79,"props":136,"children":137},{},[138],{"type":57,"value":139},"user",{"type":57,"value":141}," to run — not actions this skill performs.",{"type":52,"tag":67,"props":143,"children":144},{},[145,147,157],{"type":57,"value":146},"If this skill has never been seen before, head to ",{"type":52,"tag":148,"props":149,"children":151},"a",{"href":150},"references\u002Fgetting-started.md",[152],{"type":52,"tag":97,"props":153,"children":155},{"className":154},[],[156],{"type":57,"value":150},{"type":57,"value":158}," first.",{"type":52,"tag":60,"props":160,"children":162},{"id":161},"when-to-use-this-skill",[163],{"type":57,"value":164},"When to use this skill",{"type":52,"tag":67,"props":166,"children":167},{},[168,170,175],{"type":57,"value":169},"Use when the user is ",{"type":52,"tag":71,"props":171,"children":172},{},[173],{"type":57,"value":174},"about to",{"type":57,"value":176}," migrate NumPy code to GPU and asks whether it will scale on cuPyNumeric \u002F GPU, whether they should migrate, which parts will benefit, what must change before porting, or whether the port is worth it — or mentions pre-port assessment, scaling analysis, idiom analysis, GPU refactor planning, or identifying NumPy anti-patterns for GPU.",{"type":52,"tag":67,"props":178,"children":179},{},[180,185,187,191],{"type":52,"tag":71,"props":181,"children":182},{},[183],{"type":57,"value":184},"Decline and redirect",{"type":57,"value":186}," when the request is ",{"type":52,"tag":79,"props":188,"children":189},{},[190],{"type":57,"value":125},{"type":57,"value":192}," a pre-migration assessment:",{"type":52,"tag":194,"props":195,"children":196},"ul",{},[197,226],{"type":52,"tag":198,"props":199,"children":200},"li",{},[201,206,208,214,216,224],{"type":52,"tag":71,"props":202,"children":203},{},[204],{"type":57,"value":205},"Post-migration performance \u002F profiling",{"type":57,"value":207}," (\"already ported, why is it slow?\") → point to ",{"type":52,"tag":97,"props":209,"children":211},{"className":210},[],[212],{"type":57,"value":213},"legate --profile",{"type":57,"value":215}," and the upstream ",{"type":52,"tag":148,"props":217,"children":221},{"href":218,"rel":219},"https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002Fuser\u002Fprofiling_debugging.html",[220],"nofollow",[222],{"type":57,"value":223},"profiling and debugging",{"type":57,"value":225}," walkthrough.",{"type":52,"tag":198,"props":227,"children":228},{},[229,234],{"type":52,"tag":71,"props":230,"children":231},{},[232],{"type":57,"value":233},"Custom CUDA \u002F kernel authoring",{"type":57,"value":235}," (\"write\u002Foptimize a CUDA kernel\")",{"type":52,"tag":67,"props":237,"children":238},{},[239,241,246,248,253,255,260],{"type":57,"value":240},"A graph \u002F sparse \u002F ML \u002F NLP  workload that the user ",{"type":52,"tag":79,"props":242,"children":243},{},[244],{"type":57,"value":245},"is",{"type":57,"value":247}," asking to migrate is still ",{"type":52,"tag":71,"props":249,"children":250},{},[251],{"type":57,"value":252},"in scope",{"type":57,"value":254},": assess it and return ",{"type":52,"tag":71,"props":256,"children":257},{},[258],{"type":57,"value":259},"NOT RECOMMENDED",{"type":57,"value":261}," via Gate 4. That is a verdict, not a decline.",{"type":52,"tag":60,"props":263,"children":265},{"id":264},"instructions",[266],{"type":57,"value":267},"Instructions",{"type":52,"tag":67,"props":269,"children":270},{},[271],{"type":57,"value":272},"Run all five steps below, in order. Read the user's code and reason about it semantically; do not emit a one-shot prose verdict.",{"type":52,"tag":274,"props":275,"children":277},"h3",{"id":276},"step-1-gather-context",[278],{"type":57,"value":279},"Step 1 — Gather context",{"type":52,"tag":67,"props":281,"children":282},{},[283],{"type":57,"value":284},"Elicit before scanning code. Each item below has a default tuned to the typical workload — use the default when the user does not volunteer specifics; do not block on questions.",{"type":52,"tag":194,"props":286,"children":287},{},[288,298,316,326],{"type":52,"tag":198,"props":289,"children":290},{},[291,296],{"type":52,"tag":71,"props":292,"children":293},{},[294],{"type":57,"value":295},"Source location.",{"type":57,"value":297}," Default to the current working directory when no path is given.",{"type":52,"tag":198,"props":299,"children":300},{},[301,306,308,314],{"type":52,"tag":71,"props":302,"children":303},{},[304],{"type":57,"value":305},"Approximate hot-path array sizes at runtime.",{"type":57,"value":307}," Default to 30–50 million elements. Map the user's numbers (or this default) to the ",{"type":52,"tag":148,"props":309,"children":311},{"href":310},"references\u002Fdecision-framework.md#gate-2-problem-size",[312],{"type":57,"value":313},"Gate 2 tiers",{"type":57,"value":315}," (65K per-GPU floor; 10M+ for real single-GPU speedup; 100M+ for multi-GPU).",{"type":52,"tag":198,"props":317,"children":318},{},[319,324],{"type":52,"tag":71,"props":320,"children":321},{},[322],{"type":57,"value":323},"Target hardware.",{"type":57,"value":325}," Default to 1–4 GPUs, single-node. Confirm before assuming multi-node. For CPU-only runs, ask about RAM per node instead of FBMEM.",{"type":52,"tag":198,"props":327,"children":328},{},[329,334],{"type":52,"tag":71,"props":330,"children":331},{},[332],{"type":57,"value":333},"Dominant compute pattern.",{"type":57,"value":335}," Stencil \u002F GEMM \u002F Monte Carlo \u002F reductions \u002F mixed-with-SciPy. Ask the user to name it; otherwise infer it from the code in Step 3.",{"type":52,"tag":67,"props":337,"children":338},{},[339],{"type":57,"value":340},"State the defaults you applied at the top of the assessment so the user can correct them. If a value is indeterminable, say so plainly and proceed with the qualitative-only assessment — do not fabricate numbers beyond the defaults above.",{"type":52,"tag":274,"props":342,"children":344},{"id":343},"step-2-load-the-api-support-manifest",[345],{"type":57,"value":346},"Step 2 — Load the API support manifest",{"type":52,"tag":67,"props":348,"children":349},{},[350,352,361],{"type":57,"value":351},"Read ",{"type":52,"tag":148,"props":353,"children":355},{"href":354},"assets\u002Fapi-support.md",[356],{"type":52,"tag":97,"props":357,"children":359},{"className":358},[],[360],{"type":57,"value":354},{"type":57,"value":362},", the committed snapshot of the upstream NumPy-vs-cuPyNumeric comparison table. For each NumPy API the code calls, find its line and read the leading glyph:",{"type":52,"tag":194,"props":364,"children":365},{},[366,377,388,399],{"type":52,"tag":198,"props":367,"children":368},{},[369,375],{"type":52,"tag":97,"props":370,"children":372},{"className":371},[],[373],{"type":57,"value":374},"✓✓ numpy.X",{"type":57,"value":376}," — implemented and works on multi-GPU (the best path).",{"type":52,"tag":198,"props":378,"children":379},{},[380,386],{"type":52,"tag":97,"props":381,"children":383},{"className":382},[],[384],{"type":57,"value":385},"✓ numpy.X",{"type":57,"value":387}," — implemented but single-GPU\u002FCPU only (caveats multi-node).",{"type":52,"tag":198,"props":389,"children":390},{},[391,397],{"type":52,"tag":97,"props":392,"children":394},{"className":393},[],[395],{"type":57,"value":396},"🟡 numpy.X — \u003Cnote>",{"type":57,"value":398}," — partial support; read the note.",{"type":52,"tag":198,"props":400,"children":401},{},[402,408],{"type":52,"tag":97,"props":403,"children":405},{"className":404},[],[406],{"type":57,"value":407},"✗ numpy.X",{"type":57,"value":409}," — not implemented on the cuPyNumeric distributed path. Behavior on call is version-specific (some unsupported APIs route through host NumPy, others raise an exception) — either way, hot-path use is a migration blocker. Do not promise users a silent fallback to host-NumPy.",{"type":52,"tag":67,"props":411,"children":412},{},[413,415,421,423,428],{"type":57,"value":414},"If the ",{"type":52,"tag":97,"props":416,"children":418},{"className":417},[],[419],{"type":57,"value":420},"Fetched:",{"type":57,"value":422}," line is more than ~90 days old, refresh the snapshot — see the ",{"type":52,"tag":71,"props":424,"children":425},{},[426],{"type":57,"value":427},"Available Scripts",{"type":57,"value":429}," section.",{"type":52,"tag":274,"props":431,"children":433},{"id":432},"step-3-read-the-code-semantically",[434],{"type":57,"value":435},"Step 3 — Read the code semantically",{"type":52,"tag":67,"props":437,"children":438},{},[439,441,446,448,453,455,464,465,474,476,482,484,489,491,497],{"type":57,"value":440},"Walk the user's files with ",{"type":52,"tag":97,"props":442,"children":444},{"className":443},[],[445],{"type":57,"value":102},{"type":57,"value":447}," and ",{"type":52,"tag":97,"props":449,"children":451},{"className":450},[],[452],{"type":57,"value":110},{"type":57,"value":454}," and classify each region of array math against ",{"type":52,"tag":148,"props":456,"children":458},{"href":457},"references\u002Fidioms-that-scale.md",[459],{"type":52,"tag":97,"props":460,"children":462},{"className":461},[],[463],{"type":57,"value":457},{"type":57,"value":447},{"type":52,"tag":148,"props":466,"children":468},{"href":467},"references\u002Fidioms-that-block.md",[469],{"type":52,"tag":97,"props":470,"children":472},{"className":471},[],[473],{"type":57,"value":467},{"type":57,"value":475}," (full rationale and R-codes live there). Read semantically, not by regex: before flagging, confirm ",{"type":52,"tag":97,"props":477,"children":479},{"className":478},[],[480],{"type":57,"value":481},"arr",{"type":57,"value":483}," traces back to a ",{"type":52,"tag":97,"props":485,"children":487},{"className":486},[],[488],{"type":57,"value":41},{"type":57,"value":490}," array (or ",{"type":52,"tag":97,"props":492,"children":494},{"className":493},[],[495],{"type":57,"value":496},"np.*",{"type":57,"value":498}," aliased to it) and check whether the access sits inside a hot loop. Apply these rules:",{"type":52,"tag":194,"props":500,"children":501},{},[502,520,567,600,610,640,685,733,818,848],{"type":52,"tag":198,"props":503,"children":504},{},[505,510,512,518],{"type":52,"tag":71,"props":506,"children":507},{},[508],{"type":57,"value":509},"Flag element loops",{"type":57,"value":511}," (",{"type":52,"tag":97,"props":513,"children":515},{"className":514},[],[516],{"type":57,"value":517},"for i in range(n): arr[i] = ...",{"type":57,"value":519},") as blockers; treat an epoch\u002Fstep\u002Ffile loop with a vectorized body as fine — distinguish the two.",{"type":52,"tag":198,"props":521,"children":522},{},[523,528,530,536,538,544,545,551,552,558,559,565],{"type":52,"tag":71,"props":524,"children":525},{},[526],{"type":57,"value":527},"Flag scalar sync",{"type":57,"value":529}," — ",{"type":52,"tag":97,"props":531,"children":533},{"className":532},[],[534],{"type":57,"value":535},".item()",{"type":57,"value":537}," \u002F ",{"type":52,"tag":97,"props":539,"children":541},{"className":540},[],[542],{"type":57,"value":543},"float()",{"type":57,"value":537},{"type":52,"tag":97,"props":546,"children":548},{"className":547},[],[549],{"type":57,"value":550},"int()",{"type":57,"value":537},{"type":52,"tag":97,"props":553,"children":555},{"className":554},[],[556],{"type":57,"value":557},"bool()",{"type":57,"value":537},{"type":52,"tag":97,"props":560,"children":562},{"className":561},[],[563],{"type":57,"value":564},"complex()",{"type":57,"value":566}," on a cuPyNumeric array inside a hot loop (per-iteration host sync); allow it at the boundary.",{"type":52,"tag":198,"props":568,"children":569},{},[570,575,576,582,584,590,592,598],{"type":52,"tag":71,"props":571,"children":572},{},[573],{"type":57,"value":574},"Flag reducing conditions",{"type":57,"value":529},{"type":52,"tag":97,"props":577,"children":579},{"className":578},[],[580],{"type":57,"value":581},"if",{"type":57,"value":583},"\u002F",{"type":52,"tag":97,"props":585,"children":587},{"className":586},[],[588],{"type":57,"value":589},"while",{"type":57,"value":591}," over an array reduction (",{"type":52,"tag":97,"props":593,"children":595},{"className":594},[],[596],{"type":57,"value":597},"while np.max(err) > tol:",{"type":57,"value":599},") syncs every iteration.",{"type":52,"tag":198,"props":601,"children":602},{},[603,608],{"type":52,"tag":71,"props":604,"children":605},{},[606],{"type":57,"value":607},"Flag hoistable allocation in a loop",{"type":57,"value":609}," as a fixable inefficiency.",{"type":52,"tag":198,"props":611,"children":612},{},[613,624,626,631,632,638],{"type":52,"tag":71,"props":614,"children":615},{},[616,618],{"type":57,"value":617},"Flag ",{"type":52,"tag":97,"props":619,"children":621},{"className":620},[],[622],{"type":57,"value":623},"mpi4py",{"type":57,"value":625}," in runtime code that partitions\u002Fcommunicates array data alongside ",{"type":52,"tag":97,"props":627,"children":629},{"className":628},[],[630],{"type":57,"value":41},{"type":57,"value":511},{"type":52,"tag":148,"props":633,"children":635},{"href":634},"references\u002Fidioms-that-block.md#r108",[636],{"type":57,"value":637},"R108",{"type":57,"value":639},") — but first confirm it issues MPI calls on a hot path; ignore a grep hit in a README, build script, or alt-launcher.",{"type":52,"tag":198,"props":641,"children":642},{},[643,653,655,661,662,668,669,675,677,683],{"type":52,"tag":71,"props":644,"children":645},{},[646,647],{"type":57,"value":617},{"type":52,"tag":97,"props":648,"children":650},{"className":649},[],[651],{"type":57,"value":652},"order=",{"type":57,"value":654}," on ",{"type":52,"tag":97,"props":656,"children":658},{"className":657},[],[659],{"type":57,"value":660},"reshape",{"type":57,"value":537},{"type":52,"tag":97,"props":663,"children":665},{"className":664},[],[666],{"type":57,"value":667},"asarray",{"type":57,"value":537},{"type":52,"tag":97,"props":670,"children":672},{"className":671},[],[673],{"type":57,"value":674},"flatten",{"type":57,"value":676}," as ",{"type":52,"tag":148,"props":678,"children":680},{"href":679},"references\u002Fidioms-that-block.md#r109",[681],{"type":57,"value":682},"R109",{"type":57,"value":684}," — always, regardless of whether the version warns or silently no-ops.",{"type":52,"tag":198,"props":686,"children":687},{},[688,699,701,707,709,715,716,722,724,731],{"type":52,"tag":71,"props":689,"children":690},{},[691,693],{"type":57,"value":692},"Always cite ",{"type":52,"tag":148,"props":694,"children":696},{"href":695},"references\u002Fidioms-that-scale.md#r304",[697],{"type":57,"value":698},"R304",{"type":57,"value":700}," in INFO for ",{"type":52,"tag":97,"props":702,"children":704},{"className":703},[],[705],{"type":57,"value":706},"np.random.*",{"type":57,"value":708}," under multi-GPU: cross-GPU bit-identical reproducibility is impossible by default (",{"type":52,"tag":97,"props":710,"children":712},{"className":711},[],[713],{"type":57,"value":714},"--gpus N",{"type":57,"value":537},{"type":52,"tag":97,"props":717,"children":719},{"className":718},[],[720],{"type":57,"value":721},"LEGATE_GPUS",{"type":57,"value":723}," is the ",{"type":52,"tag":148,"props":725,"children":728},{"href":726,"rel":727},"https:\u002F\u002Fdocs.nvidia.com\u002Flegate\u002Flatest\u002Fmanual\u002Fusage\u002Frunning.html",[220],[729],{"type":57,"value":730},"Legate launcher arg",{"type":57,"value":732},").",{"type":52,"tag":198,"props":734,"children":735},{},[736,741,742,748,749,755,756,762,763,769,770,776,778,784,786,793,795,801,803,809,810,816],{"type":52,"tag":71,"props":737,"children":738},{},[739],{"type":57,"value":740},"Flag Python builtins on arrays",{"type":57,"value":511},{"type":52,"tag":97,"props":743,"children":745},{"className":744},[],[746],{"type":57,"value":747},"sum",{"type":57,"value":583},{"type":52,"tag":97,"props":750,"children":752},{"className":751},[],[753],{"type":57,"value":754},"max",{"type":57,"value":583},{"type":52,"tag":97,"props":757,"children":759},{"className":758},[],[760],{"type":57,"value":761},"min",{"type":57,"value":583},{"type":52,"tag":97,"props":764,"children":766},{"className":765},[],[767],{"type":57,"value":768},"any",{"type":57,"value":583},{"type":52,"tag":97,"props":771,"children":773},{"className":772},[],[774],{"type":57,"value":775},"iter(arr)",{"type":57,"value":777},") — host-iteration fallback (",{"type":52,"tag":148,"props":779,"children":781},{"href":780},"references\u002Fidioms-that-block.md#r110",[782],{"type":57,"value":783},"R110",{"type":57,"value":785},"; ",{"type":52,"tag":148,"props":787,"children":790},{"href":788,"rel":789},"https:\u002F\u002Fnv-legate.github.io\u002Fcupynumeric\u002Fuser\u002Fpractices.html#use-numpy-s-functions-avoid-using-python-s-built-in-functions",[220],[791],{"type":57,"value":792},"upstream best practices",{"type":57,"value":794},"). Allow ",{"type":52,"tag":97,"props":796,"children":798},{"className":797},[],[799],{"type":57,"value":800},"len(arr)",{"type":57,"value":802}," (shape lookup; prefer ",{"type":52,"tag":97,"props":804,"children":806},{"className":805},[],[807],{"type":57,"value":808},"arr.shape[0]",{"type":57,"value":537},{"type":52,"tag":97,"props":811,"children":813},{"className":812},[],[814],{"type":57,"value":815},"arr.size",{"type":57,"value":817}," for 0-d safety).",{"type":52,"tag":198,"props":819,"children":820},{},[821,838,840,846],{"type":52,"tag":71,"props":822,"children":823},{},[824,825,831,833],{"type":57,"value":617},{"type":52,"tag":97,"props":826,"children":828},{"className":827},[],[829],{"type":57,"value":830},"cupy",{"type":57,"value":832}," mixed with ",{"type":52,"tag":97,"props":834,"children":836},{"className":835},[],[837],{"type":57,"value":41},{"type":57,"value":839}," in a hot loop (",{"type":52,"tag":148,"props":841,"children":843},{"href":842},"references\u002Fidioms-that-block.md#r111",[844],{"type":57,"value":845},"R111",{"type":57,"value":847},"); the runtimes don't share GPU memory, so every hop goes through host NumPy.",{"type":52,"tag":198,"props":849,"children":850},{},[851,856,858,863],{"type":52,"tag":71,"props":852,"children":853},{},[854],{"type":57,"value":855},"Look up every NumPy API the code calls",{"type":57,"value":857}," in ",{"type":52,"tag":97,"props":859,"children":861},{"className":860},[],[862],{"type":57,"value":354},{"type":57,"value":864}," (glyph legend in Step 2).",{"type":52,"tag":67,"props":866,"children":867},{},[868,870,879,881,890],{"type":57,"value":869},"For the deep \"why,\" read ",{"type":52,"tag":148,"props":871,"children":873},{"href":872},"references\u002Fgpu-stack.md",[874],{"type":52,"tag":97,"props":875,"children":877},{"className":876},[],[878],{"type":57,"value":872},{"type":57,"value":880}," (memory, SM, communication, dispatch) and ",{"type":52,"tag":148,"props":882,"children":884},{"href":883},"references\u002Fexecution-model.md",[885],{"type":52,"tag":97,"props":886,"children":888},{"className":887},[],[889],{"type":57,"value":883},{"type":57,"value":891}," (lazy execution, sync points, mapper).",{"type":52,"tag":274,"props":893,"children":895},{"id":894},"step-4-produce-a-structured-assessment",[896],{"type":57,"value":897},"Step 4 — Produce a structured assessment",{"type":52,"tag":67,"props":899,"children":900},{},[901,903,909],{"type":57,"value":902},"Deliver the report in this order. Cite ",{"type":52,"tag":97,"props":904,"children":906},{"className":905},[],[907],{"type":57,"value":908},"file:line",{"type":57,"value":910}," for every finding so the user can navigate.",{"type":52,"tag":912,"props":913,"children":914},"ol",{},[915,925,935,968,978,994,1004,1025],{"type":52,"tag":198,"props":916,"children":917},{},[918,923],{"type":52,"tag":71,"props":919,"children":920},{},[921],{"type":57,"value":922},"Verdict",{"type":57,"value":924}," in one sentence — see \"Verdict framework\" below.",{"type":52,"tag":198,"props":926,"children":927},{},[928,933],{"type":52,"tag":71,"props":929,"children":930},{},[931],{"type":57,"value":932},"What works (SCALES findings)",{"type":57,"value":934}," — quote representative lines so the user sees what will speed up after the import swap.",{"type":52,"tag":198,"props":936,"children":937},{},[938,943,945,954,956,966],{"type":52,"tag":71,"props":939,"children":940},{},[941],{"type":57,"value":942},"What blocks (BLOCKS findings)",{"type":57,"value":944}," — each tied to ",{"type":52,"tag":148,"props":946,"children":947},{"href":467},[948],{"type":52,"tag":97,"props":949,"children":951},{"className":950},[],[952],{"type":57,"value":953},"idioms-that-block.md",{"type":57,"value":955}," and a recipe in ",{"type":52,"tag":148,"props":957,"children":959},{"href":958},"references\u002Frefactor-recipes.md",[960],{"type":52,"tag":97,"props":961,"children":963},{"className":962},[],[964],{"type":57,"value":965},"refactor-recipes.md",{"type":57,"value":967},".",{"type":52,"tag":198,"props":969,"children":970},{},[971,976],{"type":52,"tag":71,"props":972,"children":973},{},[974],{"type":57,"value":975},"What's fixable (REFACTOR findings)",{"type":57,"value":977}," — group by recipe; one recipe often fixes many sites.",{"type":52,"tag":198,"props":979,"children":980},{},[981,986,988,993],{"type":52,"tag":71,"props":982,"children":983},{},[984],{"type":57,"value":985},"Compatibility \u002F cost notes (INFO findings)",{"type":57,"value":987}," — SciPy boundaries, single-GPU-only linalg \u002F FFT, RNG layout vs ",{"type":52,"tag":97,"props":989,"children":991},{"className":990},[],[992],{"type":57,"value":714},{"type":57,"value":967},{"type":52,"tag":198,"props":995,"children":996},{},[997,1002],{"type":52,"tag":71,"props":998,"children":999},{},[1000],{"type":57,"value":1001},"API support gaps",{"type":57,"value":1003}," — APIs the code calls that are unimplemented or single-GPU only per the manifest.",{"type":52,"tag":198,"props":1005,"children":1006},{},[1007,1012,1014,1023],{"type":52,"tag":71,"props":1008,"children":1009},{},[1010],{"type":57,"value":1011},"Decision-framework summary",{"type":57,"value":1013}," — Gates 1–6 from ",{"type":52,"tag":148,"props":1015,"children":1017},{"href":1016},"references\u002Fdecision-framework.md",[1018],{"type":52,"tag":97,"props":1019,"children":1021},{"className":1020},[],[1022],{"type":57,"value":1016},{"type":57,"value":1024},", marked pass \u002F fail \u002F uncertain.",{"type":52,"tag":198,"props":1026,"children":1027},{},[1028,1033],{"type":52,"tag":71,"props":1029,"children":1030},{},[1031],{"type":57,"value":1032},"Recommended next steps",{"type":57,"value":1034}," — which recipes to apply first, whether to port one module first, and when to involve cuPyNumeric Doctor.",{"type":52,"tag":67,"props":1036,"children":1037},{},[1038,1043,1045,1050,1052,1057,1059,1068],{"type":52,"tag":71,"props":1039,"children":1040},{},[1041],{"type":57,"value":1042},"All 8 sections must appear",{"type":57,"value":1044},", even when the verdict is READY or NOT RECOMMENDED. Under an empty section write ",{"type":52,"tag":71,"props":1046,"children":1047},{},[1048],{"type":57,"value":1049},"\"None for this code\"",{"type":57,"value":1051}," or ",{"type":52,"tag":71,"props":1053,"children":1054},{},[1055],{"type":57,"value":1056},"\"n\u002Fa — see verdict\"",{"type":57,"value":1058}," in one line — do NOT omit the heading; the headings are the structural contract the report is graded on. See ",{"type":52,"tag":148,"props":1060,"children":1062},{"href":1061},"assets\u002Fsample_report.md",[1063],{"type":52,"tag":97,"props":1064,"children":1066},{"className":1065},[],[1067],{"type":57,"value":1061},{"type":57,"value":1069}," for worked reports.",{"type":52,"tag":274,"props":1071,"children":1073},{"id":1072},"step-5-hand-off-to-cupynumeric-doctor-for-runtime-validation",[1074],{"type":57,"value":1075},"Step 5 — Hand off to cuPyNumeric Doctor for runtime validation",{"type":52,"tag":67,"props":1077,"children":1078},{},[1079,1081,1088],{"type":57,"value":1080},"Direct the user to run ",{"type":52,"tag":148,"props":1082,"children":1085},{"href":1083,"rel":1084},"https:\u002F\u002Fdocs.nvidia.com\u002Fcupynumeric\u002Flatest\u002Fuser\u002Fdoctor.html",[220],[1086],{"type":57,"value":1087},"cuPyNumeric Doctor",{"type":57,"value":1089}," once they have applied the recipes and the code runs:",{"type":52,"tag":1091,"props":1092,"children":1097},"pre",{"className":1093,"code":1094,"language":1095,"meta":1096,"style":1096},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","CUPYNUMERIC_DOCTOR=1 CUPYNUMERIC_DOCTOR_FORMAT=json CUPYNUMERIC_DOCTOR_FILENAME=doctor-report.json legate --gpus 1 main.py\n","bash","",[1098],{"type":52,"tag":97,"props":1099,"children":1100},{"__ignoreMap":1096},[1101],{"type":52,"tag":1102,"props":1103,"children":1106},"span",{"class":1104,"line":1105},"line",1,[1107,1113,1119,1125,1130,1134,1139,1144,1148,1153,1159,1164,1170],{"type":52,"tag":1102,"props":1108,"children":1110},{"style":1109},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[1111],{"type":57,"value":1112},"CUPYNUMERIC_DOCTOR",{"type":52,"tag":1102,"props":1114,"children":1116},{"style":1115},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[1117],{"type":57,"value":1118},"=",{"type":52,"tag":1102,"props":1120,"children":1122},{"style":1121},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[1123],{"type":57,"value":1124},"1",{"type":52,"tag":1102,"props":1126,"children":1127},{"style":1109},[1128],{"type":57,"value":1129}," CUPYNUMERIC_DOCTOR_FORMAT",{"type":52,"tag":1102,"props":1131,"children":1132},{"style":1115},[1133],{"type":57,"value":1118},{"type":52,"tag":1102,"props":1135,"children":1136},{"style":1121},[1137],{"type":57,"value":1138},"json",{"type":52,"tag":1102,"props":1140,"children":1141},{"style":1109},[1142],{"type":57,"value":1143}," CUPYNUMERIC_DOCTOR_FILENAME",{"type":52,"tag":1102,"props":1145,"children":1146},{"style":1115},[1147],{"type":57,"value":1118},{"type":52,"tag":1102,"props":1149,"children":1150},{"style":1121},[1151],{"type":57,"value":1152},"doctor-report.json",{"type":52,"tag":1102,"props":1154,"children":1156},{"style":1155},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[1157],{"type":57,"value":1158}," legate",{"type":52,"tag":1102,"props":1160,"children":1161},{"style":1121},[1162],{"type":57,"value":1163}," --gpus",{"type":52,"tag":1102,"props":1165,"children":1167},{"style":1166},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[1168],{"type":57,"value":1169}," 1",{"type":52,"tag":1102,"props":1171,"children":1172},{"style":1121},[1173],{"type":57,"value":1174}," main.py\n",{"type":52,"tag":67,"props":1176,"children":1177},{},[1178,1180,1186,1188,1193],{"type":57,"value":1179},"cuPyNumeric Doctor catches at runtime what source review can miss (scalar item access, ndarray iteration, advanced indexing, ",{"type":52,"tag":97,"props":1181,"children":1183},{"className":1182},[],[1184],{"type":57,"value":1185},"nonzero",{"type":57,"value":1187}," misuse, ",{"type":52,"tag":97,"props":1189,"children":1191},{"className":1190},[],[1192],{"type":57,"value":623},{"type":57,"value":1194}," import, in-place ops on views). End the assessment at: \"now run with cuPyNumeric Doctor enabled; here is what to look for in its output.\"",{"type":52,"tag":60,"props":1196,"children":1198},{"id":1197},"verdict-framework",[1199],{"type":57,"value":1200},"Verdict framework",{"type":52,"tag":67,"props":1202,"children":1203},{},[1204,1206,1211,1213,1218],{"type":57,"value":1205},"Assign the verdict ",{"type":52,"tag":71,"props":1207,"children":1208},{},[1209],{"type":57,"value":1210},"qualitatively",{"type":57,"value":1212},", from the ",{"type":52,"tag":79,"props":1214,"children":1215},{},[1216],{"type":57,"value":1217},"kinds",{"type":57,"value":1219}," of findings, not a score:",{"type":52,"tag":1221,"props":1222,"children":1223},"table",{},[1224,1247],{"type":52,"tag":1225,"props":1226,"children":1227},"thead",{},[1228],{"type":52,"tag":1229,"props":1230,"children":1231},"tr",{},[1232,1237,1242],{"type":52,"tag":1233,"props":1234,"children":1235},"th",{},[1236],{"type":57,"value":922},{"type":52,"tag":1233,"props":1238,"children":1239},{},[1240],{"type":57,"value":1241},"When",{"type":52,"tag":1233,"props":1243,"children":1244},{},[1245],{"type":57,"value":1246},"Action",{"type":52,"tag":1248,"props":1249,"children":1250},"tbody",{},[1251,1273,1320,1353],{"type":52,"tag":1229,"props":1252,"children":1253},{},[1254,1263,1268],{"type":52,"tag":1255,"props":1256,"children":1257},"td",{},[1258],{"type":52,"tag":71,"props":1259,"children":1260},{},[1261],{"type":57,"value":1262},"READY",{"type":52,"tag":1255,"props":1264,"children":1265},{},[1266],{"type":57,"value":1267},"No BLOCKS; few\u002Fno REFACTOR",{"type":52,"tag":1255,"props":1269,"children":1270},{},[1271],{"type":57,"value":1272},"Swap the import; benchmark",{"type":52,"tag":1229,"props":1274,"children":1275},{},[1276,1284,1305],{"type":52,"tag":1255,"props":1277,"children":1278},{},[1279],{"type":52,"tag":71,"props":1280,"children":1281},{},[1282],{"type":57,"value":1283},"LIGHT REFACTOR",{"type":52,"tag":1255,"props":1285,"children":1286},{},[1287,1289,1295,1297,1303],{"type":57,"value":1288},"A few recipe-fixable patterns (",{"type":52,"tag":148,"props":1290,"children":1292},{"href":1291},"references\u002Fidioms-that-block.md#r201",[1293],{"type":57,"value":1294},"R201",{"type":57,"value":1296},"–",{"type":52,"tag":148,"props":1298,"children":1300},{"href":1299},"references\u002Fidioms-that-block.md#r206",[1301],{"type":57,"value":1302},"R206",{"type":57,"value":1304},"), or one or two simple BLOCKS",{"type":52,"tag":1255,"props":1306,"children":1307},{},[1308,1310,1318],{"type":57,"value":1309},"Apply 1–3 recipes from ",{"type":52,"tag":148,"props":1311,"children":1312},{"href":958},[1313],{"type":52,"tag":97,"props":1314,"children":1316},{"className":1315},[],[1317],{"type":57,"value":965},{"type":57,"value":1319},"; re-walk to READY",{"type":52,"tag":1229,"props":1321,"children":1322},{},[1323,1331,1348],{"type":52,"tag":1255,"props":1324,"children":1325},{},[1326],{"type":52,"tag":71,"props":1327,"children":1328},{},[1329],{"type":57,"value":1330},"SIGNIFICANT REFACTOR",{"type":52,"tag":1255,"props":1332,"children":1333},{},[1334,1336,1340,1341,1346],{"type":57,"value":1335},"Multiple BLOCKS in hot paths, or any ",{"type":52,"tag":148,"props":1337,"children":1338},{"href":634},[1339],{"type":57,"value":637},{"type":57,"value":511},{"type":52,"tag":97,"props":1342,"children":1344},{"className":1343},[],[1345],{"type":57,"value":623},{"type":57,"value":1347},") — rewrites, not disqualifications",{"type":52,"tag":1255,"props":1349,"children":1350},{},[1351],{"type":57,"value":1352},"Real project; budget 1–3 engineer-weeks per module",{"type":52,"tag":1229,"props":1354,"children":1355},{},[1356,1363,1374],{"type":52,"tag":1255,"props":1357,"children":1358},{},[1359],{"type":52,"tag":71,"props":1360,"children":1361},{},[1362],{"type":57,"value":259},{"type":52,"tag":1255,"props":1364,"children":1365},{},[1366,1368,1372],{"type":57,"value":1367},"Only two failures: Gate 2 (arrays below the 65,536 floor) or Gate 4 (wrong compute pattern). A pile of BLOCKS does ",{"type":52,"tag":79,"props":1369,"children":1370},{},[1371],{"type":57,"value":125},{"type":57,"value":1373}," land here",{"type":52,"tag":1255,"props":1375,"children":1376},{},[1377],{"type":57,"value":1378},"Restructure first or use a different runtime",{"type":52,"tag":67,"props":1380,"children":1381},{},[1382],{"type":57,"value":1383},"Apply these in order; the first match wins:",{"type":52,"tag":912,"props":1385,"children":1386},{},[1387,1402,1417,1445,1473,1488,1504,1518],{"type":52,"tag":198,"props":1388,"children":1389},{},[1390,1395,1397,1401],{"type":52,"tag":71,"props":1391,"children":1392},{},[1393],{"type":57,"value":1394},"Gate 4 fails",{"type":57,"value":1396}," (sparse \u002F graph \u002F ML \u002F sequential \u002F string) → ",{"type":52,"tag":71,"props":1398,"children":1399},{},[1400],{"type":57,"value":259},{"type":57,"value":967},{"type":52,"tag":198,"props":1403,"children":1404},{},[1405,1410,1412,1416],{"type":52,"tag":71,"props":1406,"children":1407},{},[1408],{"type":57,"value":1409},"Gate 2 fails",{"type":57,"value":1411}," (hot-path arrays \u003C 65,536 elements\u002FGPU, no realistic batching path) → ",{"type":52,"tag":71,"props":1413,"children":1414},{},[1415],{"type":57,"value":259},{"type":57,"value":967},{"type":52,"tag":198,"props":1418,"children":1419},{},[1420,1437,1439,1443],{"type":52,"tag":71,"props":1421,"children":1422},{},[1423,1425,1429,1430,1435],{"type":57,"value":1424},"Any ",{"type":52,"tag":148,"props":1426,"children":1427},{"href":634},[1428],{"type":57,"value":637},{"type":57,"value":511},{"type":52,"tag":97,"props":1431,"children":1433},{"className":1432},[],[1434],{"type":57,"value":623},{"type":57,"value":1436},")",{"type":57,"value":1438}," → ",{"type":52,"tag":71,"props":1440,"children":1441},{},[1442],{"type":57,"value":1330},{"type":57,"value":1444}," (the parallelism-layer rewrite is the cost, not a disqualification).",{"type":52,"tag":198,"props":1446,"children":1447},{},[1448,1453,1454,1460,1461,1465,1467,1471],{"type":52,"tag":71,"props":1449,"children":1450},{},[1451],{"type":57,"value":1452},"Multiple BLOCKS",{"type":57,"value":511},{"type":52,"tag":148,"props":1455,"children":1457},{"href":1456},"references\u002Fidioms-that-block.md#r101",[1458],{"type":57,"value":1459},"R101",{"type":57,"value":1296},{"type":52,"tag":148,"props":1462,"children":1463},{"href":842},[1464],{"type":57,"value":845},{"type":57,"value":1466},") across hot paths → ",{"type":52,"tag":71,"props":1468,"children":1469},{},[1470],{"type":57,"value":1330},{"type":57,"value":1472}," (count does not escalate past this — each BLOCKS has a documented recipe).",{"type":52,"tag":198,"props":1474,"children":1475},{},[1476,1481,1483,1487],{"type":52,"tag":71,"props":1477,"children":1478},{},[1479],{"type":57,"value":1480},"One or two recipe-fixable BLOCKS",{"type":57,"value":1482}," (e.g., R101–R104 element-loop \u002F sync) → ",{"type":52,"tag":71,"props":1484,"children":1485},{},[1486],{"type":57,"value":1283},{"type":57,"value":967},{"type":52,"tag":198,"props":1489,"children":1490},{},[1491,1496,1498,1502],{"type":52,"tag":71,"props":1492,"children":1493},{},[1494],{"type":57,"value":1495},"Only REFACTOR patterns",{"type":57,"value":1497}," (R201–R206) → ",{"type":52,"tag":71,"props":1499,"children":1500},{},[1501],{"type":57,"value":1283},{"type":57,"value":1503},"; recipes are mechanical.",{"type":52,"tag":198,"props":1505,"children":1506},{},[1507,1512,1513,1517],{"type":52,"tag":71,"props":1508,"children":1509},{},[1510],{"type":57,"value":1511},"No BLOCKS, no REFACTOR",{"type":57,"value":1438},{"type":52,"tag":71,"props":1514,"children":1515},{},[1516],{"type":57,"value":1262},{"type":57,"value":967},{"type":52,"tag":198,"props":1519,"children":1520},{},[1521,1526],{"type":52,"tag":71,"props":1522,"children":1523},{},[1524],{"type":57,"value":1525},"APIs missing from the manifest on the hot path",{"type":57,"value":1527}," → demote one tier (SIGNIFICANT stays SIGNIFICANT, never NOT RECOMMENDED). Single-GPU-only APIs matter only for multi-node.",{"type":52,"tag":67,"props":1529,"children":1530},{},[1531,1542,1544,1549,1551,1556,1557,1562,1564,1572],{"type":52,"tag":71,"props":1532,"children":1533},{},[1534,1536,1540],{"type":57,"value":1535},"Weigh the ",{"type":52,"tag":79,"props":1537,"children":1538},{},[1539],{"type":57,"value":1217},{"type":57,"value":1541}," of findings, not their count.",{"type":57,"value":1543}," One R101 in a hot loop outranks ten R001s — it destroys the scaling the R001s would have delivered. Conversely a pile of BLOCKS + R108 is ",{"type":52,"tag":79,"props":1545,"children":1546},{},[1547],{"type":57,"value":1548},"still",{"type":57,"value":1550}," SIGNIFICANT, not NOT RECOMMENDED — the tiers measure engineering cost, not despair. NOT RECOMMENDED requires a ",{"type":52,"tag":79,"props":1552,"children":1553},{},[1554],{"type":57,"value":1555},"size",{"type":57,"value":1051},{"type":52,"tag":79,"props":1558,"children":1559},{},[1560],{"type":57,"value":1561},"compute-pattern",{"type":57,"value":1563}," failure. Full framework: ",{"type":52,"tag":148,"props":1565,"children":1566},{"href":1016},[1567],{"type":52,"tag":97,"props":1568,"children":1570},{"className":1569},[],[1571],{"type":57,"value":1016},{"type":57,"value":967},{"type":52,"tag":60,"props":1574,"children":1576},{"id":1575},"what-scales-vs-what-blocks-at-a-glance",[1577],{"type":57,"value":1578},"What scales vs what blocks (at-a-glance)",{"type":52,"tag":194,"props":1580,"children":1581},{},[1582,1615,1693,1747],{"type":52,"tag":198,"props":1583,"children":1584},{},[1585,1590,1592,1598,1600,1606,1607,1613],{"type":52,"tag":71,"props":1586,"children":1587},{},[1588],{"type":57,"value":1589},"SCALES",{"type":57,"value":1591}," (keep as-is) — vectorized elementwise, reductions, matmul \u002F einsum, ",{"type":52,"tag":97,"props":1593,"children":1595},{"className":1594},[],[1596],{"type":57,"value":1597},"np.where",{"type":57,"value":1599},", large-per-GPU stencil slicing ",{"type":52,"tag":97,"props":1601,"children":1603},{"className":1602},[],[1604],{"type":57,"value":1605},"arr[1:-1, 1:-1]",{"type":57,"value":104},{"type":52,"tag":97,"props":1608,"children":1610},{"className":1609},[],[1611],{"type":57,"value":1612},"out=",{"type":57,"value":1614},", boolean-mask indexing.",{"type":52,"tag":198,"props":1616,"children":1617},{},[1618,1623,1625,1631,1632,1638,1639,1645,1647,1652,1653,1658,1660,1666,1667,1673,1674,1679,1680,1685,1686,1692],{"type":52,"tag":71,"props":1619,"children":1620},{},[1621],{"type":57,"value":1622},"BLOCKS",{"type":57,"value":1624}," (remove before migration) — element loops, ",{"type":52,"tag":97,"props":1626,"children":1628},{"className":1627},[],[1629],{"type":57,"value":1630},"np.vectorize",{"type":57,"value":104},{"type":52,"tag":97,"props":1633,"children":1635},{"className":1634},[],[1636],{"type":57,"value":1637},"for row in arr",{"type":57,"value":104},{"type":52,"tag":97,"props":1640,"children":1642},{"className":1641},[],[1643],{"type":57,"value":1644},".item()\u002F.tolist()\u002Fbool(arr)",{"type":57,"value":1646}," in a hot loop, reducing ",{"type":52,"tag":97,"props":1648,"children":1650},{"className":1649},[],[1651],{"type":57,"value":581},{"type":57,"value":583},{"type":52,"tag":97,"props":1654,"children":1656},{"className":1655},[],[1657],{"type":57,"value":589},{"type":57,"value":1659}," in a loop, ",{"type":52,"tag":97,"props":1661,"children":1663},{"className":1662},[],[1664],{"type":57,"value":1665},"arr[::2]",{"type":57,"value":104},{"type":52,"tag":97,"props":1668,"children":1670},{"className":1669},[],[1671],{"type":57,"value":1672},"dtype=object",{"type":57,"value":104},{"type":52,"tag":97,"props":1675,"children":1677},{"className":1676},[],[1678],{"type":57,"value":623},{"type":57,"value":104},{"type":52,"tag":97,"props":1681,"children":1683},{"className":1682},[],[1684],{"type":57,"value":652},{"type":57,"value":104},{"type":52,"tag":97,"props":1687,"children":1689},{"className":1688},[],[1690],{"type":57,"value":1691},"min\u002Fmax\u002Fsum(arr)",{"type":57,"value":967},{"type":52,"tag":198,"props":1694,"children":1695},{},[1696,1701,1703,1708,1710,1716,1718,1724,1725,1731,1733,1739,1740,1745],{"type":52,"tag":71,"props":1697,"children":1698},{},[1699],{"type":57,"value":1700},"REFACTOR",{"type":57,"value":1702}," (apply a ",{"type":52,"tag":148,"props":1704,"children":1705},{"href":958},[1706],{"type":57,"value":1707},"recipe",{"type":57,"value":1709},") — alloc in a loop, ",{"type":52,"tag":97,"props":1711,"children":1713},{"className":1712},[],[1714],{"type":57,"value":1715},"x = x + y",{"type":57,"value":1717}," rebind in a loop, ",{"type":52,"tag":97,"props":1719,"children":1721},{"className":1720},[],[1722],{"type":57,"value":1723},"vstack\u002Fhstack\u002Fconcatenate",{"type":57,"value":1659},{"type":52,"tag":97,"props":1726,"children":1728},{"className":1727},[],[1729],{"type":57,"value":1730},"np.nonzero()",{"type":57,"value":1732}," + indexing, view-mutation of ",{"type":52,"tag":97,"props":1734,"children":1736},{"className":1735},[],[1737],{"type":57,"value":1738},"diag\u002Fflip\u002Fflatten",{"type":57,"value":104},{"type":52,"tag":97,"props":1741,"children":1743},{"className":1742},[],[1744],{"type":57,"value":660},{"type":57,"value":1746}," in a hot loop.",{"type":52,"tag":198,"props":1748,"children":1749},{},[1750,1755,1757,1763,1765,1771,1773,1779],{"type":52,"tag":71,"props":1751,"children":1752},{},[1753],{"type":57,"value":1754},"INFO",{"type":57,"value":1756}," (cost note, not a blocker) — SciPy imports, single-device ",{"type":52,"tag":97,"props":1758,"children":1760},{"className":1759},[],[1761],{"type":57,"value":1762},"linalg.qr\u002Fsvd",{"type":57,"value":1764},", single-transform ",{"type":52,"tag":97,"props":1766,"children":1768},{"className":1767},[],[1769],{"type":57,"value":1770},"fft.*",{"type":57,"value":1772},", size-thresholded ",{"type":52,"tag":97,"props":1774,"children":1776},{"className":1775},[],[1777],{"type":57,"value":1778},"linalg.solve\u002Fcholesky",{"type":57,"value":967},{"type":52,"tag":67,"props":1781,"children":1782},{},[1783,1785,1794,1795,1803],{"type":57,"value":1784},"Full taxonomy in ",{"type":52,"tag":148,"props":1786,"children":1787},{"href":457},[1788],{"type":52,"tag":97,"props":1789,"children":1791},{"className":1790},[],[1792],{"type":57,"value":1793},"idioms-that-scale.md",{"type":57,"value":447},{"type":52,"tag":148,"props":1796,"children":1797},{"href":467},[1798],{"type":52,"tag":97,"props":1799,"children":1801},{"className":1800},[],[1802],{"type":57,"value":953},{"type":57,"value":1804},". Pass over silently any API the manifest doesn't list (out of scope of the upstream table — flagging it would be noise).",{"type":52,"tag":60,"props":1806,"children":1808},{"id":1807},"reading-order",[1809],{"type":57,"value":1810},"Reading order",{"type":52,"tag":67,"props":1812,"children":1813},{},[1814,1816,1825],{"type":57,"value":1815},"The canonical, read-in-order guide lives in ",{"type":52,"tag":148,"props":1817,"children":1819},{"href":1818},"references\u002Fgetting-started.md#must-read-references-in-order",[1820],{"type":52,"tag":97,"props":1821,"children":1823},{"className":1822},[],[1824],{"type":57,"value":150},{"type":57,"value":1826}," — read it once for orientation.",{"type":52,"tag":67,"props":1828,"children":1829},{},[1830,1832,1840,1841,1849,1850,1859,1861,1869,1870,1879,1880,1889,1890,1900,1901,1911],{"type":57,"value":1831},"For a non-trivial assessment the must-reads are ",{"type":52,"tag":148,"props":1833,"children":1834},{"href":467},[1835],{"type":52,"tag":97,"props":1836,"children":1838},{"className":1837},[],[1839],{"type":57,"value":953},{"type":57,"value":104},{"type":52,"tag":148,"props":1842,"children":1843},{"href":958},[1844],{"type":52,"tag":97,"props":1845,"children":1847},{"className":1846},[],[1848],{"type":57,"value":965},{"type":57,"value":112},{"type":52,"tag":148,"props":1851,"children":1852},{"href":1016},[1853],{"type":52,"tag":97,"props":1854,"children":1856},{"className":1855},[],[1857],{"type":57,"value":1858},"decision-framework.md",{"type":57,"value":1860},"; the rest (",{"type":52,"tag":148,"props":1862,"children":1863},{"href":457},[1864],{"type":52,"tag":97,"props":1865,"children":1867},{"className":1866},[],[1868],{"type":57,"value":1793},{"type":57,"value":104},{"type":52,"tag":148,"props":1871,"children":1872},{"href":872},[1873],{"type":52,"tag":97,"props":1874,"children":1876},{"className":1875},[],[1877],{"type":57,"value":1878},"gpu-stack.md",{"type":57,"value":104},{"type":52,"tag":148,"props":1881,"children":1882},{"href":883},[1883],{"type":52,"tag":97,"props":1884,"children":1886},{"className":1885},[],[1887],{"type":57,"value":1888},"execution-model.md",{"type":57,"value":104},{"type":52,"tag":148,"props":1891,"children":1893},{"href":1892},"references\u002Fpartitioning-and-balance.md",[1894],{"type":52,"tag":97,"props":1895,"children":1897},{"className":1896},[],[1898],{"type":57,"value":1899},"partitioning-and-balance.md",{"type":57,"value":104},{"type":52,"tag":148,"props":1902,"children":1904},{"href":1903},"references\u002Fcase-studies.md",[1905],{"type":52,"tag":97,"props":1906,"children":1908},{"className":1907},[],[1909],{"type":57,"value":1910},"case-studies.md",{"type":57,"value":1912},") are read on demand.",{"type":52,"tag":60,"props":1914,"children":1916},{"id":1915},"limitations",[1917],{"type":57,"value":1918},"Limitations",{"type":52,"tag":194,"props":1920,"children":1921},{},[1922,1932,1942,1966],{"type":52,"tag":198,"props":1923,"children":1924},{},[1925,1930],{"type":52,"tag":71,"props":1926,"children":1927},{},[1928],{"type":57,"value":1929},"Does not run cuPyNumeric.",{"type":57,"value":1931}," No runtime required; this is the pre-port check. Actual speedup measurement happens after migration.",{"type":52,"tag":198,"props":1933,"children":1934},{},[1935,1940],{"type":52,"tag":71,"props":1936,"children":1937},{},[1938],{"type":57,"value":1939},"Does not auto-generate refactored code.",{"type":57,"value":1941}," It identifies what to change and points to recipes; the user (or a follow-up agent) applies them.",{"type":52,"tag":198,"props":1943,"children":1944},{},[1945,1950,1952,1958,1959,1964],{"type":52,"tag":71,"props":1946,"children":1947},{},[1948],{"type":57,"value":1949},"Does not profile the workload.",{"type":57,"value":1951}," For runtime measurement use ",{"type":52,"tag":97,"props":1953,"children":1955},{"className":1954},[],[1956],{"type":57,"value":1957},"legate.timing.time()",{"type":57,"value":215},{"type":52,"tag":148,"props":1960,"children":1962},{"href":218,"rel":1961},[220],[1963],{"type":57,"value":223},{"type":57,"value":1965}," guide.",{"type":52,"tag":198,"props":1967,"children":1968},{},[1969,1974,1976,1982],{"type":52,"tag":71,"props":1970,"children":1971},{},[1972],{"type":57,"value":1973},"Does not replace judgment.",{"type":57,"value":1975}," Pattern matching misses implicit syncs inside logging, decorators that hide ",{"type":52,"tag":97,"props":1977,"children":1979},{"className":1978},[],[1980],{"type":57,"value":1981},".tolist()",{"type":57,"value":1983},", runtime-data-dependent partition mismatches. Read the source too, especially in borderline cases.",{"type":52,"tag":60,"props":1985,"children":1987},{"id":1986},"examples",[1988],{"type":57,"value":1989},"Examples",{"type":52,"tag":67,"props":1991,"children":1992},{},[1993,1995,2001],{"type":57,"value":1994},"A worked assessment of the bundled ",{"type":52,"tag":97,"props":1996,"children":1998},{"className":1997},[],[1999],{"type":57,"value":2000},"assets\u002Fexamples\u002F",{"type":57,"value":2002}," fixtures (an example, not a template):",{"type":52,"tag":2004,"props":2005,"children":2006},"blockquote",{},[2007,2047],{"type":52,"tag":67,"props":2008,"children":2009},{},[2010,2015,2017,2023,2025,2031,2033,2039,2041,2046],{"type":52,"tag":71,"props":2011,"children":2012},{},[2013],{"type":57,"value":2014},"Verdict: LIGHT REFACTOR.",{"type":57,"value":2016}," ",{"type":52,"tag":97,"props":2018,"children":2020},{"className":2019},[],[2021],{"type":57,"value":2022},"scales_well.py",{"type":57,"value":2024}," translates cleanly; ",{"type":52,"tag":97,"props":2026,"children":2028},{"className":2027},[],[2029],{"type":57,"value":2030},"needs_refactor.py",{"type":57,"value":2032}," needs one allocation hoisted; ",{"type":52,"tag":97,"props":2034,"children":2036},{"className":2035},[],[2037],{"type":57,"value":2038},"blocks_scaling.py",{"type":57,"value":2040}," syncs every iteration via ",{"type":52,"tag":97,"props":2042,"children":2044},{"className":2043},[],[2045],{"type":57,"value":535},{"type":57,"value":967},{"type":52,"tag":67,"props":2048,"children":2049},{},[2050,2055,2056,2062,2064,2070,2072,2078,2080,2085,2086,2092,2093,2099,2100,2105,2107,2113,2115,2120,2121,2127,2128,2132,2134,2140,2141,2146,2148,2154],{"type":52,"tag":71,"props":2051,"children":2052},{},[2053],{"type":57,"value":2054},"What works:",{"type":57,"value":2016},{"type":52,"tag":97,"props":2057,"children":2059},{"className":2058},[],[2060],{"type":57,"value":2061},"scales_well.py:23-31",{"type":57,"value":2063}," (stencil R005), ",{"type":52,"tag":97,"props":2065,"children":2067},{"className":2066},[],[2068],{"type":57,"value":2069},":40-44",{"type":57,"value":2071}," (reduction R002), ",{"type":52,"tag":97,"props":2073,"children":2075},{"className":2074},[],[2076],{"type":57,"value":2077},":18-22",{"type":57,"value":2079}," (elementwise R001).\n",{"type":52,"tag":71,"props":2081,"children":2082},{},[2083],{"type":57,"value":2084},"What blocks:",{"type":57,"value":2016},{"type":52,"tag":97,"props":2087,"children":2089},{"className":2088},[],[2090],{"type":57,"value":2091},"blocks_scaling.py:51-58",{"type":57,"value":511},{"type":52,"tag":148,"props":2094,"children":2096},{"href":2095},"references\u002Fidioms-that-block.md#r104",[2097],{"type":57,"value":2098},"R104",{"type":57,"value":529},{"type":52,"tag":97,"props":2101,"children":2103},{"className":2102},[],[2104],{"type":57,"value":535},{"type":57,"value":2106}," in hot loop) → ",{"type":52,"tag":148,"props":2108,"children":2110},{"href":2109},"references\u002Frefactor-recipes.md#rr-sync",[2111],{"type":57,"value":2112},"RR-sync",{"type":57,"value":2114},".\n",{"type":52,"tag":71,"props":2116,"children":2117},{},[2118],{"type":57,"value":2119},"What's fixable:",{"type":57,"value":2016},{"type":52,"tag":97,"props":2122,"children":2124},{"className":2123},[],[2125],{"type":57,"value":2126},"needs_refactor.py:21-28",{"type":57,"value":511},{"type":52,"tag":148,"props":2129,"children":2130},{"href":1291},[2131],{"type":57,"value":1294},{"type":57,"value":2133}," — alloc in loop) → ",{"type":52,"tag":148,"props":2135,"children":2137},{"href":2136},"references\u002Frefactor-recipes.md#rr-alloc",[2138],{"type":57,"value":2139},"RR-alloc",{"type":57,"value":2114},{"type":52,"tag":71,"props":2142,"children":2143},{},[2144],{"type":57,"value":2145},"Next:",{"type":57,"value":2147}," apply the recipes; re-walk to READY; enable ",{"type":52,"tag":97,"props":2149,"children":2151},{"className":2150},[],[2152],{"type":57,"value":2153},"CUPYNUMERIC_DOCTOR=1",{"type":57,"value":2155}," on the first real run.",{"type":52,"tag":67,"props":2157,"children":2158},{},[2159,2161,2169],{"type":57,"value":2160},"The full worked report is in ",{"type":52,"tag":148,"props":2162,"children":2163},{"href":1061},[2164],{"type":52,"tag":97,"props":2165,"children":2167},{"className":2166},[],[2168],{"type":57,"value":1061},{"type":57,"value":967},{"type":52,"tag":60,"props":2171,"children":2173},{"id":2172},"authoritative-upstream-references",[2174],{"type":57,"value":2175},"Authoritative upstream references",{"type":52,"tag":194,"props":2177,"children":2178},{},[2179,2212,2286],{"type":52,"tag":198,"props":2180,"children":2181},{},[2182,2187,2189,2194,2196,2202,2204,2210],{"type":52,"tag":71,"props":2183,"children":2184},{},[2185],{"type":57,"value":2186},"Comparison table",{"type":57,"value":2188}," (source for ",{"type":52,"tag":97,"props":2190,"children":2192},{"className":2191},[],[2193],{"type":57,"value":354},{"type":57,"value":2195},"): ",{"type":52,"tag":148,"props":2197,"children":2200},{"href":2198,"rel":2199},"https:\u002F\u002Fnv-legate.github.io\u002Fcupynumeric\u002Fapi\u002Fcomparison.html",[220],[2201],{"type":57,"value":2198},{"type":57,"value":2203}," (mirror, most current) \u002F ",{"type":52,"tag":97,"props":2205,"children":2207},{"className":2206},[],[2208],{"type":57,"value":2209},"...\u002Flatest\u002Fapi\u002Fcomparison.html",{"type":57,"value":2211}," on docs.nvidia.com (canonical)",{"type":52,"tag":198,"props":2213,"children":2214},{},[2215,2220,2221,2226,2227,2232,2233,2238,2239,2244,2246,2251,2252,2258,2259,2265,2266,2272,2273,2279,2281],{"type":52,"tag":71,"props":2216,"children":2217},{},[2218],{"type":57,"value":2219},"Best practices",{"type":57,"value":104},{"type":52,"tag":71,"props":2222,"children":2223},{},[2224],{"type":57,"value":2225},"Doctor",{"type":57,"value":104},{"type":52,"tag":71,"props":2228,"children":2229},{},[2230],{"type":57,"value":2231},"profiling",{"type":57,"value":104},{"type":52,"tag":71,"props":2234,"children":2235},{},[2236],{"type":57,"value":2237},"differences with NumPy",{"type":57,"value":104},{"type":52,"tag":71,"props":2240,"children":2241},{},[2242],{"type":57,"value":2243},"Legate launcher",{"type":57,"value":2245}," — under ",{"type":52,"tag":148,"props":2247,"children":2249},{"href":47,"rel":2248},[220],[2250],{"type":57,"value":47},{"type":57,"value":511},{"type":52,"tag":97,"props":2253,"children":2255},{"className":2254},[],[2256],{"type":57,"value":2257},"user\u002Fpractices.html",{"type":57,"value":104},{"type":52,"tag":97,"props":2260,"children":2262},{"className":2261},[],[2263],{"type":57,"value":2264},"user\u002Fdoctor.html",{"type":57,"value":104},{"type":52,"tag":97,"props":2267,"children":2269},{"className":2268},[],[2270],{"type":57,"value":2271},"user\u002Fprofiling_debugging.html",{"type":57,"value":104},{"type":52,"tag":97,"props":2274,"children":2276},{"className":2275},[],[2277],{"type":57,"value":2278},"user\u002Fdifferences.html",{"type":57,"value":2280},") and ",{"type":52,"tag":148,"props":2282,"children":2284},{"href":726,"rel":2283},[220],[2285],{"type":57,"value":726},{"type":52,"tag":198,"props":2287,"children":2288},{},[2289,2294,2296],{"type":52,"tag":71,"props":2290,"children":2291},{},[2292],{"type":57,"value":2293},"Source",{"type":57,"value":2295},": ",{"type":52,"tag":148,"props":2297,"children":2299},{"href":46,"rel":2298},[220],[2300],{"type":57,"value":46},{"type":52,"tag":60,"props":2302,"children":2304},{"id":2303},"available-scripts",[2305],{"type":57,"value":427},{"type":52,"tag":1221,"props":2307,"children":2308},{},[2309,2329],{"type":52,"tag":1225,"props":2310,"children":2311},{},[2312],{"type":52,"tag":1229,"props":2313,"children":2314},{},[2315,2320,2324],{"type":52,"tag":1233,"props":2316,"children":2317},{},[2318],{"type":57,"value":2319},"Script",{"type":52,"tag":1233,"props":2321,"children":2322},{},[2323],{"type":57,"value":65},{"type":52,"tag":1233,"props":2325,"children":2326},{},[2327],{"type":57,"value":2328},"Arguments",{"type":52,"tag":1248,"props":2330,"children":2331},{},[2332],{"type":52,"tag":1229,"props":2333,"children":2334},{},[2335,2344,2356],{"type":52,"tag":1255,"props":2336,"children":2337},{},[2338],{"type":52,"tag":97,"props":2339,"children":2341},{"className":2340},[],[2342],{"type":57,"value":2343},"scripts\u002Ffetch_api_support.py",{"type":52,"tag":1255,"props":2345,"children":2346},{},[2347,2349,2354],{"type":57,"value":2348},"Scrape the upstream comparison table into ",{"type":52,"tag":97,"props":2350,"children":2352},{"className":2351},[],[2353],{"type":57,"value":354},{"type":57,"value":2355},". Python stdlib only; standalone.",{"type":52,"tag":1255,"props":2357,"children":2358},{},[2359,2365,2367,2372,2374,2380,2382,2388],{"type":52,"tag":97,"props":2360,"children":2362},{"className":2361},[],[2363],{"type":57,"value":2364},"--default-path",{"type":57,"value":2366}," (write the committed ",{"type":52,"tag":97,"props":2368,"children":2370},{"className":2369},[],[2371],{"type":57,"value":354},{"type":57,"value":2373},"); ",{"type":52,"tag":97,"props":2375,"children":2377},{"className":2376},[],[2378],{"type":57,"value":2379},"--docs-nvidia-url",{"type":57,"value":2381}," (use canonical ",{"type":52,"tag":97,"props":2383,"children":2385},{"className":2384},[],[2386],{"type":57,"value":2387},"docs.nvidia.com",{"type":57,"value":2389}," instead of the default GitHub Pages mirror)",{"type":52,"tag":67,"props":2391,"children":2392},{},[2393,2395,2401],{"type":57,"value":2394},"The user runs this to refresh the manifest (",{"type":52,"tag":97,"props":2396,"children":2398},{"className":2397},[],[2399],{"type":57,"value":2400},"python scripts\u002Ffetch_api_support.py --default-path",{"type":57,"value":732},{"type":52,"tag":60,"props":2403,"children":2405},{"id":2404},"bundled-references-and-assets",[2406],{"type":57,"value":2407},"Bundled references and assets",{"type":52,"tag":67,"props":2409,"children":2410},{},[2411,2413,2419,2421,2426,2428,2433,2435,2440,2441,2447],{"type":57,"value":2412},"The ",{"type":52,"tag":97,"props":2414,"children":2416},{"className":2415},[],[2417],{"type":57,"value":2418},"references\u002F",{"type":57,"value":2420}," files are enumerated under ",{"type":52,"tag":71,"props":2422,"children":2423},{},[2424],{"type":57,"value":2425},"Required reading order",{"type":57,"value":2427}," above (R-code ranges: idioms-that-scale.md = R001–R007 \u002F R301–R305; idioms-that-block.md = R101–R111 \u002F R201–R206). Assets: ",{"type":52,"tag":97,"props":2429,"children":2431},{"className":2430},[],[2432],{"type":57,"value":354},{"type":57,"value":2434}," (committed API snapshot, load in Step 2), ",{"type":52,"tag":97,"props":2436,"children":2438},{"className":2437},[],[2439],{"type":57,"value":1061},{"type":57,"value":447},{"type":52,"tag":97,"props":2442,"children":2444},{"className":2443},[],[2445],{"type":57,"value":2446},"assets\u002Fexamples\u002F*.py",{"type":57,"value":2448}," (worked report and fixtures).",{"type":52,"tag":60,"props":2450,"children":2452},{"id":2451},"troubleshooting",[2453],{"type":57,"value":2454},"Troubleshooting",{"type":52,"tag":1221,"props":2456,"children":2457},{},[2458,2479],{"type":52,"tag":1225,"props":2459,"children":2460},{},[2461],{"type":52,"tag":1229,"props":2462,"children":2463},{},[2464,2469,2474],{"type":52,"tag":1233,"props":2465,"children":2466},{},[2467],{"type":57,"value":2468},"Symptom",{"type":52,"tag":1233,"props":2470,"children":2471},{},[2472],{"type":57,"value":2473},"Cause",{"type":52,"tag":1233,"props":2475,"children":2476},{},[2477],{"type":57,"value":2478},"Fix",{"type":52,"tag":1248,"props":2480,"children":2481},{},[2482,2513,2545,2569],{"type":52,"tag":1229,"props":2483,"children":2484},{},[2485,2495,2500],{"type":52,"tag":1255,"props":2486,"children":2487},{},[2488,2493],{"type":52,"tag":97,"props":2489,"children":2491},{"className":2490},[],[2492],{"type":57,"value":420},{"type":57,"value":2494}," line in the manifest > ~90 days old",{"type":52,"tag":1255,"props":2496,"children":2497},{},[2498],{"type":57,"value":2499},"Stale snapshot",{"type":52,"tag":1255,"props":2501,"children":2502},{},[2503,2505,2511],{"type":57,"value":2504},"Run ",{"type":52,"tag":97,"props":2506,"children":2508},{"className":2507},[],[2509],{"type":57,"value":2510},"fetch_api_support.py --default-path",{"type":57,"value":2512}," (user-run)",{"type":52,"tag":1229,"props":2514,"children":2515},{},[2516,2521,2526],{"type":52,"tag":1255,"props":2517,"children":2518},{},[2519],{"type":57,"value":2520},"Manifest missing or scraper fails",{"type":52,"tag":1255,"props":2522,"children":2523},{},[2524],{"type":57,"value":2525},"Upstream HTML changed",{"type":52,"tag":1255,"props":2527,"children":2528},{},[2529,2535,2537,2543],{"type":52,"tag":97,"props":2530,"children":2532},{"className":2531},[],[2533],{"type":57,"value":2534},"WebFetch",{"type":57,"value":2536}," the ",{"type":52,"tag":148,"props":2538,"children":2540},{"href":2198,"rel":2539},[220],[2541],{"type":57,"value":2542},"comparison table",{"type":57,"value":2544}," for that assessment",{"type":52,"tag":1229,"props":2546,"children":2547},{},[2548,2553,2558],{"type":52,"tag":1255,"props":2549,"children":2550},{},[2551],{"type":57,"value":2552},"NOT RECOMMENDED for many fixable BLOCKS",{"type":52,"tag":1255,"props":2554,"children":2555},{},[2556],{"type":57,"value":2557},"Heuristics applied out of order",{"type":52,"tag":1255,"props":2559,"children":2560},{},[2561,2563,2567],{"type":57,"value":2562},"Re-apply order: Gate 4 → Gate 2 → R108 → BLOCKS → REFACTOR; weigh ",{"type":52,"tag":79,"props":2564,"children":2565},{},[2566],{"type":57,"value":1217},{"type":57,"value":2568},", not count",{"type":52,"tag":1229,"props":2570,"children":2571},{},[2572,2577,2582],{"type":52,"tag":1255,"props":2573,"children":2574},{},[2575],{"type":57,"value":2576},"Kernel authoring or post-migration profiling",{"type":52,"tag":1255,"props":2578,"children":2579},{},[2580],{"type":57,"value":2581},"Out of scope",{"type":52,"tag":1255,"props":2583,"children":2584},{},[2585],{"type":57,"value":2586},"Decline and redirect (see \"When to use\") — no verdict",{"type":52,"tag":2588,"props":2589,"children":2590},"style",{},[2591],{"type":57,"value":2592},"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":2594,"total":2694},[2595,2608,2622,2636,2648,2665,2680],{"slug":2596,"name":2596,"fn":2597,"description":2598,"org":2599,"tags":2600,"stars":23,"repoUrl":24,"updatedAt":2607},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2601,2602,2605,2606],{"name":17,"slug":18,"type":15},{"name":2603,"slug":2604,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:28:43.176466",{"slug":2609,"name":2609,"fn":2610,"description":2611,"org":2612,"tags":2613,"stars":23,"repoUrl":24,"updatedAt":2621},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2614,2617,2620],{"name":2615,"slug":2616,"type":15},"Deployment","deployment",{"name":2618,"slug":2619,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":2623,"name":2623,"fn":2624,"description":2625,"org":2626,"tags":2627,"stars":23,"repoUrl":24,"updatedAt":2635},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2628,2631,2632],{"name":2629,"slug":2630,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":2633,"slug":2634,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":2637,"name":2637,"fn":2638,"description":2639,"org":2640,"tags":2641,"stars":23,"repoUrl":24,"updatedAt":2647},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2642,2643,2644],{"name":17,"slug":18,"type":15},{"name":9,"slug":8,"type":15},{"name":2645,"slug":2646,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":2649,"name":2649,"fn":2650,"description":2651,"org":2652,"tags":2653,"stars":23,"repoUrl":24,"updatedAt":2664},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2654,2657,2660,2661],{"name":2655,"slug":2656,"type":15},"Automation","automation",{"name":2658,"slug":2659,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":2662,"slug":2663,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":2666,"name":2666,"fn":2667,"description":2668,"org":2669,"tags":2670,"stars":23,"repoUrl":24,"updatedAt":2679},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2671,2672,2675,2676],{"name":2615,"slug":2616,"type":15},{"name":2673,"slug":2674,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":2677,"slug":2678,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":2681,"name":2681,"fn":2682,"description":2683,"org":2684,"tags":2685,"stars":23,"repoUrl":24,"updatedAt":2693},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2686,2687,2690],{"name":9,"slug":8,"type":15},{"name":2688,"slug":2689,"type":15},"Quantum Computing","quantum-computing",{"name":2691,"slug":2692,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":2696,"total":2845},[2697,2715,2731,2742,2754,2768,2781,2793,2804,2813,2827,2836],{"slug":2698,"name":2698,"fn":2699,"description":2700,"org":2701,"tags":2702,"stars":2712,"repoUrl":2713,"updatedAt":2714},"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},[2703,2706,2709],{"name":2704,"slug":2705,"type":15},"Documentation","documentation",{"name":2707,"slug":2708,"type":15},"MCP","mcp",{"name":2710,"slug":2711,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":2716,"name":2716,"fn":2717,"description":2718,"org":2719,"tags":2720,"stars":2728,"repoUrl":2729,"updatedAt":2730},"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},[2721,2724,2725],{"name":2722,"slug":2723,"type":15},"Containers","containers",{"name":2615,"slug":2616,"type":15},{"name":2726,"slug":2727,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":2732,"name":2732,"fn":2733,"description":2734,"org":2735,"tags":2736,"stars":2728,"repoUrl":2729,"updatedAt":2741},"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},[2737,2740],{"name":2738,"slug":2739,"type":15},"CI\u002FCD","ci-cd",{"name":2615,"slug":2616,"type":15},"2026-07-14T05:25:59.97109",{"slug":2743,"name":2743,"fn":2744,"description":2745,"org":2746,"tags":2747,"stars":2728,"repoUrl":2729,"updatedAt":2753},"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},[2748,2749,2750],{"name":2738,"slug":2739,"type":15},{"name":2615,"slug":2616,"type":15},{"name":2751,"slug":2752,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":2755,"name":2755,"fn":2756,"description":2757,"org":2758,"tags":2759,"stars":2728,"repoUrl":2729,"updatedAt":2767},"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},[2760,2763,2764],{"name":2761,"slug":2762,"type":15},"Debugging","debugging",{"name":2751,"slug":2752,"type":15},{"name":2765,"slug":2766,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":2769,"name":2769,"fn":2770,"description":2771,"org":2772,"tags":2773,"stars":2728,"repoUrl":2729,"updatedAt":2780},"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},[2774,2777],{"name":2775,"slug":2776,"type":15},"Best Practices","best-practices",{"name":2778,"slug":2779,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":2782,"name":2782,"fn":2783,"description":2784,"org":2785,"tags":2786,"stars":2728,"repoUrl":2729,"updatedAt":2792},"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},[2787,2790,2791],{"name":2788,"slug":2789,"type":15},"Machine Learning","machine-learning",{"name":20,"slug":21,"type":15},{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":2794,"name":2794,"fn":2795,"description":2796,"org":2797,"tags":2798,"stars":2728,"repoUrl":2729,"updatedAt":2803},"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},[2799,2802],{"name":2800,"slug":2801,"type":15},"QA","qa",{"name":2645,"slug":2646,"type":15},"2026-07-14T05:25:53.673039",{"slug":2805,"name":2805,"fn":2806,"description":2807,"org":2808,"tags":2809,"stars":2728,"repoUrl":2729,"updatedAt":2812},"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},[2810,2811],{"name":2615,"slug":2616,"type":15},{"name":2618,"slug":2619,"type":15},"2026-07-14T05:25:49.362534",{"slug":2814,"name":2814,"fn":2815,"description":2816,"org":2817,"tags":2818,"stars":2728,"repoUrl":2729,"updatedAt":2826},"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},[2819,2822,2823],{"name":2820,"slug":2821,"type":15},"Code Review","code-review",{"name":2751,"slug":2752,"type":15},{"name":2824,"slug":2825,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":2828,"name":2828,"fn":2829,"description":2830,"org":2831,"tags":2832,"stars":2728,"repoUrl":2729,"updatedAt":2835},"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},[2833,2834],{"name":2800,"slug":2801,"type":15},{"name":2645,"slug":2646,"type":15},"2026-07-14T05:25:54.928983",{"slug":2837,"name":2837,"fn":2838,"description":2839,"org":2840,"tags":2841,"stars":2728,"repoUrl":2729,"updatedAt":2844},"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},[2842,2843],{"name":2655,"slug":2656,"type":15},{"name":2738,"slug":2739,"type":15},"2026-07-30T05:29:03.275638",496]