ROCm

Given a TheRock nightly build (URL or run-id), return the rocm-systems pin_sha used in that build

Check whether a given rocm-systems commit is included in a specific TheRock nightly build

Find the first TheRock nightly build that includes a given rocm-systems commit

Reviews Pull Requests or local diffs with an 8-agent fan-out covering static analysis, dead code, code smells + quality (naming, complexity, single-responsibility, magic numbers), language rules (C++/Python/CMake), architecture, simplification, performance…

Walk through a PR review interactively, one finding at a time. Generate review via pr-review, then for each issue present analysis + proposed inline comment, let user accept/edit/skip, accumulate into a PENDING GitHub review, submit at end.

Use when user wants to list, analyze, review, or summarize GitHub PR comments on a pull request number or URL

Use when architectural decisions involve competing quality attributes (performance vs modifiability, availability vs consistency, security vs usability), when the user says "tradeoff analysis", "ATAM", "quality attributes", "what are we giving up", or when a…

Use when moving computation from runtime to compile time — precomputed lookup tables, compile-time constants, type-based template dispatch, or zero-cost conditional branching with if constexpr; use when a function or value could be evaluated before the binary…

Diagnoses why rocprofv3 or ROCm GPU profiling does not work on a machine, by running `rocprofv3 --doctor` and turning its report into ordered, root-cause fixes. Use when rocprofv3 fails, crashes, hangs, finds no GPU, or will not start; when an error such as…

Records a failing HIP or ROCm GPU workload into an HRR archive that can be replayed and debugged later, and checks that archive is readable before anyone is asked to send it. Use when an AMD GPU workload crashes, hangs, aborts with a memory access fault, or…

Recommends which rocprofiler-systems runtime profiling preset (balanced, trace-gpu, trace-hpc, sys-trace, workload-trace, etc.) to pass to `rocprof-sys-run --preset=<name>` or `rocprof-sys-sample --preset=<name>` based on the user's workload and profiling…

Answers questions about what rocprofiler-systems configuration options exist and how to use them — e.g. "what hardware counters are available on my system", "what ROCm APIs can I trace", "what settings control sampling/output", "how do I persist my…

Guides instrumentation with `rocprof-sys-instrument` - choosing runtime instrumentation vs binary rewrite (`-o`), including or excluding functions and modules with regex options, previewing what will be instrumented with a dry run (`--simulate`,…

Diagnoses why a compute kernel on an AMD GPU is slow, instruction by instruction, with Advanced Thread Trace (rocprofv3 --att) and the rocprof-trace-decoder Python API. Use when a kernel runs below its expected throughput and you need the instructions and…

Check and generate changelog entries for amd-smi. Use when: reviewing PRs for changelog updates, generating release notes, checking CHANGELOG.md compliance.

Profiles applications on AMD GPUs with rocprofv3 and explains the results: application tracing (hot kernels, memory copies, idle GPU, Perfetto timelines), hardware counters (occupancy, cache hit rate, HBM bandwidth, memory- or compute-bound), Advanced Thread…

Predict and analyze LDS access conflicts with an exact rocKE architecture profile and a serializable semantic result. Use when examining lane-to-LDS byte-address mappings, checking broadcasts or conflict multiplicity, comparing supported gfx profiles, or…

Render and explain MMA (MFMA/WMMA) layout visualizations at any scale — the A/B/C tee, register files, logical tiles, derived/final-C, C store distribution, and LDS view. Answers questions from "how does this distribution map a logical tile to registers?" and…

Design and author a GPU kernel with the rocKE tiling API, end-to-end. Dispatches the Tiling Kernel Architect (lead) who consults MMA/LDS/Profiling specialists and drives layout-viz / bank-conflict / build / profile. Any algorithm — the stage classes are…

Create or extend a hipDNN generic-kernel-ingestor integration: graph/kernel contracts, descriptors, native hooks, packaging, host checks, exact-engine device correctness and final corpus coverage.

Run, read, and extend hipDNN's cross-provider integration test suite (dnn-providers/integration-tests): bundles and sweeps, golden data and DVC, .support.json claim sidecars, per-engine TOML, tier YAML, CTest lanes. Doc-driven: loads all of its knowledge at…

Build hipDNN with providers via the repository superbuild. Faster than standalone since providers build alongside hipDNN in a single CMake invocation. On Windows, auto-runs the wheel-based ROCm setup if not already prepared.

Run tests against an existing hipDNN superbuild. Supports per-component selection (hipdnn, miopen-provider, hipblaslt-provider, hip-kernel-provider, integration-tests), unit/integration/external-integration scope, and gtest filtering. Reproduces the…

Run clang-tidy on hipDNN after a build, scoped to the translation units affected by changed files. Much faster than ENABLE_CLANG_TIDY=ON, which runs clang-tidy inline with every compile. Resolves changed headers to the translation units that include them. Use…

Advisory AI code review for a FlyDSL pull request. Catches silent-correctness bugs, JIT cache-key mistakes, architecture-capability holes, index-width overflow, unbounded buffer descriptors, layout-algebra errors, dead flags, and tests or benchmarks that…

Review a FlyDSL PR, commit, branch, kernel, or consuming module for API-stability compliance. Detect breaking changes to stable APIs, usage of unstable FlyDSL APIs, and direct upstream MLIR dialect operations. Use when asked to review API compatibility, a…

Review a FlyDSL release candidate against published releases for API stability violations, including export paths, signatures, result interfaces, semantics, and deprecation windows. Use for release compatibility reviews and breaking-change checks.

Comprehensive reference for authoring FlyDSL GPU kernels on AMD GPUs. Covers the layout algebra, tiled copy/MMA, buffer ops, loop-carried range loops, SharedAllocator (LDS), autotuning, and common patterns. Use when writing, reviewing, or understanding FlyDSL…

Comprehensive guide to optimizing GEMM (General Matrix Multiply) kernels in FlyDSL on AMD CDNA GPUs. Covers tiling strategy, LDS ping-pong double-buffer, XOR bank-conflict swizzle, A/B data prefetch pipeline, 2-stage software pipelining, MFMA instruction…

Detect per-kernel GPU resource regressions (VGPR, SGPR, register spills, scratch, static LDS) by diffing the final ISA before and after a change, using its `isa_resource_table.py` helper. Compile-only: needs no GPU and no profiler run, so it works on any…

Profile GPU kernels using rocprofv3 to collect ATT instruction-level traces, then analyze the trace data using hotspot_analyzer.py to identify top-K stall hotspots (VMEM-load, VMEM-wait, LDS/SMEM-wait, barrier, MFMA stalls) mapped back to source lines, and…

Optimize LDS (Local Data Share / shared memory) access patterns in FlyDSL GPU kernels. Diagnose bank conflicts and high lgkmcnt stalls from ATT trace data, then apply swizzle or padding layouts to eliminate conflicts. Also increase the distance between LDS…

Advisory AI code review for aiter and FlyDSL PRs. Catches perf regressions, silent correctness bugs, dispatch gate holes, and AI-generated code patterns, but never acts as a merge gate. Invoke with a PR number (optionally owner/repo#N) and, when one exists, a…

Comprehensive reference for authoring FlyDSL GPU kernels on AMD GPUs. Covers the layout algebra, tiled copy/MMA, buffer ops, loop-carried range loops, SharedAllocator (LDS), autotuning, and common patterns. Use when writing, reviewing, or understanding FlyDSL…

Clean up aiter FlyDSL kernels and shared helpers while preserving numerical behavior and performance. Use for raw-IR migrations, helper deduplication, dead-code removal, and reviews of those changes against the pinned FlyDSL API.

Reproducible validation executor for kernel PRs. Applies an explicit base-to-head patch in an isolated worktree, runs it on a verified-idle GPU, compares the same targets against base, policy-checks the test diff, and emits a head-bound…

How to add/upload tuned config CSVs under aiter/configs (incl. model_configs/) without introducing duplicate shapes, and how to find & resolve duplicate-shape collisions. Use whenever adding a model's tuned config, merging/uploading config CSVs, editing…

Standard structure for aiter op_tests under op_tests/test_*.py — @benchmark + run_perftest candidate loop, a torch reference, a final markdown summary table, a __main__ guard so the module is importable, and faithful reproduction of the real model call…

Module-level JIT build-wall optimization for opus-based aiter modules. Use when an aiter JIT module's first-call build wall is a user-visible bottleneck or when adding a new module.

Compile-time optimization guidance for HIP/C++ kernels using opus.hpp. Use when writing or reviewing OPUS kernels, analyzing compile time, reducing template instantiation overhead, or optimizing hipcc build performance.

Capture a PyTorch profiler / kineto trace from a running ATOM server for a short benchmark window. Use when the user asks for "a trace", "profiler trace", "GPU trace", or "抓 trace" for performance investigation — what kernels ran, what's on the critical path,…

Run any ATOM workload — accuracy eval (GSM8K via lm_eval), performance benchmark, concurrency sweep, offline simple_inference, or fault repro under rocm-debug-agent. Use when the user asks to "test accuracy", "测精度", "跑 GSM8K", "跑 benchmark", "test…

AI code review for ATOM PRs. ATOM consumes aiter kernels and integrates with vLLM/SGLang plugins. Reviews check perf claims, aiter cross-repo deps, model coverage, dispatch correctness, and AI-generated code patterns. Invoke with a PR number.

Identify which GPU kernel is faulting/hanging in ATOM via rocm-debug-agent (for faults/asserts) or rocgdb (for silent livelocks). debug-agent dumps wave registers + faulting PC + (with --save-code-objects) disassembled code object on memory faults /…

Coding patterns and architecture index for the ATOM LLM inference engine

Locate forward numerical bugs by dumping intermediate tensors from a target implementation and a known-good reference, then bisecting layer by layer. Also covers batch-invariance bisect (the same token at any batch position should produce a bitwise-identical…

Add a new byte-moving path to UMBP's distributed mode as a TransferEngine (GPU copy, a second RDMA transport, a file/GDS engine, a fake for tests). Covers CanHandle pair-dispatch and why engines must be disjoint, the plan-builder/Submit/Wait split, bounce…

Deploy and set up the MORI environment in a fresh Docker container or bare host: start the container, install ROCm dependencies, NIC userspace libraries (AINIC/Broadcom/Mellanox-NVIDIA), RDMA-core, and MORI itself. Use when the user asks to deploy MORI,…

Discover, diagram, and diagnose the RDMA/GPU network topology of a GPU cluster (rails, NICs, GID/RoCE config, GPU<->NIC PCIe affinity), determine whether the fabric supports cross-rail communication, and localize a failure to a fabric tier (NIC / leaf /…

Known environment issues that make MORI slow or behave oddly without being MORI bugs — currently: HIP VMM peer traffic silently falling off XGMI onto PCIe/host memory on Linux kernels built without CONFIG_DMABUF_MOVE_NOTIFY / CONFIG_PCI_P2PDMA, which hits…

Add a new storage medium to UMBP's distributed mode as a MediumBackend (HBM, SSD, CXL, a remote object store, a fake for tests). Covers the two shapes a backend can take — a PageMemorySource for anything page-addressable, or a full MediumBackend for a medium…

Write GPU kernels that move data with the CCO SDMA device API (ccoSdma: put/get/quiet/waitSignal/commit). Covers the build flags and env vars needed to turn the path on, which template arguments and coop scope to pick, how completion and signals work, and the…

Review PyTorch pull requests for code quality, test coverage, security, and backward compatibility. Use when reviewing PRs, when asked to review code changes, or when the user mentions "review PR", "code review", or "check this PR".

Triages GitHub issues by routing to oncall teams, applying labels, and closing questions. Use when processing new PyTorch issues or when asked to triage an issue.

Non-interactive wrapper over pr-review that emits JSON assessing whether a PR is ready for a human maintainer's time.

Manage tensor lifetimes across CUDA streams and decide whether record_stream is avoidable. Use when using a tensor on a stream other than the one it was allocated on, adding or reviewing Tensor.record_stream / CUDACachingAllocator::recordStream, side streams,…

Query PyTorch CI, GitHub Actions, HUD, Grafana, and infrastructure metrics. Use when users ask about CI duration, job failures, queue times, workflow trends, runner health, dashboard data, or PyTorch infrastructure metrics.

Tell me which HIP acceptance tests are missing and need to be added. Analyzes a source tree (RCCL/NCCL/rocSHMEM/NIXL/Mooncake) for how it uses the HIP/CUDA/HSA APIs, compares against an embedded knowledge base of required acceptance scenarios AND against the…

Tell me which Verbs (libibverbs / rdma-cm / mlx5dv) acceptance tests are missing and need to be added. Analyzes a source tree (RCCL/NCCL/NIXL/etc.) for how it uses the RDMA verbs API, compares against an embedded knowledge base of required acceptance…

Scaffold a new HIP or Verbs test in the HIP/Verbs Acceptance Suite and wire it into the policies. Use when the user asks to add an acceptance test, add a HIP/Verbs test, cover a new primitive, or extend the suite.

Build and run the HIP/Verbs Acceptance Suite and report the verdict. Use when the user asks to run acceptance tests, run the HIP or Verbs tests, gate a machine, smoke-test the harness, or bring the suite up on new hardware.

Summarize a HIP/Verbs Acceptance Suite report.json — overall verdict, per-category breakdown, and the gate failures with their reason/ticket. Use when the user asks to explain acceptance results, summarize a run, or see why the gate failed.

Step-by-step checklist for bumping the pinned ONNX dependency / opset in ONNX Runtime (e.g. ONNX 1.21 / opset 26 → 1.22 / opset 27). Use when integrating a new ONNX release or release-candidate, updating the cmake/deps.txt onnx pin or the cmake/external/onnx…

Trigger, re-run, and unblock ONNX Runtime CI checks on a GitHub pull request. Use this skill when a required check is stuck, missing, failed, or needs re-running. Nearly all ORT CI runs as GitHub Actions workflows; only the "Linux Android Emulator QNN CI…

Build and run ONNX Runtime WebGPU provider tests on Linux WITHOUT a real GPU, using a software Vulkan adapter (Mesa lavapipe). Use when you need to exercise WebGPU EP kernels off-Mac — the Linux webgpu CI leg is build-only, so software Vulkan is how you…

Build ONNX Runtime from source. Use this skill when asked to build, compile, or generate CMake files for ONNX Runtime.

Force a fuzz-fill coverage-gap analysis on one specific llvm/llvm-project PR, regardless of whether it's been reviewed before. Use when asked to "check PR <N> for gaps", "run gap analysis on this PR", or similar for a single named PR.

Scan recent llvm/llvm-project PRs labeled for clang frontend, codegen, driver, analysis, static analyzer, bytecode, dataflow, or modules for fuzz-fill coverage gaps, keeping the largest non-test clang/lib diffs in a shared quota, and draft local review files…

Scan recent llvm/llvm-project PRs labeled llvm:ir, llvm:analysis, or llvm:transforms for fuzz-fill coverage gaps, keeping the largest non-test lib diffs in a shared quota, and draft local review files for PRs with real uncovered lines. Use when asked to "scan…

Scan recent llvm/llvm-project PRs touching the AMDGPU/SPIRV backends for fuzz-fill coverage gaps, and draft review comments for the ones that have real uncovered lines. Use when asked to "scan PRs for gaps", "run the PR gap scan", or similar.

Analyze an HPC job from an Omnistat database using hypothesis-driven exploration, driven by the omnistat-inspect tool. Use this to diagnose why a job behaved as it did — performance bottlenecks, hardware issues, anomalies, or comparing a degraded job against…

Generate a factual summary report card for an HPC job from an Omnistat database using the single-shot omnistat-inspect JSON command. Use this for a quick, comprehensive snapshot of what a job did (stats, energy, health, data quality) without diagnosing why.…

Open and explore an Omnistat database using VictoriaMetrics. Use this when the user wants to view, query, or visualize collected Omnistat telemetry data from a user-mode job.