gfxGRAPH

Develop and extend gfxGRAPH internals — add gap bridges, fix HIP/ROCm issues, debug graph capture failures, and understand the architecture. Covers BridgedCUDAGraph, ShapeBucketPool, ConditionalGraph, native bridge, and the monkey-patch system. USE FOR: extend gfxGRAPH, add new gap bridge, fix gfxGRAPH bug, understand gfxGRAPH architecture, modify BridgedCUDAGraph, add gfxGRAPH feature, debug graph capture internals, improve gfxGRAPH performance, gfxGRAPH C++ native bridge, write gfxGRAPH tests. DO NOT USE FOR: just using gfxGRAPH in a project (use gfxgraph-integration), general ROCm issues.

Benchmark gfxGRAPH internals, run the public benchmark suite, and compare performance between Python and Rust.

Guide and resources for deployment and development using the ROCm/AMD ATOM (AiTer Optimized Model) inference backend.

Validates and benchmarks ROCm-DS component ports against their CPU baselines (e.g., pandas vs hipDF). Use this to generate parity tests, verify correctness, and measure the performance speedups of GPU-accelerated code.

Verifies whether a requested ROCm-DS workflow is officially supported, source-build feasible, or experimental on the target system.

Migrates pandas-like workflows to hipDF and cudf.pandas style acceleration, auditing for unsupported features.

Leverages GPU acceleration to process and analyze complex graph structures using hipGRAPH. Note: hipGRAPH is early access.

Utilizes the HIP Memory Manager (hipMM) for advanced GPU memory pooling, efficient allocation, and data movement.

Utilizes hipRAFT for foundational, reusable GPU-accelerated primitives like clustering, dimensionality reduction, and statistical operations.

Selects, builds, benchmarks, and validates hipVS ANN indexes and query paths for ROCm-DS workloads.

Handles installation and source compilation of ROCm-DS components via conda, PyPI, or source builds.

Guidelines and patterns for integrating multiple ROCm-DS components (hipDF, hipVS, hipGRAPH, hipRAFT) into a cohesive, end-to-end data science pipeline on AMD GPUs. Use this skill when building multi-stage workflows that require passing data efficiently between different ROCm-DS tools.

Parent overview skill for ROCm Data Science (ROCm-DS) components and SMI tools. Provides guidance on routing workloads, compatibility triage, and component-specific tasks like hipDF, hipVS, and hipMM.

Classifies and routes tasks to the appropriate ROCm-DS component (hipDF, hipVS, hipGRAPH, hipRAFT, hipMM) based on workload requirements.

Core runbook for diagnosing memory ceiling limits on AMD's ROCm toolkit natively. Details mitigation parameters and context-limiting configurations explicitly tailored for the Radeon RX 6700 XT.

Integrate gfxGRAPH into PyTorch or SGLang projects on AMD ROCm (RDNA2/gfx1030). Provides drop-in CUDA Graph parity, eager fallback, shape bucketing, and VRAM safety. USE FOR: enable CUDA graphs on AMD, integrate gfxGRAPH, configure gfxGRAPH for SGLang, fix graph capture failures on ROCm, RDNA2 CUDA graph compatibility, gfx1030 graph support, HIP graph bridge, monkey-patch torch.cuda.CUDAGraph, shape bucketing for dynamic batches. DO NOT USE FOR: developing or extending gfxGRAPH internals (use gfxgraph-development), general PyTorch ROCm issues unrelated to graph capture.

Configure or reconfigure gfxGRAPH runtime behavior, environment variables, and replay modes after installation.

Assess whether a target environment is ready for gfxGRAPH installation and surface blockers before any changes are made.

Install gfxGRAPH into a user-specified target directory or environment using the repo's supported uv-based workflow.

Test gfxGRAPH on real hardware and confirm it is functioning after install or configuration changes.