Skip to main content
Xilinx
GitHub 创作者资料

Xilinx

按仓库查看 4 个 GitHub 仓库中的 25 个已收集 skills。

已收集 skills
25
仓库
4
更新
2026年8月26日
仓库浏览

仓库与代表性 skills

debug-bo-corruption
软件开发工程师

Use when an NPU kernel passes its standalone shape test but produces NaN, garbage, or stale values when invoked as part of a larger pipeline. Common symptoms: correct first invocation but wrong on subsequent calls; correct in isolation but wrong when chained…

2026年6月22日
debug-fa-runtime-failure
软件开发工程师

Use when NPU FlashAttention hangs (`ERT_CMD_STATE_TIMEOUT`) or produces NaN at head_dim ≥ 128. Discriminates the three known root causes (compile-flag mismatch, seq-first dk_chunks bug, true L1 overflow) via a symptom-classification table and applies the…

2026年6月22日
debug-multi-launch-merge
软件开发工程师

Use when stitching kernels into a multi-launch ELF and the AIE compiler rejects the merged module (BD exhaustion, channel routing, herd shape conflict, IR validation error, DMA stride limitation). Discriminates the 6 known compile blockers via a…

2026年6月22日
deploy-new-llm
软件开发工程师

Entry point for deploying a new decoder-only LLM on AMD NPU2. Invoked by the user as `/deploy-new-llm <hf_model_id> [--name <dirname>] [--target npu2|npu1] [--dtype bf16|fp16]`. Bootstraps the per-model workspace, validates architecture is in scope, and…

2026年6月22日
opt-buffer-object-reuse
软件开发工程师

Optimization skill — reuse NPU BufferObjects across calls instead of re-allocating/re-writing them. Two mechanics in one class: (B1) per-layer weight BOs pre-loaded once and skipped via static_input_indices, and (B2) intermediate BOs the kernel overwrites,…

2026年6月22日
opt-layout-alignment
软件开发工程师

Optimization skill — choose activation layouts so consecutive kernels hand off on-device without a host-side transpose. Canonical case: seq-first (seq, n_heads·head_dim) so RMSNorm → RoPE → FlashAttention → O-proj stay seq-first, eliminating 1–4 host…

2026年6月22日
opt-merge-multi-launch-kernels
软件开发工程师

Procedural recipe for fusing multiple `air.launch` kernels into one multi-launch ELF (single XRT invocation). Invoked by phase-4-prefill-optimization and phase-5-decode-optimization to fuse kernel groups when building NEW model-specific fused ELFs…

2026年6月22日
phase-0-build-cpu-reference
软件开发工程师

Phase 0 of LLM deployment — produce `<model>_weights.py` (HF weight loader) and `<model>_cpu_helpers.py` (the few NumPy helpers production prefill/decode import), then confirm the HF bf16 reference baseline loads and runs via the shared…

2026年6月22日
已展示 8 / 15 个已收集 Skill。
aie-code-creator
未分类

Write efficient IRON Python designs and C++ AIE kernels for AMD XDNA NPUs (AIE2 / AIE2P architectures, e.g., Ryzen AI Phoenix, Hawk, Strix, Krackan Point). Use whenever the user asks about IRON, ObjectFifo, Worker/Program/Runtime, NPU programming, AIE…

2026年8月24日
aie-dataflow-opt
软件开发工程师

Guide to optimizing an AIE/IRON design's dataflow — dispatch partitioning, tile placement, overlays, weight/activation streaming strategy, and DMA bandwidth/compression — once it's already functionally correct. Distinct from aie-kernel-opt (which optimizes a…

2026年7月27日
aie-dataflow-presim
软件开发工程师

Guide to validating an AIE/IRON dataflow design in software before it ever touches hardware — a threaded ObjectFifo/Worker mock for deadlock/FIFO-depth detection, bit-exact validation against a numeric oracle, tiny isolated probes to de-risk novel mechanisms,…

2026年7月27日
aie-hw-bringup
软件开发工程师

Guide to bringing up a new IRON design on real AIE/NPU hardware for the first time — sequential block-by-block (and, for large models, dispatch-by-dispatch) bring-up against a reference, immediate output comparison against the numeric oracle/ONNX Runtime,…

2026年7月27日
aie-kernel-opt
软件开发工程师

Standalone guide to optimizing AIE / Peano-compiled kernels (INT8 conv, matmul, attention, elementwise). Covers the measure-first methodology (baseline, bit-exact gate, ablation, verify-in-.o) and the catalog of concrete levers in priority order — loop hints,…

2026年7月27日
aie-model-baseline
软件开发工程师

Guide to preparing an ML model for AIE/NPU deployment before any device or dataflow work begins — choosing/locking a quantization scheme, exporting ONNX, extracting a deployment manifest (per-op scales/zero-points/shifts/layout), and building a bit-exact…

2026年7月27日
已展示 4 / 4 个仓库
已展示全部仓库