Agent skill

hpc-python

Python HPC patterns — threading/multiprocessing/async, CUDA streams, latency hiding, PyTorch DDP. Use when writing performance-sensitive Python code, distributed training, or parallel data processing.

View SKILL.md on GitHub Repository
Stars 2
Forks 0

Install this agent skill to your Project

npx add-skill https://github.com/JoaquinCampo/Skills/tree/main/hpc-python

SKILL.md

HPC Python

Python parallelism, PyTorch DDP, and CPU/GPU latency hiding patterns.

Extension Files

File Content
parallel-python.md Threading vs multiprocessing vs asyncio decision tree, GIL rules, CUDA+fork safety
latency-hiding.md CUDA streams, double buffering, compute/comm overlap, CUDAGraphs, async checkpoint
pytorch-ddp.md DDP internals, gradient buckets, common bugs, mixed precision, DistributedSampler
preload-caching.md Three-level caching: L1 file (disk/memmap/shm), L2 function (lru_cache/dedup/index), L3 variable (buffer/GPU cache/warmup/KV cache)
torch-compile.md torch.compile modes, graph break diagnosis/fixes, reading generated Triton as starting point for hand-tuning
benchmarking.md Correct GPU timing (CUDA events, torch.utils.benchmark.Timer), warmup, common pitfalls, what to measure
dataloader.md DataLoader params (num_workers/pin_memory/prefetch_factor), dataset patterns, data formats, collation, worker issues

Quick Decision Tree

What is the bottleneck?
├─ I/O bound         → threading (ThreadPoolExecutor) or asyncio
├─ CPU bound         → multiprocessing (mp.Pool, fork BEFORE CUDA!)
├─ GPU bound         → batch inputs, don't parallelize
├─ Mixed CPU→GPU     → pipeline + CUDA streams (see latency-hiding.md)
└─ DDP communication → tune bucket_cap_mb, use model.no_sync()

Critical Rules

  1. mp.Pool BEFORE CUDA: Create multiprocessing pool before any torch.cuda call (fork+CUDA = deadlock)
  2. Never .item() in loops: Accumulate on GPU, transfer final result only
  3. pin_memory=True: Required for non_blocking=True transfers to actually be async
  4. DDP no_sync(): Use during gradient accumulation steps to avoid redundant AllReduce
  5. find_unused_parameters=True: Only when needed — it's expensive

Review Checklist (Python)

□ No .cpu()/.item()/.numpy() in hot loops?
□ DataLoader: num_workers > 0, pin_memory=True?
□ mp.Pool created before CUDA init?
□ Threading used only for I/O, not CPU-bound work?
□ DDP gradient accumulation uses no_sync()?
□ CUDA streams used for transfer/compute overlap?
□ Pre-allocated buffers reused (not created per iteration)?

Recommended Agent Skills

Expand your agent's capabilities with these related and highly-rated skills.

JoaquinCampo/Skills

sparse-retrieval-eval

Evaluate sparse retrieval models on standard IR benchmarks (BEIR, MIRACL, mMARCO). Covers all IR metrics (nDCG@k, Recall@k, MAP, MRR), dataset loading, sparse corpus encoding to CSR matrices, IDF-weighted retrieval, caching, and result interpretation. Triggers on: evaluate retrieval, BEIR benchmark, nDCG, recall@k, sparse retrieval evaluation, MIRACL evaluation, information retrieval metrics, IR evaluation, search quality metrics.

2 0
Explore
JoaquinCampo/Skills

wandb-plot

Download and generate plots from Weights & Biases runs. Use when you need to: - List projects you have access to - List runs in a W&B project - Inspect available metrics for a run - Download existing plot images from a run - Generate line plots from metric history (loss, accuracy, etc.)

2 0
Explore
JoaquinCampo/Skills

go

Go engineering best practices and idioms. This skill should be used when writing, reviewing, or refactoring Go code. Triggers on any .go file work, Go module operations, go test, go build, go run, go vet, go generate, or when the user mentions Go, golang, goroutines, channels, context, Go interfaces, Go error handling, Go testing, or Go concurrency. ALWAYS use this skill when Go code is involved, even for simple functions.

2 0
Explore
JoaquinCampo/Skills

fastapi

FastAPI best practices and conventions. Use when working with FastAPI APIs and Pydantic models for them. Keeps FastAPI code clean and up to date with the latest features and patterns, updated with new versions. Write new code or refactor and update old code.

2 0
Explore
JoaquinCampo/Skills

kvpress

kvpress (NVIDIA) KV-cache compression for HuggingFace LLMs. Use when: kvpress imports, compression_ratio, press(model) context managers, StreamingLLMPress, SnapKVPress, ExpectedAttentionPress, TOVAPress, KnormPress, KV-cache eviction, token pruning during generation, or attention sink methods.

2 0
Explore
JoaquinCampo/Skills

plantuml

Create, edit, and render PlantUML diagrams. Triggers on: architecture diagrams, flowcharts, sequence diagrams, data models, state machines, visual documentation.

2 0
Explore

Didn't find tool you were looking for?

Be as detailed as possible for better results