Cuda Streams

OrganizationPopular

pytorch

cuda-streams

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, wait_stream, or CUDA events for cross-stream ordering, or designing APIs that hand side-stream tensors to callers.

Overview

Publisher	pytorch
Repository	`pytorch`
Skill name	`cuda-streams`
Stars	104.6K
Forks	32.1K
Bundled files	Instructions only
Links	Skill source Repository Homepage

Markdown instructions
A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.
Works with any LLM
AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.
Self-contained
Everything the model needs lives in the instructions — no extra files to sync.
Open source
Published by pytorch on GitHub. Read the source before you install it.

Installation

Install the Cuda Streams AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

Open the app and go to Plugins → Skills.
Choose "Install from GitHub".
Paste the skill folder URL below and confirm.
https://github.com/pytorch/pytorch/tree/main/.agents/skills/cuda-streams
Enable the skill in any chat where you want it available.

Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills

git clone --depth 1 https://github.com/pytorch/pytorch.git /tmp/pytorch
mkdir -p .claude/skills
cp -r /tmp/pytorch/.agents/skills/cuda-streams .claude/skills/cuda-streams

Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Cuda Streams in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Cuda Streams on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Cuda Streams is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

Loaded only when it is needed
The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.
Switch models mid-chat
Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

CUDA Streams and record_stream

The caching allocator only reuses a block on the stream that allocated it. If a tensor is used on some other stream, then before it is freed you need either (a) a sync from that stream back to the creation stream (an event, or creation_stream.wait_stream(side_stream)), or (b) record_stream() / CUDACachingAllocator::recordStream(). (a) is deterministic. (b) is not: on later allocations the allocator polls the recorded streams, so whether memory gets reused depends on how far the GPU has gotten, and memory usage can blow up when the CPU runs ahead. Prefer (a) where possible.

When writing or reviewing code that needs one of these, work out which of the following cases applies:

Avoidable: the code doing the cross-stream use also controls when the tensor is freed, or can hold a reference until a point it controls. Record an event after the last side-stream use and make the creation stream wait on it before dropping the reference. Don't use record_stream. FSDP2 works this way (see docs/source/distributed.fsdp.fully_shard.md).
Unavoidable under the current API: the tensor goes to arbitrary user code, which frees it at an unknown time, and nothing we control runs in between where we could insert the sync. Here record_stream is the correct tool and shouldn't be flagged.
But we design the APIs. If the API is new or can still change, ask whether a different shape would give us a point to sync. Examples: return a handle whose wait() does the sync and releases internal references; allocate outputs on the caller's stream and sync the side stream back before returning; keep ownership of buffers internally. c10d shows both sides: collectives keep their tensors alive in the Work object until wait() syncs the streams, with no recordStream, but isend doesn't require wait(), so it falls back to recordStream (see ProcessGroupNCCL.cpp). Weigh the design cost against the determinism benefit, and point out the option rather than insisting on it.
Hard to avoid: sometimes avoiding it would take a redesign. For example, the autograd engine moves gradients between nodes whose streams come from the forward pass (torch/csrc/autograd/input_buffer.cpp). The engine has the graph and could delay frees to a point where it syncs, but arbitrary user hooks can see and hold on to gradients, so it doesn't know when the last use happens. Don't treat such cases as settled: say what information is missing for a sync point and what it would take to get it.

When record_stream is the right call, leave a comment explaining why the free point can't be controlled. See also the record_stream docstring in torch/_tensor_docs.py.

Frequently asked questions

What does the Cuda Streams AI skill do?

Why use Cuda Streams on TypingMind?

Because you install it once and use it with any model. Cuda Streams is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Cuda Streams in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/pytorch/pytorch/tree/main/.agents/skills/cuda-streams. TypingMind reads its SKILL.md and installs it as a skill you can enable per chat.

Which AI models can use Cuda Streams?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Cuda Streams?

As many as you like. As long as a model supports skills, you can use Cuda Streams with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Cuda Streams AI skill free?

It is published on GitHub by pytorch. Check the repository for licensing terms. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Watchers

NousResearch

OrganizationPopular

Poll RSS, JSON APIs, and GitHub with watermark dedup.

Karpathy Guidelines

Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define…

Create Agent

Create a custom agent (.agent.md) for a specific job.

Create Hook

Create a hook (.json) to enforce policy or automate agent lifecycle events.

Overview

Installation

Install in TypingMind

Install in another agent

Use it in TypingMind

Works with any AI model

Skill instructions

CUDA Streams and record_stream

Frequently asked questions

Related AI Skills

Set up your own AI workspace now