Inside The Machine
Part drawing · T-055

PyTorch

The main toolkit AI apps are built on; runs models on Mac GPUs.

Open the README (opens in a new tab)

  • Apple SiliconBuilt for Macs with M-series chips.

Title block: PyTorch

T-055
Part no.
T-055Sheet 55 of 268
GitHub stars
104,960Read 2026-10-21
Last activity
2026-10-21Last code push
Group
Runtimes & inferenceDrawer 05
Layer
Runtime2 of 5 in the stack
License
BSD-3-Clause
Kind
Code on GitHub
Language
Python
Install
Not tested yetCatalog entry
Gap Score
not scored yetStatus: pending test on Mac Studio
Entry checked
2026-10-21In our log 2026-10-20: New in the directory
1Install

Install: not tested yet.

We list this tool, but we have not run its install.

Not tested yet

Follow the project’s own instructions. Our agent file lists the steps we found in its README or docs, marked unverified, so an AI agent can read them with you. It uses pip, which needs a virtual environment (venv): a separate folder for one Python app, so it cannot break others.

Open the README (opens in a new tab)Steps we found (.md)
2What it is

Where it sits in the stack.

  1. 5ToolsIts hands and senses.
  2. 4MemoryWhat it remembers.
  3. 3HarnessThe app you use.
  4. 2RuntimeThis partPyTorch sits here.
  5. 1ModelThe brain you download.
Fig. 3The local AI stack, pulled apart. PyTorch sits on layer 2, the runtime layer.

In plain words

PyTorch: The main toolkit AI apps are built on; runs models on Mac GPUs.

Its layer · Runtime

The runtime loads the model into memory, runs it, and hands its answers to the apps on your computer.

Its group · Runtimes & inference

Runs open models on your own hardware. 19 tools share this drawer.

3Links and handoff
setups/pytorch.md

Hand this page to your agent.

The Agent view is this page as plain text: every field above, plus the install steps we found (not tested). An AI agent can read it and do the setup with you.

.md.json
4Same drawer

More in Runtimes & inference.

The next parts in this drawer, in part-number order. Bar: share of the most starred tool (OpenClaw, 392,108).catalog.json ·
No.PartWhat it doesLicenseGitHub starsLast activity
T-056RamaLamaPythonServe models from any source in containers.Serve models from any source in containers.MIT3,0872026-10-13
T-057SGLangPythonHigh-throughput serving for LLMs and VLMs.High-throughput serving for LLMs and VLMs.Apache-2.037,1752026-10-15
T-058vLLMPythonHigh-throughput LLM serving engine.High-throughput LLM serving engine.Apache-2.093,8632026-10-15
T-059vLLM MetalPythonvLLM plugin for Apple Silicon.vLLM plugin for Apple Silicon.Apache-2.01,8722026-10-15
T-041KoboldCppC++One-file GGUF runner with a built-in UI.One-file GGUF runner with a built-in UI.AGPL-3.012,0092026-10-15
T-042llama-swapGoHot-swaps models behind one OpenAI-style endpoint.Hot-swaps models behind one OpenAI-style endpoint.MIT5,9642026-10-15
6 of the 18 other tools in this group.All 19 in Runtimes & inference
Agent view · text/markdown · or send Accept: text/markdown
---
schema: "https://insidethemachine.ai/schema/listing.schema.json"
slug: "pytorch"
name: "PyTorch"
kind: "repo"
tier: "catalog"
category: "runtimes"
category_label: "Runtimes & inference"
stack_layer: "runtime"
homepage: "https://pytorch.org"
repo: "https://github.com/pytorch/pytorch"
license: "BSD-3-Clause"
stars: 104960
stars_as_of: "2026-10-21"
last_push: "2026-10-21"
archived: false
part_no: "T-055"
labels: ["Apple Silicon"]
tags: ["apple-silicon"]
seen_in_videos: []
gap_score: null
gap_score_status: "pending test on Mac Studio"
verified_on: "2026-10-21"
---

PyTorch 


**What:** The main toolkit AI apps are built on; runs models on Mac GPUs.

**Why:** Runs open models on your own hardware.

**Category:** Runtimes & inference · **Stack layer:** runtime

Status 


Catalog entry. Existence and repo metadata checked 2026-10-21 via the GitHub API. Install commands below are auto-extracted from the README and have NOT been executed or verified. Read the README before running anything; prefer a sandbox or venv.

Install (unverified) 


1. [macos-apple-silicon] source: Hugging Face blog 'MPS in PyTorch 2.15' (2026-10-20), release candidate; status: unverified, not executed

```bash
uv pip install torch==2.15.0 --index-url https://download.pytorch.org/whl/test/cpu
```

It uses pip, which needs a virtual environment (venv): a separate folder for one Python app.

Verify 


No verified check yet. Minimum: the install command exits 0 and the tool's own help/version flag runs (exact flag: see README).

Rollback (unverified) 


1. source: derived inverse of install command

```bash
uv pip uninstall torch
```

Links 


- Homepage: https://pytorch.org
- Repo: https://github.com/pytorch/pytorch
- README: https://github.com/pytorch/pytorch#readme

Part sheet (human page https://insidethemachine.ai/setups/pytorch) 


- Part number: T-055 (sheet 55 of 268). Part numbers follow the stack: layer, then group, then A to Z by name. Star counts never move a number.
- Group: Runtimes & inference (drawer 05 of 20, https://insidethemachine.ai/setups/c/runtimes). Layer: 2 of 5, runtime.
- GitHub (catalog.json, read 2026-10-21): 104,960 stars; last code push 2026-10-21.
- Install: Not tested yet (catalog entry).
- Entry checked: 2026-10-21. In our change log 2026-10-20: New in the directory.
- Apple Silicon: Built for Macs with M-series chips.
- Next parts in this drawer: T-056 RamaLama (https://insidethemachine.ai/setups/ramalama), T-057 SGLang (https://insidethemachine.ai/setups/sglang), T-058 vLLM (https://insidethemachine.ai/setups/vllm), T-059 vLLM Metal (https://insidethemachine.ai/setups/vllm-metal), T-041 KoboldCpp (https://insidethemachine.ai/setups/koboldcpp), T-042 llama-swap (https://insidethemachine.ai/setups/llama-swap).
- Previous part: T-054 Ollama (https://insidethemachine.ai/setups/ollama). Next part: T-056 RamaLama (https://insidethemachine.ai/setups/ramalama).