Horace He
PyTorch core maintainer and torch.compile contributor; wrote a Thinking Machines post on LLM nondeterminism
- Core Maintainer, PyTorch (open-source project)1
Horace He is a PyTorch core maintainer, and PyTorch's maintainers page lists Horace He as an emeritus maintainer of TorchInductor and the PT2 dispatcher. Horace He's homepage describes work on PyTorch and torch.compile at Meta, including GPT-Fast, and Horace He co-authored the PyTorch 2 paper (ASPLOS 2024), listed with Meta. In September 2025, Thinking Machines Lab published "Defeating Nondeterminism in LLM Inference," credited to Horace He in collaboration with others at Thinking Machines.1234
Work
PyTorch (torch.compile)Open-source code
Listed as a PyTorch core maintainer and an emeritus maintainer of TorchInductor and the PT2 dispatcher; co-author of the PyTorch 2 paper introducing TorchDynamo and TorchInductor, which implement torch.compile. The repository is under a BSD-style license.136
Defeating Nondeterminism in LLM Inference (external site: thinkingmachines.ai)Published research
Credited author, with Thinking Machines Lab, of this Connectionism post, which examines why LLM inference results are not reproducible in practice and describes batch-invariant kernels for RMSNorm, matrix multiplication, and attention.4
GPT-Fast (external site: github.com)Open-source code
Horace He's homepage cites GPT-Fast as an example of Horace He's work on PyTorch and torch.compile; the BSD-3-Clause repository provides simple, PyTorch-native transformer text generation.25
Other documented roles
Roles the sources document but that are not confirmed as current.
- Credited "in collaboration with others at Thinking Machines" on a September 2025 research post, Thinking Machines Lab4
- Worked on PyTorch and torch.compile, Meta23
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.