PyTorch hardware backends
What happened
TechCrunch reported on October 19, 2026, that AWS had released TorchNeuron.1 According to that report, PyTorch AI apps could move onto Trainium with one config line and no code rewrite.1 Hugging Face reported the same day that TorchNeuron, a PyTorch backend for AWS Trainium, had entered public beta.2 Hugging Face said Transformers, Diffusers, Accelerate, PEFT and TRL ran natively on it.2
Amazon's code and Google's release
TechCrunch reported on October 20 that Amazon was submitting code to decouple PyTorch from Nvidia hardware.3 TechCrunch also reported that day that Google had made TorchTPU, its open source PyTorch version, available.4 According to TechCrunch, apps could run on TPUs with a simple configuration change and no rewriting.4
Hugging Face reported on October 20 that it was adding native TPU support to Transformers, Diffusers and Kernels through TorchTPU.5 Hugging Face said Optimum TPU was no longer required to run its models on TPUs.5
PyTorch 2.15 for Apple silicon
Also on October 20, the Hugging Face PyTorch team published a release post on PyTorch 2.15, with a release candidate available.6 The post covered FP8 on MPS, faster MoE and linear decoding, Metal kernels and memory fixes.6 Latent Space reported on October 21 that PyTorch 2.15 replaces MPSGraph paths with Metal kernels, with 5-6.5x faster MoE decoding.7 Latent Space also reported Qwen3.6-35B-A3B running at 61.3 tokens/s on M3 Ultra.8
Latest reported
Latent Space reported on October 22 that Google had open-sourced TorchTPU for PyTorch training and serving on TPUs.9 Latent Space said SGLang and vLLM plugins add caching, speculative decoding and disaggregated serving.10
Sources for the overview
- TechCrunch: AWS releases TorchNeuron
- Hugging Face: TorchNeuron launches with Hugging Face library support
- TechCrunch: Amazon releases TorchNeuron
- TechCrunch: Google releases TorchTPU
- Hugging Face: TorchTPU brings native TPU support to Hugging Face
- Hugging Face: PyTorch 2.15 brings MPS FP8 and faster decoding
- Latent Space: PyTorch 2.15 Metal kernels for Apple silicon
- Latent Space: PyTorch 2.15 Metal kernel gains
- Latent Space: Google open-sources TorchTPU
- Latent Space: Google open-sources TorchTPU
The stories
- · PyTorch 2.15 Apple silicon
- · TorchNeuron launch
- · TorchTPU launch