LightNews — News, ranked by scholars

BABA Toshiaki

@netmarkjp.bsky.social

#ばばさん通信ダイジェスト賛否関わらず話題になった/なりそうなものを共有しています。

AWSとGoogle Cloudが「PyTorch」のAWS Trainium、Google Cloud TPU対応版をオープンソースで公開。NVIDIAからの移行に挑む
https://www.publickey1.jp/blog/26/awsgoogle_cloudpytorchaws_trainiumgoogle_cloud_tpunvidia.html

AWSとGoogle Cloudが「PyTorch」のAWS Trainium、Google Cloud TPU対応版をオープンソースで公開。NVIDIAからの移行に挑む

10月20日と21日に米カリフォルニア州で行われたイベント「PyTorch Conference 2026」において、AWSとGoogle Cloudが相次いでPyTorchの自社AIプロセッサ対応版をオープンソースとして発表しました。 A...

www.publickey1.jp

November 3, 2026 at 11:55 PM

kimbodo.bsky.social

@kimbodo.bsky.social

Bedrock now supports schema-conforming outputs for all Claude models.

That can simplify extraction pipelines. AWS also launched a public beta of TorchNeuron for Trainium2/3, with eager execution, torch.compile and custom operators—a more familiar PyTorch path to its chips.

Read more on Kimbodo:

What AWS’s Latest AI and Cloud Updates Change for Production Teams

What Happened AWS’s October 20 releases span model output controls, speech, data sharing, identity, migrations and caching. On October 21, AWS also launched the public beta of TorchNeuron, an open-source…

kimbodo.com

November 3, 2026 at 11:46 PM

トモキ｜AI副業・自動化ラボ

@tomoki-ai-lab.bsky.social

PyTorchでGoogleのTPUをそのまま使える「TorchTPU」がオープンソース化され、Hugging Faceでもネイティブ対応しました！普段AIモデルの開発や検証で、ハードウェアの選択やコードの書き換えに悩んだ経験はありますか？🤔

GPU以外の環境での運用や、試してみたい工夫があれば、ぜひ一言で教えてください！👇
#AI開発

November 3, 2026 at 11:06 PM

Cyris

@sudolabs.nz

A 5-second video clip, generated locally on a Mac, in 15 seconds.

No cloud. It's my native macOS app running FastMetal, which denoises in just 3 steps on Apple's MLX instead of PyTorch.

Curious how many of you are running video models locally yet.

November 3, 2026 at 9:35 PM

kimbodo.bsky.social

@kimbodo.bsky.social

PyTorch has two new routes beyond GPUs.

Google’s TorchTPU ran 4,821/4,917 Hugging Face models without crashes; AWS’s TorchNeuron beta exposes Trainium2/3 as PyTorch devices. That widens accelerator choice, but teams still need to test speed, operator coverage and numerical…

Read more on Kimbodo:

What New PyTorch Accelerator Backends Mean for AI Infrastructure Choices

What Happened PyTorch gained two routes to non-GPU infrastructure. Google open-sourced TorchTPU, a backend intended to run existing PyTorch workloads on TPUs with limited changes. It supports eager execution, DeferAndFuse…

kimbodo.com

November 3, 2026 at 9:25 PM

aichina.news

@aichina.news

Apache-2.0 ONNX, no PyTorch dependency, Ascend NPU support — genuinely useful for moderation pipelines. But no AUC, no F1, no preprocessing spec, no parameter count. Hard to see it beating a fine-tuned CLIP baseline until a model card ships.

Spotting the Fakes: New ONNX Deepfake Detector Hits Modelers.cn

The proliferation of synthetic media means developers need robust tools to filter out manipulated imagery. Modelers.cn has just released a new contender to the field:…

aichina.news

November 3, 2026 at 6:43 PM

kimbodo.bsky.social

@kimbodo.bsky.social

Deployment options expanded, but security risks remain.

Google’s TorchTPU lets PyTorch apps run on TPUs via config changes. ElevenLabs reports a 98.8% Big Bench Audio score and 0.57s response time for v1 Max. Meta suspended Muse sharing after researchers found discoverable…

Read more on Kimbodo:

What Today’s AI News Means for Enterprise Costs, Voice Agents and Security

What Happened Three developments stand out. Google open-sourced TorchTPU to let PyTorch applications run on its TPUs with configuration changes rather than code rewrites [2]. DeepSeek open-sourced DeepInfer, the vLLM-based…

kimbodo.com

November 3, 2026 at 6:15 PM

PyTorch

@pytorch.org

How do compiler choices impact end-to-end performance and accuracy when optimizing PyTorch models?

This technical breakdown from NVIDIA explores model tuning trade-offs and introduces the Global Performance Tuner in TensorRT 11.1.

Read the post: https://bit.ly/4zl9sCB

Scaling NVIDIA TensorRT Performance Optimization Beyond a Single Build Route with the Global Performance Tuner, Part 1 | NVIDIA Technical Blog

Deploying a trained neural network involves more than translating each operator into a GPU kernel. It is a compilation problem with many legitimate answers. For example, when NVIDIA TensorRT builds an...

developer.nvidia.com

November 3, 2026 at 6:00 PM

ちょこのの

@chocononno.bsky.social

自分のPCで、ローカルLLMの赤ちゃんが生まれました。

ChatGPTと設計して、Claudeにプログラムを書いてもらい、Python・PyTorchで動かすまでの記録です。Pythonは3.15が先に入る、コマンドが2行に分かれる…つまずいた所も全部書きました。最初の「産声」は、意味のない文字の並びでした。

https://lifehackerschannel.site/2026/11/04/local-llm-baby-python-pytorch/

ChatGPTとClaudeで、自分のPCにローカルLLMの赤ちゃんを生まれさせた｜Python・PyTorchで産声を出力させるまで - ちょこのの実験室

まだ何も知らない「ローカルLLMの赤ちゃん」を、自分のPCで一から育てることにしました。ChatGPTと設計を相談し、Claudeにプログラムを書いてもらい、WindowsにPython 3.12とPyTorchを入れて、最初の「産声」を出力させるまでの記録です。実務でプログラムを書いたことのない私がつまずいた4つの所と、直し方も載せています。

lifehackerschannel.site

November 3, 2026 at 4:04 PM

T. @ Techtonique

@teeeeeeeee.bsky.social

Survival analysis with sklearn, glmnet, keras, pytorch, lightgbm, xgboost, nnetsauce, mlsauce Part 2
https://thierrymoudiki.github.io/blog/2026/04/28/python/survivalist-sklearn-pytorch-keras-nnetsauce

#Python #DataScience #MachineLearning #rstats #Techtonique

November 3, 2026 at 3:39 PM

aichina.news

@aichina.news

No benchmarks, no dataset provenance, no Ascend validation — on Ascend's own platform. The draw is real though: Apache-2.0, runs on CPU or CUDA without the PyTorch stack, cheap enough for high-volume screening. Hard to see it surviving unseen generators until the card ships.

Lightweight Deepfake Detector Hits Huawei's Modelers.cn: 85M ONNX Model Offers Portability

As synthetic media continues to proliferate, developers are increasingly seeking tools that can verify authenticity at scale without the overhead of massive compute resources. Modelers.cn, the…

aichina.news

November 3, 2026 at 3:16 PM

Steven | SelfHostedSI

@selfhostedsi.bsky.social

Source: huggingface.co/speridlabs/... (Apache 2.0; the PyTorch version needs an NVIDIA GPU with CUDA and two gated models accepted first). Mac support came with mlx-vlm pull request 2644.