Mastodawn

I just published our next #LLVM #Meetup in #Darmstadt (Germany) on Wed. 30th July, starting 7pm.
We will have Lukas Sommer from #Codeplay talk about "Compiling Machine Learning Models with #PyTorch 2.0 and #Triton"

RSVP at https://www.meetup.com/llvm-social-darmstadt/events/308590919

#Compilers #Software

LLVM/Clang Social Darmstadt, Wed, Jul 30, 2025, 7:00 PM | Meetup

Come and join us to share your experiences and interest in LLVM/Clang and related tools. **Agenda** 18.45 Doors Open 19.00 Welcome 19.05 "Compiling Machine Learning Models

Meetup

Ahmet Akkoç 2d ago

👀Looking into the future! 📈 Just wanted to share our patient outcome prediction model which we presented at EULAR 2025.

We have designed a #forecasting #model that can follow a patient's swollen joint count over time and estimate if their
condition will improve or worsen.

ZiteLab #pytorch #sktime #opensource #personalizedmedicine #ai #ml #timeseries #rheumatology

(1/2)

GripNews 3d ago

🌘 容錯的 Llama：在 Crusoe L40S 上以每 15 秒 2000 次的合成故障進行訓練，無需檢查點
➤ 極端環境下的模型訓練：torchft 的可靠性驗證
✤ https://pytorch.org/blog/fault-tolerant-llama-training-with-2000-synthetic-failures-every-15-seconds-and-no-checkpoints-on-crusoe-l40s/
這篇文章介紹瞭如何使用 torchft 和 torchtitan 在真實環境中訓練大型語言模型 (LLM)，並在極高故障率（每 15 秒約 2000 次合成故障）下驗證其可靠性和正確性。研究人員在 Crusoe 的 300 個 L40S GPU 集羣上，成功地在沒有檢查點的情況下訓練了一個 1B 參數的 Llama 3 模型，展示了 torchft 的優勢，特別是在非傳統環境和有限網路頻寬的狀況下。
+ 這太酷了！以前從沒想過能在這麼多故障的情況下訓練大型模型。這對於需要高度可靠性
#人工智慧 #機器學習 #分散式訓練 #容錯 #PyTorch

Fault Tolerant Llama: training with 2000 synthetic failures every ~15 seconds and no checkpoints on Crusoe L40S – PyTorch

N-gated Hacker News 3d ago

🔥 Behold the #PyTorch blog masterpiece: "Fault Tolerant #Llama Training" - because who doesn't love 2000 failures every 15 seconds? 😂💥 Forget checkpoints, because llamas are clearly bred for #chaos on a Crusoe L40S! 🙄✨
https://pytorch.org/blog/fault-tolerant-llama-training-with-2000-synthetic-failures-every-15-seconds-and-no-checkpoints-on-crusoe-l40s/ #Training #FaultTolerance #MachineLearning #HackerNews #ngated

Fault Tolerant Llama: training with 2000 synthetic failures every ~15 seconds and no checkpoints on Crusoe L40S – PyTorch

Hacker News 3d ago

Fault Tolerant Llama training – PyTorch blog

https://pytorch.org/blog/fault-tolerant-llama-training-with-2000-synthetic-failures-every-15-seconds-and-no-checkpoints-on-crusoe-l40s/

#HackerNews #Fault #Tolerant #Llama #training #PyTorch #blog #machinelearning #AI #training

Fault Tolerant Llama: training with 2000 synthetic failures every ~15 seconds and no checkpoints on Crusoe L40S – PyTorch

Collabora 3d ago

Object detection & tracking with gst-python-ml, a powerful, pure python #ML framework that supports a broad range of ML vision and language models & works seamlessly with upstream #GStreamer distributions. https://www.youtube.com/watch?v=vn7p2hXUlcs #Python #PyTorch #AI #OpenSource

GStreamer Python ML - Soccer Object Detection & Tracking

YouTube

GripNews 5d ago

🌘 PyTorch 使用 None 進行重塑
➤ 遮罩技巧與 PyTorch 重塑的應用
✤ https://blog.detorch.xyz/post/2025-06-21-pytorch-reshaping-with-none.md
這篇文章探討了在 PyTorch 中使用 `None` 進行張量重塑的方法，尤其是在處理可變長度序列資料時，例如自然語言處理。作者解釋瞭如何利用廣播機制 (broadcast mechanism) 建立遮罩 (mask)，以忽略填充的無效 token，並提供了使用 `reshape` 函數作為替代方案，以提高程式碼的可讀性。
+ 這篇文章清楚地解釋了 PyTorch 中 `None` 的用法，對於初學者來說很有幫助。
+ 我一直對廣播機制感到困惑，這篇文章提供了一個很好的實際例子，讓我更容易理解。
#程式設計 #PyTorch #深度學習 #序列處理

Pytorch Reshaping With None

Collabora Jun 16

GStreamer Python ML - Soccer Object Detection & Tracking

YouTube

GripNews Jun 15

🌕 miniDiffusion：純 PyTorch 實現的 Stable Diffusion 3.5
➤ 開源 Stable Diffusion 3.5 簡化實現，助力學習與開發
✤ https://github.com/yousef-rafat/miniDiffusion
miniDiffusion 是一個用純 PyTorch 語言重新實作的 Stable Diffusion 3.5 模型，旨在用於教育、實驗和開發用途。它以最簡化的程式碼量（約 2800 行）重現了 Stable Diffusion 3.5，涵蓋了從 VAE 到 DiT 以及訓練和資料集腳本。該專案包含核心圖像生成模組、SD3 組件、訓練和推論腳本，並使用了 MIT 授權許可，方便使用者進行學習和實驗。
+ 這個專案對於想了解 Stable Diffusion 內部運作原理的人來說非常棒，程式碼簡潔明瞭，方便研究。
+ 期待這個專案能有更多優化和測試，成為一個更完善的開源圖像生成工具。
#人工智慧 #機器學習 #圖像生成 #PyTorch

GitHub - yousef-rafat/miniDiffusion: A reimplementation of Stable Diffusion 3.5 in pure PyTorch

A reimplementation of Stable Diffusion 3.5 in pure PyTorch - yousef-rafat/miniDiffusion

GitHub

N-gated Hacker News Jun 14

🚀🎨 WOW, someone spent their weekend reinventing the wheel in #PyTorch, because clearly, Stable Diffusion 3.5 wasn't stable enough! Now you can enjoy all the bugs and none of the efficiencies—because who needs progress when you have GitHub Copilot to code your mediocrity? 🤖💥
https://github.com/yousef-rafat/miniDiffusion #StableDiffusion #GitHubCopilot #softwaredevelopment #innovation #HackerNews #ngated

GitHub - yousef-rafat/miniDiffusion: A reimplementation of Stable Diffusion 3.5 in pure PyTorch

A reimplementation of Stable Diffusion 3.5 in pure PyTorch - yousef-rafat/miniDiffusion

GitHub