AI Briefing#pytorch
The latest AI and developer news about #pytorch, with the original source and a short summary.
The latest AI and developer news about #pytorch, with the original source and a short summary.
Pinterest optimized feed diversity and quality together, evolving from DPP to SSD and soft spacing.
Netflix built a flexible LLM Post-Training Framework on top of Ray, PyTorch, and vLLM.
To increase the expressiveness of ad ranking models, the team moved away from the existing Two-Tower architecture and built a new GPU-based serving architecture.
Hugging Face has released a tool that makes it easy to build and share ROCm kernels for AMD GPUs.
Rebellions and SqueezeBits held a hands-on workshop where participants ran vLLM directly on NPUs.
torch.compile recompiles when shapes change and restricts execution scope using guards.
This introduces a method to speed up model inference on ZeroGPU Spaces by up to 1.8x using PyTorch's AoT compilation.
Hugging Face has released the kernel-builder library, which supports everything from developing to deploying custom CUDA kernels.
Hugging Face has released nanoVLM, a lightweight toolkit that lets even beginners easily train Vision Language Models.
Hugging Face released TimmWrapper, which enables various computer vision models from timm to be used within the transformers ecosystem.
Explains how to visualize and analyze GPU memory usage using PyTorch's memory snapshot feature.
AMD's 5th generation EPYC CPU 'Turin' has been unveiled, improving LLM inference performance by up to 2x compared to the previous generation.
Hugging Face's distributed training framework Accelerate has announced the release of version 1.0.0.
Transformers models from Hugging Face can now be loaded and used directly in KerasHub.
FuriosaAI has unveiled **RNGD**, a highly efficient AI accelerator optimized for LLM and multimodal model inference.
This analyzes how differences in the internal precision handling methods of DeepSpeed and PyTorch FSDP affect model training outcomes.
Hugging Face has released Quanto, a general-purpose PyTorch quantization backend for the Optimum library.
Global enterprises are planning to expand AI infrastructure in 2024, with a particular focus on securing cost-efficient inference infrastructure to replace GPUs.
It introduces how to apply fp16 and use memory-efficient attention to speed up inference and reduce memory usage for the SDXL model.
By reproducing OpenAI's RLHF code, key technical details to watch out for when implementing PPO were compiled.
Coupang's ML Platform standardizes everything from notebooks to deployment to make ML development and operations fast.
It can be mixed with existing PyTorch code and improves training and inference speeds by approximately 15% on the same hardware.
Boosted CPU serving performance 4x through torchserve thread configuration and ipex pinning.
This article explains the principles of MLM and CLM, which are core to BERT and GPT2, and introduces how to build domain-specific models by directly implementing a Transformer encoder in PyTorch.