Search Results for "load"

DeepSeek-V3 is a robust Mixture-of-Experts (MoE) language model developed by DeepSeek, featuring a total of 671 billion parameters, with 37 billion activated per token. It employs Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture to enhance computational efficiency. The model introduces an auxiliary-loss-free load balancing strategy and a multi-token prediction training objective to boost performance. Trained on 14.8 trillion diverse, high-quality tokens, DeepSeek-V3 underwent supervised fine-tuning and reinforcement learning to fully realize its capabilities. Evaluations indicate that it outperforms other open-source models and rivals leading closed-source models, achieving this with a training duration of 55 days on 2,048 Nvidia H800 GPUs, costing approximately $5.58 million.

...The idea is to “pack” frames by detecting shared structure and storing differences efficiently, which can accelerate training or inference on video-like data. By reducing I/O and memory bandwidth, datasets become lighter to load while models still see the essential temporal variation. The repository demonstrates both packing and unpacking steps, making it straightforward to integrate into preprocessing pipelines. It’s useful for diffusion and generative models that learn from sequential image datasets, as well as classical pipelines that batch many related frames. ...

Step-Video-T2V is a state-of-the-art text-to-video foundation model developed to generate videos from natural-language prompts; its 30B-parameter architecture is designed to produce coherent, temporally extended video sequences — up to around 204 frames — based on input text. Under the hood it uses a compressed latent representation (a Video-VAE) to reduce spatial and temporal redundancy, and a denoising diffusion (or similar) process over that latent space to generate smooth, plausible...

...DualPipe), and how MoE / EP / parallelism strategies interact in real systems. The repository contains JSON trace files like train.json, prefill.json, decode.json, and associated assets. Users can load them into tools like Chrome tracing to inspect GPU idle times, overlapping operations, and scheduling alignment. The idea is to bring transparency to internal efficiency tradeoffs, enabling researchers to reproduce, analyze, or improve on DeepSeek’s parallelism strategies. The README explains how trace data corresponds to forward/backward chunks, settings (e.g. ...

MedicalGPT training medical GPT model with ChatGPT training pipeline, implementation of Pretraining, Supervised Finetuning, Reward Modeling and Reinforcement Learning. MedicalGPT trains large medical models, including secondary pre-training, supervised fine-tuning, reward modeling, and reinforcement learning training.