By optimizing how process memory and GPU state travel together:
🧠 Decoupled model weights via a GPU Memory Service
✂️ Shrunk checkpoints by dynamically freeing physical KV cache pages
⚡ Patched CRIU with parallel memfd restore & Linux native AIO
https://t.co/dCcsjeFTZ1
Scaling LLM inference replicas elastically on Kubernetes usually means brutal cold starts. My team and a large group of engineers across NVIDIA just launched NVIDIA Dynamo Snapshot https://t.co/dCcsjeFTZ1 to solve this—slashing gpt-oss-120b restore times down to under 5 seconds.
We’re releasing Nemotron-Labs-Diffusion - the first Tri-mode LM family (3B/8B/14B) that switches between 1⃣Autoregressive, 2⃣Diffusion, and 3⃣Self-Speculation decoding by simply changing the attention pattern/mask.
One model Three decoding modes. No extra draft models. No architecture changes. Just significantly better efficiency across different concurrency levels.
Up to 4× higher real throughput for a single user.
🤗 HF Collection: https://t.co/1zStcCCWPi, open license
🛜 Project page: https://t.co/y6TEAvLFvD
📰 Tech report: https://t.co/NSjKxEyHnT
Details below 👇
The stable release of @PyTorch 1.0 is here! A big thank you to the community that's formed around #PyTorch and to those contributing with code, feedback, or new projects. Learn how to get started here: https://t.co/GX8poSLL6o