Scaling Kimi Inference with DSpark Speculative Decoding in vLLM
See how DSpark improves Kimi-K2.6 and Kimi-K2.7-Code throughput as the speculative window scales from n=3 to n=7 in vLLM.
Filter posts by category
Optimizing GLM4-MoE for Production: 65% Faster TTFT with SGLang
As the state-of-the-art GLM 4.7 model continues to lead in coding performance, Novita AI remains committed to delivering a reliable, efficient, and production-grade GLM service to
Revolutionizing Large Language Model Inference: Speculative Decoding and Low-Precision Quantization
Learn how speculative sampling and low-precision quantization reduce costs and accelerate speed, offering practical solutions for scalable AI deployment.
Dynamic KV Cache compression based on vLLM framework
Novita AI speeds up Llama-70B loading with KV sparsity, reducing memory, computation, and I/O overhead for faster inference and minimal accuracy loss.
How to Select the Best GPU for LLM Inference: Benchmarking Insights
Discover how to select cost-effective GPUs for large model inference, focusing on performance metrics and best practices to enhance efficiency.
How KV Sparsity Achieves 1.5x Acceleration for vLLM
Boost AI inference speed with KV sparsity. Understand how it works and optimize your models for real-world applications.
Dynamic allocation of GPU resources for Kubernetes workloads
Currently, to schedule GPU Pods in Kubernetes (k8s), various extension solutions are put into action, including Device Plugin, Extended Resource, scheduler extender, scheduler fram
Dynamically Adding Port Mappings to Running Docker Containers
Port mapping is a crucial aspect of developing and deploying containerized applications. Typically, we establish a connection between a container's internal port and a port on the
GPU Container Core Binding Strategy Based on Affinity
Introduction to Optimizing CPU and GPU Performance In high-performance computing and large-scale parallel task processing, GPUs have become indispensable accelerators. To fully uti
Will Speculative Decoding Harm LLM Inference Accuracy?
Mitchell Stern et al. 2018 introduced the prototype concept of speculative decoding. This method has since been further developed and refined by various approaches, including Looka
Quantization Methods for 100X Speedup in Large Language Model Inference
Discover how selecting the best data types and optimizing GPU hardware support unlocks new pathways for spending up quantization inference.