84 related articles

NVIDIA introduces Nonuniform Tensor Parallelism, letting GPUs bear different compute loads so training can continue without checkpoint rollback during hardware failures—boosting LLM training Goodput and fault tolerance at scale.

Complete guide to deploying production-grade LLM inference on Kubernetes, covering GPU scheduling, vLLM engine selection, autoscaling, observability, and cost optimization.

A deep dive into LLM inference cost structure and profitability models—from GPU throughput, MoE architecture, and KV Cache to scale effects—revealing the business logic behind API price wars.

Moonshot AI open-sources FlashKDA, providing high-performance CUDA kernels for Kimi Delta Attention. Learn about its technical principles, performance gains, and value for long-context training and inference acceleration.

Moonshot AI open-sources FlashKDA, providing high-performance CUDA kernels for Kimi Delta Attention. Explore its technical principles, performance gains, and value for long-context training and inference.

Moonshot AI releases Kimi K3 open-weight model with 2.8T parameters and 1M token context. Our deep dive covers coding, 3D dev, agent capabilities, and safety concerns.

Google's Gemini consistently triggers Error 1076 on the 16th conversation turn, regardless of context size. Analysis points to a session state management defect, with three workarounds provided.

An in-depth look at the real daily work of data scientists, MLEs, and MLOps engineers — covering responsibilities, essential tools, and career paths to help you find your direction in AI.

Deep dive into a real-time 3D human mesh reconstruction project using a single RGB camera, built with Rust, Candle, and CUDA, achieving 55ms/frame on RTX 5080. Exploring its architecture, Metal porting plans, and applications in VTuber, AR/VR, and sports analysis.

A deep dive into a real-time 3D human mesh reconstruction project using a single RGB camera, built with Rust, Candle, and CUDA, achieving 55ms/frame on an RTX 5080.

Detailed analysis of Kimi K3 quantization deployment options, comparing q4 vs q8 storage requirements, precision trade-offs, and hardware configurations for local self-hosting.
In-Depth Analysis of the Claude Opus 5…
Deep analysis of the Claude Opus 5 elevated error rate incident, exploring LLM service reliability challenges and providing developers with practical strategies including multi-model redundancy, retry mechanisms, and graceful degradation.

Facing US chip bans and closed-source monopoly, how do China's open-source AI models keep fighting back? A deep dive into three core paths: open-source pricing games, optical interconnect positioning, and on-device scenarios.

Facing US chip bans and closed-source monopolies, how do China's open-source AI models keep striking back? A deep dive into three core paths: open-source pricing-power games, optical interconnect positioning, and edge-side use cases.

A systematic review of must-know topics for AI Application Engineer interviews: PTQ/QAT quantization, operator fusion, inference pipelines, latency/throughput analysis, and edge deployment of detection/segmentation/BEV models.

A systematic guide to must-know AI application engineer interview topics: PTQ/QAT quantization, operator fusion, inference pipelines, latency/throughput analysis, and edge deployment of detection/segmentation/BEV models.

In-depth review of Panel AI v1.1.1: second-level installation, no-public-IP networking, batch compute cluster management. Learn how enterprise AI on-premises deployment barriers are dramatically lowered.

Alibaba's Qwen3.8 challenges larger models with a 2.4T-parameter MoE architecture, claiming second only to Gemini. A deep dive into MoE mechanics, continuous updates, two-speed release strategy, and real local deployment requirements.

Benchmarking 4×V100 16G PCIe vs. 2×V100 32G SXM adapter for local LLM inference. Prefill speed, decode speed, power limits, and bandwidth bottlenecks analyzed.

Claude Opus 5 launches next week; Alibaba Qwen integrates into Apple Intelligence for Chinese users; 27B on-device model compressed to 3.8GB; open-source models narrow gap to closed-source by 3.3%.