#KV cache bandwidth
182 related articles
Tech Frontiers·2 min
LFM2.5-8B-A1B: A MoE Model with 1.5B Active Parameters Delivering 4x Its Weight Class Performance
Liquid AI releases LFM2.5-8B-A1B, a MoE model with 8B total params but only 1.5B active, matching 6B-class models in tool calling. Supports 128K context, local deployment, multilingual, with SGLang Day-0 support.
Read more →
Industry Insights·2 min
AMD MI355X Beats B200: Full-Stack Optimization Breakdown for 5% Lower TCO on DeepSeek-R1 Inference
AMD Instinct MI355X achieves 5% lower TCO than NVIDIA B200 on DeepSeek-R1 disaggregated inference via SGLang+MoRI full-stack optimization with 1.25x per-GPU throughput.
Read more →
Tech Frontiers·1 min
Cloudflare Contributes Critical KV Cache and Mooncake Fixes to SGLang
Cloudflare contributes decode KV cache offload and Mooncake recovery fixes to SGLang, resolving garbled output under high concurrency for Kimi K2.6 and enabling automatic fault recovery in distributed inference.
Read more →
Industry Insights·2 min
NVIDIA Dynamo Snapshot: A Snapshot Recovery Solution for GPU Inference Cold Start Problems
Deep dive into how NVIDIA Dynamo Snapshot reduces LLM inference cold start time from minutes to seconds via GPU state snapshot and recovery, covering Kubernetes integration and elastic inference.
Read more →
Product Reviews·3 min
Local Deployment of Qwen 3.6 27B on 4×3080Ti: Real-World Coding Test with OpenCode
Real-world test of Qwen 3.6 27B FP8 deployed on 4×3080Ti 16GB modded GPUs with OpenCode for system tool development. Covers hardware setup, inference speed, context management, and productivity gains.
Read more →
Tutorials·3 min
Decoding LLM Naming Conventions: Parameter Counts, Quantization Formats & VRAM Requirements Quick Reference
Decode LLM naming conventions, understand 32B parameters & AWQ/GGUF quantization formats, with 4-bit VRAM estimation formulas, MOE model pitfalls, and model selection by GPU tier.
Read more →
Industry Insights·2 min
NVIDIA Blackwell Sets New STAC-AI Records for Financial LLM Inference
NVIDIA Blackwell GPU sets new LLM inference records in STAC-AI financial benchmark. Explore Blackwell architecture advantages, TensorRT-LLM co-optimization, and LLM applications in trading and risk management.
Read more →
Product Reviews·1 min
Qwen 3.6 Local Deployment in Practice: 35B Model Reverse-Engineers LTE Modem, Beating Claude
Qwen 3.6 35B MoE model runs locally on MacBook, reverse-engineers LTE modem web portal by extracting login logic from obfuscated JS, outperforming Claude Sonnet while Gemma 4 fails the same task.
Read more →
Tutorials·2 min
How to Tell If Your GPU Is Out of VRAM When Running Local LLMs: A Complete Guide to VRAM Monitoring
How to tell if your GPU is out of VRAM when running local LLMs. Learn the difference between dedicated and shared GPU memory, monitor VRAM overflow via Task Manager, and use quantization and context length control to avoid OOM.
Read more →
Tutorials·2 min
Hermes + Qwen3.6 Local Deployment Tutorial: Build Your Private AI Agent at Zero Cost
Step-by-step tutorial to deploy Hermes Agent with Qwen3.6 open-source LLM locally. Covers WSL setup, model download, Telegram bot integration for a zero-cost private AI Agent.
Read more →
Tech Frontiers·3 min
GPT Image 2 Generates APP Interfaces from Scratch — A Paradigm Revolution in AI Front-End Development
GPT Image 2 generates flawless text and photorealistic APP interfaces from scratch. Combined with Codex, AI front-end development enters a paradigm revolution. Coverage includes Amap's ABOT, LLM interpretability breakthroughs, and Huang's TPU rebuttal.
Read more →
Deep Dives·3 min
NVIDIA Vera Rubin Platform Deep Dive: Solving the Scalability Challenge of Agentic AI Inference
Deep dive into NVIDIA's Vera Rubin platform Pod-level architecture and next-gen NVLink, revealing how it solves Agentic AI inference scalability bottlenecks and the industry shift from training-first to inference-first.
Read more →
Product Reviews·4 min
Cube Studio Deep Dive: Tencent's Open-Source One-Stop MLOps Platform
Deep dive into Cube Studio, Tencent Music's open-source one-stop AI platform, covering architecture design, distributed training, large model fine-tuning and inference, and domestic chip adaptation.
Read more →
Deep Dives·5 min
Hugging Face Transformers: A Comprehensive Guide to the 160K-Star Open-Source AI Framework
Comprehensive guide to Hugging Face Transformers, the 160K-star GitHub framework—covering architecture, multimodal support, quantization, and inference optimization for loading, fine-tuning, and deploying pre-trained models.
Read more →
Deep Dives·3 min
Google TPU v8t and v8i Deep Dive: The New Era of Separated Training and Inference AI Chips
Google Cloud Next unveils TPU v8t (training) and TPU v8i (inference) chips. Deep analysis of their architecture, strategic significance, and impact on AI chip competition.
Read more →
Tutorials·3 min
4-bit QLoRA Fine-tuning LLaMA 3 in Practice: A Guide to Training 8 Billion Parameter LLMs on Consumer GPUs
Learn how to fine-tune LLaMA 3 8B on consumer GPUs using QLoRA and Unsloth. Covers 4-bit quantization, LoRA adaptation, Alpaca data format, training pipeline setup, and model deployment—runs on RTX 3090.
Read more →
Product Reviews·3 min
Cube Studio: Tencent's Open-Source One-Stop AI Platform — A Deep Dive into Full MLOps Lifecycle
Deep dive into Tencent Music's open-source Cube Studio cloud-native AI platform, covering Notebook development, Pipeline orchestration, distributed training, LLM fine-tuning, inference deployment, and domestic hardware adaptation for full MLOps lifecycle.
Read more →
Product Reviews·4 min
Cube Studio: Tencent's Open-Source All-in-One AI Platform Covering the Full Pipeline from Large Model Training to Inference
Deep dive into Tencent Music's open-source Cube Studio cloud-native AI platform, covering distributed training, LLM fine-tuning, vLLM inference, VGPU virtualization, and Huawei Ascend adaptation.
Read more →
Product Reviews·4 min
Deep Dive into Cube Studio: Tencent's Open-Source All-in-One AI Platform
Deep dive into Tencent's open-source AI platform Cube Studio, covering distributed training, large model fine-tuning and inference, Pipeline orchestration, VGPU virtualization, and Huawei Ascend support for enterprise cloud-native MLOps.
Read more →
Product Reviews·4 min
Hugging Face Transformers: The Technical Architecture & Practical Guide Behind 160K Stars
Deep dive into Hugging Face Transformers: core architecture, Pipeline API, model fine-tuning, and multimodal support. A practical guide to the 160K-star AI framework.
Read more →