#FP8 mixed precision
98 related articles
·3 min
DeepSeek-V4 Flash RL Training Successfully Ported to AMD MI355X GPU: A Complete Breakdown
The Miles team and AMD announce the full port of DeepSeek-V4 Flash RL training to AMD Instinct MI355X GPUs on ROCm, boosting AIME pass@1 from 0.39 to 0.49—a milestone for compute ecosystem diversity.
Read more →
·2 min
Directly Optimizing Face Similarity: A Practical Guide to Accelerating Character LoRA Training
Ditch inefficient noise prediction and train character LoRAs by directly optimizing face similarity. Using the DRaFT differentiable reward method, training completes in 10-12 minutes on an RTX 4090, far outperforming traditional SFT. Includes open-source code.
Read more →
·6 min
Home Self-Hosting Server Setup Guide: Building a Private Cloud Ecosystem from Scratch
One 32GB server, 8 hours to build a complete home self-hosting ecosystem. This article details the selection logic behind core services like Jellyfin, Forgejo, Immich, and Ollama, plus infrastructure setup essentials for Nginx Proxy Manager and Podman Compose.
Read more →
·6 min
Alibaba Bans Claude Code: The Battle Over AI Coding Tool Security and Data Sovereignty
Alibaba bans Claude Code over security risks, sparking tech community debate. An in-depth analysis of data leakage risks with cloud AI coding tools and the rise of local AI deployment.
Read more →
·7 min
Practical RL Compute Optimization: Core Strategies to Boost CPU and GPU Utilization
An in-depth look at why CPU and GPU utilization is low in RL training, covering vectorized environment parallelism, distributed Actor-Learner architectures, GPU-side simulation (Isaac Gym/Brax), and Ray RLlib practice.
Read more →
·3 min
ComfyUI-INT4-Fast: Running Flux Models Smoothly on 6GB GPUs
ComfyUI-INT4-Fast brings W4A4 quantized inference to ComfyUI. RTX 3060 (6GB VRAM) generates 1024×1024 images in 17s. Per-layer mixed-precision routing balances speed and quality for Flux models.
Read more →
·6 min
GPT-5.6 Now Widely Available: The Era of Three-Tier Models and Multi-Agent Orchestration Arrives
OpenAI releases GPT-5.6 (SOUL/TERRA/LUNA), with Ultra mode running four agents in parallel; Meta launches Muse Spark 1.1 with million-token context; ChatGPT desktop unifies Chat, Work, and Codex.
Read more →
·4 min
Demystifying Unified Memory Architecture: How Mini PCs Run 70B Large Models
Why can a mini PC with unified memory run a 70B model while an RTX 4090 can't? A deep dive into the VRAM wall and unified memory architecture for smarter local AI hardware choices.
Read more →
·6 min
Why Is Qwen3 27B So Slow on a 4090? A Hands-On Guide to Local Inference Performance Optimization
RTX 4090 taking over 400 seconds to run Qwen3 27B inference? This article analyzes the core causes—VRAM overflow and CPU offloading drag—and offers targeted fixes.
Read more →
·3 min
From IT Support to AI Engineer: A Real NCA-GENL Certification Journey and Lessons Learned
A real NCA-GENL study journal from an IT-support-turned-AI-engineer: 50+ scenario questions, 7-week prep, and a brutal 40% on Trustworthy AI. Covers Transformer concepts, NVIDIA tools, and what actually works.
Read more →
·7 min
Gemini Live Integrates Nano Banana: A Complete Guide to Real-Time Camera, Image Generation, and Maps Integration
Google's Gemini Live now integrates the Nano Banana image generation model with Connected Apps like Google Maps, supporting real-time camera scene understanding and visualization. Free worldwide.
Read more →
·6 min
Cursor Meets Grok 4.5: A Hands-On Test of an AI Coding Assistant That's 5x Faster with Cleaner Output
A developer deeply tests Grok 4.5 High Fast in Cursor, finding it rivals Claude Opus in quality but runs 5x faster with cleaner, filler-free output. Full hands-on review and analysis.
Read more →
·5 min
Runs on Low VRAM! A Hands-On Look at the ID LoRA Face-to-Video Workflow
A face-to-video workflow built on GGUF quantized models and ID LoRA runs on just an RTX 3060 with 6GB VRAM. This article breaks down its core principles, four-step process, and how it tackles facial consistency in AI video.
Read more →
·5 min
28 Enterprise-Grade AI Agent Projects: A Complete Guide That Even Beginners Can Deploy
A collection of 28 fully reproducible enterprise-grade AI Agent projects covering code debugging, financial analysis, customer service, and multi-agent collaboration—deployable even for beginners.
Read more →
·4 min
DeepSeek-V4 Deep Dive: Million-Token Context Architecture Breakthrough and Core Innovations
Deep dive into DeepSeek-V4: 1.6T-parameter MoE, CSA+HCA hybrid attention, MHC & MUON optimizer. Inference FLOPs drop to 27% of V3.2, redefining open-source LLM SOTA.
Read more →
·7 min
Krea2 Tested: Does the FP8 vs BF16 Quality Gap Still Exist?
A Reddit user compared FP8 and BF16 precision on the Krea2 model and found almost no difference in image quality. This article analyzes the technical reasons behind the shrinking quantization gap and its real-world impact on VRAM usage and inference speed.
Read more →
·7 min
Complete Guide to Training Small Models on Kaggle's Free Compute: GPU Quotas and Practical Tips
In-depth guide to Kaggle's free-tier compute: P100/T4 GPU with 30 hours/week quota, 12-hour sessions, suitable models like CNN and BERT fine-tuning, plus tips like mixed precision and checkpointing to start deep learning at zero cost.
Read more →
·6 min
Grok 4.5 Deep Dive: Coding on Par with Top Models, at a Fraction of the Cost
A senior developer's 24-hour deep test of Grok 4.5: a 1.5T-param MoE model at $2/M input tokens, with coding benchmarks rivaling GPT-5.5. Real performance, token efficiency, and limits explained.
Read more →
·4 min
Unsloth v0.1.463-beta: Fixes Studio Crash and llama-server Permission Issue
Unsloth v0.1.463-beta fixes a Studio crash caused by access-denied errors during llama-server service discovery. Improves stability for multi-user servers and Windows environments.
Read more →
·6 min
From Video to Notes to Knowledge Base: A Practical Guide to AI-Assisted Learning Workflows
A complete AI learning workflow: batch download videos, auto-transcribe, generate structured notes with AI, then build intelligent search and Q&A via Dify. Turn scattered videos into a reusable personal knowledge base.
Read more →