LLM 教程

Learn how to accelerate local Gemma 4 31B inference on a single RTX 4090 using DFlash speculative decoding and Flash Attention against a baseline setup.

Build a Streamlit app that runs Claude Opus 4.8 with adaptive thinking, auto-scores each response with Haiku 4.5, and charts the cost-quality tradeoff.

Set up a multi-GPU Docker environment with tensor parallelism and EAGLE speculative decoding to serve Mistral Medium 3.5 128B through an OpenAI-compatible API.

Run Qwen3.6 27B on an RTX 3090 and learn how Multi-Token Prediction (MTP) with llama.cpp can boost local LLM inference by almost 2x without upgrading your GPU.

Learn how OpenAI's gpt-realtime-2, gpt-realtime-translate, and gpt-realtime-whisper differ, then test each one with working Python WebSocket code.

Learn how to run the full DeepSeek V4 Flash model on a single GPU using a modified llama.cpp build and a compatible GGUF file in this hands-on tutorial.

Learn how DeepSeek V4's three reasoning modes work, and build a Streamlit comparison app that shows when each mode actually wins on quality, speed, and cost.

Learn to fine-tune NVIDIA Nemotron-3-Nano-4B on a psychology Q&A dataset using an RTX 3090 GPU using LoRA and TRL after downloading the model from Hugging Face.

Learn how to use Alibaba's Qwen 3.6 Plus, Python, and the OpenAI SDK to build an automated invoice processing pipeline with native vision and tool calling.

Learn how to fine-tune Qwen3.6 on an H100 NVL GPU using SFT, from dataset preparation and 4-bit loading to training and evaluation.

Learn how to build an automated LLM-as-a-judge system to evaluate your RAG pipelines for faithfulness and relevance at scale and bridge the gap in AI testing.

Learn the capabilities of Anthropic’s best publicly available model, Claude Opus 4.7, and build a data science tool that can turn a chart into raw data.