Meta · Multimodal LLMs
THE MODEL LEARNING STUDIO · 01—04
阅读中文版 →Take models apart. Then run them.
Build intuition from first principles, then verify it with tensors and code. Each path includes complete bilingual chapters, working experiments and deeper reading.
Token → Tensor / Train → Align / Inspect → Evaluate
LLMs, starting with one token
Tensors, attention and training objectives, then real model I/O, retrieval, MoE and inference-time compute.
8 chapters · theory / experiments / notebook ↗Serve / Shard / Profile
LLM infrastructure, from memory to throughput
vLLM, SGLang, TensorRT-LLM, Megatron-LM and DeepSpeed, with GPU/TPU architecture, kernels, sharding and profiling.
8 chapters · theory / experiments / notebook ↗Align / Reconstruct / Ground
Multimodal understanding, from pixels to meaning
Image patches, CLIP, masked reconstruction, vision-language models and video timelines: understand what a model sees.
5 chapters · theory / experiments / notebook ↗Sequence / Transduce / Rank
HSTU, from actions to generative recommendation
Trace HSTU tensor by tensor, train a tiny sequential recommender, then examine ragged execution, candidate scoring and evaluation.
4 chapters · theory / experiments / notebook ↗Follow a question. You do not need to read everything at once.
New to language models? Begin with 01. Serving too slowly? Explore 02. Working with images and video? Start with patches and alignment in 03. Studying recommendation? Follow the action timeline in 04.
Every experiment states its execution mode: a numerical teaching model, a real pretrained browser model, or a Python/GPU notebook. Estimates are not hardware measurements.