Deep dive into every component of the Transformer, from architecture to attention

16 articles

From text representation to multimodal generation — understand how Transformers adapt to text, image, audio, and video modalities. Recommended: complete the Transformer Core Mechanisms path first.

10 articles

From data type fundamentals to cutting-edge quantization algorithms — weight quantization, KV cache quantization, and inference-time quantization

7 articles

From PagedAttention to RadixAttention, from scheduling preemption to structured output — a systematic guide to modern LLM inference engine algorithms and design philosophy.

5 articles

Automatically select the right LLM based on task complexity. Covers the full spectrum from simple classifiers to RL-based online learning, query-level to token-level routing, and single-model selection to multi-model collaboration.

8 articles

Systematic understanding of LLM evaluation: from benchmark design principles to specific benchmark deep dives, from optimization accuracy assessment to model selection decisions. Covers knowledge, reasoning, code, and agent evaluation with focus on OpenVINO toolchain and small model assessment.

10 articles

Deep dive into Ollama and llama.cpp internals — architecture, quantization, compute graphs, hardware backends, and serving infrastructure.

9 articles

Trace llama.cpp's complete C/C++ execution flow function by function. This path extends the 'Ollama + llama.cpp Deep Dive' from concepts to source-level implementation details.

8 articles

Understanding the AI software stack from inference frameworks to hardware ISA

5 articles

Deep dive into ML compiler internals: the complete journey from graph capture to optimized execution. Dual-track coverage of PyTorch 2.0 (torch.compile / TorchInductor / Triton) and MLIR (Dialect system / Progressive Lowering). Prerequisite: AI Compute Stack.

17 articles

From MDP to Policy Gradient, from RLHF to GRPO, from Reward Modeling to Test-Time Scaling — a systematic guide to how reinforcement learning drives LLM alignment, optimization, and reasoning.

8 articles

From Xe2 microarchitecture to oneDNN primitives, from SPIR-V compilation to OpenVINO graph optimization, from performance analysis to GPU+NPU co-inference — a systematic deep dive into AI inference on Intel iGPU.

12 articles

Matrices are the lingua franca of ML. This path builds four core tools (decomposition, measurement, calculus, iteration), covers classical methods (SVD, PCA, NMF) and operator analysis (PageRank, spectral clustering), then converges on modern architectures (LoRA, Efficient Attention, SSM/Mamba). The "decompose → propagate → converge" arc reveals how one mathematical tool manifests across seemingly different domains.

31 articles

Graphs are the universal modeling language for "entities + relationships." This path builds three layers of capability (structural exploration, measurement, combinatorial optimization), covers classical algorithms (BFS/DFS, shortest paths, network flow), then converges on modern methods (random graph models, probabilistic graphical models, GNNs). The "explore → measure → optimize → model" arc reveals how one set of graph tools manifests across seemingly different engineering domains.

22 articles