DL Notes

A focused knowledge base for modern deep learning systems, from hardware fundamentals to LLM serving and neural rendering.

Start Here

1. Hardware - GPU/NPU basics, memory systems, and practical GPU configuration
2. Tensor Operations - core tensor manipulations, activation functions, and CUDA Graph notes
3. AI Infra Overview - the layered structure of an end-to-end inference system
4. Inference Request Lifecycle - how one request moves through the whole online system
5. AI Infra Metrics - TTFT, TPOT, and TPS
6. Roofline Model for Inference - compute ceilings, bandwidth ceilings, and how optimizations move bottlenecks
7. Data Movement and Communication - memory hierarchy, runtime state flow, and cross-device tensor exchange
8. KV Cache - KV memory semantics, paged blocks, and prefix reuse
9. Serving Runtime - chunked prefill, admission control, and runtime-side stability
10. Parallelism - DP and TP from the perspective of memory, throughput, and communication
11. Decoding and Sampling - temperature, top-k, top-p, and min-p
12. Speculative Decoding - verification, accepted length, and per-token latency
13. Training Objective - autoregressive pre-training objective
14. Position Encoding - RoPE, M-RoPE, and TM-RoPE from sequence to multimodal space-time
15. Models - model-specific notes (Qwen3-Omni, DFlash) and practical serving commands
16. Neural Graphics - NeRF and Flow Matching foundations

Documentation Map

Systems and Infrastructure

• Tensor Operations

AI Infra

• Overview
• Hardware
• Request Lifecycle
• Metrics
• Roofline Model for Inference
• Data Movement and Communication
• KV Cache
• Serving Runtime
• Parallelism
• Decoding and Sampling
• Speculative Decoding
• Training Objective

Models

• Position Encoding
• Models

Graphics and Generative Modeling

• Neural Graphics

Scope

This site emphasizes practical understanding:

• concise theory with equations where useful
• implementation-minded notes and runnable snippets
• serving and performance considerations for real workloads