H Huaweidata

Column

Structured long-form essays, mechanism breakdowns, and interactive notes.

  1. Q, K and V feed attention computation, with retained K/V reused during generation
    QKV: Transformer Computation Atlas

    Follow Q, K and V through Transformer computation and data flow, from input processing to generation. Illustrated explanations, worked examples and exercises reveal how KV caches, attention, model architectures and speculative decoding work.

    Read
  2. K3 Architecture Illustrated
    K3 Architecture Illustrated

    Understand K3's experts, attention, and depth-wise information flow through the computation of a single token: illustrated chapters, derivations, numerical examples, and resource tradeoffs.

    Read
  3. The article’s opening screen: the headline “How does one token’s cache become a shared memory?” beside a diagram of input history, state lookup and a shared global KV
    Xiao Wang Asks, Da Wang Answers: From KV Cache to YOCO and Shared Memory

    22 questions and answers between Xiao Wang and Da Wang, with diagrams: the KV cache, chained hashes, Claude’s cache breakpoints, DeepSeek’s persistence, YOCO’s shared memory, an 890-byte conditional ledger, approximate restoration, sparse retrieval and Engram. 17 mechanism diagrams and 4 offline interactive labs.

    Read
  4. Technical flow diagram showing a question moving through tokens, context, tool calls, sampling, and output generation in an LLM
    A Question's Journey - LLM from Input to Output

    A full mechanism timeline from one prompt through context, tokenization, Transformer internals, generation, tool calls, sampling, latency, and reasoning budget.

    Read