Junxian Ma / Writings
  • DeepSeek's Choice of EP Scale: The Insight Between Prefill and Decode

    Nov 8, 2026 Technical

    EP does not change compute per GPU; it changes rows per expert, copies per token and expert load imbalance. From these three quantities, this post explains why DeepInfer uses a small EP for prefill and a large one for decode.

  • All-to-All with Unknown Receive Counts: Four Ways DeepEP Does Dispatch

    Nov 5, 2026 Technical

    MoE routing is only settled after the gate scores tokens, so the receiver does not know how much data is coming. This post compares four DeepEP dispatch designs using one set of questions and turns the differences into NIC bytes and communication time.

  • Between Push and Pull: A Brief Analysis of the Communication Schemes in DeepGEMM Mega MoE and DeepEP

    Oct 30, 2026 Technical

    DeepEP pushes, Mega MoE pulls. Starting from whether the writer knows the final address, a comparison of the two: five constraints, from scatter to gather, and how the SMs divide the work.

  • An Attempt at a Fused-Operator Strategy for DeepSeek V3.2

    Oct 13, 2026 Technical

    Fusing DeepSeek-V3.2 Indexer scoring and exact top-k into one kernel on H20: design, attribution experiments, distance to peak, and TTFT validation in vLLM.

  • Profiling Ozaki II: Why Can Preprocessing Cost More Than Matrix Multiplication?

    Oct 1, 2026 Technical

    Analyzing Ozaki II preprocessing costs on H20 through the algorithm, complexity, hardware metrics, and assembly.

  • Sparse Attention: Algorithm–Infra Co-design

    Aug 10, 2026 Technical

    How NSA creates regularity before the load, how DSA reconstructs it afterward, and why Discovery can still grow quadratically behind a sparse main kernel.

  • Thoughts at the Start of Summer

    Jun 24, 2026 Reflection

    Plans for training judgment across research, market, AI thesis, and small feedback loops.

  • Building Tolerance for the Unresolved

    Jun 15, 2026 Reflection

    A short note on gray areas, control, and learning to let things unfold.

  • Faster Iteration, Sharper Perception, Broader Research, Deeper Thinking

    May 30, 2026 Reflection

    Before N1 begins, some reflections from conversations with entrepreneurs on iteration, perception, research, and thinking.

  • Some Thoughts After Conversations

    May 26, 2026 Reflection

    Notes on entrepreneurship, problem-solving, uncertainty, and plans for the next stage.