[kevin-mii]

Fills in the AMD subsection of the Low-Latency Mode post. Its structure and chart style match the NVIDIA section in #2, so the two vendor sections read in parallel.

Changes

• AMD Low-Latency Kernels (about 650 words), in the same order as the NVIDIA section:

  • the plugin and its flag (--monokernel-backend amd);
  • the low-concurrency problem: six verify rows per request, and about 1,300 kernels per DSpark cycle;
  • bold-lead technique paragraphs: two persistent launches per layer, overlapping stages inside a launch, and the all-reduce inside the kernel;
  • the serving configuration;
  • the reported AgentX results and the observed accuracy.

  Source links are pinned to sgl-amd-mono e952b22.

• AgentX chart (amd-agentx-interactivity.svg), styled after nvidia-static-decode.svg: orange plugin off, blue plugin on, ratio labels, setup subtitle, and flag and source footnotes. It shows P90 interactivity at 1, 4 and 8 concurrent sessions on MI355X: 1.50x, 1.56x and 1.45x.

• Layer diagram (amd-mono-layer.svg): a schematic timeline of one DeepSeek-V4.1 layer on the standard path and with the plugin, checked against the kernel code.

  • K1 loads its projection weights ahead of the norm and exchanges no data across GPUs.
  • K2 runs attention, the output projection, the second mHC update, the router and the experts. The mHC mix weights and the shared expert load early, routed-expert weights load after routing, and the attention and MoE partial sums are pushed to every GPU as they are produced.

The Figure 1 placeholder is left in place for the shared cross-vendor figure, as in #2. This PR only inserts into the AMD subsection, and merging it with #2 produces no conflicts.

Sources

• AgentX numbers and accuracy: [AMD][V4.1] Mono kernel framework (--layer-backend mono) and DeepSeek-V4.1 decode layers sgl-project/sglang#43984, the in-tree prototype of these kernels (MI355X, TP4, DSpark block 5, against SGLang's AgentX baseline with the same recipe).
• Kernels per DSpark cycle and per-kernel cost: [AMD][V4.1] Trim per-cycle bookkeeping launches in DSpark decode sgl-project/sglang#43821.
• Kernel design: sgl-project/sgl-amd-mono at e952b22.

To check before publishing

• A rerun of AgentX at 1, 4 and 8 sessions with the current sgl-amd-mono would make the chart match the plugin exactly.
• Fixed acceptance at 3.51 is inferred from #43984 comparing against the AgentX baseline "with the same recipe", which uses the golden acceptance length. Confirm with the author.
• GSM8K is 0.901 with the plugin and 0.911 without it (1,319 questions). The text reports these as observed results, as the NVIDIA section does.

🤖 Generated with Claude Code