MoET in FP4 MoE Inference | Generated by AI

Home 2026.07

Question: What does “MoET” mean in the context of running v4 flash with FP4-level throughput?

Answer: I couldn’t find “MoET” as a standard, citable term in the quantization literature — it’s not a paper name like GPTQ/AWQs are. Based on your context (running a flash MoE model, using “moet” to “补精度” so FP4 throughput doesn’t tank quality), it’s almost certainly community shorthand for mixed-precision MoE compensation — the general technique class covered by things like MoEQuant, AlphaQ, and LREC:

The core idea across all of these:

  1. Quantize the bulk of weights (usually the FFN/expert matrices) to FP4/INT4 for the throughput win.
  2. Keep a small set of “sensitive” components at higher precision — router/gate weights, first/last layers, or specific outlier channels — since these dominate output quality loss.
  3. Add a lightweight error-compensation term (low-rank adapter, per-channel scale, or per-expert calibration) that’s cheap to compute but corrects most of the quantization error introduced in step 1.

That’s why you get FP4-level throughput/VRAM while accuracy holds close to bf16 — you’re not uniformly quantizing everything, you’re doing selective/mixed precision + a correction pass.

Your numbers (4 concurrent → 300 tok/s, 8 concurrent → 400 tok/s) look like a decode-bound MoE serving profile — sublinear scaling from 4→8 concurrency suggests you’re hitting memory bandwidth on the expert weights, not compute. That’s consistent with FP4 weight-only quant: weights are cheap to move, but KV cache + activation dequant become the bottleneck as batch grows.

If you can share which repo/tool actually prints “MoET” in its logs or docs (vLLM fork, SGLang, or a custom quant script), I can pull the exact definition instead of the generic pattern above — worth checking since these Chinese-community deployment tools often name things ad hoc.

References:


Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate