---
cited_by:
  - '[[thoughts/papers/kimiteam2026kimik3openfrontier#18-quant-formats|kimik3 §1.8]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/_aiter_ops.py
  - file:///Users/aarnphm/workspace/vllm/vllm/envs.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
merged: '2026-09-27'
number: 58201
source:
  - https://github.com/vllm-project/vllm/blob/3137ff07739f0cd19e06119506c6432a4a7e77ad/vllm/_aiter_ops.py
  - https://github.com/vllm-project/vllm/blob/3137ff07739f0cd19e06119506c6432a4a7e77ad/vllm/envs.py
  - https://github.com/vllm-project/vllm/blob/3137ff07739f0cd19e06119506c6432a4a7e77ad/vllm/model_executor/layers/fused_moe/experts/rocm_aiter_moe.py
  - https://github.com/vllm-project/vllm/blob/3137ff07739f0cd19e06119506c6432a4a7e77ad/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#58201 [ROCm][Kimi-K3] Make VLLM_ROCM_USE_AITER_MOE_SITUV2 select a4w4/a8w4/a16w4'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/58201
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/58201
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/58201.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
The SiTU MXFP4 MoE has three FlyDSL kernel families that differ only in activation precision; the vLLM · full dump: [[thoughts/inference/vllm-prs/58201|#58201]]

