---
cited_by:
  - '[[thoughts/papers/kimiteam2026kimik3openfrontier#18-quant-formats|kimik3 §1.8]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/quantization/mxfp4.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/quantization/online/base.py
  - file:///Users/aarnphm/workspace/vllm/vllm/config/quantization.py
merged: '2026-10-03'
number: 51274
source:
  - https://github.com/vllm-project/vllm/blob/b0e21b308352587a6fd02f72722a2e815bfd62f0/vllm/model_executor/layers/quantization/mxfp4.py
  - https://github.com/vllm-project/vllm/blob/b0e21b308352587a6fd02f72722a2e815bfd62f0/vllm/model_executor/layers/quantization/online/base.py
  - https://github.com/vllm-project/vllm/blob/b0e21b308352587a6fd02f72722a2e815bfd62f0/vllm/config/quantization.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#51274 [ROCm][Kimi-K3] Add opt-in gfx942 MXFP4-to-int4 conversion'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/51274
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/51274
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/51274.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Add an explicit `int4_per_group_32` MoE override for Kimi-K3 on gfx942. When selected, the MXFP4 expert weights are requantized once at load time and served through AITER’s replacement BF16 x packed-int4 FlyDSL path. · full dump: [[thoughts/inference/vllm-prs/51274|#51274]]

