---
cited_by:
  - '[[thoughts/papers/kimiteam2026attentionresiduals#14-attnres-attention-residuals-arxiv-260315031-kimi-team-expansion-confirmed-from-the-paper-title|kimik3 §1.4]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-08'
number: 59069
source:
  - https://github.com/vllm-project/vllm/blob/458ba2edf85bc9b7ebc0d5141f34ea658520faf5/vllm/models/kimi_k3/amd/linear.py
  - https://github.com/vllm-project/vllm/blob/458ba2edf85bc9b7ebc0d5141f34ea658520faf5/vllm/models/kimi_k3/amd/ops/attn_res.py
  - https://github.com/vllm-project/vllm/blob/458ba2edf85bc9b7ebc0d5141f34ea658520faf5/vllm/models/kimi_k3/amd/mla.py
  - https://github.com/vllm-project/vllm/blob/458ba2edf85bc9b7ebc0d5141f34ea658520faf5/vllm/models/kimi_k3/amd/kda.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#59069 [ROCm][Kimi-K3] Fuse AttnRes output with per-token FP8 input quantization'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/59069
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/59069
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/59069.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Eligible Kimi-K3 AMD AttnRes outputs currently materialize BF16 activations that the next linear immediately quantizes. Fuse that per-token FP8 conversion into AttnRes and pass native `QuantizedActivation` to supported consumers. · full dump: [[thoughts/inference/vllm-prs/59069|#59069]]

