---
cited_by:
  - '[[thoughts/inference/kimi-k3#16-speculative-decoding-mtp-dspark-eagle3dflash|kimik3 §1.6]]'
  - '[[thoughts/inference/kimi-k3-state#33-spec-decode-with-recurrent-state|kimik3 §3.3]]'
  - '[[thoughts/inference/kimi-k3-prs#4-pr-reading-order|kimik3 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/flashinfer_mla.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/worker/gpu/model_states/mamba_hybrid.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/attention/mla_attention.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/flashattn_mla.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/flashmla.py
merged: '2026-09-23'
number: 52988
source:
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/v1/attention/backends/mla/flashinfer_mla.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/models/kimi_k3/nvidia/dspark_mla.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/models/kimi_k3/nvidia/kda_metadata.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/v1/worker/gpu/model_states/mamba_hybrid.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/model_executor/layers/attention/mla_attention.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/models/dots3_note/nvidia/attention.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/v1/attention/backends/mla/flashattn_mla.py
  - https://github.com/vllm-project/vllm/blob/88aa0d287dd3abac89741bbea349350a4d49194e/vllm/v1/attention/backends/mla/flashmla.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#52988 [Spec decode] Support variable-length decode for Kimi-K3 adaptive ver'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/52988
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/52988
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/52988.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Make Kimi-K3’s MLA and KDA decode paths capturable as FULL varlen CUDA graphs so they work under adaptive DSpark verification ([#47808](https://github.com/aarnphm/aarnphm.github.io/issues/47808)), which schedules ragged per-request draft budgets and requires device-sourced query lengths. Before this… · full dump: [[thoughts/inference/vllm-prs/52988|#52988]]

