---
cited_by:
  - '[[thoughts/inference/architectures#5-four-things-called-replay|README §5]]'
  - '[[thoughts/inference/architectures#6-reading-order|README §6]]'
  - '[[thoughts/inference/deepseek-v4.1-state#34-not-replayssm|dsv41 §3.4]]'
  - '[[thoughts/inference/kimi-k3-state#33-spec-decode-with-recurrent-state|kimik3 §3.3]]'
  - '[[thoughts/inference/replayssm#3-mtp--spec-decode-52928|replayssm §3]]'
  - '[[thoughts/inference/replayssm-serving#5-pr-reading-order|replayssm §5]]'
  - '[[thoughts/inference/replayssm#unverified--gaps|replayssm § Unverified / gaps]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/mamba/ops/ssu_dispatch.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/mamba/mamba_mixer2.py
  - file:///Users/aarnphm/workspace/vllm/vllm/config/vllm.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mamba_attn.py
  - file:///Users/aarnphm/workspace/vllm/.buildkite/test_areas/engine.yaml
  - file:///Users/aarnphm/workspace/vllm/vllm/config/cache.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/mamba/mamba_utils.py
merged: '2026-09-30'
number: 52928
source:
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/model_executor/layers/mamba/ops/ssu_dispatch.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/model_executor/layers/mamba/mamba_mixer2.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/config/vllm.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/v1/attention/backends/mamba_attn.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/model_executor/warmup/replayssm_warmup.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/.buildkite/test_areas/engine.yaml
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/config/cache.py
  - https://github.com/vllm-project/vllm/blob/deb148e730225c64427f8ccf1f313cb4b33325b0/vllm/model_executor/layers/mamba/mamba_utils.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#52928 [Mamba] Add FlashInfer ReplaySSM support for MTP'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/52928
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/52928
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/52928.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
This stacked follow-up extends the native FlashInfer ReplaySSM integration from single-token decode in [#52506](https://github.com/aarnphm/aarnphm.github.io/issues/52506) to multi-token prediction (MTP) verification for Mamba2 models. · full dump: [[thoughts/inference/vllm-prs/52928|#52928]]

