---
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/replayssm#output-only-and-standard-decode|replayssm § Output only and standard…]]'
  - '[[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/v1/attention/backends/mamba_attn.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/mamba/mamba_mixer2.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/mamba/mamba_utils.py
  - file:///Users/aarnphm/workspace/vllm/vllm/config/vllm.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/models/nemotron_h.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/worker/gpu_input_batch.py
merged: '2026-07-24'
number: 48018
source:
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/model_executor/layers/mamba/ops/selective_state_update_replayssm_output_only.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/v1/attention/backends/mamba_attn.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/model_executor/layers/mamba/mamba_mixer2.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/model_executor/layers/mamba/ops/replayssm_config.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/model_executor/layers/mamba/mamba_utils.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/config/vllm.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/model_executor/models/nemotron_h.py
  - https://github.com/vllm-project/vllm/blob/866fea2b9900bf49d552c205d2eaac4716fb63ac/vllm/v1/worker/gpu_input_batch.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#48018 [Kernel] ReplaySSM: cache SSM inputs for faster Mamba2 standard decode'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/48018
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/48018
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/48018.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Everything here is opt-in. With the feature off, the decode path is byte-for-byte the baseline. · full dump: [[thoughts/inference/vllm-prs/48018|#48018]]

