---
cited_by:
  - '[[thoughts/inference/qwen3.8-serving#4-pr-reading-order-vllm-main-git-show-sha|qwen4next §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/.buildkite/test_areas/models_basic.yaml
merged: '2026-09-25'
number: 57497
source:
  - https://github.com/vllm-project/vllm/blob/267eee54bf232e4008a7e0b510f19a62e23aaaca/vllm/models/qwen4_exp/common/ngram_embedding.py
  - https://github.com/vllm-project/vllm/blob/267eee54bf232e4008a7e0b510f19a62e23aaaca/vllm/models/qwen4_exp/nvidia/ngram_embedding.py
  - https://github.com/vllm-project/vllm/blob/267eee54bf232e4008a7e0b510f19a62e23aaaca/vllm/models/qwen4_exp/amd/ple_layer.py
  - https://github.com/vllm-project/vllm/blob/267eee54bf232e4008a7e0b510f19a62e23aaaca/vllm/config/engram.py
  - https://github.com/vllm-project/vllm/blob/267eee54bf232e4008a7e0b510f19a62e23aaaca/.buildkite/test_areas/models_basic.yaml
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#57497 [Qwen4Exp][ROCm] PLE n-gram table CPU offload'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/57497
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/57497
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/57497.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Enable `VLLM_PLE_CPU_OFFLOAD` on the AMD/ROCm path: keep the Qwen4Exp PLE n-gram table in pinned host memory, read through a UVA view on the GPU. · full dump: [[thoughts/inference/vllm-prs/57497|#57497]]

