Saved in:
Bibliographic Details
Main Author: Nguyen, An Xuan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.22416
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916036276649984
author Nguyen, An Xuan
author_facet Nguyen, An Xuan
contents Hybrid language models like Jamba mix attention layers with State Space Models (SSMs), creating two memory cache types with opposite profiles: Key-Value (KV) caches grow linearly with sequence length, while SSM states stay fixed per layer. Current inference engines handle this poorly. Unified pools pad SSM states to attention page sizes, wasting up to 7.3x capacity. Static dual pools cannot adapt when prompt distributions shift between requests. We present Asymmetric Virtual Memory Paging (AVMP). The allocator separates the two cache types into physically distinct pools behind a unified virtual address space, and migrates capacity between pools when one runs out. Migration triggers only on allocation failure, keeping behavior deterministic. We evaluate AVMP across 270 synthetic cells plus 60 cells of ShareGPT trace replay on an RTX 3060 12GB. Out-of-Memory events drop 7.6% and request throughput improves 1.83x to 13.3x across synthetic workloads and 2.36x on ShareGPT. All gains hold under paired-bootstrap 95% confidence intervals. A phase-time breakdown reveals two distinct mechanisms: shorter OOM recovery on capacity-pressured workloads, and faster allocation calls on KV-heavy workloads. Implementation is pure Python; Triton integration is future work.
format Preprint
id arxiv_https___arxiv_org_abs_2605_22416
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference
Nguyen, An Xuan
Machine Learning
Distributed, Parallel, and Cluster Computing
Performance
D.4.2; B.3.2
Hybrid language models like Jamba mix attention layers with State Space Models (SSMs), creating two memory cache types with opposite profiles: Key-Value (KV) caches grow linearly with sequence length, while SSM states stay fixed per layer. Current inference engines handle this poorly. Unified pools pad SSM states to attention page sizes, wasting up to 7.3x capacity. Static dual pools cannot adapt when prompt distributions shift between requests. We present Asymmetric Virtual Memory Paging (AVMP). The allocator separates the two cache types into physically distinct pools behind a unified virtual address space, and migrates capacity between pools when one runs out. Migration triggers only on allocation failure, keeping behavior deterministic. We evaluate AVMP across 270 synthetic cells plus 60 cells of ShareGPT trace replay on an RTX 3060 12GB. Out-of-Memory events drop 7.6% and request throughput improves 1.83x to 13.3x across synthetic workloads and 2.36x on ShareGPT. All gains hold under paired-bootstrap 95% confidence intervals. A phase-time breakdown reveals two distinct mechanisms: shorter OOM recovery on capacity-pressured workloads, and faster allocation calls on KV-heavy workloads. Implementation is pure Python; Triton integration is future work.
title Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference
topic Machine Learning
Distributed, Parallel, and Cluster Computing
Performance
D.4.2; B.3.2
url https://arxiv.org/abs/2605.22416