RestoreKV-Qwen3-8B_plus

PEFT LoRA adapter + restore-token embeddings for RestoreKV (paper), a budget-matched, single-pass plug-in that recovers full-cache behavior under aggressive KV cache eviction on top of KVzip.

Base model: Qwen/Qwen3-8B — base scorer: KVzip+
adapter_model.safetensors: rank-8 LoRA (α=16) on q/k/v/o/gate/up/down proj
restore_embeddings.safetensors: 8 learned restore-token embeddings

Load via the RestoreKVPress in NVIDIA/kvpress.

License

CC BY-NC 4.0 (distilled with LongAlpaca-derived data; research / non-commercial).

Downloads last month: 10

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for higokri/RestoreKV-Qwen3-8B_plus

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Adapter

(1988)

this model

Collection including higokri/RestoreKV-Qwen3-8B_plus

RestoreKV

Collection

PEFT LoRA adapters for RestoreKV (arXiv:2608.01247): budget-matched learned restoration on top of KVzip. • 4 items • Updated 2 days ago

Paper for higokri/RestoreKV-Qwen3-8B_plus

RestoreKV: Recovering Full-Cache Behavior Under Aggressive Query-Agnostic KV Cache Eviction

Paper • 2608.01247 • Published 7 days ago • 11