RFC: Integrating KVMem virtual working set for 256K context on 16GB VRAM (Working prototype ready in fork)
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 8/100
- Issue 类型
- 功能
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- cpp
调研方向
This is an RFC, not a scoped task: it asks the maintainer to choose between upstreaming a ~36K LOC KVMem working-set engine and keeping it as a downstream fork, and the 27-comment thread is where that decision is being made. To follow it, read the fork branch beellama-kvmem and the referenced ggml-cuda/gated_delta_net.cu Blackwell fix, then the thread for the maintainer's answer. Done means an architectural decision plus, at most, a targeted PR such as the GDN CUDA fix; nothing here is a first contribution.
由索引模型根据 Issue 内容生成。
描述
RFC: Integrating KVMem virtual working set for 256K context on 16GB VRAM (Working prototype ready in fork)
Context & Motivation
Large context windows (128K–256K) on 27B+ models like Qwen3.8-27B (and hybrid architectures with Gated Delta Net) are traditionally difficult to run on consumer hardware (16GB VRAM cards like RTX 4080 / RTX 5060 Ti).
While BeeLlama's native KVarN (and SWA) provides substantial cache savings (~2.67x with kvarn6), pushing beyond 100K–128K on 16GB VRAM still pushes right up against physical memory limits.
We have successfully integrated the KVMem virtual working set engine (arXiv:2609.04852) into BeeLlama v0.4.6, maintaining full compatibility with BeeLlama's core optimizations while allowing true 256K context streaming.
A fully working implementation, benchmarks, and Windows binaries are available in the fork:
👉 Repository & Branch: Andgihat/beellama.cpp (branch: beellama-kvmem)
👉 Release & Binaries: v0.4.6-kvmem
How it works
- Host RAM Virtual Working Set: The complete 262,144 (256K) context lives in pinned Host RAM (DDR5, ~8–12 GB).
- Fixed GPU Slot Pool: Only an active working set (e.g. Budget $B = 36K$ + Reserve $R = 16K$) sits in VRAM (~14.6 GB total with weights, zero OOM).
- Graph Recall & Cache Quantization: Blocks (128 tokens) are staged in/out over PCIe on the fly using Hadamard-rotated quantized slots (
q4_0/q8_0). - Blackwell sm_120 Fix: In the process, we resolved a GDN warp reduction bug in
ggml-cuda/gated_delta_net.cu(lanes_per_colstride and pointer offset checkkeep_rs = K > 1), restoring 100% correct mathematics for hybrid Qwen3.8 on RTX 50-series GPUs.
Benchmark Results (Qwen3.8-27B-UD-Q3_K_XL on RTX 5060 Ti 16GB)
Tested on real text documents from 10K to 256K context:
- Prompt Processing (Prefill @ 10K+ tokens):
915.84 tokens/s(10,196 tokens processed in 11.13s with-b 2048 -ub 1024). - Decode Speed:
35.6 - 40.3 tokens/swith built-in MTP (--spec-type draft-mtp). - MTP Acceptance Rate:
76.0%(consistent 2.46 tokens/round). - RAM Usage @ 10K: Only 187 MB with
q4_0cache. - VRAM Headroom: Fits comfortably within 15.3 GB on a 16GB GPU.
Question for Maintainer (@Anbeeld)
Because the KVMem integration is relatively substantial (~36K LOC including tests, standalone llama-kvmem-server, and runtime adapters), we wanted to check with you first:
- Upstream PR: Would you be interested in having KVMem integrated directly into upstream BeeLlama (e.g. as an optional feature / server target)?
- Specialized Fork: Or would you prefer we maintain this as a specialized downstream fork / companion project (
beellama-kvmem), and only submit targeted PRs (such as the Blackwell GDN CUDA fixes ingated_delta_net.cu) to upstream?
Happy to adapt the changes or submit targeted PRs depending on your architectural preferences. Thank you for the awesome work on BeeLlama!
- 主要语言
- C++
- 星标
- 1.2k
- 派生
- 80
- 平均合并
- 4 天 20 小时
- 30 天内合并 PR
- 2
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
Anbeeld/beellama.cpp 的其他 Issue
-
[BUG] KVarN assertion failure in build_inp_mem_hybrid_iswa() due to uninitialized self_kvarn_rot未关闭
难度 2/5 1-3 小时 新手友好度 68/100
Anbeeld/beellama.cpp#205 ·
-
难度 3/5 半天 新手友好度 58/100
Anbeeld/beellama.cpp#206 ·
-
难度 3/5 1-2 天 新手友好度 57/100
Anbeeld/beellama.cpp#204 ·
-
难度 4/5 3-5 天 新手友好度 45/100
Anbeeld/beellama.cpp#203 · 2 条评论 ·
-
难度 4/5 3-5 天 新手友好度 45/100
Anbeeld/beellama.cpp#202 · 1 条评论 ·
查看 Anbeeld/beellama.cpp 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 72/100
CachyOS/packageinstaller#144 ·
维护者通常 1 天内回复
-
documentation
难度 1/5 1 小时以内 新手友好度 82/100
TomerAberbach/profiler-md#951 ·
维护者通常 1 天内回复
-
[Bug] Default key routing (JavaStringHash) picks a different partition than the Java client for about half of keys可能已有人在做 @0z5a 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 82/100
apache/pulsar-client-go#1575 ·
维护者通常 2 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
pocoproject/poco#5764 ·
维护者通常 2 天内回复
-
bug cpp-core-guidelines low-priority
难度 2/5 1-3 小时 新手友好度 62/100
NatronGitHub/openfx-supportext#164 ·
维护者通常 1 天内回复