Feature Request: fixed VRAM allocation pool

未关闭
#115 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

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
22/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
冷清
技术栈
cpp

调研方向

The request centres on the CUDA backend's buffer allocation, so start with ggml_backend_cuda_buffer_type_alloc_buffer in ggml/src/ggml-cuda.cu and trace where cudaMalloc is called for model weights and the KV cache. The issue proposes a fixed pool sized as a fraction of VRAM, but the design (how to size it, how to handle fragmentation and freeing) is unsettled and the code in the issue is an unverified sketch. Done would need a maintainer-agreed design before any implementation.

由索引模型根据 Issue 内容生成。

描述

enhancement
Prerequisites
  • I am running the latest code. Mention the version if possible as well.
  • I carefully followed the README.md.
  • I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
  • I reviewed the Discussions, and have a new and useful enhancement to share.
Feature Description

It would be nice to be able to set the value of preallocated VRAM in percents like for vLLM (ex. 0.9 = 90%) and to use preallocated pool in this mode, so that extra software never influence beellama work.

Motivation

The problem: whatever manually crafted settings for beellama and model are made, the VRAM consumption is not fully controlled, because VRAM is allocated during context growth and beellama work and because other GUI applications (browsers, electron-based projects) also continue to consume extra VRAM. Currently Firefox can easily consume up to 1.5 GB vram and begin with 100 MB initially, so generation speed just drops to almost zero from unpredictable moment.

Possible Implementation

LLM based suggestions:

Method 2: Code Modification (Hard Physical Reservation Pool)

If you want llama.cpp to allocate a single massive contiguous block of VRAM (e.g., 20 GB out of 24 GB) via cudaMalloc at program start, and have GGML consume slices of that block instead of calling cudaMalloc dynamically:

Step 1: Create a Fixed Global Pool

In ggml/src/ggml-cuda.cu (or ggml-cuda.cpp), declare a static memory pool pointer and allocate it during CUDA backend initialization:

// Static preallocated pool
static void* g_cuda_vram_pool = nullptr;
static size_t g_vram_pool_offset = 0;
static size_t g_vram_pool_size = 0;

void init_cuda_vram_pool(size_t reserve_bytes) {
    if (g_cuda_vram_pool != nullptr) return;

    // Physically allocate the entire pool at start
    cudaError_t err = cudaMalloc(&g_cuda_vram_pool, reserve_bytes);
    if (err == cudaSuccess) {
        g_vram_pool_size = reserve_bytes;
        printf("CUDA: Preallocated %zx bytes of VRAM pool successfully.\n", reserve_bytes);
    } else {
        printf("CUDA: Failed to preallocate VRAM pool: %s\n", cudaGetErrorString(err));
    }
}

Step 2: Override GGML's CUDA Buffer Allocation Function

In ggml-cuda.cu, locate the buffer allocation function used by the CUDA backend: ggml_backend_cuda_buffer_type_alloc_buffer.

Replace the underlying cudaMalloc call inside that function so it returns an offset pointer into g_cuda_vram_pool:

// Simplistic bump-allocator replacement inside CUDA backend
static void* pool_alloc(size_t size) {
    // Maintain 256-byte alignment required by CUDA
    size_t aligned_size = (size + 255) & ~255; 

    if (g_vram_pool_offset + aligned_size > g_vram_pool_size) {
        printf("CUDA Pool Error: Out of preallocated VRAM!\n");
        return nullptr;
    }

    void* ptr = (uint8_t*)g_cuda_vram_pool + g_vram_pool_offset;
    g_vram_pool_offset += aligned_size;
    return ptr;
}

// In your buffer allocation routine:
// Replace cudaMalloc(&ptr, size) with:
ptr = pool_alloc(size);

Step 3: Call Initialization in llama_backend_init

Call init_cuda_vram_pool(target_size_bytes) inside your initialization function (e.g., llama_backend_init) before loading model layers or allocating context buffers.

Result

When llama.cpp starts up, it immediately issues a single cudaMalloc call for your requested size. All model weights, KV cache matrices, and scratch workspaces allocated during token generation will be carved directly out of this existing buffer. No further calls to cudaMalloc or cudaFree will touch the driver or OS.

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

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

Anbeeld/beellama.cpp 的其他 Issue

查看 Anbeeld/beellama.cpp 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。