init repo
This commit is contained in:
33
.memory/cvpr2026-3d-vision.md
Normal file
33
.memory/cvpr2026-3d-vision.md
Normal file
@@ -0,0 +1,33 @@
|
||||
---
|
||||
name: CVPR 2026 3D Vision Foundation Models 研究
|
||||
description: CVPR 2026 基础视觉大模型在三维视觉领域的热门开源项目调研,含 G2VLM/VLM-3R/TALO/Flow3r,待 eGPU+NVIDIA 卡到位后本地运行
|
||||
type: project
|
||||
---
|
||||
|
||||
## 项目信息
|
||||
|
||||
- **路径**:`~/Projects/research/20260319-cvpr2026-3d-vision/`
|
||||
- **端口**:4070(`python3 -m http.server 4070`)
|
||||
- **内容**:单文件 HTML 研究展示页
|
||||
|
||||
## 调研项目列表
|
||||
|
||||
| 项目 | Stars | 机构 | 最低 VRAM | GitHub |
|
||||
|------|-------|------|-----------|--------|
|
||||
| G2VLM (2B) | 291 | 上海AI Lab/UCLA/浙大 | ~8GB | InternRobotics/G2VLM |
|
||||
| VLM-3R (7B) | 363 | UT Austin/Meta/UCLA | ~16GB | VITA-Group/VLM-3R |
|
||||
| TALO | 56 | — | ~12GB | Xian-Bei/TALO |
|
||||
| Flow3r | 60 | — | ~8GB | Kidrauh/flow3r |
|
||||
|
||||
## 硬件需求
|
||||
|
||||
- 全部依赖 CUDA 12.1+、flash-attention、triton(NVIDIA 专属)
|
||||
- macOS 不支持 NVIDIA eGPU 驱动,Apple Silicon 也不支持外置 Linux+eGPU
|
||||
- **推荐**:RTX 4090(24GB VRAM)覆盖全部项目
|
||||
- **待定**:用户 Mac 型号(Intel vs Apple Silicon),决定 eGPU 方案或独立 Linux 小主机方案
|
||||
|
||||
## 可行方案
|
||||
|
||||
1. **eGPU + Linux**(仅 Intel Mac):eGPU 盒子 + RTX 4090 + Ubuntu
|
||||
2. **独立 Linux 小主机**(最稳,不限 Mac 型号):mini-ITX + RTX 4090 + Ubuntu + SSH
|
||||
3. **云 GPU**(零投入):Colab Pro+ / RunPod / Vast.ai
|
||||
3
.memory/worklog.json
Normal file
3
.memory/worklog.json
Normal file
@@ -0,0 +1,3 @@
|
||||
{
|
||||
"entries": []
|
||||
}
|
||||
Reference in New Issue
Block a user