为什么值得关注

该仓库填补了消费级 Blackwell 硬件官方支持不完整的空白，提供了真实世界的问题和解决方案，且每个条目均带有日期和版本号，任何人都可以复现。

适合谁使用

RTX 50 系列显卡用户
在消费级硬件上进行本地推理或微调的 AI 开发者
希望在单卡上运行大模型的爱好者与研究人员
调试 Blackwell 上 PyTorch/CUDA 兼容性的工程师

典型使用场景

快速诊断并修复 sm_120 上 torch 内核不匹配的问题
优化单张 32GB 显卡上大模型的推理吞吐量与内存占用
量化模型（FP4、AWQ、GGUF）以适配显存并测量真实性能
使用正确的 CUDA 和工具链设置从源码构建 AI 框架

项目优势

所有结论均有日期、版本号，并使用开放代码和数据可复现
覆盖 torch、推理、量化、源码构建、多模态等广泛实用问题
提供简单的 5 分钟快速测试，可在深入构建前验证环境
完全基于真实可购买的显卡，无需依赖云端

使用前须知

仅针对单张 RTX 5090，未涉及多 GPU 或数据中心场景
条目可能随软件版本更新而过时；‘验证于’行提供了可信度参考
并非所有 GPU 的通用指南，仅解决消费级 Blackwell (sm_120) 的问题

README 快速开始

the sm_120 field guide

AI on consumer Blackwell. the fixes and footguns from one RTX 5090, measured honestly.

every entry here is something I hit running modern AI on an RTX 50-series card: the error message, the mechanism behind it, the fix that worked, and the date + versions it was verified on. if I didn't run it, it isn't here. if you own a 50-series card (sm_120), this is the debugging session you get to skip.

most AI infrastructure targets datacenter cards. consumer Blackwell is close enough to look supported and different enough to break in quiet, expensive ways: wheels that ship without kernels, and headline speedups that assume an H200. this guide maps where it breaks and what to do about it.

the machine

part	spec
GPU	RTX 5090 32GB (Blackwell, sm_120)
CPU	AMD Ryzen 5 9600 (6c/12t)
RAM	64GB DDR5-5600
OS	Ubuntu Server 26.04 LTS, headless

one card you can actually buy. no rack, no cloud fallback.

the rules

reproducible. open code, open data, on a machine I specify. the raw benchmark data behind these entries is public.
honest. every claim is dated and versioned. entries go stale as stacks move; the Verified: line tells you how much to trust each one.
sovereign. if a result needs the cloud, it doesn't count here.

start here: the 5-minute kill test

before building any repo's stack on a 50-series card, test only its torch pin:

uv venv /tmp/probe && source /tmp/probe/bin/activate
uv pip install "torch==" --index-url https://download.pytorch.org/whl/cu124
python -c 'import torch; x = torch.ones(2, device="cuda"); print(x + x)'

if that fails with CUDA error: no kernel image is available for execution on the device, stop. torch wheels below 2.7 (the cu124/cu126 era) ship no sm_120 kernels, and nothing built on top of that pin can fix it. reimplement on torch ≥ 2.7 + cu128 instead. old-transformers-on-new-torch is salvageable; old-torch never is. full entry in torch, kernels & training stacks.

the guide

80 entries across seven files, newest first in each.

torch, kernels & training stacks · 7 entries

torch pins below 2.7 ship no sm_120 kernels: kill a doomed stack in 5 minutes
custom trust_remote_code models inherit transformers version skew: pin the OLD transformers, keep the NEW torc

项目描述

AI on consumer Blackwell — a field guide for the RTX 50-series (sm_120). Fixes, footguns, and honest measurement from one RTX 5090.

sm120-field-guide