PyTorch-HighSierra-CUDA-Revival

KO | EN

gitlite — search

#javascript #typescript #rust #deep-learning #docker #macos #cli #kubernetes #vue #awesome-list #pytorch #security #mcp #openai

PyTorch-HighSierra-CUDA-Revival

(★ 27)

Revives the last functional CUDA deep-learning stack on macOS High Sierra. NVIDIA CUDA 10.2 + cuDNN 7.6.5 + PyTorch 1.7.0 fully working on GTX Pascal GPUs — a discontinued ecosystem resurrected.

PyTorch-HighSierra-CUDA-Revival 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

---

# 🔥 PyTorch High Sierra CUDA Revival  
## **The Final Fully GPU-Accelerated Deep Learning Stack for macOS**

This project restores full CUDA compute support on macOS High Sierra using NVIDIA CUDA 10.2, cuDNN 7.6.5, and a custom‑built PyTorch 1.7.0. 
A completely discontinued ecosystem — resurrected and preserved.

## 📚 Quick Docs Menu
- Installation: `docs/installation.md`
- Troubleshooting: `docs/troubleshooting.md`
- Compatibility Matrix: `docs/compatibility_matrix.md`
- Architecture Overview: `docs/architecture_overview.md`
- Rebuild From Source: `docs/rebuild_from_source.md`
- Performance Benchmarks: `docs/performance_benchmarks.md`
- Historical Notes: `docs/historical_notes.md`
- Environment Specs: `docs/environment_specs.md`
- Examples: `examples/`

# 📌 Background & Motivation

CUDA on macOS is **dead**.  
Apple killed NVIDIA support.  
NVIDIA discontinued macOS drivers.  
PyTorch removed CUDA for macOS entirely.

But High Sierra (10.13.6) is the **last remaining macOS** that can still run:

- CUDA 10.2  
- cuDNN 7.6.5  
- NVIDIA WebDrivers  
- Pascal GPUs at full power

This repository preserves the *only functioning* deep learning GPU environment for macOS.

This is not just a wheel —  
It is a **technical museum piece** that still works.

---

# 🖥️ System Compatibility Matrix

| Component | Version | Status |
|----------|---------|--------|
| macOS | **10.13.6 High Sierra** | ✔ Supported |
| NVIDIA Drivers | **WebDriver 387/387+** | ✔ Required |
| CUDA Toolkit | **10.2** | ✔ Fully Compatible |
| cuDNN | **7.6.5** | ✔ Final Version for macOS |
| GPU Architecture | **Pascal (SM_61)** | ✔ Optimal |
| PyTorch | **1.7.0a0 (Custom CUDA Build)** | ✔ Working |
| Python | **3.8.x** | ✔ Compatible |

---

# 🧩 Supported GPUs (All WebDriver-Compatible)

### **Pascal (Full CUDA Support)**
- GeForce GTX 1050 / 1050 Ti  
- GeForce GTX 1060  
- GeForce GTX 1070  
- GeForce GTX 1080  
- GeForce GTX 1080 Ti  
- Titan X / Titan Xp

### **Kepler (Works but slower)**
- GT 710 / GT 730  
- GTX 770 / 780 / 780 Ti

### **Official WebDriver Mac GPUs**
#### GeForce 600 Series  
- GTX 680

#### GeForce 200 Series  
- GTX 285

#### GeForce 100 Series  
- GT 120

#### GeForce 8 Series  
- 8800 GT

#### Quadro Series  
- Quadro K5000 (Mac)  
- Quadro 4000 (Mac)

#### Quadro FX Series  
- FX 4800  
- FX 5600

**If WebDriver works → CUDA PyTorch works.**

---

# 📚 Documentation & Preservation Index

- Architecture overview: `docs/architecture_overview.md`
- Troubleshooting (WebDriver, CUDA, cuDNN, sm_61, nvcc, SIP): `docs/troubleshooting.md`
- Performance benchmarks (matmul, transformer, kernels): `docs/performance_benchmarks.md`
- Compatibility matrix (GPU list & tiers): `docs/compatibility_matrix.md`
- Rebuild from source (wheel reproduction): `docs/rebuild_from_source.md`
- Historical notes (macOS CUDA deprecation timeline): `docs/historical_notes.md`

Preservation artifacts: `archive/` (build logs, NVCC flags, CMake cache, wheel hashes, CUDA lib hashes, test outputs)

Environment reproducibility: `environment/` (pip JSON, YAML, full CUDA/cuDNN dumps, system specs)

Schemas for CI validation: `schemas/`

---

# 🚀 Installation

### 1️⃣ Download & Install (Release Asset)

Direct wheel download:

`https://github.com/careunix/PyTorch-HighSierra-CUDA-Revival/releases/download/v0.1.0/torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl`

Option A — install directly from URL:
```bash
pip install "https://github.com/careunix/PyTorch-HighSierra-CUDA-Revival/releases/download/v0.1.0/torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl"
```

Option B — download then install:
```bash
curl -L -o torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl \
  https://github.com/careunix/PyTorch-HighSierra-CUDA-Revival/releases/download/v0.1.0/torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl
pip install ./torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl
```

### 2️⃣ Integrity & Signature Verification

SHA256 should be:
```text
38da4acfe780a041b1f73f67c66efcdb37e9773615446f6a02ed2586f3cff9c7
```

Check locally:
```bash
shasum -a 256 torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl | awk '{print $1}'
```

Verify GPG signature (public key is in repo):
```bash
# Import public key (from repo)
gpg --import archive/keys/careunix.pub.asc

# Or fetch raw key without cloning
curl -L -o careunix.pub.asc \
  https://raw.githubusercontent.com/careunix/PyTorch-HighSierra-CUDA-Revival/main/archive/keys/careunix.pub.asc
gpg --import careunix.pub.asc

# Verify detached signature (from repo)
curl -L -o SIGNATURE.asc \
  https://raw.githubusercontent.com/careunix/PyTorch-HighSierra-CUDA-Revival/main/wheel/SIGNATURE.asc
gpg --verify SIGNATURE.asc torch-1.7.0a0-cp38-cp38-macosx_10_13_x86_64.whl
```

### 3️⃣ Verify CUDA functionality
```python
import torch
print("CUDA:", torch.cuda.is_available())
print("GPU:", torch.cuda.get_device_name(0))
```

Expected:
```text
CUDA: True
GPU: GeForce GTX 1060
```

---
## 🔧 Verified Test Environment

```text
$ nvcc --version
Cuda compilation tools, release 10.2, V10.2.89

$ python --version
Python 3.8.20

$ conda --version
conda 23.5.2

$ python -m pip --version
pip 24.2 (Python 3.8)
```

This wheel was built & validated inside the environment above. Any system matching these versions (or close variants) will work.

---
# 📦 Python Package Environment — Verified Working Set  
### *PyTorch CUDA 10.2 Build on macOS High Sierra (careunix)*

This section documents the **exact Python package list** used during the successful compilation of:

- PyTorch **1.7.0a0** (custom CUDA build)
- CUDA **10.2**
- cuDNN **7.6.5**
- Python **3.8.20**
- Conda **23.5.2**

Environment name: **gpt2env**  
Primary developer: **careunix**

These versions are confirmed **100% compatible** and form a *fully reproducible build environment*.

---

## 📌 Export Your Own List  
You can generate the same format with:

```bash
python -m pip list --format=json > pip_environment.json
```

---

## ✔ Verified Working Package Set  
*(Direct JSON output from careunix’s functional build environment)*

```json
[
  {"name": "accelerate", "version": "1.0.1"},
  {"name": "autocommand", "version": "2.2.2"},
  {"name": "backports.tarfile", "version": "1.2.0"},
  {"name": "Brotli", "version": "1.0.9"},
  {"name": "certifi", "version": "2024.8.30"},
  {"name": "cffi", "version": "1.17.1"},
  {"name": "charset-normalizer", "version": "3.3.2"},
  {"name": "click", "version": "8.1.8"},
  {"name": "filelock", "version": "3.16.1"},
  {"name": "future", "version": "0.18.3"},
  {"name": "huggingface-hub", "version": "0.0.12"},
  {"name": "idna", "version": "3.7"},
  {"name": "importlib_metadata", "version": "8.0.0"},
  {"name": "importlib_resources", "version": "6.4.0"},
  {"name": "inflect", "version": "7.3.1"},
  {"name": "jaraco.collections", "version": "5.1.0"},
  {"name": "jaraco.context", "version": "5.3.0"},
  {"name": "jaraco.functools", "version": "4.0.1"},
  {"name": "jaraco.text", "version": "3.12.1"},
  {"name": "joblib", "version": "1.4.2"},
  {"name": "mkl-fft", "version": "1.3.8"},
  {"name": "mkl-random", "version": "1.2.4"},
  {"name": "mkl-service", "version": "2.4.0"},
  {"name": "more-itertools", "version": "10.3.0"},
  {"name": "numpy", "version": "1.23.5"},
  {"name": "packaging", "version": "25.0"},
  {"name": "pip", "version": "24.2"},
  {"name": "platformdirs", "version": "4.2.2"},
  {"name": "pycparser", "version": "2.21"},
  {"name": "PySocks", "version": "1.7.1"},
  {"name": "PyYAML", "version": "6.0.2"},
  {"name": "regex", "version": "2024.11.6"},
  {"name": "requests", "version": "2.32.3"},
  {"name": "sacremoses", "version": "0.1.1"},
  {"name": "setuptools", "version": "75.1.0"},
  {"name": "six", "version": "1.16.0"},
  {"name": "tokenizers", "version": "0.10.3"},
  {"name": "tomli", "version": "2.0.1"},
  {
    "name": "torch",
    "version": "1.7.0a0",
    "editable_project_location": "/Users/careunix/pytorch"
  },
  {"name": "tqdm", "version": "4.67.1"},
  {"name": "transformers", "version": "4.8.2"},
  {"name": "typeguard", "version": "4.3.0"},
  {"name": "typing_extensions", "version": "4.11.0"},
  {"name": "urllib3", "version": "2.2.3"},
  {"name": "wheel", "version": "0.44.0"},
  {"name": "zipp", "version": "3.19.2"}
]
```

---

## 🏁 Notes
- This list ensures your build can be identically recreated.
- This project is part of the *CUDA Revival* effort led by **careunix**.
- If you fork the main repo, keep this section for scientific reproducibility.

---

# 🧪 Tests & Validation

- Run all tests and capture outputs: `bash scripts/run_all_tests.sh`
- Export environment to JSON + YAML: `bash scripts/export_env.sh`
- System probe (GPU, driver, CUDA, cuDNN → JSON): `python scripts/system_probe.py --out environment/system_specs.json`
- Validate wheel integrity and tags: `python scripts/validate_whl.py`

CI pipelines: link checks, style (black + flake8), CPU fallback tests, JSON schema validation.

---

# ⚡ GPU Stress Test (cuBLAS)

```python
import torch
a = torch.randn(10000, 10000, device="cuda")
b = torch.randn(10000, 10000, device="cuda")
c = torch.mm(a, b)
torch.cuda.synchronize()
print("OK:", c.shape)
```

Confirms:
- GPU kernels  
- Full BLAS acceleration  
- CUDA stability

---

# 🤖 GPT‑2 Medium Inference Example

```python
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch

device = "cuda"
tokenizer = GPT2Tokenizer.from_pretrained("gpt2-medium")
model = GPT2LMHeadModel.from_pretrained("gpt2-medium").to(device)

inp = tokenizer.encode("Once upon a time", return_tensors="pt").to(device)
out = model.generate(inp, max_length=60)
print(tokenizer.decode(out[0]))
```

---

# 🧠 CUDA Libraries Included

### Core GPU Libraries
- libcudart  
- libcublas  
- libcufft  
- libcurand  
- libcusolver  
- libcusparse  
- nvrtc  
- nvToolsExt

### Deep Learning
- cuDNN 7.6.5  
- MKLDNN  
- QNNPACK  
- NNPACK

---

# 🔧 Build Configuration (Verified)

```text
CUDA version: 10.2
cuDNN version: 7.6.5
GPU arch: sm_61
USE_CUDA: ON
USE_CUDNN: ON
USE_MKLDNN: ON
USE_NNPACK: ON
USE_QNNPACK: ON
Build Type: Release
```

---

# 🧩 CPU‑Only Mode

If CUDA is not available:
- The wheel still imports normally  
- PyTorch works fully (CPU backend)  
- No configuration needed

This makes the package safe for **any High Sierra installation**.

---

# 🧬 Project Purpose

- Archive & preserve the last NVIDIA macOS CUDA stack  
- Keep High Sierra CUDA systems alive  
- Support researchers using Pascal GPUs  
- Provide a fully reproducible environment  
- Document a disappeared ecosystem

This repository is **digital preservation**.

---

# 🤝 Contributing

You may contribute via:

- GPU compatibility reports  
- Benchmarks  
- Documentation improvements  
- Mirrors / backups / alternative builds

---

# 📜 License
MIT License

---

Preserving the final NVIDIA‑powered AI stack ever possible on macOS. 
 Pascal lives. care unix;)

PyTorch-HighSierra-CUDA-Revival

PyTorch-HighSierra-CUDA-Revival 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

PyTorch-HighSierra-CUDA-Revival

PyTorch-HighSierra-CUDA-Revival 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README