SETUP_GUIDE.md

# `yw_div` conda environment — H100 / CUDA 12.6 setup guide

This guide reproduces the conda environment **`yw_div`** on a target server.

- **Source machine:** Linux x86_64, NVIDIA RTX A6000, CUDA 12.6, Python 3.11.12
- **Target machine:** Linux x86_64, NVIDIA H100, CUDA 12.6 (driver ≥ 555/560)
- **Core stack:** Python 3.11.12, PyTorch 2.7.0+cu126, torchvision 0.22.0+cu126, torchaudio 2.7.0+cu126, Triton 3.3.0, NumPy 2.2.5

The PyTorch wheels for `cu126` ship their own CUDA 12.6 runtime, cuDNN 9.5, NCCL 2.26, etc. — you do **not** need a system-wide CUDA toolkit, only an NVIDIA driver that supports CUDA 12.6.

---

## 0. Prerequisites on the target server

```bash
# Verify the driver supports CUDA 12.6 (driver >= 555). H100 should show CC 9.0.
nvidia-smi
# Expect: "CUDA Version: 12.6" (or higher) in the top-right of the table.

# Conda (Miniconda/Anaconda/Mambaforge) must be installed.
conda --version    # any modern conda (>= 23.x) is fine
```

If conda is not installed:

```bash
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /tmp/miniconda.sh
bash /tmp/miniconda.sh -b -p "$HOME/miniconda3"
source "$HOME/miniconda3/etc/profile.d/conda.sh"
conda init bash   # then restart your shell
```

---

## 1. Files in this bundle

| File | Purpose |
|---|---|
| `SETUP_GUIDE.md` | This guide. |
| `setup_yw_div.sh` | One-shot install script (recommended). |
| `requirements_torch.txt` | Pinned PyTorch wheels (must use the cu126 index). |
| `requirements_pip.txt` | All other Python packages (PyPI). |
| `environment_full.yml` | Full conda export from the source machine (reference only). |
| `environment_from_history.yml` | Minimal conda spec (reference only). |
| `requirements_full.txt` | Raw `pip freeze` from the source machine (reference only). |

Copy the entire `yw_div_env/` directory to the target server (e.g. via `scp -r`).

---

## 2. Quick install (recommended)

```bash
cd /path/to/yw_div_env
bash setup_yw_div.sh
```

The script performs:

1. `conda create -n yw_div python=3.11.12 pip` (channel: `conda-forge`)
2. Pin baseline `pip / setuptools / wheel`
3. Install PyTorch stack from `https://download.pytorch.org/whl/cu126`
4. Install the rest of the packages from `requirements_pip.txt`
5. Print a sanity check (torch version, CUDA visible, device names)

To use a different env name: `ENV_NAME=my_env bash setup_yw_div.sh`.

Activate after install:

```bash
conda activate yw_div
```

---

## 3. Manual install (step by step)

If you'd rather run the steps yourself:

```bash
# 1. Create the env
conda create -y -n yw_div -c conda-forge python=3.11.12 pip
conda activate yw_div

# 2. Pin pip toolchain
pip install --upgrade pip==25.0.1 setuptools==75.8.2 wheel==0.45.1

# 3. PyTorch + torchvision + torchaudio (CUDA 12.6 wheels)
pip install --index-url https://download.pytorch.org/whl/cu126 \
    torch==2.7.0+cu126 \
    torchvision==0.22.0+cu126 \
    torchaudio==2.7.0+cu126

# 4. Everything else
pip install -r requirements_pip.txt
```

---

## 4. Verification

```bash
conda activate yw_div
python - <<'PY'
import torch, torchvision, torchaudio, triton, numpy
print("torch       :", torch.__version__)
print("torchvision :", torchvision.__version__)
print("torchaudio  :", torchaudio.__version__)
print("triton      :", triton.__version__)
print("numpy       :", numpy.__version__)
print("cuda build  :", torch.version.cuda)
print("cudnn       :", torch.backends.cudnn.version())
print("nccl        :", torch.cuda.nccl.version())
print("cuda avail  :", torch.cuda.is_available())
for i in range(torch.cuda.device_count()):
    name = torch.cuda.get_device_name(i)
    cc   = torch.cuda.get_device_capability(i)
    print(f"  gpu[{i}]  {name}  cc={cc}")
# Tiny H100 sanity op
x = torch.randn(4096, 4096, device="cuda", dtype=torch.bfloat16)
y = x @ x
torch.cuda.synchronize()
print("matmul ok   :", y.shape, y.dtype, y.device)
PY
```

Expected on an H100 node:

- `cuda build  : 12.6`
- `cuda avail  : True`
- `gpu[0]  NVIDIA H100 ...  cc=(9, 0)`

---

## 5. Notes & troubleshooting

- **`+cu126` is required.** Do not `pip install torch` without the index URL — that pulls a CPU build and silently breaks GPU code.
- **Driver too old.** If `nvidia-smi` reports `CUDA Version: 12.5` or lower, ask the admin to upgrade the driver to one that supports 12.6 (≥ 555.x). The toolkit on disk doesn't matter; the driver does.
- **H100 + bf16 / FP8.** `torch==2.7.0+cu126` already supports `torch.bfloat16` and the H100 transformer engine paths via cuDNN 9.5. No extra steps required.
- **flash-attn / xformers.** Not in this snapshot. If your project needs them, install separately, e.g. `pip install flash-attn --no-build-isolation` (requires `nvcc`/CUDA dev toolkit) or use a prebuilt wheel matching torch 2.7 + cu126.
- **Conda channel pollution.** The base env on the source machine includes some conda-only packages (`conda-build`, `libmambapy`, `lief`, …) that are not installed in `yw_div`. They are conda-runtime utilities, not project dependencies.
- **Re-running the script** is safe to re-attempt installs but will fail at step 1 if the env already exists. Either remove it (`conda env remove -n yw_div`) or pass a different `ENV_NAME=`.
- **HF cache / wandb.** `huggingface_hub`, `wandb`, `tensorboard` are pre-installed; configure tokens (`huggingface-cli login`, `wandb login`) before first run.

---

## 6. Removing the env

```bash
conda env remove -n yw_div
```