diff --git a/.github/pull_request_template.md b/.github/pull_request_template.md
index 69e0c00e..97639b23 100644
--- a/.github/pull_request_template.md
+++ b/.github/pull_request_template.md
@@ -11,18 +11,11 @@ bash setup_conda_env.sh
 conda activate nvsubquadratic
 ```
 
-## Pre-commit
-
-Run locally before pushing:
-
-```bash
-pre-commit install
-pre-commit run --all-files
-```
-
 ## Test plan
 
-<!-- How did you test this? Bulleted checklist preferred. -->
+- [ ] `pre-commit run --all-files` passes (`pre-commit install` if not yet set up).
+- [ ] Existing tests pass (`pytest tests/`).
+- [ ] New tests added, or explain why not needed:
 
 ## Documentation checklist
 
diff --git a/README.md b/README.md
index 4ae37ceb..1ba6d20a 100644
--- a/README.md
+++ b/README.md
@@ -37,6 +37,20 @@ nvSubquadratic provides a **high-level PyTorch interface** that depends on the *
 
 ## Installation
 
+### PyPI
+
+```bash
+pip install nvsubquadratic
+```
+
+Requires Python 3.11+ and a CUDA 12-compatible GPU. Optional extras:
+
+```bash
+pip install "nvsubquadratic[quack]"   # fused RMSNorm kernel (Hopper/Blackwell only)
+pip install "nvsubquadratic[dali]"    # NVIDIA DALI for ImageNet/Well data pipelines (~400 MB)
+pip install "nvsubquadratic[all]"         # all extras
+```
+
 ### Package Manager
 
 This project uses **pip** with `pyproject.toml` for dependency management. A `Pipfile.lock` is maintained for nSpect security scanning compliance.
diff --git a/docs/getting_started.md b/docs/getting_started.md
index 1c82bb3d..2bc51596 100644
--- a/docs/getting_started.md
+++ b/docs/getting_started.md
@@ -16,7 +16,16 @@ pure-PyTorch path automatically.
 
 ## Install
 
-The recommended developer setup is conda:
+For users who just want to use the library:
+
+```bash
+pip install nvsubquadratic                   # core library
+pip install "nvsubquadratic[dali]"           # + NVIDIA DALI for ImageNet/Well examples
+pip install "nvsubquadratic[quack]"          # + fused RMSNorm kernel (Hopper/Blackwell only)
+pip install "nvsubquadratic[all]"            # all extras
+```
+
+For contributors or developer setup, the recommended path is conda:
 
 ```bash
 bash setup_conda_env.sh
diff --git a/docs/index.rst b/docs/index.rst
index c0026864..29237985 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -81,13 +81,19 @@ package tour.
 Installation
 ------------
 
-The package is installed from source:
+.. code-block:: bash
+
+    pip install nvsubquadratic
+
+Optional extras:
 
 .. code-block:: bash
 
-    pip install -e .
+    pip install "nvsubquadratic[quack]"       # fused RMSNorm kernel (Hopper/Blackwell only)
+    pip install "nvsubquadratic[dali]"        # NVIDIA DALI for ImageNet/Well data pipelines
+    pip install "nvsubquadratic[all]"         # all extras
 
-To enable the optional fused RMSNorm kernel on Hopper / Blackwell GPUs:
+For development (editable install from source):
 
 .. code-block:: bash
 
diff --git a/pyproject.toml b/pyproject.toml
index befa3cb6..52113d99 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -58,7 +58,6 @@ dependencies = [
     "huggingface_hub",
     "subquadratic-ops-torch-cu12>=0.2.0",
     "timm==1.0.24",
-    "nvidia-dali-cuda120==1.53.0",
 ]
 
 [project.urls]
@@ -71,6 +70,10 @@ Issues = "https://github.com/NVIDIA-BioNeMo/nvSubquadratic/issues"
 # Optional: use QuACK fused RMSNorm kernel (Hopper/Blackwell only: H100, B200, B300).
 # On Ampere and older GPUs the kernel is not available; RMSNorm uses PyTorch fallback.
 quack = ["quack-kernels"]
+# Optional: NVIDIA DALI data-loading pipeline, used by ImageNet and The Well examples.
+# Not required for the core library; ~400 MB download.
+dali = ["nvidia-dali-cuda120==1.53.0"]
+all = ["quack-kernels", "nvidia-dali-cuda120==1.53.0"]
 
 [tool.setuptools.dynamic]
 version = { file = "VERSION" }