diff --git a/.gitignore b/.gitignore
index 27cac26..40e96d1 100644
--- a/.gitignore
+++ b/.gitignore
@@ -8,10 +8,10 @@ renders
 videos
 edited
 **/__pycache__
-submodules/simple-knn
-submodules/tiny-cuda-nn-fp32
-submodules/diff-surfel-rasterization
-submodules/diff-surfel-rasterization_debug
+submodules/
+*.egg-info/
+uv.lock
+.venv/
 internal/dataparsers/colmap_cluster_dataparser.py
 internal/utils/data_sampler.py
 tools/merge_new_img.py
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..9c7b9b6
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,150 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+CityGaussian is a large-scale scene reconstruction system using 3D/2D Gaussian Splatting. Built on [Gaussian Lightning](https://github.com/yzslab/gaussian-splatting-lightning) with PyTorch Lightning.
+
+## Build & Installation
+
+**Important**: This project requires CUDA-compiled submodules. For CUDA 12.x (GTX 40-series and newer), the submodules need `<cstdint>` includes in header files.
+
+**Always use `uv` for package management. Never use `pip` or `pip install` directly - always use `uv add` or `uv pip install`.**
+
+```bash
+# Create environment
+uv venv --python 3.9
+source .venv/bin/activate
+
+# Install PyTorch matching your CUDA version
+# For CUDA 11.8:
+uv add torch==2.0.1 torchvision torchaudio --index https://download.pytorch.org/whl/cu118
+# For CUDA 12.4:
+uv add torch torchvision torchaudio --index https://download.pytorch.org/whl/cu124
+
+# Install base requirements
+uv add -r requirements.txt
+
+# Install CityGaussian specific packages
+uv add -r requirements/CityGS.txt
+
+# Build CUDA submodules (requires --no-build-isolation for torch access during build)
+uv pip install -e ./submodules/diff-gaussian-rasterization --no-build-isolation
+uv pip install -e ./submodules/diff-surfel-rasterization-city --no-build-isolation
+uv pip install -e ./submodules/diff-trim-gaussian-rasterization --no-build-isolation
+uv pip install -e ./submodules/simple-knn --no-build-isolation
+
+# Fix setuptools version (newer versions removed pkg_resources needed by lightning)
+uv pip install "setuptools<70"
+```
+
+## Common Commands
+
+### Training Pipeline
+```bash
+# 1. Train coarse model
+python main.py fit --config configs/citygsv2_lfls_coarse_sh2.yaml -n experiment_name
+
+# 2. Partition scene and assign data
+python utils/partition_citygs.py --config_path configs/citygsv2_lfls_sh2_trim.yaml --force
+
+# 3. Parallel fine-tuning
+python utils/train_citygs_partitions.py -n experiment_name
+
+# 4. Merge checkpoints
+python utils/merge_citygs_ckpts.py outputs/experiment_name
+```
+
+### Evaluation & Rendering
+```bash
+# Test/evaluate model
+python main.py test --config outputs/experiment_name/config.yaml --save_val --test_speed
+
+# Mesh extraction
+python utils/gs2d_mesh_extraction.py outputs/experiment_name --voxel_size 0.05
+
+# Render video
+python render.py model.pth --camera-path-filename camera.json --output-path output.mp4
+
+# Web viewer
+gs-viewer model.pth --port 8080
+```
+
+### Testing
+```bash
+python -m unittest discover tests -v
+python -m unittest tests.vanilla_gaussian_model_test -v
+```
+
+### Quick Installation Test
+```bash
+# Download test dataset (Tanks & Temples truck scene, ~180MB)
+wget "https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/tandt_db.zip" -O data/tandt.zip
+unzip data/tandt.zip -d data/
+ln -s images data/tandt/truck/images_2
+
+# Train for 3000 steps (~2 min on RTX 4080)
+python main.py fit --data.path data/tandt/truck \
+    --data.parser.class_path Colmap \
+    --data.parser.init_args.down_sample_factor 2 \
+    --data.parser.init_args.down_sample_rounding_mode ceil \
+    --trainer.max_steps 3000 -n test_truck
+
+# View result
+gs-viewer outputs/test_truck/checkpoints/*.ckpt --port 8080
+```
+
+## Architecture
+
+### Entry Points
+- `main.py` → `internal/entrypoints/gspl.py` - Main CLI (fit/val/test/predict)
+- `render.py` - Video rendering with camera paths
+- `viewer.py` - Web-based model viewer
+
+### Core Modules (internal/)
+- `gaussian_splatting.py` - Main LightningModule for training
+- `models/` - Gaussian model variants (vanilla, 2D, mip, deformable, appearance)
+- `renderers/` - Pluggable rendering backends (vanilla, gsplat, partition_lod)
+- `dataset.py` - DataModule with async caching for large datasets
+- `dataparsers/` - Dataset loaders (Colmap, Blender, NSVF, Nerfies)
+- `density_controllers/` - Gaussian densification strategies
+- `callbacks.py` - Training callbacks (SaveGaussian, etc.)
+
+### Submodules
+- `diff-gaussian-rasterization` → `diff_gaussian_rasterization` - Standard 3DGS rasterizer
+- `diff-surfel-rasterization-city` → `diff_trim_surfel_rasterization` - Modified surfel rasterizer for street views
+- `diff-trim-gaussian-rasterization` → `diff_trim_gaussian_rasterization` - Trimmed Gaussian variant
+- `simple-knn` → `simple_knn` - KNN implementation
+
+## Configuration
+
+YAML configs in `/configs/` define model, data, renderer, and optimizer settings. Override via CLI:
+```bash
+python main.py fit --config configs/base.yaml --trainer.max_steps 30000 --data.parser.image_scale 2
+```
+
+Output structure: `outputs/<experiment_name>/config.yaml`, `checkpoints/`, `cameras.json`
+
+## Key Patterns
+
+- **Lightning-based**: Training uses LightningModule/DataModule pattern
+- **Modular renderers**: Swap renderers via config (vanilla, gsplat, feature_3dgs)
+- **Scene partitioning**: Large scenes split into blocks for parallel training
+- **Caching**: Async image caching to manage memory on large datasets
+
+## Allowed Commands
+
+The following commands are pre-approved and can be run without confirmation:
+
+- `python -c "import ..."` - Python import tests
+- `python main.py --help` - CLI help
+- `python main.py fit/validate/test/predict ...` - Training and evaluation
+- `python -m unittest ...` - Running tests
+- `uv add ...` - Installing packages
+- `uv pip install ... --no-build-isolation` - Building CUDA extensions
+- `gs-viewer ...` - Web viewer
+- `wget ...` - Download files
+- `curl ...` - Download files
+- `gdown ...` - Download from Google Drive
+- `unzip ...` - Extract archives
diff --git a/doc/installation.md b/doc/installation.md
index dbcaedb..b563f6a 100644
--- a/doc/installation.md
+++ b/doc/installation.md
@@ -9,6 +9,19 @@ cd CityGaussian
 
 ### B. Create virtual environment
 
+#### Option 1: Using uv (Recommended)
+
+```bash
+# install uv if not already installed
+curl -LsSf https://astral.sh/uv/install.sh | sh
+
+# create virtual environment
+uv venv --python 3.9
+source .venv/bin/activate
+```
+
+#### Option 2: Using conda
+
 ```bash
 # create virtual environment
 conda create -yn gspl python=3.9 pip
@@ -16,24 +29,79 @@ conda activate gspl
 ```
 
 ### C. Install PyTorch
-* Tested on `PyTorch==2.0.1`
-* You must install the one match to the version of your nvcc (nvcc --version)
-* For CUDA 11.8
+* Tested on `PyTorch==2.0.1` and `PyTorch==2.5.1`
+* You must install the one matching your CUDA version (check with `nvcc --version`)
+
+#### For CUDA 11.8
 
-  ```bash
-  pip install -r requirements/pyt201_cu118.txt
-  ```
+```bash
+# Using uv
+uv pip install torch==2.0.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
+
+# Using conda/pip
+pip install -r requirements/pyt201_cu118.txt
+```
+
+#### For CUDA 12.4+ (RTX 40-series and newer)
+
+```bash
+# Using uv
+uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
+
+# Using conda/pip
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
+```
 
 ### D. Install requirements
 
 ```bash
+# Using uv
+uv pip install -r requirements.txt
+
+# Using conda/pip
 pip install -r requirements.txt
 ```
+
 For pose and 3DGS joint optimization, please directly jump to this [document](../doc/vggt_x.md).
 
 ### E. Install additional package for CityGaussian
 
 ```bash
+# Using uv
+uv pip install -r requirements/CityGS.txt
+
+# Using conda/pip
 pip install -r requirements/CityGS.txt
 ```
-Note that here we use modified version of Trim2DGS rasterizer, so as to resolve [impulse noise problem](https://github.com/hbb1/2d-gaussian-splatting/issues/174) under street views. This version also avoids interference from out-of-view surfels.
\ No newline at end of file
+
+Note that here we use modified version of Trim2DGS rasterizer, so as to resolve [impulse noise problem](https://github.com/hbb1/2d-gaussian-splatting/issues/174) under street views. This version also avoids interference from out-of-view surfels.
+
+### F. Build CUDA submodules
+
+For CUDA 12.x, you need to build submodules with `--no-build-isolation` to allow torch access during compilation:
+
+```bash
+# Using uv (recommended for CUDA 12.x)
+uv pip install -e ./submodules/diff-gaussian-rasterization --no-build-isolation
+uv pip install -e ./submodules/diff-surfel-rasterization-city --no-build-isolation
+uv pip install -e ./submodules/diff-trim-gaussian-rasterization --no-build-isolation
+uv pip install -e ./submodules/simple-knn --no-build-isolation
+
+# Using conda/pip
+pip install -e ./submodules/diff-gaussian-rasterization
+pip install -e ./submodules/diff-surfel-rasterization-city
+pip install -e ./submodules/diff-trim-gaussian-rasterization
+pip install -e ./submodules/simple-knn
+```
+
+### G. Fix setuptools (if using PyTorch Lightning)
+
+Newer setuptools versions (v70+) removed `pkg_resources` which PyTorch Lightning requires:
+
+```bash
+# Using uv
+uv pip install "setuptools>=61.0,<70"
+
+# Using conda/pip
+pip install "setuptools>=61.0,<70"
+```
diff --git a/pyproject.toml b/pyproject.toml
index 4999dad..9f71605 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -9,6 +9,9 @@ include = ["internal*", "utils*"]
 name = "gaussian-splatting-lightning"
 dynamic = ["version"]
 requires-python = ">=3.8"
+dependencies = [
+    "setuptools>=61.0,<70",
+]
 
 [project.scripts]
 gs-fit = "internal.entrypoints.gspl:cli_fit"