diff --git a/log.md b/log.md
new file mode 100644
index 0000000..b02bd8d
--- /dev/null
+++ b/log.md
@@ -0,0 +1,178 @@
+# How2comm Inference Log
+
+**Date:** 2026-03-14
+**Goal:** Run pretrained model inference on V2XSet and OPV2V datasets.
+
+---
+
+## Environment
+
+- OS: Linux (Ubuntu), CUDA 12.8 on system
+- Conda environment: `How2comm` at `/data/common/conda_envs_shared/yitonglin/How2comm`
+- Python: 3.9
+- PyTorch: 2.5.1+cu124
+- spconv: spconv-cu124
+
+---
+
+## Prerequisites
+
+### 1. Pretrained Models
+
+Download pretrained models from Google Drive or Baidu Disk (links in README.md).
+
+Place files as follows:
+```
+v2xvit/logs/how2comm/
+    config.yaml
+    net_epoch36.pth          # OPV2V pretrained model
+
+v2xvit/logs/how2comm_v2xset/
+    config.yaml
+    net_epoch32.pth          # V2XSet pretrained model
+```
+
+### 2. Datasets
+
+Place datasets at the repo root:
+```
+opv2v/
+    train/
+    validate/
+    test/
+
+v2xset/
+    train/
+    validate/
+    test/
+```
+
+The `validate_dir` in the config files should point to the `test/` split:
+- `v2xvit/logs/how2comm/config.yaml` → `validate_dir: /data/opv2v/test`
+- `v2xvit/logs/how2comm_v2xset/config.yaml` → `validate_dir: /home/yitonglin/OpenCOOD/centerformer/How2comm/v2xset/test`
+
+---
+
+## Setup Steps
+
+### Step 1: Install PyTorch with CUDA support （PyTorch 是深度学习框架：张量计算，自动求导，神经网络搭建，模型训练，模型推理）
+
+The environment must have a CUDA-enabled PyTorch. CPU-only PyTorch will fail.
+
+```sh
+pip install --force-reinstall torch==2.5.1+cu124 torchvision==0.20.1+cu124 \
+    --index-url https://download.pytorch.org/whl/cu124
+```
+
+Verify:
+```sh
+python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
+# Expected: 2.5.1+cu124 12.4 True
+```
+
+### Step 2: Install spconv （spconv 是 3D 点云任务常见的稀疏卷积库，主要用于 3D 点云、体素化 voxel、BEV 检测这类任务，cumm是spconv 底层依赖的 CUDA/C++ 内核生成和算子支持库。如果import 的是 spconv，但运行时如果缺底层依赖，就会报：ModuleNotFoundError: No module named 'cumm'）
+
+```sh
+pip install spconv-cu124
+```
+
+### Step 3: Install other dependencies （icecream: 调试打印工具; open3d: 点云处理/可视化相关依赖)
+
+```sh
+pip install icecream
+pip install open3d
+```
+
+### Step 4: Install the v2xvit package
+
+From the repo root:
+```sh
+pip install -e . --config-settings editable_mode=compat
+```
+
+### Step 5: Build MultiScaleDeformableAttention (MSDA) extension （Attention：注意力机制
+Deformable：可变形/稀疏采样式注意力； MultiScale：多尺度，该项目中用于特征融合，BEV/多车特征交互，backbone/neck/head 中某个检测模块）
+
+This requires CUDA-enabled PyTorch (Step 1 must be done first).
+
+```sh
+python /path/to/centerformer/det3d/models/ops/setup.py build install
+```
+
+> **Note:** Replace `/path/to/centerformer` with the actual path to the centerformer repo on your machine. In our setup it is `/home/yitonglin/OpenCOOD/centerformer`.
+
+### Step 6: Build box_overlaps Cython extension （两个框的重叠程度）
+
+```sh
+python v2xvit/utils/setup.py build_ext --inplace
+```
+
+### Step 7: Fix NumPy 2.0 incompatibility
+
+`np.Inf` was removed in NumPy 2.0. Patch `v2xvit/utils/pcd_utils.py` line 189:
+
+```python
+# Change:
+minimum = np.Inf
+# To:
+minimum = np.inf
+```
+
+This has already been applied in the current codebase.
+
+---
+
+## Running Inference
+
+Because `libc10.so` (PyTorch's C++ library) is not in the default library path, you must set `LD_LIBRARY_PATH` and `PYTHONPATH` when running inference.
+
+Use the following command template:
+
+```sh
+conda run -p /data/common/conda_envs_shared/yitonglin/How2comm bash -c \
+  "export LD_LIBRARY_PATH=/data/common/conda_envs_shared/yitonglin/How2comm/lib/python3.9/site-packages/torch/lib:\$LD_LIBRARY_PATH && \
+   PYTHONPATH=/home/yitonglin/OpenCOOD/centerformer/How2comm:\$PYTHONPATH \
+   python v2xvit/tools/inference.py --model_dir <MODEL_DIR> --eval_epoch <EPOCH>"
+```
+
+### V2XSet
+
+```sh
+conda run -p /data/common/conda_envs_shared/yitonglin/How2comm bash -c \
+  "export LD_LIBRARY_PATH=/data/common/conda_envs_shared/yitonglin/How2comm/lib/python3.9/site-packages/torch/lib:\$LD_LIBRARY_PATH && \
+   PYTHONPATH=/home/yitonglin/OpenCOOD/centerformer/How2comm:\$PYTHONPATH \
+   python v2xvit/tools/inference.py --model_dir v2xvit/logs/how2comm_v2xset --eval_epoch 32"
+```
+
+**Result:**
+```
+Epoch: 32 | AP @0.3: 0.8412 | AP @0.5: 0.8178 | AP @0.7: 0.6494 | comm_rate: 0.139198
+```
+
+### OPV2V
+
+```sh
+conda run -p /data/common/conda_envs_shared/yitonglin/How2comm bash -c \
+  "export LD_LIBRARY_PATH=/data/common/conda_envs_shared/yitonglin/How2comm/lib/python3.9/site-packages/torch/lib:\$LD_LIBRARY_PATH && \
+   PYTHONPATH=/home/yitonglin/OpenCOOD/centerformer/How2comm:\$PYTHONPATH \
+   python v2xvit/tools/inference.py --model_dir v2xvit/logs/how2comm --eval_epoch 36"
+```
+
+**Result:**
+```
+Epoch: 36 | AP @0.3: 0.8735 | AP @0.5: 0.8573 | AP @0.7: 0.7138 | comm_rate: 0.142713
+```
+
+---
+
+## Common Errors & Fixes
+
+| Error | Cause | Fix |
+|-------|-------|-----|
+| `ModuleNotFoundError: No module named 'torch'` | Wrong conda env active | Use `conda run -p ...` as shown above |
+| `ModuleNotFoundError: No module named 'cumm'` | spconv not installed | `pip install spconv-cu124` |
+| `ModuleNotFoundError: No module named 'MultiScaleDeformableAttention'` | MSDA extension not built | Run Step 5 above |
+| `NotImplementedError: Cuda is not availabel` | PyTorch installed without CUDA | Run Step 1 to reinstall with CUDA |
+| `AttributeError: np.Inf was removed in NumPy 2.0` | NumPy 2.0 incompatibility | Apply Step 7 patch |
+| `ImportError: libc10.so: cannot open shared object file` | torch lib not in `LD_LIBRARY_PATH` | Use the full command with `LD_LIBRARY_PATH` as shown above |
+| `ModuleNotFoundError: No module named 'v2xvit.models.comm_modules'` | Stale site-packages install | Run Step 4 with `PYTHONPATH` set as shown above |