From 8b163d01556784d51593d3d2b6d0a7d50d81b26b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=B4=B9=E6=94=BF=E8=81=AA?= <3300949577@qq.com>
Date: Wed, 17 Jul 2024 02:22:29 -0500
Subject: [PATCH] Update README.md
---
README.md | 15 ++++++++++++++-
1 file changed, 14 insertions(+), 1 deletion(-)
diff --git a/README.md b/README.md
index f5d1e86..200d3d4 100644
--- a/README.md
+++ b/README.md
@@ -1,4 +1,4 @@
-## Scaling Diffusion Transformers with Mixture of Experts
Official PyTorch Implementation
+## Scaling Diffusion Transformers with Mixture of Experts
Official PyTorch Implementation
[](https://arxiv.org/abs/2407.11633)
@@ -32,6 +32,19 @@ torchrun --nnodes=1 --nproc_per_node=N train.py \
--global-batch-size 256
```
+For multiple node training, we solve the [bug](https://github.com/facebookresearch/DiT/blob/main/train.py#L149) at original DiT repository, and you can run with 8 nodes as:
+```bash
+torchrun --nnodes=8 \
+ --node_rank=0 \
+ --nproc_per_node=8 \
+ --master_addr="10.0.0.0" \
+ --master_port=1234 \
+ train.py \
+ --model DiT-B/2 \
+ --global-batch-size 1024 \
+ --data-path /path/to/imagenet/train
+```
+
### 2. Inference