【cherry-pick】add hunyuanvideo readme & fix bug in attention_processor.py (#1232) (#1274)

Author: kamiLight

ppdiffusers/examples/HunyuanVideo/README.md

@@ -0,0 +1,24 @@
+# HunyuanVideo视频生成
+
+HunyuanVideo是由腾讯开发的13B参数量的开源视频生成模型，能够生成高质量，高动态的视频。本仓库提供了HunyuanVideo的paddle实现，目前仅支持推理。
+
+## 快速开始
+### 环境准备
+若曾使用PaddlePaddle主页build_paddle_env.sh脚本安装PaddlePaddle，请根据本身cuda版本手动更新版本[Installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html)。
+
+更新diffusers：在ppdiffusers目录下运行以下命令:
+```shell
+python install -e .
+```
+
+## 推理示例
+
+### 硬件要求
+* 硬件要求：至少要求50G以上显存
+
+### 文本到视频
+
+运行如下脚本，生成视频
+```shell
+python ppdiffusers/examples/inference/text_to_video_generation-hunyuan_video.py
+```

ppdiffusers/examples/inference/text_to_video_generation-hunyuan_video.py

@@ -29,7 +29,10 @@
     transformer = transformer,
     text_encoder = text_encoder,
     tokenizer = tokenizer,
-    paddle_dtype=paddle.float16)
+    paddle_dtype=paddle.float16,
+    map_location="cpu")
+pipe.vae.enable_tiling()
+pipe.vae.enable_slicing()
 prompt = 'A cat walks on the grass, realistic.'
 output = pipe(
     prompt=prompt,

ppdiffusers/ppdiffusers/models/transformer_hunyuan_video.py

@@ -139,9 +139,6 @@ def __call__(
         if attention_mask.dtype == paddle.bool:
             L, S = query.shape[-2], key.shape[-2]
             attention_mask_tmp = paddle.zeros([1,1,L, S], dtype=query.dtype)
-            # attention_mask_tmp = paddle.zeros_like(attention_mask).to(query.dtype)
-            # attention_mask_tmp[attention_mask==True] = 0
-            # attention_mask_tmp[attention_mask==False] = float("-inf")
             attention_mask_tmp = attention_mask_tmp.masked_fill(attention_mask.logical_not(), float("-inf"))
             attention_mask = attention_mask_tmp
 
@@ -153,17 +150,6 @@ def __call__(
             dropout_p=0.0,
             is_causal=False,
         ).transpose([0,2,1,3])
-        # hidden_states = query
-        # import pdb;pdb.set_trace()
-
-        # hidden_states = scaled_dot_product_attention_paddle(
-        #     query=query,
-        #     key=key,
-        #     value=value,
-        #     attn_mask=attention_mask,
-        #     dropout_p=0.0,
-        #     is_causal=False,
-        # )
 
         hidden_states = hidden_states.transpose(
             perm=dim2perm(hidden_states.ndim, 1, 2)