Fix OpenSora 1.2 training

examples/opensora_hpcai/README.md

@@ -159,9 +159,18 @@ Other useful documents and links are listed below.
 ## Installation
 
 1. Install MindSpore according to the [official instructions](https://www.mindspore.cn/install).
-    For Ascend devices, please install **CANN driver C18 (0705)** from [here](https://repo.mindspore.cn/ascend/ascend910/20240705/) and install **MindSpore 2.3** from [here](https://www.mindspore.cn/install).
+    For Ascend devices, please install [CANN8.0.RC2.beta1](https://www.hiascend.com/developer/download/community/result?module=cann&cann=8.0.RC2.beta1) and install [MindSpore 2.3.1](https://www.mindspore.cn/install).
+    > To reduce compilation time and training time, you may install MindSpore2.4-20240904 from [here](https://repo.mindspore.cn/mindspore/mindspore/version/202409/20240904/master_20240904010023_67b5df247045f509c4ca2169bac6a551291a3111_newest/unified/aarch64/)
 
-2. Install requirements
+    You may check your versions by running the following commands. The default installation path of CANN is usually  `/usr/local/Ascend/ascend-toolkit` unless you specify a custom one.
+
+    ```bash
+    cat /usr/local/Ascend/ascend-toolkit/latest/version.cfg  
+
+    python -c "import mindspore;mindspore.set_context(device_target='Ascend');mindspore.run_check()"
+    ```
+
+3. Install requirements
 ```bash
 pip install -r requirements.txt
 ```
@@ -517,9 +526,11 @@ bucket_config:
   "2048": { 1: [ 0.01, 5 ] }
 ```
 
+Knowing that the optimal bucket config can varies from device to device, we have tuned and provided bucket config that are more balanced on Ascend + MindSpore in `configs/opensora-v1-2/train/{stage}_ms.yaml`. You may use them for better training performance.
+
 More details on the bucket configuration can be found in [Multi-resolution Training with Buckets](./docs/quick_start.md#4-multi-resolution-training-with-buckets-opensora-v11-and-above).
 
-Then you can launch the dynamic training task following the previous section. An example running script is `scripts/run/run_train_os1.2_stage2.sh`.
+The instruction for launching the dynamic training task is smilar to the previous section. An example running script is `scripts/run/run_train_os1.2_stage2.sh`.
 
 
 ### Open-Sora 1.1
@@ -622,11 +633,13 @@ Here ✅ means that the data is seen during training, and 🆗 means although no
 
 We evaluate the training performance of Open-Sora v1.2 on the MixKit dataset with high-resolution videos (1080P, duration 12s to 100s). The results are as follows.
 
-| Model       | Context      | jit_level | Precision | BS | NPUs | Size (TxHxW) | Train T. (s/step) |
-|:------------|:-------------|:--------|:---------:|:--:|:----:|:----------------------:|:-----------------:|
-| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3](https://www.mindspore.cn/install) |    O1  |    BF16   |  1 |  8   |       51x720x1280      |        **14.60**       |
-| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3.1(0726)](https://repo.mindspore.cn/mindspore/mindspore/version/202407/20240726/master_20240726220021_4c913fb116c83b9ad28666538483264da8aebe8c_newest/unified/)  |    O1  |    BF16   |  1 |  8   |       Stage 2 Dyn.     |        **33.10**       |
-| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3.1(0726)](https://repo.mindspore.cn/mindspore/mindspore/version/202407/20240726/master_20240726220021_4c913fb116c83b9ad28666538483264da8aebe8c_newest/unified/)  |    O1  |    BF16   |  1 |  8   |       Stage 3 Dyn.     |        **37.7**       |
+| Model       | Context      | jit_level | Precision | BS | NPUs | Size (TxHxW) | Train T. (s/step) |  config |
+|:------------|:-------------|:--------|:---------:|:--:|:----:|:----------------------:|:-----------------:|:-----------------:|
+| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3](https://www.mindspore.cn/install) |    O1  |    BF16   |  1 |  8   |       51x720x1280      |        **14.60**       | [yaml](configs/opensora-v1-2/train/train_720x1280x51.yaml)   |
+| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3.1(0726)](https://repo.mindspore.cn/mindspore/mindspore/version/202407/20240726/master_20240726220021_4c913fb116c83b9ad28666538483264da8aebe8c_newest/unified/)  |    O1  |    BF16   |  1 |  8   |       Stage 2 Dyn.     |        **33.10**       | [yaml](configs/opensora-v1-2/train/train_stage2.yaml)   |
+| STDiT3-XL/2 | D910\*-[C18](https://repo.mindspore.cn/ascend/ascend910/20240705/)-[MS2.3.1(0726)](https://repo.mindspore.cn/mindspore/mindspore/version/202407/20240726/master_20240726220021_4c913fb116c83b9ad28666538483264da8aebe8c_newest/unified/)  |    O1  |    BF16   |  1 |  8   |       Stage 3 Dyn.     |        **34**       | [yaml](configs/opensora-v1-2/train/train_stage3.yaml)   |
+
+
 > Context: {G:GPU, D:Ascend}{chip type}-{CANN version}-{mindspore version}; "Dyn." is short for dynamic shape.
 
 Note that the step time of dynamic training can be influenced by the resolution and duration distribution of the source videos. Training performance is under optimization.

examples/opensora_hpcai/configs/opensora-v1-2/train/train_stage2_ms.yaml

@@ -0,0 +1,84 @@
+# model
+model_version: v1.2
+pretrained_model_path:  PATH_TO_YOUR_MODEL
+model_max_length: 300
+freeze_y_embedder: True
+
+noise_scheduler: rflow
+sample_method: logit-normal
+use_timestep_transform: True
+
+vae_type: OpenSoraVAE_V1_2
+vae_checkpoint: models/OpenSora-VAE-v1.2/model.ckpt
+vae_dtype: bf16
+vae_micro_batch_size: 4
+vae_micro_frame_size: 17  # keep it unchanged for the best results
+
+enable_flash_attention: True
+use_recompute: True
+
+# data
+num_parallel_workers: 2
+num_workers_dataset: 2
+prefetch_size: 2
+max_rowsize: 256
+
+# mindspore params, refer to https://www.mindspore.cn/docs/zh-CN/r2.3.1/api_python/mindspore/mindspore.set_context.html
+max_device_memory: "59GB"
+jit_level: "O1"
+manual_pad: True
+
+# precision
+amp_level: "O2"
+dtype: bf16
+loss_scaler_type: static
+init_loss_scale: 1
+
+# training hyper-params
+scheduler: "constant"
+start_learning_rate: 1.e-4
+end_learning_rate: 1.e-4
+# warmup_steps: 1000
+
+clip_grad: True
+max_grad_norm: 1.0
+use_ema: True
+# ema_decay: 0.99  # default 0.9999 gives better result in our experiments
+
+optim: "adamw_re"
+optim_eps: 1e-15
+weight_decay: 0.
+
+# epochs: 2
+train_steps: 23000
+ckpt_save_steps:  500
+
+mask_ratios:
+  random: 0.005
+  interpolate: 0.002
+  quarter_random: 0.007
+  quarter_head: 0.002
+  quarter_tail: 0.002
+  quarter_head_tail: 0.002
+  image_random: 0.0
+  image_head: 0.22
+  image_tail: 0.005
+  image_head_tail: 0.005
+
+bucket_config:
+  # Structure: "resolution": { num_frames: [ keep_prob, batch_size ] }
+  # Setting [ keep_prob, batch_size ] to [ 0.0, 0 ] forces longer videos into smaller resolution buckets
+  "144p": { 1: [ 1.0, 475 ], 51: [ 1.0, 40 ], 102: [ [ 1.0, 0.33 ], 20 ], 204: [ [ 1.0, 0.1 ], 10 ], 408: [ [ 1.0, 0.1 ], 6 ] }
+  "256": { 1: [ 0.4, 297 ], 51: [ 0.5, 24 ], 102: [ [ 0.5, 0.33 ], 12 ], 204: [ [ 0.5, 1.0 ], 6 ], 408: [ [ 0.5, 1.0 ], 2 ] }
+  "240p": { 1: [ 0.3, 297 ], 51: [ 0.4, 16 ], 102: [ [ 0.4, 0.33 ], 8 ], 204: [ [ 0.4, 1.0 ], 4 ], 408: [ [ 0.4, 1.0 ], 2 ] }
+  "360p": { 1: [ 0.5, 141 ], 51: [ 0.15, 6 ], 102: [ [ 0.3, 0.5 ], 3 ], 204: [ [ 0.3, 1.0 ], 2 ], 408: [ [ 0.5, 0.5 ], 1 ] }
+  "512": { 1: [ 0.4, 141 ], 51: [ 0.15, 6 ], 102: [ [ 0.2, 0.4 ], 3 ], 204: [ [ 0.2, 1.0 ], 1 ], 408: [ [ 0.4, 0.5 ], 1 ] }
+  "480p": { 1: [ 0.5, 89 ], 51: [ 0.2, 5 ], 102: [ 0.2, 2 ], 204: [ 0.1, 1 ] }
+  "720p": { 1: [ 0.1, 36 ], 51: [ 0.03, 1 ] }
+  "1024": { 1: [ 0.1, 36 ], 51: [ 0.02, 1 ] }
+  "1080p": { 1: [ 0.01, 5 ] }
+  "2048": { 1: [ 0.01, 5 ] }
+
+
+# ---------- Validation ----------
+validate: False

examples/opensora_hpcai/configs/opensora-v1-2/train/train_stage3_ms.yaml

@@ -0,0 +1,83 @@
+# model
+model_version: v1.2
+pretrained_model_path: PATH_TO_YOUR_MODEL
+model_max_length: 300
+freeze_y_embedder: True
+
+noise_scheduler: rflow
+sample_method: logit-normal
+use_timestep_transform: True
+
+vae_type: OpenSoraVAE_V1_2
+vae_checkpoint: models/OpenSora-VAE-v1.2/model.ckpt
+vae_dtype: bf16
+vae_micro_batch_size: 4
+vae_micro_frame_size: 17  # keep it unchanged for the best results
+
+enable_flash_attention: True
+use_recompute: True
+
+# data
+num_parallel_workers: 2
+num_workers_dataset: 2
+prefetch_size: 2
+max_rowsize: 256
+
+# precision
+amp_level: "O2"
+dtype: bf16
+loss_scaler_type: static
+init_loss_scale: 1
+
+# mindspore params, refer to https://www.mindspore.cn/docs/zh-CN/r2.3.1/api_python/mindspore/mindspore.set_context.html
+max_device_memory: "59GB"
+jit_level: "O1"
+manual_pad: True
+
+# training hyper-params
+scheduler: "constant"
+start_learning_rate: 1.e-4
+end_learning_rate: 1.e-4
+warmup_steps: 1000
+
+clip_grad: True
+max_grad_norm: 1.0
+use_ema: True
+# ema_decay: 0.99
+
+optim: "adamw_re"
+optim_eps: 1e-15
+weight_decay: 0.
+
+# epochs: 15
+train_steps: 15000
+ckpt_save_steps:  500
+
+mask_ratios:
+  random: 0.01
+  interpolate: 0.002
+  quarter_random: 0.002
+  quarter_head: 0.002
+  quarter_tail: 0.002
+  quarter_head_tail: 0.002
+  image_random: 0.0
+  image_head: 0.22
+  image_tail: 0.005
+  image_head_tail: 0.005
+
+bucket_config:
+  # Structure: "resolution": { num_frames: [ keep_prob, batch_size ] }
+  # Setting [ keep_prob, batch_size ] to [ 0.0, 0 ] forces longer videos into smaller resolution buckets
+  "144p": {1: [1.0, 475], 51: [1.0, 51], 102: [1.0, 27], 204: [1.0, 13], 408: [1.0, 6]}
+  "256": {1: [1.0, 297], 51: [0.5, 20], 102: [0.5, 10], 204: [0.5, 6], 408: [[0.5, 0.5], 2]}
+  "240p": {1: [1.0, 297], 51: [0.5, 20], 102: [0.5, 10], 204: [0.5, 5], 408: [[0.5, 0.4], 2]}
+  "360p": {1: [1.0, 141], 51: [0.5, 8], 102: [0.5, 4], 204: [0.5, 2], 408: [[0.5, 0.3], 1]}
+  "512": {1: [1.0, 141], 51: [0.5, 8], 102: [0.5, 4], 204: [0.5, 2], 408: [[0.5, 0.2], 1]}
+  "480p": {1: [1.0, 89], 51: [0.5, 5], 102: [0.5, 2], 204: [[0.5, 0.5], 1], 408: [0.0, 0]}
+  "720p": {1: [0.3, 36], 51: [0.2, 2], 102: [0.1, 1], 204: [0.0, 0]}
+  "1024": {1: [0.3, 36], 51: [0.1, 2], 102: [0.1, 1], 204: [0.0, 0]}
+  "1080p": {1: [0.1, 5]}
+  "2048": {1: [0.05, 5]}
+
+# ---------- Validation ----------
+validate: False

examples/opensora_hpcai/opensora/datasets/video_dataset_refactored.py

@@ -10,7 +10,6 @@
 
 import cv2
 import numpy as np
-from decord import VideoReader
 from tqdm import tqdm
 
 import mindspore as ms
@@ -252,6 +251,8 @@ def _get_item(self, idx: int) -> Tuple[Any, ...]:
 
         else:
             if self.video_backend == "decord":
+                from decord import VideoReader
+
                 reader = VideoReader(video_path)
                 min_length = self._min_length
                 video_length = len(reader)

examples/opensora_hpcai/opensora/models/layers/blocks.py

@@ -5,7 +5,7 @@
 import numpy as np
 
 import mindspore as ms
-from mindspore import Parameter, Tensor, nn, ops
+from mindspore import Parameter, Tensor, mint, nn, ops
 from mindspore.common.initializer import initializer
 
 from mindone.models.modules.flash_attention import FLASH_IS_AVAILABLE, MSFlashAttention
@@ -26,9 +26,7 @@ def __init__(self, hidden_size, eps=1e-6):
         self.variance_epsilon = eps
 
     def construct(self, hidden_states: Tensor):
-        variance = hidden_states.pow(2).mean(-1, keep_dims=True)
-        hidden_states = hidden_states * ops.rsqrt(variance + self.variance_epsilon)
-        return self.gamma * hidden_states
+        return ops.rms_norm(hidden_states, self.gamma, self.variance_epsilon)[0]
 
 
 class Attention(nn.Cell):
@@ -325,10 +323,12 @@ def __init__(self, normalized_shape, eps=1e-5, elementwise_affine: bool = True,
         else:
             self.gamma = ops.ones(normalized_shape, dtype=dtype)
             self.beta = ops.zeros(normalized_shape, dtype=dtype)
-        self.layer_norm = ops.LayerNorm(-1, -1, epsilon=eps)
 
     def construct(self, x: Tensor):
-        x, _, _ = self.layer_norm(x, self.gamma, self.beta)
+        normalized_shape = x.shape[-1:]
+        # mint layer_norm fuses the operations in layer normorlization and it's faster than ops.LayerNorm
+        x = mint.nn.functional.layer_norm(x, normalized_shape, self.gamma, self.beta, self.eps)
+
         return x
 
 

examples/opensora_hpcai/opensora/models/layers/operation_selector.py

@@ -56,6 +56,8 @@ def get_repeat_interleave_op():
         # provide better performance for static shape in graph mode
         return ops.repeat_interleave
     else:
+        # FIXME: check overflow for v2
+        # return repeat_interleave_ext_v2
         return repeat_interleave_ext
 
 

examples/opensora_hpcai/opensora/models/stdit/stdit3.py

@@ -35,6 +35,8 @@
 
 
 class STDiT3Block(nn.Cell):
+    # to reduce compilation time
+    @ms.lazy_inline(policy="front")
     def __init__(
         self,
         hidden_size,
@@ -100,9 +102,10 @@ def construct(
         )
 
         # modulate (attention)
-        x_m = t2i_modulate(self.norm1(x), shift_msa, scale_msa)
+        norm1 = self.norm1(x)
+        x_m = t2i_modulate(norm1, shift_msa, scale_msa)
         # frames mask branch
-        x_m_zero = t2i_modulate(self.norm1(x), shift_msa_zero, scale_msa_zero)
+        x_m_zero = t2i_modulate(norm1, shift_msa_zero, scale_msa_zero)
         x_m = t_mask_select(frames_mask, x_m, x_m_zero, T, S)
 
         # attention
@@ -128,9 +131,10 @@ def construct(
         x = x + self.cross_attn(x, y, mask)
 
         # modulate (MLP)
-        x_m = t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)
+        norm2 = self.norm2(x)
+        x_m = t2i_modulate(norm2, shift_mlp, scale_mlp)
         # frames mask branch
-        x_m_zero = t2i_modulate(self.norm2(x), shift_mlp_zero, scale_mlp_zero)
+        x_m_zero = t2i_modulate(norm2, shift_mlp_zero, scale_mlp_zero)
         x_m = t_mask_select(frames_mask, x_m, x_m_zero, T, S)
 
         # MLP

examples/opensora_hpcai/opensora/models/vae/modules.py

@@ -2,6 +2,7 @@
 
 import numpy as np
 
+# import mindspore as ms
 from mindspore import nn, ops
 
 _logger = logging.getLogger(__name__)
@@ -177,6 +178,7 @@ def make_attn(in_channels, attn_type="vanilla"):
 
 # used in vae
 class Encoder(nn.Cell):
+    # @ms.lazy_inline()
     def __init__(
         self,
         *,
@@ -274,18 +276,18 @@ def construct(self, x):
         temb = None
 
         # downsampling
-        hs = [self.conv_in(x)]
+        hs = self.conv_in(x)
         for i_level in range(self.num_resolutions):
             for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](hs[-1], temb)
+                h = self.down[i_level].block[i_block](hs, temb)
                 if len(self.down[i_level].attn) > 0:
                     h = self.down[i_level].attn[i_block](h)
-                hs.append(h)
+                hs = h
             if i_level != self.num_resolutions - 1:
-                hs.append(self.down[i_level].downsample(hs[-1]))
+                hs = self.down[i_level].downsample(hs)
 
         # middle
-        h = hs[-1]
+        h = hs
         h = self.mid.block_1(h, temb)
         h = self.mid.attn_1(h)
         h = self.mid.block_2(h, temb)
@@ -297,8 +299,8 @@ def construct(self, x):
         return h
 
 
-# used in vae
 class Decoder(nn.Cell):
+    # @ms.lazy_inline()
     def __init__(
         self,
         *,

examples/opensora_hpcai/opensora/models/vae/vae_temporal.py

@@ -150,6 +150,7 @@ def get_activation_fn(activation):
 class Encoder(nn.Cell):
     """Encoder Blocks."""
 
+    # @ms.lazy_inline()
     def __init__(
         self,
         in_out_channels=4,
@@ -260,6 +261,7 @@ def construct(self, x):
 class Decoder(nn.Cell):
     """Decoder Blocks."""
 
+    # @ms.lazy_inline()
     def __init__(
         self,
         in_out_channels=4,

examples/opensora_hpcai/opensora/schedulers/rectified_flow.py

@@ -72,6 +72,7 @@ def __call__(
                 mask_t = frames_mask * self.num_timesteps
                 x0 = z.copy()
                 x_noise = self.scheduler.add_noise(x0, ops.randn_like(x0), t)
+                # x_noise = self.scheduler.add_noise(x0, ms.Tensor(np.random.randn(*x0.shape), dtype=ms.float32), t)
 
                 model_kwargs["frames_mask"] = mask_t_upper = mask_t >= t.unsqueeze(1)
                 mask_add_noise = (mask_t_upper * (1 - noise_added)).astype(dtype.bool_)