Skip to content

[MovieGen] update code to support MindSpore 2.7.0 #1362

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Oct 18, 2025
Merged

[MovieGen] update code to support MindSpore 2.7.0 #1362

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
76b0ddc
[MovieGen] update code to support MindSpore 2.7.0
hadipash Oct 13, 2025
04dbdec
fix
hadipash Oct 13, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 13 additions & 13 deletions examples/moviegen/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,9 +63,9 @@ this project!

<div align="center">

| MindSpore | Ascend Driver | Firmware | CANN toolkit/kernel |
|:---------:|:-------------:|:-----------:|:-------------------:|
| 2.5.0 | 24.0.0 | 7.5.0.3.220 | 8.0.0.beta1 |
| MindSpore | Ascend Driver | Firmware | CANN toolkit/kernel |
|:----------------:|:-------------:|:-----------:|:-------------------:|
| 2.5.0 and higher | 24.0.0 | 7.5.0.3.220 | 8.0.0.beta1 |

</div>

Expand Down Expand Up @@ -256,23 +256,23 @@ For more details, check `scripts/tae/encode_tae.sh` or run `scripts/inference_ta

### Performance

Experiments were conducted on Ascend Atlas 800T A2 machines using MindSpore 2.3.1 in Graph mode.
Experiments were conducted on Ascend Atlas 800T A2 machines using MindSpore 2.7.0 in Graph mode.

> [!NOTE]
> We trained all the models using BF16 precision and JIT level `O1`.
> To disable Flash Attention recompute, set `model.not_recompute_fa` to `True`.

| Model | Cards | Stage | Batch size | Resolution | Compile time | Recompute | Gradient Acc | ZeRO | Sequence Parallel | TAE Cache | Time (s/step) | Config |
|:-----:|:-----:|:---------:|:---------------------:|:--------------------:|:------------:|:------------------------:|:------------:|:----:|:-----------------:|:---------:|:-------------:|:--------------------------------------------------------------:|
| 30B | 8 | 1 (T2I) | 10 | 256x455 | 4m 40s | ON | 1 | 3 | No | Yes | 3.37 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 30B | 8 | 2 (T2V) | Video: 1 | 256x256x455 | 7m 40s | ON | 1 | 3 | 8 shards | Yes | 2.58 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 30B | 8 | 3 (T2V) | Video: 1 | 256x576x1024 | 7m 40s | ON | 1 | 3 | 8 shards | Yes | 31.9 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |
| 5B | 8 | 1 (T2I) | 10 | 256px | 2m 30s | OFF | 1 | 3 | No | Yes | 0.64 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 5B | 8 | 2 (T2I/V) | Image: 1<br/>Video: 1 | 256px<br/>256 frames | 2m 40s | OFF | 5 | 2 | No | Yes | 1.54 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 5B | 8 | 3 (T2I/V) | Image: 1<br/>Video: 1 | 768px<br/>256 frames | 4m | ON | 5 | 2 | No | Yes | 82.8 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |
| 1B | 8 | 1 (T2I) | 10 | 256px | 1m 40s | OFF | 1 | No | No | Yes | 0.34 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 1B | 8 | 2 (T2I/V) | Image: 1<br/>Video: 1 | 256px<br/>256 frames | 1m 40s | OFF | 5 | No | No | Yes | 0.55 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 1B | 8 | 3 (T2I/V) | Image: 1<br/>Video: 1 | 768px<br/>256 frames | 2m | ON<br/>(No FA recompute) | 5 | No | No | Yes | 22.1 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |
| 30B | 8 | 1 (T2I) | 10 | 256x455 | 3m 10s | ON | 1 | 3 | No | Yes | 3.62 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 30B | 8 | 2 (T2V) | Video: 1 | 256x256x455 | 4m 10s | ON | 1 | 3 | 8 shards | Yes | 2.57 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 30B | 8 | 3 (T2V) | Video: 1 | 256x576x1024 | 4m 50s | ON | 1 | 3 | 8 shards | Yes | 29.7 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |
| 5B | 8 | 1 (T2I) | 10 | 256px | 2m 30s | OFF | 1 | 3 | No | Yes | 0.65 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 5B | 8 | 2 (T2I/V) | Image: 1<br/>Video: 1 | 256px<br/>256 frames | 1m 30s | ON<br/>(No FA recompute) | 5 | 2 | No | Yes | 1.76 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 5B | 8 | 3 (T2I/V) | Image: 1<br/>Video: 1 | 768px<br/>256 frames | 2m 30s | ON | 5 | 2 | No | Yes | 66.2 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |
| 1B | 8 | 1 (T2I) | 10 | 256px | 1m 15s | OFF | 1 | No | No | Yes | 0.33 | [stage1_t2i_256px.yaml](configs/train/stage1_t2i_256px.yaml) |
| 1B | 8 | 2 (T2I/V) | Image: 1<br/>Video: 1 | 256px<br/>256 frames | 1m 20s | OFF | 5 | No | No | Yes | 0.56 | [stage2_t2iv_256px.yaml](configs/train/stage2_t2iv_256px.yaml) |
| 1B | 8 | 3 (T2I/V) | Image: 1<br/>Video: 1 | 768px<br/>256 frames | 1m 40s | ON<br/>(No FA recompute) | 5 | No | No | Yes | 22.1 | [stage3_t2iv_768px.yaml](configs/train/stage3_t2iv_768px.yaml) |

### Validation During Training

Expand Down
2 changes: 1 addition & 1 deletion examples/moviegen/scripts/inference_tae.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,7 @@ def encode(args, tae: TemporalAutoencoder, save_dir: Path, rank_id: int, device_
mean = np.transpose(mean, (0, 2, 1, 3, 4))
std = np.transpose(std, (0, 2, 1, 3, 4))

for m, s, path in zip(mean, std, samples[1].tolist()):
for m, s, path in zip(mean, std, samples[1].asnumpy().tolist()):
out_path = save_dir / path
out_path.parent.mkdir(parents=True, exist_ok=True)
np.savez(out_path.with_suffix(".npz"), latent_mean=m, latent_std=s)
Expand Down
260 changes: 260 additions & 0 deletions examples/moviegen/scripts/moviegen/train_perf_bench.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,260 @@
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

# log level
export GLOG_v=2

MODE=0

# ---------------------------- 1B ----------------------------

# Stage 1
output_dir=output/test/stage1_t2i_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8200 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage1_t2i_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name llama-1B \
--model.recompute_every_nth_block "" \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents_images \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob 0 \
--dataset.deterministic_sample=True \
--dataloader.batch_size=10 \
--valid.dataset "" \
--train.ema "" \
--train.save.ckpt_save_policy=latest_k \
--train.output_path "$output_dir" \
--train.steps=500

echo "Completed 1B stage 1: $output_dir"
rm -rf "$output_dir"/ckpt

# Stage 2
output_dir=output/test/stage2_t2iv_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8220 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage2_t2iv_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name llama-1B \
--model.recompute_every_nth_block "" \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob=0 \
--dataset.deterministic_sample=True \
--train.ema "" \
--train.settings.gradient_accumulation_steps=5 \
--train.output_path "$output_dir" \
--train.steps=300

echo "Completed 1B stage 2: $output_dir"
rm -rf "$output_dir"/ckpt

# Stage 3
output_dir=output/test/stage3_t2iv_768px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8230 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage3_t2iv_768px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name llama-1B \
--model.not_recompute_fa True \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/high_tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob=0 \
--dataset.deterministic_sample=True \
--train.ema "" \
--train.settings.gradient_accumulation_steps=5 \
--train.output_path "$output_dir" \
--train.steps=30

echo "Completed 1B stage 3: $output_dir"
rm -rf "$output_dir"/ckpt


# ---------------------------- 5B ----------------------------

# Stage 1
output_dir=output/test/stage1_t2i_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8210 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage1_t2i_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--train.settings.zero_stage 3 \
--model.recompute_every_nth_block "" \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents_images \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob 0 \
--dataset.deterministic_sample=True \
--dataloader.batch_size=10 \
--valid.dataset "" \
--train.ema "" \
--train.save.ckpt_save_policy=latest_k \
--train.output_path "$output_dir" \
--train.steps=300

echo "Completed 5B stage 1: $output_dir"
find "$output_dir" -name '*.ckpt' -type f -delete

# Stage 2
output_dir=output/test/stage2_t2iv_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8220 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage2_t2iv_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--train.settings.zero_stage 2 \
--model.not_recompute_fa True \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob=0 \
--dataset.deterministic_sample=True \
--train.ema "" \
--train.settings.gradient_accumulation_steps=5 \
--train.output_path "$output_dir" \
--train.steps=200

echo "Completed 5B stage 2: $output_dir"
rm -rf "$output_dir"/ckpt

# Stage 3
output_dir=output/test/stage3_t2iv_768px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8230 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage3_t2iv_768px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--train.settings.zero_stage 2 \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/high_tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob=0 \
--dataset.deterministic_sample=True \
--train.ema "" \
--train.settings.gradient_accumulation_steps=5 \
--train.output_path "$output_dir" \
--train.steps=10

echo "Completed 5B stage 3: $output_dir"
rm -rf "$output_dir"/ckpt


# ---------------------------- 30B ----------------------------

# Stage 1
output_dir=output/test/stage1_t2i_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8210 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage1_t2i_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name=llama-30B \
--train.settings.zero_stage 3 \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents_images \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob 0 \
--dataset.deterministic_sample=True \
--dataloader.batch_size=10 \
--valid.dataset "" \
--train.optimizer.name adamw_re \
--train.ema "" \
--train.save.ckpt_save_policy=latest_k \
--train.output_path "$output_dir" \
--train.steps=100

echo "Completed 30B stage 1: $output_dir"
find "$output_dir" -name '*.ckpt' -type f -delete

# Stage 2
output_dir=output/test/stage2_t2iv_256px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8220 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage2_t2iv_256px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name=llama-30B \
--train.settings.zero_stage 3 \
--train.sequence_parallel.shards 8 \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob 0 \
--dataset.sample_n_frames 32 \
--dataset.deterministic_sample=True \
--dataloader.batch_size 1 \
--train.optimizer.name adamw_re \
--train.ema "" \
--train.output_path "$output_dir" \
--train.steps=100

echo "Completed 30B stage 2: $output_dir"
find "$output_dir" -name '*.ckpt' -type f -delete

# Stage 3
output_dir=output/test/stage3_t2iv_768px/$(date +"%Y.%m.%d-%H.%M.%S")

msrun --bind_core=True --master_port=8230 --worker_num=8 --local_worker_num=8 --log_dir="$output_dir" --join=True \
python scripts/train.py \
--config configs/train/stage3_t2iv_768px.yaml \
--env.mode $MODE \
--env.max_device_memory 59GB \
--env.distributed True \
--model.name=llama-30B \
--train.settings.zero_stage 3 \
--train.sequence_parallel.shards 8 \
--dataset.csv_path ../../../datasets/mixkit-100videos/video_caption_train_updated.csv \
--dataset.video_folder ../../../datasets/mixkit-100videos/mixkit \
--dataset.tae_latent_folder ../../../datasets/mixkit-100videos/high_tae_latents \
--dataset.text_emb_folder.ul2 ../../../datasets/mixkit-100videos/ul2_emb_300 \
--dataset.text_emb_folder.byt5 ../../../datasets/mixkit-100videos/byt5_emb_100 \
--dataset.text_drop_prob=0 \
--dataset.sample_n_frames 32 \
--dataset.deterministic_sample=True \
--dataloader.batch_size 1 \
--train.optimizer.name adamw_re \
--train.ema "" \
--train.output_path "$output_dir" \
--train.steps=10

echo "Completed 30B stage 3: $output_dir"
find "$output_dir" -name '*.ckpt' -type f -delete