Mindspeed Mm Generative

MindSpeed-MM Generative Model Training

This Skill guides users through the end-to-end training pipeline for multimodal generative models (text-to-video, text-to-image) on Huawei Ascend NPU.

Prerequisites

Critical: Do NOT use bash scripts/install.sh for generative models. Official MindSpeed-MM docs state install.sh only fully supports Qwen3/Qwen3.5. For Wan2.1 and other generative models, follow the manual install flow below (matches examples/wan2.1/README.md).

Step P1: Clone repositories

git clone https://gitcode.com/Ascend/MindSpeed-MM.git /root/workspace/MindSpeed-MM
git clone https://github.com/NVIDIA/Megatron-LM.git /root/workspace/Megatron-LM
cd /root/workspace/Megatron-LM && git checkout core_v0.12.1
cp -r megatron /root/workspace/MindSpeed-MM/
cd /root/workspace/MindSpeed-MM

Step P2: Install PyTorch + torch_npu

Always use pre-built wheels from Ascend PyTorch Releases. Do not use pip install torch==2.7.1 — aarch64 wheels on PyPI are unreliable.

# Download matching wheels from https://gitcode.com/Ascend/pytorch/releases
# Replace cp310 with cp311 if using Python 3.11
pip install torch-2.7.1-cp310-cp310-manylinux_2_28_aarch64.whl
pip install torch_npu-2.7.1rc1-cp310-cp310-manylinux_2_28_aarch64.whl
# For x86_64: same naming pattern with _x86_64 suffix

# Required by torch_npu and other components:
pip install numpy pyyaml scipy attrs decorator psutil

Step P3: Install MindSpeed

git clone https://gitcode.com/Ascend/MindSpeed.git /root/workspace/MindSpeed
cd /root/workspace/MindSpeed
git checkout 93c45456c7044bacddebc5072316c01006c938f9
pip install -r requirements.txt
pip install -e .
cd /root/workspace/MindSpeed-MM

Step P4: Install MindSpeed-MM base dependencies

# Installs base stack from pyproject.toml: transformers==4.57.0, diffusers==0.30.3, peft==0.7.1, etc.
pip install -e .

Step P5: Activate CANN environment

source /usr/local/Ascend/cann/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh

Step P6: Smoke test

python3 -c "
import torch, torch_npu
assert torch.npu.is_available(), 'NPU not available'
print(f'NPU count: {torch_npu.npu.device_count()}')
import transformers, diffusers, mindspeed
print(f'transformers={transformers.__version__}, diffusers={diffusers.__version__}')
"

Then proceed to Step 0 below for model-specific dependencies (diffusers upgrade, decord, etc.) — Prerequisites alone is NOT enough for any generative model.

For detailed troubleshooting and alternative install paths, see mindspeed-mm-env-setup.

Supported Models

Sub-task descriptions:

• t2v (text-to-video): Generate video from text, the most fundamental generation task
• i2v (image-to-video): Generate video from image + text, requires a reference image and different model weights
• v2v (video-to-video): Video style transfer
• flf2v (first-last-frame-to-video): Generate intermediate video from the first and last frames

How to Train Any Generative Model

For ANY generative model, use the Model Registry to look up the exact backend, entry script, converter, and feature extraction script. Then:

1. Check examples/<model_name>/README.md for model-specific instructions
2. Read the shell script to identify: entry script, config files, backend pattern
3. Check if the model needs feature extraction (not all do — see registry)
4. Check if the model needs weight conversion (diffusers models use HF weights directly)
5. Modify config files with your data/weight paths
6. Launch the shell script

Backend Patterns

MindSpeed-MM uses four different training backends depending on the model:

Feature Extraction Matrix

Not all models require feature extraction. Models that train on raw data skip this step entirely.

Weight Converter Reference

Key Difference from VLM

The generative model training pipeline includes an additional feature extraction step (for models that need it):

Data Preparation → Feature Extraction (VAE + TextEncoder) → Training → Inference

VLMs (Vision-Language Models) train directly from raw image/video inputs, whereas generative models first use a VAE to encode video into latent space features and a text encoder to encode text into vectors, then train a diffusion model in the feature space. Some newer models (wan2.2, ltx2, diffusers models) skip feature extraction and train on raw data directly.

Quick Start: Wan2.1-1.3B T2V (End-to-End, Flagship Example)

Using Wan2.1-1.3B text-to-video as the flagship example (Megatron backend with feature extraction). This section walks through the complete training pipeline. To adapt for other models, refer to the "How to Train Any Generative Model" section above and the "Adaptation Notes" section below.

Step 0: Install Model-Specific Dependencies

MANDATORY: Wan2.1 requires diffusers==0.33.1. The base environment ships diffusers==0.30.3, which does NOT contain AutoencoderKLWan and will fail at feature extraction. You must upgrade before proceeding.

# Wan2.1 — upgrade diffusers (MANDATORY)
pip install diffusers==0.33.1

# IMPORTANT: After upgrading diffusers, verify core dependencies haven't been broken
pip install numpy==1.26.0 pandas==2.0.3  # Pin back if upgraded
python3 -c "import numpy, pandas, sklearn; print('Dependencies OK')"

# Video decoder (DecordVideo is the official default for Wan2.1)
# x86 architecture:
pip install decord==0.6.0
# ARM (aarch64) — build decord from source (apt-get may need retries due to network):
apt-get update || (sleep 5 && apt-get update)  # retry if first attempt times out
apt-get install -y cmake build-essential \
    libavcodec-dev libavformat-dev libavutil-dev libswscale-dev libavfilter-dev
git clone --recursive https://github.com/dmlc/decord.git /tmp/decord
cd /tmp/decord && mkdir build && cd build
cmake .. -DUSE_CUDA=0 -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)
cd /tmp/decord/python && pip install .
# Note: libavfilter-dev is required — without it cmake fails with "FFMPEG_LIBAVFILTER-NOTFOUND"
# Verify: python3 -c "import decord; print(decord.__version__)"

Different models have significantly different dependency versions. See references/per-model-deps.md for details. It is strongly recommended to create a separate container for each model to avoid dependency conflicts.

Step 1: Weight Download and Conversion

Download the Wan-AI/Wan2.1-T2V-1.3B-Diffusers weights from HuggingFace.

Weight conversion (HuggingFace format -> MindSpeed-MM format):

mm-convert WanConverter hf_to_mm \
  --cfg.source_path ./weights/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/ \
  --cfg.target_path ./weights/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/

When using Pipeline Parallelism (PP), you need to additionally specify the layer splitting scheme:

mm-convert WanConverter hf_to_mm \
  --cfg.source_path ./weights/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/ \
  --cfg.target_path ./weights/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/ \
  --cfg.target_parallel_config.pp_layers '[[7,8,8,7]]'

Why LOAD_PATH uses transformer/ (parent dir): mm-convert hf_to_mm writes latest_checkpointed_iteration.txt in transformer/ and stores weights in transformer/release/. Megatron checkpoint loading reads latest_checkpointed_iteration.txt to resolve the actual subdirectory automatically. Set LOAD_PATH to .../transformer/, not .../transformer/release/. For detailed information on weight conversion, refer to the mindspeed-mm-weight-prep Skill.

Step 2: Dataset Preparation

Dataset directory structure:

<dataset>/
├── data.json       # Video-text pair metadata
└── videos/
    ├── video0001.mp4
    └── video0002.mp4

data.json format:

[
    {
        "path": "videos/video0001.mp4",
        "cap": "A cat playing with a ball in a garden.",
        "num_frames": 81,
        "fps": 24,
        "resolution": {"height": 480, "width": 832}
    },
    {
        "path": "videos/video0002.mp4",
        "cap": "Ocean waves hitting the shore at sunset.",
        "num_frames": 81,
        "fps": 24,
        "resolution": {"height": 480, "width": 832}
    }
]

Field descriptions:

• path: Relative path to the video file (relative to the dataset root directory)
• cap: Video description text
• num_frames: Number of frames (Wan2.1 typically uses 81 frames)
• fps: Frame rate
• resolution: Resolution (height x width)

Then edit examples/wan2.1/feature_extract/data.txt, with each line in the format:

<dataset_root_directory>,<data.json_path>

Example:

/home/dataset/wan_training,/home/dataset/wan_training/data.json

Step 3: Feature Extraction

Docker users: If your container was not created with --ipc=host or --shm-size=16g, DataLoader workers will crash with Bus error. Set --num-workers 0 in the script as a workaround.

Feature extraction is a step unique to generative models. This step runs the VAE and text encoder on NPU to pre-encode raw video and text into feature vectors, so that training can directly use features without repeated encoding/decoding.

Configuration files (3 files need to be modified):

1. examples/wan2.1/feature_extract/model_t2v.json: Set the from_pretrained paths for the VAE and text encoder
2. examples/wan2.1/feature_extract/data.json: Set num_frames, max_height, max_width, and the tokenizer's from_pretrained path
3. mindspeed_mm/tools/tools.json: Set sorafeature.save_path, the output directory for extracted features

Launch feature extraction:

bash examples/wan2.1/feature_extract/feature_extraction.sh

Underlying call: torchrun ... mindspeed_mm/tools/feature_extraction/get_wan_feature.py

After extraction completes, the output structure is:

./sora_features/               ← sorafeature.save_path (feature dataset root)
├── data.jsonl                 ← file paths are RELATIVE: "features/test_0000.pt"
└── features/
    ├── test_0000.pt
    └── test_0001.pt

Path rule (important): data_folder in feature_data.json must point to the save_path root (e.g., ./sora_features/), NOT ./sora_features/features/. The data.jsonl already includes features/ in its relative paths. Setting data_folder to the features/ subdirectory causes double-nesting (features/features/...) and data loading failures.

For detailed feature extraction configuration, see references/feature-extraction.md.

Step 4: Training Configuration

After feature extraction is complete, configure the training script.

Update data.txt: Edit examples/wan2.1/1.3b/t2v/data.txt to point to the feature dataset root (e.g., ./sora_features).

Key configuration files:

• feature_data.json: Dataset configuration. Set data_folder to sorafeature.save_path root (e.g., ./sora_features/), dataset_type: "feature", and tokenizer path
• pretrain_model.json: Model architecture configuration, defines the diffusion scheduler and predictor (wandit) architecture parameters

Training script examples/wan2.1/1.3b/t2v/pretrain.sh key parameters:

# Path configuration
LOAD_PATH="./weights/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/"
SAVE_PATH="./output/wan2.1-1.3b-t2v/"

# Parallelism strategy
TP=1   # Tensor Parallelism
PP=1   # Pipeline Parallelism
VP=1   # Virtual Pipeline
CP=1   # Context Parallelism
MBS=1  # Micro Batch Size

# Memory optimization (key parameters)
--recompute-granularity full
--recompute-method block
--recompute-num-layers 20

# Distributed optimizer
--use-distributed-optimizer
--overlap-grad-reduce
--overlap-param-gather

Step 5: Launch Training

Docker users: If your container was not created with --ipc=host or --shm-size=16g, DataLoader workers will crash with Bus error. Set --num-workers 0 in the script as a workaround.

bash examples/wan2.1/1.3b/t2v/pretrain.sh

Entry script: pretrain_sora.py

Training checkpoints are saved in the directory configured by SAVE_PATH.

Step 6: Inference Generation (Optional)

After training completes, you can use the inference script to generate video from text prompts:

bash examples/wan2.1/1.3b/t2v/inference.sh

Entry script: inference_sora.py

Adaptation Notes for Other Models

Megatron Backend Models

These models follow the same pipeline as Wan2.1 (feature extraction -> training via pretrain_sora.py):

• CogVideoX: examples/cogvideox/, uses CogVideoConverter, feature extraction via get_sora_feature.py
• OpenSoraPlan 1.3: examples/opensoraplan1.3/, uses OpenSoraPlanConverter (hf_to_mm), feature extraction via get_sora_feature.py
• OpenSoraPlan 1.5: examples/opensoraplan1.5/, uses OpenSoraPlanConverter (source_to_mm), no feature extraction (trains on raw data)
• OpenSora 2.0: examples/opensora2.0/
• HunyuanVideo: examples/hunyuanvideo/, uses HunyuanVideoConverter, feature extraction via get_hunyuan_feature.py
• StepVideo: examples/stepvideo/, uses StepVideoConverter, feature extraction via get_sora_feature.py
• Lumina: uses LuminaConverter, feature extraction via get_lumina_feature.py

Megatron+FSDP2 Backend Models

These models use pretrain_sora.py with --use-torch-fsdp2 and require CUDA_DEVICE_MAX_CONNECTIONS=2:

• Wan2.2: examples/wan2.2/, uses WanConverter, NO feature extraction needed (trains on raw data). Additional deps: pip install -r examples/wan2.2/requirements.txt (diffusers==0.35.1, peft==0.17.1)
• HunyuanVideo 1.5: examples/hunyuanvideo_1.5/, uses HunyuanVideoConverter. Additional deps: pip install -r examples/hunyuanvideo_1.5/requirements.txt (transformers==4.57.1, diffusers==0.35.0)
• VACE: examples/vace/, uses VACEConverter, feature extraction via get_vace_feature.py

FSDP2-Native Backend Models

• LTX2: Uses mindspeed_mm/fsdp/train/trainer.py directly, NO feature extraction needed, no Megatron dependency

Accelerate+DeepSpeed Backend Models (Diffusers)

These models use accelerate launch with DeepSpeed, NOT pretrain_sora.py. They use HuggingFace weights directly (no converter needed) and do NOT need feature extraction:

• FLUX: examples/diffusers/flux/, text-to-image
• SD3: examples/diffusers/sd3/, text-to-image
• SDXL: examples/diffusers/sdxl/, text-to-image
• Sana: examples/diffusers/sana/, text-to-image
• HiDream: examples/diffusers/hidream/, text-to-image

Diffusers models have a completely different training flow from Megatron models. Read the shell scripts in examples/diffusers/<model>/ to understand the specific launch command and config structure.

Post-Training

MindSpeed-MM supports the following post-training optimization methods:

Directory Structure Reference

MindSpeed-MM/
├── pretrain_sora.py                 # Megatron/FSDP2 training entry point
├── inference_sora.py                # Inference entry point
├── posttrain_sora_dpo.py            # DPO post-training entry point
├── posttrain_flux_dancegrpo.py      # GRPO post-training entry point
├── mindspeed_mm/
│   ├── fsdp/train/trainer.py        # FSDP2-native trainer (used by ltx2)
│   └── tools/
│       ├── tools.json               # Global config (sorafeature.save_path for feature output)
│       └── feature_extraction/
│           ├── get_wan_feature.py        # Wan2.1 feature extraction
│           ├── get_hunyuan_feature.py    # HunyuanVideo feature extraction
│           ├── get_sora_feature.py       # CogVideoX, StepVideo, OpenSoraPlan
│           ├── get_lumina_feature.py     # Lumina feature extraction
│           └── get_vace_feature.py       # VACE feature extraction
└── examples/
    ├── wan2.1/                      # Megatron backend
    │   ├── feature_extract/
    │   │   ├── feature_extraction.sh
    │   │   ├── model_t2v.json
    │   │   ├── data.json
    │   │   └── data.txt
    │   ├── 1.3b/t2v/
    │   │   ├── pretrain.sh
    │   │   ├── inference.sh
    │   │   ├── data.txt
    │   │   ├── feature_data.json
    │   │   └── pretrain_model.json
    │   └── 14b/t2v/
    ├── wan2.2/                      # Megatron+FSDP2 backend
    ├── hunyuanvideo/                # Megatron backend
    ├── hunyuanvideo_1.5/            # Megatron+FSDP2 backend
    ├── cogvideox/                   # Megatron backend
    ├── opensoraplan1.3/             # Megatron backend
    ├── opensoraplan1.5/             # Megatron backend
    ├── opensora2.0/                 # Megatron backend
    ├── stepvideo/                   # Megatron backend
    ├── vace/                        # Megatron+FSDP2 backend
    ├── ltx2/                        # FSDP2-native backend
    └── diffusers/                   # Accelerate+DeepSpeed backend
        ├── flux/
        ├── sd3/
        ├── sdxl/
        ├── sana/
        └── hidream/

FAQ

Q: Feature extraction reports OOM -- Reduce max_height/max_width or num_frames in data.json, or reduce --nproc_per_node in torchrun.

Q: Training loss does not converge -- Check that LOAD_PATH points to transformer/ (not model root), weight conversion completed successfully, and dataset_type in feature_data.json is "feature".

Q: Inference generates all-black or noisy video -- Confirm the inference script loads the correct checkpoint path and sufficient training steps have been completed.

Q: mm-convert command not found -- Run pip install -e . to install MindSpeed-MM. mm-convert is a console entry point registered during installation. Q: Feature extraction cannot read videos -- Install decord: pip install decord==0.6.0 on x86, or compile from source on ARM. Q: Communication_Error_Bind_IP_Port -- Stale process holding port. Kill with lsof -i :29500 + kill -9 <PID>, or change MASTER_PORT.

References

• Model Delta Cards - Execution checklists for non-flagship models (Wan2.2, HunyuanVideo, CogVideoX, Diffusers/FLUX)
• Per-Model Dependencies - Dependency versions and installation instructions for each generative model
• Feature Extraction Guide - Feature extraction configuration files, workflow, and troubleshooting
• Model Registry - Complete lookup table for all models
• MindSpeed-MM Repository