diff --git a/docs/megatron/models/dit/README.md b/docs/megatron/models/dit/README.md index 98e556e..b5290dc 100644 --- a/docs/megatron/models/dit/README.md +++ b/docs/megatron/models/dit/README.md @@ -82,7 +82,7 @@ Done ### 🐳 Build Container -Please follow the instructions in the [container](https://github.com/NVIDIA-NeMo/DFM#-built-your-own-container) section of the main README. +Please follow the instructions in the [container](https://github.com/NVIDIA-NeMo/DFM#build-your-own-container) section of the main README. --- @@ -106,12 +106,12 @@ The model architecture can be customized through parameters such as `num_layers` First, copy the example config file and update it with your own settings: ```bash -cp examples/megatron/recipes/dit/conf/dit_pretrain_example.yaml examples/megatron/recipes/dit/conf/my_config.yaml +cp examples/megatron/recipes/dit/conf/dit_pretrain.yaml examples/megatron/recipes/dit/conf/my_config.yaml # Edit my_config.yaml to set: # - model.vae_cache_folder: Path to VAE cache folder # - dataset.path: Path to your dataset folder # - checkpoint.save and checkpoint.load: Path to checkpoint folder -# - train.global_batch_size: Set to match be divisible by NUM_GPUs +# - train.global_batch_size: Set to be divisible by NUM_GPUs # - logger.wandb_exp_name: Your experiment name ``` diff --git a/examples/README.md b/examples/README.md index 3e7d62a..a8c4943 100644 --- a/examples/README.md +++ b/examples/README.md @@ -1,3 +1,18 @@ -# Examples +# DFM Examples -Training and inference examples for dtensor and megatron models +Collection of examples and recipes for training and inference using the Data Foundation Model (DFM) framework. + +## Quick Start + +If you are new to DFM, start with the **[Automodel Examples](automodel/)** for high-level API usage. + +## Examples by Category + +| Category | Description | Key Examples | +|----------|-------------|--------------| +| **[Automodel](automodel/)** | High-level API for seamless training and inference | [Wan 2.1 Fine-tuning](automodel/README.md), [Pretraining](automodel/pretrain/), [Generation](automodel/generate/) | +| **[Megatron](megatron/)** | Advanced recipes and configurations using Megatron-Core | [DiT Recipes](megatron/recipes/dit/), [Wan Recipes](megatron/recipes/wan/) | + +## Support + +For issues or questions, please open a GitHub issue or refer to the main documentation. diff --git a/examples/automodel/README.md b/examples/automodel/README.md index 791bf66..a3f3f13 100644 --- a/examples/automodel/README.md +++ b/examples/automodel/README.md @@ -1,3 +1,17 @@ +# Automodel Examples + +High-level API examples for training, fine-tuning, and generating with DFM models. + +## Supported Tasks + +| Task | Directory | Available Examples | +|------|-----------|-------------------| +| **Fine-tuning** | **[finetune](finetune/)** | • **[Wan 2.1 T2V](finetune/finetune.py)**: Fine-tuning with Flow Matching
• **[Multi-node](finetune/wan2_1_t2v_flow_multinode.yaml)**: Distributed training config | +| **Generation** | **[generate](generate/)** | • **[Generate](generate/wan_generate.py)**: Run inference with Wan 2.1
• **[Validate](generate/wan_validate.py)**: Run validation loop | +| **Pre-training** | **[pretrain](pretrain/)** | • **[Wan 2.1 T2V](pretrain/pretrain.py)**: Pre-training from scratch | + +--- + # Diffusion Model Fine-tuning with Automodel Backend Train diffusion models with distributed training support using NeMo Automodel and flow matching. diff --git a/examples/megatron/README.md b/examples/megatron/README.md index ba59237..476d14a 100644 --- a/examples/megatron/README.md +++ b/examples/megatron/README.md @@ -1,3 +1,17 @@ # Megatron Examples -Recipes and configuration overrides for megatron training. +Advanced recipes and configuration overrides for training models using the Megatron-Core backend. + +## Available Model Recipes + +| Recipe | Key Scripts | Description | +|--------|-------------|-------------| +| **[DiT](recipes/dit/README.md)** | • [Pretrain](recipes/dit/pretrain_dit_model.py)
• [Inference](recipes/dit/inference_dit_model.py) | Diffusion Transformer (DiT) training on butterfly dataset | +| **[Wan](recipes/wan/README.md)** | • [Pretrain](recipes/wan/pretrain_wan.py)
• [Inference](recipes/wan/inference_wan.py) | Wan 2.1 model pre-training and inference | + +## Directory Structure + +| Directory | Description | +|-----------|-------------| +| **[recipes](recipes/)** | Source code and scripts for the models above | +| **[override_configs](override_configs/)** | Configuration overrides for customizing parallelism (TP/CP/SP) | diff --git a/examples/megatron/override_configs/README.md b/examples/megatron/override_configs/README.md index 6731a25..3512a67 100644 --- a/examples/megatron/override_configs/README.md +++ b/examples/megatron/override_configs/README.md @@ -1,3 +1,13 @@ -# Override Configs +# Configuration Overrides -Parallelism configuration overrides for different CP/TP/SP sizes. +Collection of YAML configuration files used to override default settings in Megatron training recipes. These are typically used for specifying parallelization strategies (Tensor Parallelism, Context Parallelism, Sequence Parallelism) or data configurations. + +## Files + +| File | Description | +|------|-------------| +| `wan_pretrain_sample_data.yaml` | Sample data configuration for Wan pre-training. | + +## Usage + +These configs can be passed to the training script arguments to override defaults. diff --git a/examples/megatron/recipes/README.md b/examples/megatron/recipes/README.md index 7ed591b..9531fad 100644 --- a/examples/megatron/recipes/README.md +++ b/examples/megatron/recipes/README.md @@ -1,3 +1,10 @@ -# Recipe +# Model Recipes -Training recipes for Wan2.1 pretraining, finetuning, and weight verification. +Collection of end-to-end training recipes for specific model architectures. + +## Available Recipes + +| Recipe | Description | +|--------|-------------| +| **[DiT](dit/)** | Diffusion Transformer (DiT) training on butterfly dataset | +| **[Wan](wan/)** | Wan 2.1 model pre-training and inference | diff --git a/examples/megatron/recipes/wan/README.md b/examples/megatron/recipes/wan/README.md new file mode 100644 index 0000000..f3fba44 --- /dev/null +++ b/examples/megatron/recipes/wan/README.md @@ -0,0 +1,13 @@ +# Wan Recipes + +Recipes for training and inferencing Wan models using Megatron-Core. + +## Files + +- `pretrain_wan.py`: Main pre-training script. +- `inference_wan.py`: Inference script. +- `prepare_energon_dataset_wan.py`: Dataset preparation utility. + +## Performance Testing + +See **[Performance Test Guide](README_perf_test.md)** for details on running performance benchmarks on different hardware (H100, GB200, etc.).