fixlint

Signed-off-by: Pablo Garay <pagaray@nvidia.com>

Author: pablo-garay

nemo_vfm/diffusion/data/readme.rst

@@ -1,4 +1,4 @@
-Preparing Image / Video Megatron Energon WebDataset with Cosmos Tokenizer
+Preparing Image / Video Megatron Energon WebDataset with Cosmos Tokenizer
 ===========================
 
 This script is an example on preparing a WebDataset for an image / video + text dataset using distributed processing with the Cosmos Tokenizer. It processes each sample by generating a **continuous** image / video latent using the Cosmos video tokenizer and a T5 embedding from the text caption. Then, the processed data is stored in a WebDataset-compatible format.
@@ -22,5 +22,5 @@ Usage
 
 3. **Running the Script**:
    To run the script on 8 GPUs, use the following command:
-   
+
    ``bash torchrun --nproc_per_node=8 nemo/collections/diffusion/data/prepare_energon_dataset.py``

nemo_vfm/diffusion/readme.rst

@@ -1,10 +1,10 @@
-Diffusion Training Framework
+Diffusion Training Framework
 =============
 
 Overview
 --------
 
-The NeMo Diffusion Training Framework provides a scalable training platform for diffusion models with transformer backbones.  Our new features streamline the training process, allowing developers to efficiently train state-of-the-art models with ease. 
+The NeMo Diffusion Training Framework provides a scalable training platform for diffusion models with transformer backbones.  Our new features streamline the training process, allowing developers to efficiently train state-of-the-art models with ease.
 
 
 Some of the features we currently support include:
@@ -23,13 +23,13 @@ We support image diffusion training. Video training incoming.
 +---------------------------+------------------+
 | Parallelism               | Status           |
 +===========================+==================+
-| FSDP                      | ✅ Supported     |
+| FSDP                      | ✅ Supported     |
 +---------------------------+------------------+
-| CP+TP+SP+distopt          | ✅ Supported     |
+| CP+TP+SP+distopt          | ✅ Supported     |
 +---------------------------+------------------+
-| CP+TP+SP+PP+distopt       | ✅ Supported     |
+| CP+TP+SP+PP+distopt       | ✅ Supported     |
 +---------------------------+------------------+
-| CP+TP+SP+FSDP             | 🕒 Coming Soon   |
+| CP+TP+SP+FSDP             | 🕒 Coming Soon   |
 +---------------------------+------------------+
 
 
@@ -44,11 +44,11 @@ We support image diffusion training. Video training incoming.
 +--------------+-------------------+-----------------+
 | Model Size   | Modality          | Status          |
 +==============+===================+=================+
-| DiT 30B+     | 256px image       | ✅ Supported    |
+| DiT 30B+     | 256px image       | ✅ Supported    |
 +--------------+-------------------+-----------------+
-| DiT 30B+     | 256px image+video | 🕒 Coming Soon  |
+| DiT 30B+     | 256px image+video | 🕒 Coming Soon  |
 +--------------+-------------------+-----------------+
-| DiT 30B+     | 768px image+video | 🕒 Coming Soon  |
+| DiT 30B+     | 768px image+video | 🕒 Coming Soon  |
 +--------------+-------------------+-----------------+
 
 
@@ -58,7 +58,7 @@ Energon Dataloader for Webscale Dataloading
 Webscale Dataloading
 ^^^^^^^^^^^^^^^^^^^^
 
-Megatron-Energon is an optimized multi-modal dataloader for large-scale deep learning with Megatron. Energon allows for distributed loading of large training training data for multi-modal model training. Energon allows for blending many datasets together and distributing the dataloading workflow across multiple cluster nodes/processes while ensuring reproducibility and resumability. 
+Megatron-Energon is an optimized multi-modal dataloader for large-scale deep learning with Megatron. Energon allows for distributed loading of large training training data for multi-modal model training. Energon allows for blending many datasets together and distributing the dataloading workflow across multiple cluster nodes/processes while ensuring reproducibility and resumability.
 
 Dataloader Checkpointing
 ^^^^^^^^^^^^^^^^^^^^^^^^
@@ -84,7 +84,7 @@ Our dataloader provides support for mixed image-video training by using the NeMo
 
 Model and Data Parallelism
 --------------------------
-NeMo provides support for training models using tensor parallelism, sequence parallelism, pipeline parallelism, and context parallelism. To support pipeline parallelism with conditional diffusion training, we duplicate the conditional embeddings across the pipeline stages, and perform an all-reduce during the backward pass. This approach uses more compute, but it has a lower communication cost than sending the conditional embeddings through different pipeline stages. 
+NeMo provides support for training models using tensor parallelism, sequence parallelism, pipeline parallelism, and context parallelism. To support pipeline parallelism with conditional diffusion training, we duplicate the conditional embeddings across the pipeline stages, and perform an all-reduce during the backward pass. This approach uses more compute, but it has a lower communication cost than sending the conditional embeddings through different pipeline stages.
 
 .. image:: assets/pipeline_conditioning.png
    :alt: Conditioning mechanism for pipeline parallelism
@@ -116,14 +116,14 @@ this will generate a folder a tar files. .pth contains image/video latent repres
 .. code-block:: bash
 
    shard_000.tar
-   ├── samples/sample_0000.pth
-   ├── samples/sample_0000.pickle
-   ├── samples/sample_0000.json
-   ├── samples/sample_0001.pth
-   ├── samples/sample_0001.pickle
-   ├── samples/sample_0001.json
-   └── ...
-   shard_001.tar   
+   ├── samples/sample_0000.pth
+   ├── samples/sample_0000.pickle
+   ├── samples/sample_0000.json
+   ├── samples/sample_0001.pth
+   ├── samples/sample_0001.pickle
+   ├── samples/sample_0001.json
+   └── ...
+   shard_001.tar
 
 The following is a sample command to prepare prepare webdataset into energon dataset:
 

nemo_vfm/diffusion/vae/readme.rst

@@ -1,4 +1,4 @@
-============================
+============================
 Pretraining Variational AutoEncoder
 ============================
 
@@ -49,11 +49,11 @@ Data Preparation
     .. code-block:: bash
 
         000000.tar
-        ├── 1.jpg
-        ├── 2.jpg
+        ├── 1.jpg
+        ├── 2.jpg
         000001.tar
-        ├── 3.jpg
-        ├── 4.jpg
+        ├── 3.jpg
+        ├── 4.jpg
 
 2. next we need to index the webdataset with `energon <https://nvidia.github.io/Megatron-Energon/>`_. navigate to the dataset directory and run the following command:
 
@@ -64,7 +64,7 @@ Data Preparation
 3. then select dataset type `ImageWebdataset` and specify the type `jpg`. Below is an example of the interactive setup:
 
     .. code-block:: bash
-        
+
         Found 2925 tar files in total. The first and last ones are:
         - 000000.tar
         - 002924.tar

nemo_vfm/physicalai/Cosmos/cosmos1/models/autoregressive/nemo/post_training/tokenizer/README.md

@@ -1,4 +1,4 @@
-# Cosmos Autoregressive-based World Foundation Models: Changing the Video Tokenizer
+# Cosmos Autoregressive-based World Foundation Models: Changing the Video Tokenizer
 
 Learn how to post-train Cosmos Autoregressive-based World Foundation Models (WFMs) using the NVIDIA NeMo Framework when swapping out the original tokenizer for a new one. This recipe provides a conceptual overview of how to adapt your existing model to handle a different discrete video (DV) tokenizer, ensuring high-quality video generation under a new compression setting.
 
@@ -49,20 +49,20 @@ The NeMo Framework supports the following Cosmos Autoregressive (AR) models for
 
 ## Why Change Tokenizers?
 
-Cosmos Autoregressive models are typically trained on a specific tokenizer configuration (e.g., 8×16×16). If you’d like to reduce patch size or change compression (e.g., to 4×8×8), you can post-train the existing weights so that the model effectively aligns its internal representations with the new token embeddings—without re-training the tokenizer.
+Cosmos Autoregressive models are typically trained on a specific tokenizer configuration (e.g., 8×16×16). If you’d like to reduce patch size or change compression (e.g., to 4×8×8), you can post-train the existing weights so that the model effectively aligns its internal representations with the new token embeddings—without re-training the tokenizer.
 
 ## Tutorial: Finetuning Cosmos-4B on 10k Videos with a New Tokenizer
 
 In this tutorial, we will:
-- Take a model originally trained on an 8×16×16 tokenizer.
-- Post-train it on a 4×8×8 tokenizer.
+- Take a model originally trained on an 8×16×16 tokenizer.
+- Post-train it on a 4×8×8 tokenizer.
 - Demonstrate using a sample dataset of 10 (in production, we recommend using 10k videos from a distribution similar to that of pretraining).
 
 ### 1. Calculate Sequence Lengths
 
-- **Original Sequence Length:** 12,800 tokens (33 frames * 640 px width * 1024 px height, when tokenized with an 8×16×16 tokenizer becomes ⌈33/8⌉ * (640/16) * (1024/16) = 5 * 64 * 80 = 12,800 tokens)
-- **New Sequence Length:** 12,800 tokens (17 frames * 320 px width * 512 px height, when tokenized with a 4×8×8 tokenizer becomes ⌈17/8⌉ * (320/4) * (512/8) = 5 * 80 * 64 = 12,800 tokens)
-- For other resolutions or frame counts, recalculate your maximum tokens to ensure you do not exceed the model’s capacity.
+- **Original Sequence Length:** 12,800 tokens (33 frames * 640 px width * 1024 px height, when tokenized with an 8×16×16 tokenizer becomes ⌈33/8⌉ * (640/16) * (1024/16) = 5 * 64 * 80 = 12,800 tokens)
+- **New Sequence Length:** 12,800 tokens (17 frames * 320 px width * 512 px height, when tokenized with a 4×8×8 tokenizer becomes ⌈17/8⌉ * (320/4) * (512/8) = 5 * 80 * 64 = 12,800 tokens)
+- For other resolutions or frame counts, recalculate your maximum tokens to ensure you do not exceed the model’s capacity.
 
 ```bash
 export WIDTH=512
@@ -137,15 +137,15 @@ torchrun --nproc-per-node=$NUM_GPUS cosmos1/models/autoregressive/nemo/post_trai
 ```
 
 - Adjust parameters such as `--max_steps`, `--global_batch_size`, and `--lr` to suit your needs.
-- Ensure that `--model_path` matches the checkpoint originally trained on an 8×16×16 tokenizer.
+- Ensure that `--model_path` matches the checkpoint originally trained on an 8×16×16 tokenizer.
 - For an explanation on other configuration options, please see [the general post-training tutorial](https://github.com/NVIDIA/Cosmos/blob/main/cosmos1/models/autoregressive/nemo/post_training/README.md#configuration-options).
 
 ### 5. Monitor Quality
 
 After training, monitor checkpoints and generate sample outputs. Please keep in mind that you will have to
 disable the diffusion decoder, as this is not yet supported for non-standard tokenizers.
 
-1. **Checkpoints:**  
+1. **Checkpoints:**
    Checkpoints are saved under `./experiments/example_log_dir/default/checkpoints/`. For example:
    ```bash
    epoch=1-step=49
@@ -154,7 +154,7 @@ disable the diffusion decoder, as this is not yet supported for non-standard tok
    ```
    Choose a checkpoint (e.g., `epoch=1-step=99`) for inference.
 
-2. **Run Inference:**  
+2. **Run Inference:**
    Generate sample video outputs:
    ```bash
    # Set the checkpoint directory
@@ -177,7 +177,7 @@ disable the diffusion decoder, as this is not yet supported for non-standard tok
        --tokenizer_compression_factor $TOKENIZER_COMPRESSION_FACTOR
    ```
 
-3. **(Optional) Evaluate with TokenBench:**  
+3. **(Optional) Evaluate with TokenBench:**
    To compute quantitative metrics (e.g., PSNR, SSIM), generate a large number of videos using the script above and place them into the `evals`
    folder of the checkpoint. Please ensure that the input filename in the original folder (e.g. `my-video.mp4` matches the output filename in the
    `evals` folder). This is because PSNR and SSIM compare video quality against a reference, and our script will match the output video with
@@ -245,7 +245,7 @@ python cosmos1/models/autoregressive/nemo/post_training/video2world_prepare_data
 
 ### 3. Training/Finetuning for Video2World
 
-Fine-tune a Video2World–specific checkpoint (or any compatible Cosmos AR model) on your new data. For example, using the 5B Video2World model:
+Fine-tune a Video2World–specific checkpoint (or any compatible Cosmos AR model) on your new data. For example, using the 5B Video2World model:
 
 ```bash
 torchrun --nproc-per-node=$NUM_GPUS cosmos1/models/autoregressive/nemo/post_training/video2world_finetuning.py \

nemo_vfm/physicalai/Cosmos/cosmos1/models/diffusion/nemo/post_training/README.md

@@ -1,4 +1,4 @@
-# Cosmos Diffusion-based World Foundation Models: NeMo Framework User Guide
+# Cosmos Diffusion-based World Foundation Models: NeMo Framework User Guide
 
 Learn how to [post-train](#post-train) Cosmos Diffusion-based World Foundation Models (WFMs) using the [NVIDIA NeMo Framework](https://docs.nvidia.com/nemo-framework/user-guide/latest/overview.html) for your custom Physical AI tasks by following this guide.
 
@@ -439,7 +439,7 @@ srun \
    torchrun --nproc_per_node=8 cosmos1/models/diffusion/nemo/post_training/multicamera.py  --factory cosmos_multicamera_diffusion_7b_text2world_overfit   --yes   trainer.max_steps=1000     optim.config.lr=1e-6
    ```
 
-4. (Optional) Run Largescale Full Post-Training  
+4. (Optional) Run Largescale Full Post-Training
    You will need to prepare you own data similar with the overfit dataset in step 3. Change XDG_CACHE_HOME to point to the custom dataset
    ```bash
    torchrun --nproc_per_node=8 cosmos1/models/diffusion/nemo/post_training/multicamera.py --yes   trainer.max_steps=1000     optim.config.lr=1e-6
@@ -461,26 +461,26 @@ srun \
    # --factory cosmos_multicamera_diffusion_7b_image2world_finetune_w_traj
    ```
 
-For different post-training recipes make sure to change the autoresume directory so that checkpoints generated from recipe A are not used for recipe B. Modify 'recipe.trainer.callbacks.dirpath' and 'recipe.resume.resume_from_directory'.  
+For different post-training recipes make sure to change the autoresume directory so that checkpoints generated from recipe A are not used for recipe B. Modify 'recipe.trainer.callbacks.dirpath' and 'recipe.resume.resume_from_directory'.
 
-'recipe.trainer.val_check_interval' and 'recipe.trainer.limit_val_batches' can be set to 0 to disable validation during post-training.  
+'recipe.trainer.val_check_interval' and 'recipe.trainer.limit_val_batches' can be set to 0 to disable validation during post-training.
 
 Under `HF_HOME`, there should be a `hub` folder, which contains the base checkpoints. Under `multicamera` folder, there should be view embeddings, `mkdir -p $HF_HOME/multicamera && symlink -s cosmos1/models/diffusion/nemo/post_training/multicamera/*.pt $HF_HOME/multicamera`
 ```
 hub/
-├── models--nvidia--Cosmos-1.0-Diffusion-7B-Text2World/
-├── models--nvidia--Cosmos-1.0-Diffusion-7B-Video2World/
-└── models--nvidia--Cosmos-1.0-Tokenizer-CV8x8x8/
+├── models--nvidia--Cosmos-1.0-Diffusion-7B-Text2World/
+├── models--nvidia--Cosmos-1.0-Diffusion-7B-Video2World/
+└── models--nvidia--Cosmos-1.0-Tokenizer-CV8x8x8/
 
 multicamera/
-├── video_camera_embeddings_v0_camera_front_tele_30fov.pt
-├── video_camera_embeddings_v0_camera_rear_tele_30fov.pt
-├── video_camera_embeddings_v0_camera_front_wide_120fov.pt
-├── video_neg_prompt_embeddings_v0.pt
-├── video_camera_embeddings_v0_camera_cross_left_120fov.pt
-├── video_camera_embeddings_v0_camera_cross_right_120fov.pt
-├── video_camera_embeddings_v0_camera_rear_left_70fov.pt
-└── video_camera_embeddings_v0_camera_rear_right_70fov.pt
+├── video_camera_embeddings_v0_camera_front_tele_30fov.pt
+├── video_camera_embeddings_v0_camera_rear_tele_30fov.pt
+├── video_camera_embeddings_v0_camera_front_wide_120fov.pt
+├── video_neg_prompt_embeddings_v0.pt
+├── video_camera_embeddings_v0_camera_cross_left_120fov.pt
+├── video_camera_embeddings_v0_camera_cross_right_120fov.pt
+├── video_camera_embeddings_v0_camera_rear_left_70fov.pt
+└── video_camera_embeddings_v0_camera_rear_right_70fov.pt
 ```
 
 ## DiT Video2World Camera-Control Post-Training