Important

NeMo 2.0 is an experimental feature and currently released in the dev container only: nvcr.io/nvidia/nemo:dev. Please refer to NeMo 2.0 overview for information on getting started.

NeVA (LLaVA)

Originating from LLaVA (Large Language and Vision Assistant), NeVA is a groundbreaking addition to the NeMo Multimodal ecosystem. This model seamlessly integrates large language-centric models (like NVGPT or Llama2) with a vision encoder, and is trained with machine-generated multimodal language-image instruction-following data. Building on the foundation set by LLaVA, NeVA further enhances training by leveraging features of the NeMo LLM framework such as model parallelism, activation checkpointing, AMP O2, Flash Attention, and more. While traditional language models have been primarily focused on textual processing, NeVA boldly adopts a holistic approach, bridging visual and linguistic comprehension.

New in NeVA

Here are some of the new features introduced in NeVA.

Pipeline Parallelism Support: Pipeline parallelism is added in NeVA. When employing pipeline parallelism, the vision encoder (if loaded from Hugging Face) will duplicate on GPUs where the pipeline parallelism rank equals 0.

Loading ViT from HF
DP0
   PP rank 0
      TP rank 0 (if HF, ViT)
      TP rank 1 (if HF, ViT)
   PP rank 1
      TP rank 0
      TP rank 1
DP1
   PP rank 0
      TP rank 0 (if HF, ViT)
      TP rank 1 (if HF, ViT)
   PP rank 1
      TP rank 0
      TP rank 1

Loading ViT from .nemo
DP0
   PP rank 0
      TP rank 0 (if NeMo, ViT TP rank 0)
      TP rank 1 (if NeMo, ViT TP rank 1)
   PP rank 1
      TP rank 0
      TP rank 1
DP1
   PP rank 0
      TP rank 0 (if NeMo, ViT TP rank 0)
      TP rank 1 (if NeMo, ViT TP rank 1)
   PP rank 1
      TP rank 0
      TP rank 1

Mixtral Support with Expert Parallelism: NeVA now supports Mixtral as a foundation LLM with Expert Parallelism.
SigLIP Encoder from Hugging Face: NeVA has been updated to use the SigLIP encoder, which can be sourced from Hugging Face. You can now configure this feature by setting model.mm_cfg.vision_encoder.from_pretrained='google/siglip-so400m-patch14-384'. Support for integrating the SigLIP module into the NeMo Multimodal framework is underway and will be available soon.
Distributed Checkpoint Format Support: NeVA now supports a distributed checkpoint format. You no longer need to manually change checkpoint partitions from pretraining to tuning or from training to inference. You can now load checkpoints in any new TP/PP combinations. For example, pretrain with TP=8, PP=1, and then fine-tune with TP=8, PP=2. All newly created checkpoints will automatically be saved in a distributed format. This process requires no changes to the existing configuration.
Sequence Packing: Refer to the documentation at NeVA Sequence Packing.
NeVA now supports a broader range of language models as backend, including:
- LLama 2 & 3
- Mistral & Mixtral
- Nemotron-4

Feature	Training	Inference
Data parallelism	Yes	N/A
Tensor parallelism	Yes	Yes
Pipeline parallelism	Yes	No
Sequence parallelism	Yes	No
Activation checkpointing	Yes	No
FP32/TF32	Yes	Yes (FP16 enabled by default)
AMP/FP16	No	Yes
AMP/BF16	Yes	No
BF16 O2	Yes	No
TransformerEngine/FP8	Yes	No
Multi-GPU	Yes	Yes
Multi-Node	Yes	Yes
Inference deployment	N/A	NVIDIA Triton supported
SW stack support	Slurm DeepOps/Base Command Manager/Base Command Platform	Slurm DeepOps/Base Command Manager/Base Command Platform
NVfuser	No	N/A
Distributed Optimizer	No	N/A
TorchInductor	No	N/A
Flash Attention	Yes	N/A