{ "cells": [ { "cell_type": "markdown", "id": "b29a4b72-31bb-4268-9598-2cd2b6f7475e", "metadata": { "collapsed": false, "jupyter": { "outputs_hidden": false } }, "source": [ "# NeVA Training / Inference Tutorial\n", "\n", "### Note:\n", "Currently, this notebook must be run in a NeMo container. An example command to launch the container:\n", "\n", "```\n", "docker run --gpus all -it --rm -v :/opt/NeMo --shm-size=8g \\\n", " -p 8888:8888 --ulimit memlock=-1 --ulimit \\\n", " stack=67108864 \n", "```\n", "\n", "## Introduction\n", "\n", "In NeMo Framework 24.07 release, we introduce a wide variety of new features for training NeVA. This notebook illustrates the new features within NeVA such as: Mistral + Mixtral Support, Token Compression and SigLIP support.\n", "\n", "\n", "\n", "\n", "## Latest Enhancements and Expanded Capabilities\n", "\n", "NeVA has expanded its capabilities with support for multiple GPT architectures, vision encoders, and new modalities. Here's an overview of the latest additions:\n", "\n", "### 1. Expanded Language Model Support\n", "\n", "NeVA now supports a broader range of language models, including:\n", "\n", "* LLama 2 & 3\n", "* Mistral & Mixtral\n", "* Nemotron-3 & 4\n", "\n", "#### Integrating LLaMA or Mistral Models\n", "\n", "To utilize LLaMA or Mistral models as your base language model in NeVA:\n", "\n", "1. Download the model from Hugging Face.\n", "2. Convert the model from HF to '.nemo' format using the appropriate conversion script:\n", " - LLama 7b & 70b: [convert_llama_hf_to_nemo.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/checkpoint_converters/convert_llama_hf_to_nemo.py)\n", " - Mistral 7b: [convert_mistral_7b_hf_to_nemo.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/checkpoint_converters/convert_mistral_7b_hf_to_nemo.py)\n", " - Mixtral 8x7b: [convert_mixtral_hf_to_nemo.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/checkpoint_converters/convert_mixtral_hf_to_nemo.py)\n", "\n", "\n", "### 2. Enhanced Vision Encoder Options\n", "\n", "NeVA now supports both CLIP and SigLIP as vision encoder options. Configuration examples:\n", "\n", "#### CLIP Configuration\n", "\n", "```\n", "model.mm_cfg.vision_encoder.from_pretrained = 'openai/clip-vit-large-patch14-336'\n", "model.mm_cfg.vision_encoder.hidden_size = 1024\n", "model.mm_cfg.vision_encoder.from_hf = True\n", "model.mm_cfg.vision_encoder.crop_size = [336, 336]\n", "model.mm_cfg.vision_encoder.class_token_length = 1\n", "```\n", "\n", "#### SigLIP Configuration\n", "\n", "```\n", "model.mm_cfg.vision_encoder.from_pretrained = 'google/siglip-so400m-patch14-384'\n", "model.mm_cfg.vision_encoder.hidden_size = 1152\n", "model.mm_cfg.vision_encoder.from_hf = True\n", "model.mm_cfg.vision_encoder.crop_size = [384, 384]\n", "model.mm_cfg.vision_encoder.class_token_length = 0\n", "```\n", "\n", "### 3. Token Fusion Method from [VILA](https://github.com/NVlabs/VILA)\n", "\n", "NeVA introduces an additional multimodal adapter design, which fuses tokens and reduces the number of tokens needed to represent an image by 4x.\n", "\n", "To utilize this new token fusion method, modify `model.mm_cfg.mm_mlp_adapter_type` to `mlp_downsample`.\n", "\n", "### 4. Video Modality Support\n", "\n", "NeVA now supports video processing by representing video as multiple image frames. Configure as follows:\n", "\n", "```\n", "model.data.media_type = 'video'\n", "model.data.num_frames = 5\n", "model.data.splice_single_frame = None\n", "model.data.video_folder = '/path/to/videos'\n", "```\n", "\n", "For more detailed documentation on VideoNeVA, refer to the [official documentation](https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/multimodal/mllm/video_neva.html).\n", "\n", "\n", "## This tutorial will guide you through the following topics:\n", "1. Prepare pre-requisites for NeVA training\n", "2. Training a NeVA model with an MoE Mixtral model, utilize token compression and replace CLIP with SigLIP as your vision encoder.\n", "3. Performing inference with the trained model\n", "\n", "## Datasets\n", "\n", "Please refer to [NeMo User Guide](https://docs.nvidia.com/nemo-framework/user-guide/latest/multimodalmodels/multimodallanguagemodel/neva/dataprep.html#prepare-pretraining-and-fine-tuning-datasets) for preparing NeVA dataset for pretrain and fine-tuning.\n", "\n", "\n", "### Pre-Training Dataset\n", "\n", "The pre-training dataset is open-sourced from the LLaVA implementation and can be downloaded [here](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain). The dataset consists of a 558K subset of the LAION-CC-SBU dataset with BLIP captions.\n", "\n", "The associated images for pretraining can be downloaded via HuggingFace [here](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/blob/main/images.zip).\n", "\n", "### Instruction Tuning Dataset\n", "\n", "The instruction tuning annotations are sourced from the LLaVA implementation and are available [here](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_v1_5_mix665k.json).\n", "\n", "The associated images for the mixture instruction tuning annotations can be found [here](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#visual-instruction-tuning). After extracting, the data should be formatted as follows:\n", "\n", "```\n", " images\n", " ├── coco\n", " │ └── train2017\n", " ├── gqa\n", " │ └── images\n", " ├── ocr_vqa\n", " │ └── images\n", " ├── textvqa\n", " │ └── train_images\n", " └── vg\n", " ├── VG_100K\n", " └── VG_100K_2\n", "```\n", "\n", "After downloading all below datasets for pretraining and instruction tuning, please put data folder at `/workspace/datasets`. Your dataset directory should look something similar to:\n", "\n", "```\n", "LLaVA-Pretrain-LCS-558K\n", "├── blip_laion_cc_sbu_558k.json\n", "├── images\n", "LLaVA-Instruct-mixture\n", "├── llava_v1_5_mix665k.json\n", "└── images\n", " └── ...\n", "```\n", "\n", "## Setting up Checkpoint and Tokenizer\n", "\n", "In this notebook, we first need to convert the Vicuna 1.5 checkpoint into the .nemo format. Meanwhile, special tokens must be incorporated into the tokenizer for NeVA training. After downloading language models from Hugging Face, ensure you also fetch the corresponding tokenizer model. Using the 7B-chat model as a reference." ] }, { "cell_type": "code", "execution_count": null, "id": "6d80adff-bd3a-40e0-9441-684328ec7596", "metadata": { "vscode": { "languageId": "plaintext" } }, "outputs": [], "source": [ "! mkdir -p /workspace/checkpoints\n", "\n", "# Download Mixtral Instruct 0.1 checkpoint from HF\n", "! git clone https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1 /workspace/checkpoints/mixtral-8x7b-instruct\n", "\n", "# Convert checkpoint\n", "! python /opt/NeMo/scripts/checkpoint_converters/convert_mixtral_hf_to_nemo.py \\\n", " --input_name_or_path /workspace/checkpoints/mixtral-8x7b-instruct \\\n", " --output_path /workspace/checkpoints/mixtral-8x7b-instruct.nemo\n", "\n", "# Unpacking the NeMo checkpoint\n", "#When working with large NeMo checkpoints, it's efficient to unpack them beforehand. This avoids repeated unpacking every time the model is loaded, saving time and resources. Once unpacked, you can use the checkpoint in the same way as before.\n", "\n", "! mkdir mixtral_nemo_unpacked\n", "! tar -xvf /workspace/checkpoints/mixtral-8x7b-instruct.nemo -C mixtral_nemo_unpacked\n", "\n", "# Prepare tokenizer\n", "! cd /opt && git clone https://github.com/google/sentencepiece.git && \\\n", " cd sentencepiece && \\\n", " mkdir build && \\\n", " cd build && \\\n", " cmake .. && \\\n", " make && \\\n", " make install && \\\n", " ldconfig && \\\n", "cd /opt/sentencepiece/src/ && protoc --python_out=/opt/NeMo/scripts/tokenizers/ sentencepiece_model.proto && \\\n", "export PYTHONPATH=$PYTHONPATH:/opt/NeMo/scripts/tokenizers\n", "\n", "! python /opt/NeMo/scripts/tokenizers/add_special_tokens_to_sentencepiece.py \\\n", "--input_file /workspace/checkpoints/mixtral-8x7b-instruct/tokenizer.model \\\n", "--output_file /workspace/checkpoints/mixtral-8x7b-instruct/tokenizer_neva.model \\\n", "--is_userdefined \\\n", "--tokens \"\" \"\" \"\" \"\" \\\n", " \"\" \"\" \"\" \"\"\n" ] }, { "cell_type": "markdown", "id": "6b619e0a", "metadata": {}, "source": [ "## Training\n", "\n", "### Feature Alignment Pre-Training\n", "\n", "We provide a set of scripts for pre-training and fine-tuning which can be kicked off with CLI flags defining specified arguments. \n", "\n", "An example of a pre-training script execution (note the scripts will only perform 100 steps with a small micro batch size, this is not a full training):" ] }, { "cell_type": "code", "execution_count": null, "id": "9d16b0b9", "metadata": { "vscode": { "languageId": "plaintext" } }, "outputs": [], "source": [ "! torchrun --nproc_per_node=4 /opt/NeMo/examples/multimodal/multimodal_llm/neva/neva_pretrain.py \\\n", "--config-path=/opt/NeMo/examples/multimodal/multimodal_llm/neva/conf \\\n", "--config-name=neva_mixtral_config.yaml \\\n", "++cluster_type=BCP \\\n", "trainer.precision=bf16 \\\n", "trainer.num_nodes=1 \\\n", "trainer.devices=4 \\\n", "trainer.val_check_interval=50 \\\n", "trainer.limit_val_batches=5 \\\n", "trainer.log_every_n_steps=1 \\\n", "trainer.max_steps=100 \\\n", "model.micro_batch_size=2 \\\n", "model.global_batch_size=4 \\\n", "model.tensor_model_parallel_size=2 \\\n", "model.expert_model_parallel_size=2 \\\n", "model.pipeline_model_parallel_size=1 \\\n", "model.mcore_gpt=True \\\n", "model.transformer_engine=True \\\n", "model.data.data_path=/workspace/datasets/LLaVA-Pretrain-LCS-558K/blip_laion_cc_sbu_558k.json \\\n", "model.data.image_folder=/workspace/datasets/LLaVA-Pretrain-LCS-558K/images \\\n", "model.tokenizer.library=sentencepiece \\\n", "model.tokenizer.model=\"/workspace/checkpoints/mixtral-8x7b-instruct/tokenizer_neva.model\" \\\n", "model.data.num_workers=0 \\\n", "model.mm_cfg.llm.from_pretrained=/workspace/checkpoints/mixtral_nemo_unpacked \\\n", "model.mm_cfg.llm.model_type=mistral \\\n", "model.data.conv_template=mistral \\\n", "model.mm_cfg.vision_encoder.from_pretrained=\"google/siglip-so400m-patch14-384\" \\\n", "model.mm_cfg.vision_encoder.crop_size=\"[384,384]\" \\\n", "model.mm_cfg.vision_encoder.class_token_length=0 \\\n", "model.mm_cfg.mm_mlp_adapter_type=\"mlp_downsample\" \\\n", "model.mm_cfg.vision_encoder.from_hf=True \\\n", "model.optim.name=\"fused_adam\" \\\n", "exp_manager.create_checkpoint_callback=True \\\n", "exp_manager.create_wandb_logger=False" ] }, { "cell_type": "markdown", "id": "f24ee70d-3025-47f6-8571-295b024c3e05", "metadata": {}, "source": [ "**Note**: To initialize training a model from scratch rather than from a pretrained checkpoint, you may specify `null` instead of a path in the CLI arguments.\n", "\n", "### Image-Language Pair Instruction Fine-Tuning\n", "\n", "Fine-tuning can also be run from within the container via a similar command leveraging the `neva_finetune.py` script. We leverage the checkpoint saved from pretrain step to further finetune it, given by `model.restore_from_path=/workspace/nemo_experiments/nemo_neva/checkpoints/nemo_neva.nemo`.\n", "\n", "An example of an image-text pair instruction tuning script execution (note the scripts will only perform 1000 steps with a small micro batch size, this is not a full training):\n", "\n", "Note: For running SFT on multiple nodes on a Slurm cluster, replace the `torchrun --nproc_per_node=8` with `python`. Mixtral 8x7B has a minimum requirement of TP8PP2 for finetuning, which will require two nodes. You can refer to the [NeMo Framework Launcher](https://github.com/NVIDIA/NeMo-Framework-Launcher/blob/main/launcher_scripts/conf/fine_tuning/neva/mixtral_8x7b_instruct.yaml) for detailed instructions on setting up and running multimodal fine-tuning for Mixtral." ] }, { "cell_type": "code", "execution_count": null, "id": "97963224", "metadata": { "vscode": { "languageId": "plaintext" } }, "outputs": [], "source": [ "python /opt/NeMo/examples/multimodal/multimodal_llm/neva/neva_finetune.py \\\n", "--config-path=/opt/NeMo/examples/multimodal/multimodal_llm/neva/conf \\\n", "--config-name=neva_mixtral_config.yaml \\\n", "++cluster_type=BCP \\\n", "trainer.precision=bf16 \\\n", "trainer.num_nodes=1 \\\n", "trainer.devices=4 \\\n", "trainer.val_check_interval=50 \\\n", "trainer.limit_val_batches=5 \\\n", "trainer.log_every_n_steps=1 \\\n", "trainer.max_steps=100 \\\n", "model.micro_batch_size=2 \\\n", "model.global_batch_size=4 \\\n", "model.tensor_model_parallel_size=8 \\\n", "model.expert_model_parallel_size=1 \\\n", "model.pipeline_model_parallel_size=2 \\\n", "model.mcore_gpt=True \\\n", "model.transformer_engine=True \\\n", "model.restore_from_path=/workspace/nemo_experiments/nemo_neva/checkpoints/nemo_neva.nemo \\\n", "model.data.data_path=/workspace/datasets/LLaVA-Instruct-mixture/llava_v1_5_mix665k.json \\\n", "model.data.image_folder=/workspace/datasets/LLaVA-Instruct-mixture/images \\\n", "model.tokenizer.library=sentencepiece \\\n", "model.tokenizer.model=/workspace/checkpoints/mixtral-8x7b-instruct/tokenizer_neva.model \\\n", "model.data.num_workers=0 \\\n", "model.mm_cfg.llm.from_pretrained=/workspace/checkpoints/mixtral_nemo_unpacked \\\n", "model.mm_cfg.llm.freeze=False \\\n", "model.mm_cfg.llm.model_type=mistral \\\n", "model.data.conv_template=mistral \\\n", "model.mm_cfg.vision_encoder.from_pretrained=\"google/siglip-so400m-patch14-384\" \\\n", "model.mm_cfg.vision_encoder.crop_size=\"[384,384]\" \\\n", "model.mm_cfg.vision_encoder.class_token_length=0 \\\n", "model.mm_cfg.mm_mlp_adapter_type=\"mlp_downsample\" \\\n", "model.mm_cfg.vision_encoder.from_hf=True \\\n", "model.optim.name=\"fused_adam\" \\\n", "exp_manager.create_checkpoint_callback=True \\\n", "exp_manager.name=\"nemo_neva_finetune\" \\\n", "exp_manager.create_wandb_logger=False" ] }, { "cell_type": "markdown", "id": "5235639a", "metadata": {}, "source": [ "### Running Inference\n", "\n", "NeVA inference via the NeMo Framework can be quickly spun up via the NeMo Launcher and a few modifications to use the default NeVA inference config file.\n", "\n", "Inference can be run with a similar command leveraging the provided inference script `neva_evaluation.py` within the container. You can unpack the finetuned checkpoint as mentioned earlier in the tutorial to load the model checkpoint faster.\n", "\n", "An example of an inference script execution:" ] }, { "cell_type": "code", "execution_count": null, "id": "ee0156ea", "metadata": { "vscode": { "languageId": "plaintext" } }, "outputs": [], "source": [ "! echo '{\"image\": \"RTX4080.png\", \"prompt\": \"\\nCan you describe this image?\"}' > sample.jsonl\n", "! mkdir images && wget https://assets.nvidia.partners/images/png/TUF_Gaming_GeForce_RTX_4080_SUPER_OC_edition_packaging_with_card__12419.png --output-document=images/RTX4080.png\n", "! torchrun --nproc_per_node=1 /opt/NeMo/examples/multimodal/multimodal_llm/neva/neva_evaluation.py \\\n", "tensor_model_parallel_size=4 \\\n", "pipeline_model_parallel_size=1 \\\n", "neva_model_file=/workspace/nemo_neva_finetune/checkpoints/mixtral_neva.nemo \\\n", "trainer.devices=4 \\\n", "trainer.precision=bf16 \\\n", "prompt_file=sample.jsonl \\\n", "inference.media_base_path=images \\\n", "output_file=output.jsonl \\\n", "inference.temperature=0.2 \\\n", "inference.tokens_to_generate=256" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.11.6" } }, "nbformat": 4, "nbformat_minor": 5 }