Duplicate from nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16

Co-authored-by: Yoshi Suhara <suhara@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,514 @@
+---
+library_name: transformers
+license: other
+license_name: nvidia-nemotron-open-model-license
+license_link: >-
+  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
+pipeline_tag: text-generation
+language:
+  - en
+  - es
+  - fr
+  - de
+  - ja
+  - it
+  - pt
+  - zh
+  - ar
+  - da
+  - ko
+  - nl
+  - pl
+  - ru
+  - sv
+  - th
+tags:
+- nvidia
+- pytorch
+datasets:
+  - nvidia/Nemotron-Pretraining-Code-v1
+  - nvidia/Nemotron-CC-v2
+  - nvidia/Nemotron-Pretraining-SFT-v1
+  - nvidia/Nemotron-CC-Math-v1
+  - nvidia/Nemotron-Pretraining-Code-v2
+  - nvidia/Nemotron-Pretraining-Specialized-v1
+  - nvidia/Nemotron-CC-v2.1
+  - nvidia/Nemotron-CC-Code-v1
+  - nvidia/Nemotron-Pretraining-Dataset-sample
+track_downloads: true
+---
+
+# NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
+<div align="center" style="line-height: 1;">
+<a href="https://build.nvidia.com/nvidia/nemotron-3-nano-30b-a3b" target="_blank" style="margin: 2px;">
+    <img alt="Chat" src="https://img.shields.io/badge/🤖Chat-Nemotron_3_Nano-536af5?color=76B900&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+</a>
+<a href="https://arxiv.org/abs/2512.20848" target="_blank" style="margin: 2px;">
+    <img alt="Chat" src="https://img.shields.io/badge/📝Paper-Read Now!-536af5?color=76B900&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+</a>
+<a href="https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets" target="_blank" style="margin: 2px;">
+    <img alt="Pre-Training Datasets" src="https://img.shields.io/badge/🗄️_Pre--Training_Datasets-Available_Here-76B900?logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+</a>
+</div>
+<div align="center" style="line-height: 1;">
+  <a href="https://developer.nvidia.com/nemotron" target="_blank" style="margin: 2px;">
+    <img alt="Homepage" src="https://img.shields.io/badge/🏠Nemotron Developer Page-Learn More Here!-536af5?color=76B900&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+<a href="https://discord.gg/9xpKQtVvrk" target="_blank" style="margin: 2px;">
+    <img alt="Homepage" src="https://img.shields.io/badge/Discord-NVIDIA%20AI%20Developer-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+<div align="center" style="line-height: 1;">
+  <a href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/" style="margin: 2px;">
+    <img alt="License" src="https://img.shields.io/badge/License-NVIDIA Nemotron Open Model License-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+![](./accuracy_chart.svg)
+
+## Model Overview
+
+**Model Developer:** NVIDIA Corporation
+
+**Model Dates:**
+
+September 2025 \- December 2025
+
+**Data Freshness:**
+
+The pre-training data has a cutoff date of June 25, 2025.
+
+## Description
+
+NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 is a base large language model (LLM) trained from scratch by NVIDIA, with the next token prediction loss. It provides a good starting point for instruction fine-tuning.
+
+This model is ready for commercial use.
+
+### What is Nemotron?
+
+NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
+
+## License/Terms of Use
+
+GOVERNING TERMS: Use of this model is governed by the [NVIDIA Nemotron Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/).
+
+## Base Benchmark Evaluations
+
+We evaluated our model on the following benchmarks:
+
+| Task | Qwen3 30B-A3B-Base | NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 |
+| :---- | :---- | :---- |
+| **General Knowledge** | | |
+| MMLU (5-shot, acc) | **81.07** | 78.56 |
+| MMLU-Pro (5-shot, CoT EM) | 61.71 | **65.05** |
+| AGIEval-En (3/5-shot, CoT acc) | 63.12 | **68.32** |
+| **Code** | | |
+| HumanEval (0-shot) | 70.73 | **78.05** |
+| MBPP-Sanitized (3-shot) | 73.15 | **75.49** |
+| **Math** | | |
+| GSM8K (8-shot, acc) | 89.01 | **92.34** |
+| MATH (4-shot, acc) | 61.14 | **82.88** |
+| MATH-500 (4-shot, avg@32) | 55.08 | **78.63** |
+| **Commonsense Understanding** | | |
+| ARC-Challenge (25-shot, acc_norm) | **94.45** | 91.89 |
+| HellaSwag (10-shot, acc_norm) | 83.14 | **85.56** |
+| OpenBookQA (0-shot, acc_norm) | 44.80 | **46.20** |
+| PIQA (0-shot, acc_norm) | 81.01 | **84.33** |
+| WinoGrande (5-shot, acc) | 78.22 | **79.64** |
+| **Reading Comprehension** | | |
+| RACE (0-shot, acc) | **90.05** | 88.04 |
+| **Multilingual** | | |
+| MMLU Global Lite (5-shot, avg acc) | **76.84** | 74.47 |
+| MGSM (8-shot, avg acc) | 82.53 | **83.00** |
+| **Long Context** | | |
+| RULER (64K, 0-shot, acc) | 63.55 | **87.50** |
+| RULER (128K, 0-shot, acc) | 60.69 | **82.92** |
+| RULER (256K, 0-shot, acc) |   Not Supported   | **75.44** |
+| RULER (512K, 0-shot, acc) |   Not Supported   | **70.56** |
+
+All evaluation results were collected via [Nemo Evaluator SDK](https://github.com/NVIDIA-NeMo/Evaluator) and [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness). The open source container on LM Evaluation Harness packaged via NVIDIA's Nemo Evaluator SDK used for evaluations can be found [here](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/eval-factory/containers/lm-evaluation-harness). A reproducibility tutorial along with all configs can be found in [Nemo Evaluator SDK examples](https://github.com/NVIDIA-NeMo/Evaluator/tree/main/packages/nemo-evaluator-launcher/examples/nemotron/nano-v3-reproducibility.md).
+
+### Deployment Geography: Global
+
+### Use Case
+
+This model is intended for developers and researchers building instruction-following LLMs.
+
+Supported languages include: English, Spanish, French, German, Japanese, Italian, Chinese, Arabic, Hebrew, Hindi, Korean, Czech, Danish, Dutch, Finnish, Polish, Portuguese, Thai, Swedish, and Russian.
+
+### Release Date: 
+
+December 15, 2025 via [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16)
+
+## Reference(s)
+
+* [NVIDIA Nemotron 3 model family on Hugging Face](https://huggingface.co/collections/nvidia/nvidia-nemotron-v3)
+* [NVIDIA Nemotron 2 model family on Hugging Face](https://huggingface.co/collections/nvidia/nvidia-nemotron-v2)
+* [NVIDIA Nemotron 3 White Paper](https://arxiv.org/abs/2512.20856)
+
+## Model Architecture
+
+- **Architecture Type:** Mamba2-Transformer Hybrid Mixture of Experts (MoE)
+- **Network Architecture:** Nemotron Hybrid MoE
+
+- **Number of model parameters:** 30B
+
+## Training Methodology
+
+Stage 1: Pre-Training  
+* [NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16) model was pre-trained using crawled and synthetic code, math, science, and general knowledge data. All datasets are disclosed in the [Training, Testing, and Evaluation Datasets](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16#training-testing-and-evaluation-datasets) section of this document. Major portions of the pre-training corpus are released in the [Nemotron-Pre-Training-Datasets](https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets) collection.
+* Software used for pre-training: [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+
+The end-to-end training recipe is available in the [NVIDIA Nemotron Developer Repository](https://github.com/NVIDIA-NeMo/Nemotron). Evaluation results can be replicated using the [NeMo Evaluator SDK](https://github.com/NVIDIA-NeMo/Evaluator). More details on the datasets and synthetic data generation methods can be found in the technical report [NVIDIA Nemotron 3 Nano](https://arxiv.org/abs/2512.20848).
+
+## Input
+
+- **Input Type(s):** Text
+
+- **Input Format(s):** String
+
+- **Input Parameters:** One-Dimensional (1D): Sequences
+
+- **Maximum input size:** 128K tokens
+
+- **Other Properties Related to Input:**
+Supported languages include: English, Spanish, French, German, Japanese, Italian, Chinese, Arabic, Hebrew, Hindi, Korean, Czech, Danish, Dutch, Finnish, Polish, Portuguese, Thai, Swedish, and Russian.
+
+
+## Output
+
+- **Output Type(s):** Text
+
+- **Output Format:** String
+
+- **Output Parameters:** One-Dimensional (1D): Sequences
+
+- **Maximum output size:** 128K tokens
+
+
+Our AI models are designed and optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
+
+## Software Integration
+
+- Runtime Engine(s): NeMo 25.11.01
+- Supported Hardware Microarchitecture Compatibility: NVIDIA H100-80GB, NVIDIA A100
+- Operating System(s): Linux
+
+
+The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
+
+### Use it with Transformers
+
+The snippet below shows how to use this model with Huggingface Transformers (tested on version 4.57.3). We recommend using NeMo Framework 25.11.01 to ensure all required libraries are available.
+
+Please note that the model supports up to a 1M context size, although the default context size in the Hugging Face configuration is 256k due to higher VRAM requirements.
+
+```
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+model_name = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    torch_dtype=torch.bfloat16,
+    trust_remote_code=True,
+    device_map="auto"
+)
+
+prompt = "The capital of France is"
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+
+outputs = model.generate(
+    **inputs,
+    max_new_tokens=32,
+    do_sample=False,
+    eos_token_id=tokenizer.eos_token_id
+)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## Model Version(s)
+
+- v1.0
+
+# Training, Testing, and Evaluation Datasets
+
+**Data Modality:**  Text  
+**The total size:**  10,648,823,153,919 Tokens  
+**Total number of datasets:** 141  
+**Dataset partition:** *Training \[100%\], testing \[0%\], validation \[0%\]*  
+**Time period for training data collection:** 2013 to May 1, 2025  
+**Time period for testing data collection:** 2013 to May 1, 2025  
+**Time period for validation data collection:** 2013 to May 1, 2025  
+**Data Collection Method by dataset:** Hybrid: Automated, Human, Synthetic
+
+NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 is pre-trained on a large corpus of high-quality curated and synthetically-generated data. It is trained in the English language, as well as 19 other languages and 43 programming languages. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracy. The model was trained for approximately 25 trillion tokens.
+
+Alongside the model, we release our final [pre-training](https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets) data, as outlined in this section. For ease of analysis, there is a sample set that is ungated. For all remaining code, math and multilingual data, gating and approval is required, and the dataset is permissively licensed for model training purposes.
+
+More details on the datasets and synthetic data generation methods can be found in the technical report [NVIDIA Nemotron 3 Nano](https://arxiv.org/abs/2512.20848).
+
+## Public dataset
+
+| Dataset | Collection Period |
+| :---- | :---- |
+| [GSM8K](https://github.com/openai/grade-school-math) | 4/23/2025 |
+| [CC-NEWS](https://commoncrawl.org/blog/news-dataset-available) | 4/23/2025 |
+| [Common Crawl](https://commoncrawl.org/) | 4/23/2025 |
+| [Wikimedia](https://dumps.wikimedia.org/) | 4/23/2025 |
+| [Bespoke-Stratos-17k](https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-17k) | 4/23/2025 |
+| [tigerbot-kaggle-leetcodesolutions-en-2k](https://huggingface.co/datasets/TigerResearch/tigerbot-kaggle-leetcodesolutions-en-2k) | 4/23/2025 |
+| [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) | 4/23/2025 |
+| [APIGen Function-Calling](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) | 4/23/2025 |
+| [LMSYS-Chat-1M](https://huggingface.co/datasets/lmsys/lmsys-chat-1m) | 4/23/2025 |
+| [Open Textbook Library \- CC BY-SA & GNU subset](https://open.umn.edu/opentextbooks/textbooks/) and [OpenStax \- CC BY-SA subset](https://openstax.org/) | 4/23/2025 |
+| [Advanced Reasoning Benchmark](https://github.com/TheDuckAI/arb), [tigerbot-kaggle-leetcodesolutions-en-2k](https://huggingface.co/datasets/TigerResearch/tigerbot-kaggle-leetcodesolutions-en-2k), [PRM800K](https://github.com/openai/prm800k), and [SciBench](https://github.com/mandyyyyii/scibench) | 4/23/2025 |
+| [FineWeb-2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) | 4/23/2025 |
+| [Court Listener](https://www.courtlistener.com/help/api/bulk-data/) | Legacy Download |
+| [peS2o](https://huggingface.co/datasets/allenai/peS2o) | Legacy Download |
+| [OpenWebMath](https://huggingface.co/datasets/open-web-math/open-web-math) | Legacy Download |
+| [BioRxiv](https://www.biorxiv.org/tdm) | Legacy Download |
+| [PMC Open Access Subset](https://pmc.ncbi.nlm.nih.gov/tools/openftlist/) | Legacy Download |
+| [OpenWebText2](https://openwebtext2.readthedocs.io/en/latest/) | Legacy Download |
+| [Stack Exchange Data Dump](https://archive.org/details/stackexchange) | Legacy Download |
+| [PubMed Abstracts](https://github.com/thoppe/The-Pile-PubMed) | Legacy Download |
+| [NIH ExPorter](https://exporter.nih.gov/ExPORTER_Catalog.aspx) | Legacy Download |
+| [arXiv](https://info.arxiv.org/help/bulk_data/index.html) | Legacy Download |
+| [BigScience Workshop Datasets](https://github.com/bigscience-workshop/bigscience/tree/master/train/tr11-176B-ml#datasets) | Legacy Download |
+| [Reddit Dataset](https://files.pushshift.io/reddit/) | Legacy Download |
+| [SEC's Electronic Data Gathering, Analysis, and Retrieval (EDGAR)](https://www.sec.gov/search-filings) | Legacy Download |
+| [Advanced Mathematical Problem Solving](https://github.com/hendrycks/math?tab=readme-ov-file) | Legacy Download |
+| [MathPile](https://github.com/GAIR-NLP/MathPile/) | Legacy Download |
+| [NuminaMath CoT](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT) | Legacy Download |
+| [PMC Article](https://pmc.ncbi.nlm.nih.gov/tools/textmining/) | Legacy Download |
+| [FLAN](https://github.com/google-research/FLAN) | Legacy Download |
+| [Advanced Reasoning Benchmark](https://github.com/TheDuckAI/arb) | Legacy Download |
+| [SciBench](https://github.com/mandyyyyii/scibench) | Legacy Download |
+| [WikiTableQuestions](https://huggingface.co/datasets/wikitablequestions) | Legacy Download |
+| [FinQA](https://finqasite.github.io/) | Legacy Download |
+| [Riddles](https://github.com/crawsome/riddles) | Legacy Download |
+| [Problems in Elementary Mathematics for Home Study](https://archive.org/details/AntonovVygodskyNikitinSankinProblemsInElementaryMathematicsForHomeStudyMir1982) | Legacy Download |
+| [MedMCQA](https://huggingface.co/datasets/openlifescienceai/medmcqa) | Legacy Download |
+| [Cosmos QA](https://huggingface.co/datasets/allenai/cosmos_qa) | Legacy Download |
+| [MCTest](https://huggingface.co/datasets/sagnikrayc/mctest) | Legacy Download |
+| [AI2's Reasoning Challenge](https://huggingface.co/datasets/ai2_arc) | Legacy Download |
+| [OpenBookQA](https://github.com/allenai/OpenBookQA) | Legacy Download |
+| [MMLU Auxiliary Train](https://huggingface.co/datasets/cais/mmlu/viewer/all/auxiliary_train) | Legacy Download |
+| [social-chemestry-101](https://huggingface.co/datasets/tasksource/social-chemestry-101) | Legacy Download |
+| [Moral Stories](https://huggingface.co/datasets/demelin/moral_stories) | Legacy Download |
+| [The Common Pile v0.1](https://huggingface.co/common-pile) | Legacy Download |
+| [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath) | Legacy Download |
+| [MegaMath](https://huggingface.co/datasets/LLM360/MegaMath) | Legacy Download |
+| [MegaMath](https://huggingface.co/datasets/LLM360/MegaMath) | Legacy Download |
+| [MultiverseMathHard](https://huggingface.co/datasets/Nexusflow/MultiverseMathHard) | 10/2/2025 |
+| [News Commentary](https://opus.nlpl.eu/News-Commentary.php) | 10/2/2025 |
+| [Essential-Web](https://huggingface.co/datasets/EssentialAI/essential-web-v1.0) | 10/2/2025 |
+| [finepdfs](https://huggingface.co/datasets/HuggingFaceFW/finepdfs) | 10/2/2025 |
+| [HotpotQA](https://huggingface.co/hotpot_qa/datasets) | 10/2/2025 |
+| [SQuAD2.0](https://rajpurkar.github.io/SQuAD-explorer/) | 10/2/2025 |
+| [NLTK Words Lists](https://www.nltk.org/nltk_data/) | 10/2/2025 |
+
+## Private Non-publicly Accessible Datasets of Third Parties
+
+| Dataset |
+| :---- |
+| Global Regulation |
+| TAUS Translation Memory |
+| Scale HLE |
+| HackerRank Coding |
+
+## Private Non-publicly Accessible Datasets by NVIDIA
+
+| Dataset |
+| :---- |
+| Machine Translation of STEM data using [Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) |
+
+## Crawled and Scraped from Online Sources by NVIDIA
+
+The English Common Crawl data was downloaded from the Common Crawl Foundation (see their FAQ for details on their crawling) and includes the snapshots CC-MAIN-2013-20 through CC-MAIN-2025-13. The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper. Additionally, we extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC.
+
+The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API. Each crawl was operated in accordance with the rate limits set by its respective source, either GitHub or S3. We collect raw source code and subsequently remove any having a license which does not exist in our permissive-license set (for additional details, refer to the [technical report](https://arxiv.org/abs/2512.20848)).
+
+| Dataset | Modality | Dataset Size | Collection Period | Collecting Organisation |
+| :---- | :---- | :---- | :---- | :---- |
+| English Common Crawl | Text | 3.36T | 4/8/2025 | NVIDIA Advanced Deep Learning Research |
+| English Common Crawl 1.1 | Text | Not disclosed | 10/2/2025 | NVIDIA Advanced Deep Learning Research |
+| Multilingual Common Crawl | Text | 812.7B | 5/1/2025 | NVIDIA Advanced Deep Learning Research |
+| GitHub Crawl | Text | 747.4B | 4/29/2025 | NVIDIA Advanced Deep Learning Research |
+
+## NVIDIA-Sourced Synthetic Datasets
+
+| Dataset | Modality | Dataset Size | Seed Dataset | Model(s) used for generation |
+| :---- | :---- | :---- | :---- | :---- |
+| Synthetic Art of Problem Solving from DeepSeek-R1 | Text | 40B | [Art of Problem Solving](https://artofproblemsolving.com/company); [American Mathematics Competitions 8](https://artofproblemsolving.com/wiki/index.php/AMC_8_Problems_and_Solutions); [American Mathematics Competitions 10](https://artofproblemsolving.com/wiki/index.php/AMC_10_Problems_and_Solutions); | [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) |
+| Synthetic Moral Stories and Social Chemistry from Mixtral-8x22B-v0.1 | Text | 327M | [social-chemestry-101](https://huggingface.co/datasets/tasksource/social-chemestry-101); [Moral Stories](https://huggingface.co/datasets/demelin/moral_stories) | [Mixtral-8x22B-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-v0.1) |
+| Synthetic Social Sciences seeded with OpenStax from DeepSeek-V3, Mixtral-8x22B-v0.1, and Qwen2.5-72B | Text | 83.6M | [OpenStax \- CC BY-SA subset](https://openstax.org/) | [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3); [Mixtral-8x22B-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-v0.1); [Qwen2.5-72B](https://huggingface.co/Qwen/Qwen2.5-72B) |
+| Synthetic Health Sciences seeded with OpenStax from DeepSeek-V3, Mixtral-8x22B-v0.1, and Qwen2.5-72B | Text | 9.7M | [OpenStax \- CC BY-SA subset](https://openstax.org/) | [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3); [Mixtral-8x22B-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-v0.1); [Qwen2.5-72B](https://huggingface.co/Qwen/Qwen2.5-72B) |
+| Synthetic STEM seeded with OpenStax, Open Textbook Library, and GSM8K from DeepSeek-R1, DeepSeek-V3, DeepSeek-V3-0324, and Qwen2.5-72B | Text | 175M | [OpenStax \- CC BY-SA subset](https://openstax.org/); [GSM8K](https://github.com/openai/grade-school-math); [Open Textbook Library \- CC BY-SA & GNU subset](https://open.umn.edu/opentextbooks/textbooks/) | [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1), [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3); [DeepSeek-V3-0324](https://huggingface.co/deepseek-ai/DeepSeek-V3-0324); [Qwen2.5-72B](https://huggingface.co/Qwen/Qwen2.5-72B) |
+| [Nemotron-PrismMath](https://huggingface.co/datasets/nvidia/Nemotron-PrismMath) | Text | 4.6B | [Big-Math-RL-Verified](https://huggingface.co/datasets/SynthLabsAI/Big-Math-RL-Verified); [OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k) | [Qwen2.5-0.5B-instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct), [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct); [DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) |
+| Synthetic Question Answering Data from Papers and Permissible Books from Qwen2.5-72B-Instruct | Text | 350M | [arXiv](https://info.arxiv.org/help/bulk_data/index.html); [National Institutes of Health ExPorter](https://www.nih.gov/); [BioRxiv](https://www.biorxiv.org/tdm); [PMC Article](https://pmc.ncbi.nlm.nih.gov/tools/textmining/); [USPTO Backgrounds](https://data.uspto.gov/apis/transition-guide/bdss#pats); [peS2o](https://huggingface.co/datasets/allenai/peS2o); Global Regulation; [CORE](https://core.ac.uk/documentation/dataset); [PG-19](https://github.com/google-deepmind/pg19); [DOAB CC BY & CC BY-SA subset](https://www.doabooks.org/en); [NDLTD](https://ndltd.org/thesis-resources/global-etd-search/) | [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) |
+| Synthetic Rephrased [Math Data from Common Crawl](https://huggingface.co/datasets/nvidia/Nemotron-MIND) from phi-4 | Text | 73B | [Common Crawl](https://commoncrawl.org/latest-crawl) | [phi-4](https://huggingface.co/microsoft/phi-4) |
+| Synthetic Math Data from Common Crawl 4plus | Text | 52.3B | [Common Crawl](https://commoncrawl.org/latest-crawl) | [phi-4](https://huggingface.co/microsoft/phi-4) |
+| Synthetic Math Data from Common Crawl 3 | Text | 80.9B | [Common Crawl](https://commoncrawl.org/latest-crawl) | [phi-4](https://huggingface.co/microsoft/phi-4) |
+| Synthetic AGIEval seeded with AQUA-RAT, LogiQA, and AR-LSAT from DeepSeek-V3 and DeepSeek-V3-0324 | Text | 4.0B | [AQUA-RAT](https://huggingface.co/datasets/deepmind/aqua_rat); [LogiQA](https://huggingface.co/datasets/lucasmccabe/logiqa); [AR-LSAT](https://github.com/zhongwanjun/AR-LSAT) | [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3); [DeepSeek-V3-0324](https://huggingface.co/deepseek-ai/DeepSeek-V3-0324) |
+| Synthetic AGIEval seeded with AQUA-RAT, LogiQA, and AR-LSAT from Qwen3-30B-A3B | Text | 4.2B | [AQUA-RAT](https://huggingface.co/datasets/deepmind/aqua_rat); [LogiQA](https://huggingface.co/datasets/lucasmccabe/logiqa); [AR-LSAT](https://github.com/zhongwanjun/AR-LSAT) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Art of Problem Solving from Qwen2.5-32B-Instruct, Qwen2.5-Math-72B, Qwen2.5-Math-7B, and Qwen2.5-72B-Instruct | Text |  | [Art of Problem Solving](https://artofproblemsolving.com/company); [American Mathematics Competitions 8](https://artofproblemsolving.com/wiki/index.php/AMC_8_Problems_and_Solutions); [American Mathematics Competitions 10](https://artofproblemsolving.com/wiki/index.php/AMC_10_Problems_and_Solutions); [GSM8K](https://github.com/openai/grade-school-math); [PRM800K](https://github.com/openai/prm800k) | [Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct); [Qwen2.5-Math-72B](https://huggingface.co/Qwen/Qwen2.5-Math-72B); [Qwen2.5-Math-7B](https://huggingface.co/Qwen/Qwen2.5-Math-7B); [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) |
+| Synthetic MMLU Auxiliary Train from DeepSeek-R1 | Text | 0.5B | [MMLU Auxiliary Train](https://huggingface.co/datasets/cais/mmlu/viewer/all/auxiliary_train) | [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) |
+| Synthetic Long Context Continued Post-Training Data from Papers and Permissible Books from Qwen2.5-72B-Instruct | Text |  | [arXiv](https://info.arxiv.org/help/bulk_data/index.html); [National Institutes of Health ExPorter](https://www.nih.gov/); [BioRxiv](https://www.biorxiv.org/tdm); [PMC Article](https://pmc.ncbi.nlm.nih.gov/tools/textmining/); [USPTO Backgrounds](https://data.uspto.gov/apis/transition-guide/bdss#pats); [peS2o](https://huggingface.co/datasets/allenai/peS2o); Global Regulation; [CORE](https://core.ac.uk/documentation/dataset); [PG-19](https://github.com/google-deepmind/pg19); [DOAB CC BY & CC BY-SA subset](https://www.doabooks.org/en); [NDLTD](https://ndltd.org/thesis-resources/global-etd-search/) | [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) |
+| Synthetic Common Crawl from Qwen3-30B-A3B and Mistral-Nemo-12B-Instruct | Text | 415.8B | [Common Crawl](https://commoncrawl.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B); [Mistral-NeMo-12B-Instruct](https://huggingface.co/nvidia/Mistral-NeMo-12B-Instruct) |
+| Synthetic Multilingual Data from Common Crawl from Qwen3-30B-A3B | Text |  | [Common Crawl](https://commoncrawl.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Multilingual Data from Wikimedia from Qwen3-30B-A3B | Text |  | [Wikimedia](https://dumps.wikimedia.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Math Data from Wikimedia from Nemotron-4-340B-Instruct | Text |  | \- | [Nemotron-4-340B-Instruct](https://huggingface.co/nvidia/Nemotron-4-340B-Instruct) |
+| Synthetic Common Crawl Code from phi-4 | Text | 427.9B | [Common Crawl](https://commoncrawl.org/latest-crawl) | [phi-4](https://huggingface.co/microsoft/phi-4) |
+| Synthetic Scientific Coding from Qwen3-235B-A22B | Text | 1.2B | [Wikimedia](https://dumps.wikimedia.org/) | [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507) |
+| Tool Calling Data | Text | 26.2B |  | [Qwen3-235B-A22B-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507); [gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b) |
+| Synthetic Essential-Web from QwQ-32B | Text | 28.1B | [Essential-Web](https://huggingface.co/datasets/EssentialAI/essential-web-v1.0) |  [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) |
+| Translated Synthetic Crawl | Text | 389.9B | [Common Crawl](https://commoncrawl.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Translated Synthetic Wikipedia | Text | 7.9B | [Wikimedia](https://dumps.wikimedia.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Long Context from Qwen3-235B-A22B-Instruct-2507 | Text | Undisclosed | [CORE](https://core.ac.uk/documentation/dataset); [PG-19](https://github.com/google-deepmind/pg19); [DOAB CC BY & CC BY-SA subset](https://www.doabooks.org/en); [NDLTD](https://ndltd.org/thesis-resources/global-etd-search/) | [Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507) |
+| Synthetic Search STEM OPENQ from DeepSeek-R1-0528 | Text | Undisclosed | - | [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic MCQ from Qwen2.5-32B-Instruct and DeepSeek-R1-0528 | Text | Undisclosed | - | [Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct); [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Offline Search MCQA HLE from DeepSeek-R1-0528 | Text | Undisclosed | - | [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Offline Search MCQA GPQA from Qwen3-235B-A22B and DeepSeek-R1-0528 | Text | Undisclosed | - | [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B); [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Human Preference from QwQ-32B, Qwen3-30B-A3B, Qwen3-235B-A22B, Qwen3-235B-A22B-Instruct-2507, Mistral-Small-3.1-24B-Instruct-2503, Mistral-Small-3.2-24B-Instruct-2506, MiniMax-M1-80k, MiniMax-M1-40k, Kimi-K2-Instruct, DeepSeek-V3-0324, DeepSeek-R1-0528 | Text | Undisclosed | - | [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B); [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B); [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B); [Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507); [Mistral-Small-3.1-24B-Instruct-2503](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503); [Mistral-Small-3.2-24B-Instruct-2506](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506); [MiniMax-M1-80k](https://huggingface.co/MiniMaxAI/MiniMax-M1-80k); [MiniMax-M1-40k](https://huggingface.co/MiniMaxAI/MiniMax-M1-40k); [Kimi-K2-Instruct](https://huggingface.co/moonshotai/Kimi-K2-Instruct); [DeepSeek-V3-0324](https://huggingface.co/deepseek-ai/DeepSeek-V3-0324); [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Code from Qwen3-32B | Text | Undisclosed | English Common Crawl; English Common Crawl 1.1 | [Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Synthetic OpenCodeReasoning from DeepSeek-R1 | Text | Undisclosed | [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning) | [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) |
+| Synthetic LIMO from DeepSeek-R1-0528 | Text | Undisclosed | [LIMO](https://huggingface.co/datasets/GAIR/LIMO) | [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic SCP from DeepSeek-R1-0528 | Text | Undisclosed | [SCP-116K](https://huggingface.co/datasets/EricLu/SCP-116K) | [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Stack Exchange from DeepSeek-R1-0528 | Text | Undisclosed | [Stack Exchange](https://archive.org/details/stackexchange) | [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+| Synthetic Common Crawl from Qwen3-30B-A3B | Text | Undisclosed | [Common Crawl](https://commoncrawl.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Wikipedia from Qwen3-30B-A3B | Text | Undisclosed | [Wikimedia](https://dumps.wikimedia.org/) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Synthetic Essential-Web from Qwen3-30B-A3B and Qwen3-235B-A22B-Thinking-2507 | Text | Undisclosed | [Essential-Web](https://huggingface.co/datasets/EssentialAI/essential-web-v1.0) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B); [Qwen3-235B-A22B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507) |
+| Synthetic Textbook Math from Qwen3-30B-A3B, Qwen3-235B-A22B, phi-4 | Text | Undisclosed | [Common Crawl](https://commoncrawl.org/); [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath) | [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B); [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B); [phi-4](https://huggingface.co/microsoft/phi-4) |
+| Synthetic Math and Code from DeepSeek-R1 and DeepSeek-R1-0528 | Text | Undisclosed | [Magicoder-Evol-Instruct-110K](https://huggingface.co/datasets/ise-uiuc/Magicoder-Evol-Instruct-110K); [opc-sft-stage2](https://huggingface.co/datasets/OpenCoder-LLM/opc-sft-stage2); [TACO](https://huggingface.co/datasets/BAAI/TACO); [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning); [OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning); [NuminaMath CoT](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT) | [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1); [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) |
+
+
+## Training Dataset
+
+| Dataset | \# of Tokens in Nemotron Nano 2 | \# of Tokens in Nemotron 3 Nano |
+| :---- | :---- | :---- |
+| English Common Crawl | 3,360,110,334,818 | 3,456,523,212,210 |
+| English Synthetic CC | 1,949,464,641,123 | 4,340,740,677,920 |
+| Crawl++ | 360,389,153,262 | 360,389,153,262 |
+| Math | 124,606,230,663 | 154,217,502,165 |
+| Synthetic Math | 73,007,767,155 | 73,007,767,155 |
+| Code | 747,409,228,724 | 1,043,856,922,136 |
+| Synthetic Code | 175,067,553,293 | 453,117,917,176 |
+| Common Crawl Code | 0 | 263,072,374,097 |
+| English Wiki | 17,349,266,926 | 17,349,266,926 |
+| Synthetic Wiki | 0 | 7,850,648,552 |
+| Books | 0 | 0 |
+| Papers | 191,586,493,365 | 191,586,493,365 |
+| PDF-to-text | 141,096,578,533 | 141,096,578,533 |
+| Code SFT | 60,025,726,817 | 102,863,752,325 |
+| STEM SFT | 272,680,426,295 | 359,826,214,274 |
+| General SFT | 6,057,478,645 | 6,057,478,645 |
+| Tool-Calling SFT | 0 | 26,244,716,867 |
+| Multilingual | 2,172,261,909,350 | 1,743,892,490,859|
+| Synthetic multilingual | 997,710,364,950 | 595,140,661,135 |
+| **Total** | **10,648,823,153,919** | **13,336,833,827,602** |
+
+We use a considerable amount of synthetic data. Out of 10.6 trillion tokens, 3,534,013,958,278 tokens are synthetically generated.
+
+We extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC. Additionally, we used data from Wikipedia and FineWeb-2 (Penedo et al., 2025\) for these fifteen languages.
+
+| Language | Total Tokens |
+| :---- | :---- |
+| Arabic | 118,056,362,726 |
+| Danish | 117,747,321,618 |
+| German | 146,613,691,781 |
+| Spanish | 469,156,575,409 |
+| French | 139,982,002,289 |
+| Italian | 298,858,370,174 |
+| Japanese | 682,755,693,336 |
+| Korean | 127,099,747,538 |
+| Dutch | 89,041,592,681 |
+| Polish | 105,356,493,147 |
+| Portuguese | 243,249,275,089 |
+| Russian | 185,314,014,057 |
+| Swedish | 74,954,953,299 |
+| Thai | 160,778,944,467 |
+| Chinese | 211,007,236,689 |
+
+We collect a total of 922,476,782,017 tokens of code in 43 different languages.
+
+| Language | Tokens |
+| :---- | :---- |
+| Assembly | 750,628,764 |
+| C | 42,657,300,868 |
+| C\# | 56,153,329,307 |
+| C++ | 67,773,701,658 |
+| CommonLisp | 263,234,672 |
+| CSS | 38,848,760,035 |
+| Cuda | 400,222,993 |
+| Dart | 3,816,960,470 |
+| Dockerfile | 474,958,084 |
+| Fortran | 1,105,049,387 |
+| Go | 8,332,419,480 |
+| Haskell | 1,294,613,669 |
+| HTML | 69,082,117,487 |
+| Java | 131,440,465,822 |
+| JavaScript | 75,573,420,861 |
+| JSON | 15,366,881,241 |
+| Julia | 621,046,949 |
+| JupyterNotebook | 2,241,893,197 |
+| Lua | 4,146,420,802 |
+| Makefile | 12,640,010,879 |
+| Markdown | 64,796,743,311 |
+| Mathematica | 320,504,225 |
+| OmniversePython | 26,946,093 |
+| Pascal | 1,625,013,876 |
+| Perl | 1,575,314,434 |
+| PHP | 61,575,339,005 |
+| Python | 126,916,727,384 |
+| R | 19,811,381,935 |
+| reStructuredText | 1,779,876,391 |
+| Ruby | 6,446,962,615 |
+| Rust | 4,438,640,533 |
+| Scala | 3,343,959,154 |
+| Shell | 18,758,779,250 |
+| SQL | 23,205,633,085 |
+| Swift | 5,976,714,881 |
+| SystemVerilog | 233,056,185 |
+| TeX | 7,347,157,527 |
+| TypeScript | 15,657,838,582 |
+| Verilog | 811,884,369 |
+| VHDL | 648,401,444 |
+| VisualBasic.NET | 1,005,680,881 |
+| XML | 12,616,779,741 |
+| YAML | 10,574,010,491 |
+
+## Evaluation Dataset
+
+* Data Collection Method by dataset: Hybrid: Human, Synthetic
+* Labeling Method by dataset: Hybrid: Automated, Human, Synthetic
+
+## Inference
+
+- Engines: HF, vLLM, TRT-LLM
+
+- Test Hardware: NVIDIA A100 80GB, H100 80GB
+
+## Ethical Considerations
+
+NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications.  When downloaded or used in accordance with our [Trustworthy AI terms of service](https://www.nvidia.com/en-us/agreements/trustworthy-ai/terms/), developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
+
+We advise against circumvention of any provided safety guardrails contained in the Model without a substantially similar guardrail appropriate for your use case. For more details: [Safety & Security](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/blob/main/safety.md).
+
+For more detailed information on ethical considerations for this model, please see the Model Card++ [Bias](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/blob/main/bias.md), [Explainability](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/blob/main/explainability.md), and [Privacy](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/blob/main/privacy.md) Subcards.
+
+Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
+
+## Citation
+
+```
+@misc{nvidia_nemotron_nano_v3_2025,
+  title  = {{Nemotron 3 Nano}: Open, Efficient Mixture-of-Experts Hybrid {Mamba}-{Transformer} Model for {Agentic} Reasoning},
+  author = {{NVIDIA}},
+  year   = {2025},
+  url    = {https://arxiv.org/abs/2512.20848},
+  note   = {Technical report}
+}
+```

bias.md

@@ -0,0 +1,9 @@
+| Field | Response |
+| :---- | :---- |
+| Participation considerations from adversely impacted groups [protected classes](https://www.senate.ca.gov/content/protected-classes) in model design and testing: | None |
+| Bias Metric (If Measured): | Not Relevant for the Base Model |
+| Which characteristic (feature) show(s) the greatest difference in performance?: | The model shows high variance in the characteristics when it is used with a high temperature. |
+| Which feature(s) have the worst performance overall? | Not Relevant for the Base Model |
+| Measures taken to mitigate against unwanted bias: | Not Applicable |
+| If using internal data, description of methods implemented in data acquisition or processing, if any, to address the prevalence of identifiable biases in the training, testing, and validation data: | The training datasets contain a large amount of synthetic data generated by LLMs. We manually curated prompts. |
+| Tools used to assess statistical imbalances and highlight patterns that may introduce bias into AI models: | These datasets, such as Common Crawl, CC-News, and Wikimedia, do not collectively or exhaustively represent all demographic groups (and proportionally therein). For instance, these datasets do not contain explicit mentions of demographic classes such as age, gender, or ethnicity in over 85% of samples. In the subset where such terms are present, Common Crawl and CC-News contain notable representational skews—for example, references to "male" significantly outnumber those to "female," and mentions of "White" are the most frequent among ethnic identifiers. To mitigate these imbalances, we recommend considering evaluation techniques such as bias audits, fine-tuning with demographically balanced datasets, and mitigation strategies like counterfactual data augmentation to align with the desired model behavior. This evaluation used a 3,000-sample subset per dataset, identified as the optimal threshold for maximizing embedder accuracy, and includes outputs from uncalibrated embedders; as such, certain limitations may exist in the reliability of the embedding. |

config.json

@@ -0,0 +1,69 @@
+{
+  "architectures": [
+    "NemotronHForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_nemotron_h.NemotronHConfig",
+    "AutoModel": "modeling_nemotron_h.NemotronHForCausalLM",
+    "AutoModelForCausalLM": "modeling_nemotron_h.NemotronHForCausalLM"
+  },
+  "bos_token_id": 1,
+  "chunk_size": 128,
+  "conv_kernel": 4,
+  "dtype": "bfloat16",
+  "eos_token_id": 2,
+  "expand": 2,
+  "head_dim": 128,
+  "hidden_dropout": 0.0,
+  "hidden_size": 2688,
+  "hybrid_override_pattern": "MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME",
+  "initializer_range": 0.02,
+  "intermediate_size": 1856,
+  "layer_norm_epsilon": 1e-05,
+  "mamba_head_dim": 64,
+  "mamba_hidden_act": "silu",
+  "mamba_num_heads": 64,
+  "mamba_proj_bias": false,
+  "max_position_embeddings": 262144,
+  "mlp_bias": false,
+  "mlp_hidden_act": "relu2",
+  "model_type": "nemotron_h",
+  "moe_intermediate_size": 1856,
+  "moe_shared_expert_intermediate_size": 3712,
+  "n_group": 1,
+  "n_groups": 8,
+  "n_routed_experts": 128,
+  "n_shared_experts": 1,
+  "norm_eps": 1e-05,
+  "norm_topk_prob": true,
+  "num_attention_heads": 32,
+  "num_experts_per_tok": 6,
+  "num_hidden_layers": 52,
+  "num_key_value_heads": 2,
+  "num_logits_to_keep": 1,
+  "pad_token_id": 0,
+  "partial_rotary_factor": 1.0,
+  "rescale_prenorm_residual": true,
+  "residual_in_fp32": false,
+  "rope_theta": 10000,
+  "routed_scaling_factor": 2.5,
+  "sliding_window": null,
+  "ssm_state_size": 128,
+  "tie_word_embeddings": false,
+  "time_step_floor": 0.0001,
+  "time_step_limit": [
+    0.0,
+    Infinity
+  ],
+  "time_step_max": 0.1,
+  "time_step_min": 0.001,
+  "topk_group": 1,
+  "transformers_version": "4.57.1",
+  "use_bias": false,
+  "use_cache": true,
+  "use_conv_bias": true,
+  "use_mamba_kernels": true,
+  "vocab_size": 131072
+}

configuration_nemotron_h.py

@@ -0,0 +1,262 @@
+# coding=utf-8
+# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""NemotronH model configuration"""
+
+import re
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronHConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`NemotronHModel`]. It is used to instantiate a
+    NemotronH model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the NemotronH-v0.1 model.
+
+    [todo](todo)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 131072):
+            Vocabulary size of the NemotronH model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`NemotronHModel`]
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
+            model has a output word embedding layer.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 21504):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 52):
+            Number of hidden layers in the Transformer encoder.
+        hybrid_override_pattern (`str`, *optional*, defaults to `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
+            The pattern of the hybrid model. The pattern is a string of characters where each character represents M: Mamba2, *: Attention, -: MLP
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each attention head.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used.
+        mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
+            The non-linear activation function in the MLP layers.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in attention layers.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in MLP layers.
+        use_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the model.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        residual_in_fp32 (`bool`, *optional*, defaults to `False`):
+            Whether or not residuals should be in `float32`. If set to `False` residuals will keep the same `dtype` as the rest of the model.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
+            Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
+            integer value, only last `num_logits_to_keep` logits will be calculated.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        sliding_window (`int`, *optional*, defaults to None):
+            Sliding window attention window size.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        hidden_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the hidden states.
+        use_mamba_kernels (`bool`, *optional*, defaults to `True`):
+            Flag indicating whether or not to use the fast mamba kernels. These are available only if `mamba-ssm` and
+            `causal-conv1d` are installed, and the mamba modules are running on a CUDA device.
+        ssm_state_size (`int`, *optional*, defaults to 128):
+            The dimension of the mamba state space latents.
+        mamba_num_heads (`int`, *optional*, defaults to 128):
+            Number of heads in Mamba layers.
+        mamba_n_groups (`int`, *optional*, defaults to 8):
+            Number of groups in Mamba layers.
+        mamba_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of each Mamba head.
+        mamba_d_conv (`int`, *optional*, defaults to 4):
+            The size of the mamba convolution kernel.
+        mamba_expand (`int`, *optional*, defaults to 2):
+            Expanding factor used to determine the mamba intermediate size.
+        mamba_hidden_act (`str`, *optional*, defaults to "silu"):
+            The non-linear activation function in the Mamba layers.
+        mamba_dt_min (`float`, *optional*, defaults to 0.001):
+            Minimum value for the time step in Mamba.
+        mamba_dt_max (`float`, *optional*, defaults to 0.1):
+            Maximum value for the time step in Mamba.
+        mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
+            Limits for the time step in Mamba.
+        mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
+            Floor value for time step initialization in Mamba.
+        mamba_conv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to use bias in the convolution layer of the mamba mixer block.
+        mamba_proj_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the input and output projections of the mamba mixer block.
+        mamba_chunk_size (`int`, *optional*, defaults to 256):
+            Size of chunks for Mamba processing.
+        rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the pre-normalization residual connections.
+    """
+
+    model_type = "nemotron_h"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=131072,
+        tie_word_embeddings=False,
+        hidden_size=4096,
+        intermediate_size=21504,
+        num_hidden_layers=52,
+        hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
+        num_attention_heads=32,
+        head_dim=128,
+        num_key_value_heads=8,  # nemo: num_query_groups
+        mlp_hidden_act="relu2",
+        attention_bias=False,
+        mlp_bias=False,
+        use_bias=False,
+        initializer_range=0.02, # nemo: init_method_std
+        layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon
+        residual_in_fp32=False,  #  Megatron Core default value
+        use_cache=True,
+        num_logits_to_keep=1,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        sliding_window=None,
+        max_position_embeddings=4096,
+        attention_dropout=0.0,
+        hidden_dropout=0.0, # * ADDED
+        use_mamba_kernels=True,
+        ssm_state_size=128, # mamba_state_size
+        mamba_num_heads=128,
+        mamba_n_groups=8,  # nemo: mamba_ssm_ngroups = num_heads
+        mamba_head_dim=64,
+        mamba_d_conv=4,
+        mamba_expand=2,
+        mamba_hidden_act="silu",
+        mamba_dt_min=0.001,
+        mamba_dt_max=0.1,
+        mamba_dt_limit=(0.0, float("inf")),
+        mamba_dt_init_floor=1e-4,
+        mamba_conv_bias=True,
+        mamba_proj_bias=False,
+        mamba_chunk_size=128,
+        rescale_prenorm_residual=True,
+        n_routed_experts=8,
+        n_shared_experts=1,
+        moe_intermediate_size=7688,
+        moe_shared_expert_intermediate_size=7688,
+        num_experts_per_tok=2,
+        routed_scaling_factor=1.0,
+        n_group=1,
+        topk_group=1,
+        norm_topk_prob=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.tie_word_embeddings = tie_word_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hybrid_override_pattern = hybrid_override_pattern
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim
+        self.sliding_window = sliding_window
+        self.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.hidden_dropout = hidden_dropout
+
+        # Validate hybrid_override_pattern
+        # M: Mamba2, *: Attention, -: MLP
+        assert len(self.hybrid_override_pattern) == self.num_hidden_layers, "hybrid_override_pattern must have the same length as num_hidden_layers"
+        assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), "hybrid_override_pattern must only contain characters 'M', '*', or '-'"
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.mlp_hidden_act = mlp_hidden_act
+        self.attention_bias = attention_bias
+        self.mlp_bias = mlp_bias
+        self.use_bias = use_bias
+        self.initializer_range = initializer_range
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.residual_in_fp32 = residual_in_fp32
+
+        self.use_cache = use_cache
+        self.num_logits_to_keep = num_logits_to_keep
+
+        self.use_mamba_kernels = use_mamba_kernels
+        self.n_groups = mamba_n_groups
+        self.mamba_head_dim = mamba_head_dim
+        self.ssm_state_size = ssm_state_size
+        self.mamba_num_heads = mamba_num_heads
+        self.conv_kernel = mamba_d_conv
+        self.expand = mamba_expand
+        self.mamba_hidden_act = mamba_hidden_act
+        self.time_step_min = mamba_dt_min
+        self.time_step_max = mamba_dt_max
+        self.time_step_limit = mamba_dt_limit
+        self.time_step_floor = mamba_dt_init_floor
+        self.use_conv_bias = mamba_conv_bias
+        self.mamba_proj_bias = mamba_proj_bias
+        self.chunk_size = mamba_chunk_size
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+        self.n_routed_experts = n_routed_experts
+        self.n_shared_experts = n_shared_experts
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.routed_scaling_factor = routed_scaling_factor
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.norm_topk_prob = norm_topk_prob
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def layers_block_type(self):
+        return [
+            "mamba" if self.hybrid_override_pattern[i] == "M" else
+            "attention" if self.hybrid_override_pattern[i] == "*" else
+            "mlp" if self.hybrid_override_pattern[i] == "-" else "moe"
+            for i in range(self.num_hidden_layers)]

explainability.md

@@ -0,0 +1,14 @@
+| Field | Response |
+| :---- | :---- |
+| Intended Task/Domain: | Text generation, reasoning, and chat |
+| Model Type: | Text-to-text Mamba2-Transformer Hybrid |
+| Intended Users: | Generative AI creators working with conversational AI models and image content. |
+| Output: | Text |
+| Tools used to evaluate datasets to identify synthetic data and ensure data authenticity. | We used a Gemma-3 4B-based filtering model fine-tuned on [Nemotron Content Safety Dataset v2](https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0) to ensure the quality of synthetic data. |
+| Describe how the model works: | Generates text by predicting the next word or token based on the context provided in the input sequence using multiple self-attention layers. |
+| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
+| Technical Limitations & Mitigation: | This model performs particularly well in instruction following regimes, as such may be strongly influenced by untrusted inputs and should be paired with appropriate guardrails and data filtering to better align use-case behaviors when exposed to such data. |
+| Verified to have met prescribed NVIDIA quality standards: | Yes |
+| Performance Metrics: | Accuracy, Throughput, and User-side throughput |
+| Potential Known Risks: | The model was optimized explicitly for instruction following and as such may be influenced by untrusted inputs (prompt injection, indirect prompt injection, jailbreaking, web search, etc.) as a result of its instruction tuning that may degrade safety alignment and other training efforts. This model should be paired with additional guardrails and data filtering to limit exposure to instructions from malicious sources. Bypassing of safety alignment, system guardrails, and filters may allow harmful outcomes up to and including remote code execution in some agentic systems when effective security controls are not in place. The model was trained on data that contains toxic language and societal biases originally crawled from the internet. Therefore, the model may generate and amplify harmful, biased, or otherwise unsafe content reinforcing these biases and return toxic responses especially when prompted with toxic prompts. The model may also generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.  The model may exhibit self-anthropomorphism (e.g., displaying human-like characteristics in dialogue, such as expressing preferences and emotions). In integrated system contexts, the model could potentially be exploited to access or disclose information beyond the model’s intended permissions or scope of operation.|
+| Licensing: | GOVERNING TERMS: Use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). |

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    2,
+    11
+  ],
+  "pad_token_id": 0,
+  "transformers_version": "4.57.1"
+}

model-00001-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6a453e622cb76b6791a2f1c60aa7c8f97dce304d0e215e34b88e47bc937a68b
+size 4991210024

model-00002-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7362f9e3cfc3bd23c5c2c2f6f520914c6da6711e27eca221a396287bbb9e9b5
+size 4992601016

model-00003-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6068ce876022895ca36b7ba553e11c381628c7cb8365d8dce7fa377278bf2b83
+size 4992601368

model-00004-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:932dbd5beca60ced569b408d1dcab91765b23a4f88f2a82bc0aec3df354a14c9
+size 4995692800

model-00005-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57ec79cc6d9147831104e9c063a6c8ca61d6c7084a6960fdc617476465070ec7
+size 4980553832

model-00006-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e93dcfee76290ec6f8d530e344b390067197bda2cb5151bc161df917888aeaa
+size 4989423536

model-00007-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eca33a8f41bc9c43f5e267e25c76d667436c53dcabbb6ca2e95e4075570cc4a2
+size 4992601488

model-00008-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:813a4811ee783e8be46e4cd4b49e185ee990f9dac40934cd3fd8b127eaca3ecd
+size 4992601520

model-00009-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac57992ddfba2ae131aa666a7f59e7310d3ca81cde270953b2c846489a30c099
+size 4995692800

model-00010-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9a73f676d1ef6892cf039ee7d156c7678ea78e19275398497b9fcadb971fec0
+size 4992601520

model-00011-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4153d01cdbe92c1b0d1cc5247ed028c38477d5cd662c4fc214c06265f73e4551
+size 4995692800

model-00012-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18e68fd3a50df00b240615acf4fd3f9a821a1aff70dfe41446b1c3b60859c12b
+size 4995692816

model-00013-of-00013.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdf84d933727aadd1aa3519276ea9ef090e74cf857bfa447bcbe6efdcd404dbf
+size 3249723504

modeling_nemotron_h.py

@@ -0,0 +1,1740 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch NemotronH model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import DynamicCache  # we need __iter__ and __len__ of pkv
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+)
+from transformers.utils.import_utils import (
+    is_causal_conv1d_available,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    is_mamba_2_ssm_available,    
+)
+from .configuration_nemotron_h import NemotronHConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.mamba.modeling_mamba2.modeling_mamba2.py with MAMBA2->NEMOTRONH,Mamba2->NemotronH
+# For Mamba2 components Mamba2->NemotronHMamba2
+if is_mamba_2_ssm_available():
+    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
+else:
+    mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined, selective_state_update = None, None, None
+
+try:
+    #from mamba_ssm.ops.triton.layernorm_gated import RMSNorm as RMSNormGated
+    from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn
+except ImportError:
+    raise ImportError("mamba-ssm is required by the Mamba model but cannot be imported")
+
+if is_causal_conv1d_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+else:
+    causal_conv1d_update, causal_conv1d_fn = None, None
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+is_fast_path_available = all(
+    (
+        selective_state_update,
+        mamba_chunk_scan_combined,
+        mamba_split_conv1d_scan_combined,
+        causal_conv1d_fn,
+        causal_conv1d_update,
+    )
+)
+
+
+_CHECKPOINT_FOR_DOC = "nvidia/Nemotron-H-56B-Base-8K"
+_CONFIG_FOR_DOC = "NemotronHConfig"
+
+
+# Helper methods for segment sum computation
+
+
+def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
+    """
+    Padding x tensor with `pad_size` on the seq_len dim (dim=1)
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)
+
+    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)
+
+
+def reshape_into_chunks(input_tensor, pad_size, chunk_size):
+    """
+    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
+    simultaneously splitting it into chunk sequences.
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
+    input_tensor = pad_tensor_by_size(input_tensor, pad_size)
+
+    if len(input_tensor.shape) == 3:
+        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
+        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
+    else:
+        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
+        return input_tensor.reshape(
+            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
+        )
+
+
+def segment_sum(input_tensor):
+    """
+    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
+    """
+    chunk_size = input_tensor.size(-1)
+    # 1. expand input tensor to have an additional dimension and repeat along that dimension
+    # [..., chunk_size] -> [..., chunk_size, chunk_size]
+    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
+    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
+    input_tensor = input_tensor.masked_fill(~mask, 0)
+    # 3. compute actual cumsum
+    tensor_segsum = torch.cumsum(input_tensor, dim=-2)
+
+    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
+    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
+    return tensor_segsum
+
+
+def apply_mask_to_padding_states(hidden_states, attention_mask):
+    """
+    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
+    """
+    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+        dtype = hidden_states.dtype
+        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+    return hidden_states
+
+# Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+class HybridMambaAttentionDynamicCache(DynamicCache):
+    """
+    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
+    (which has a constant shape regardless of seq_len).
+
+    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
+    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
+    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
+    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
+    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
+    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
+    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
+    """
+
+    def __init__(self, config, batch_size, dtype=torch.float16, device=None):
+        super().__init__()
+        self.dtype = dtype
+        self.hybrid_override_pattern = config.hybrid_override_pattern
+        self.has_previous_state = False  # only used by mamba
+        intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        ssm_state_size = config.ssm_state_size
+        conv_kernel_size = config.conv_kernel
+        self.conv_states = []
+        self.ssm_states = []
+        self.transformer_layers = []
+        for i in range(config.num_hidden_layers):
+            if self.hybrid_override_pattern[i] == "M":
+                # Mamba layer
+                self.conv_states += [
+                    torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
+                ]
+                self.ssm_states += [
+                    torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
+                ]
+            else:
+                # Attention or MLP layer
+                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.transformer_layers.append(i)
+
+        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Update the cache
+        if self.key_cache[layer_idx].shape[-1] == 0:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+        else:
+            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
+            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
+
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        for layer_idx in range(len(self.key_cache)):
+            device = self.key_cache[layer_idx].device
+            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.value_cache[layer_idx].device
+            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
+
+            device = self.conv_states[layer_idx].device
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
+            device = self.ssm_states[layer_idx].device
+            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
+        if len(self.key_cache) <= layer_idx:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    @classmethod
+    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
+        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
+
+    # Copied from modeling_mamba2.py
+    def update_conv_state(
+        self, layer_idx: int, new_conv_state: torch.Tensor, cache_init: bool = False
+    ) -> torch.Tensor:
+        if cache_init:
+            self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
+        else:
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
+            self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
+        return self.conv_states[layer_idx]
+
+    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
+        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
+        return self.ssm_states[layer_idx]
+
+    def reset(self):
+        self.conv_states.zero_()
+        self.ssm_states.zero_()
+
+class MambaRMSNormGated(torch.nn.Module):
+    def __init__(self, hidden_size, group_size, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+        self.group_size = group_size
+
+    # jan28b version
+    def forward(self, hidden_states, gate=None):
+        return rmsnorm_fn(x=hidden_states,
+                          weight=self.weight,
+                          bias=None, # No bias
+                          z=gate,
+                          eps=self.variance_epsilon,
+                          group_size=self.group_size,
+                          norm_before_gate=False
+        )
+
+class NemotronHMamba2Mixer(nn.Module):
+    """
+    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
+    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
+    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
+    and is why Mamba is called **selective** state spaces)
+    """
+
+    def __init__(self, config: NemotronHConfig, layer_idx: int):
+        super().__init__()
+        self.num_heads = config.mamba_num_heads
+        self.hidden_size = config.hidden_size
+        self.ssm_state_size = config.ssm_state_size
+        self.conv_kernel_size = config.conv_kernel
+        self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
+        self.layer_idx = layer_idx
+        self.use_conv_bias = config.use_conv_bias
+        self.activation = config.mamba_hidden_act
+        self.act = ACT2FN[config.mamba_hidden_act]
+
+        self.layer_norm_epsilon = config.layer_norm_epsilon
+
+        self.n_groups = config.n_groups
+        self.head_dim = config.mamba_head_dim
+        self.chunk_size = config.chunk_size
+
+        self.time_step_limit = config.time_step_limit
+        self.time_step_min = config.time_step_min
+        self.time_step_max = config.time_step_max
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=config.use_conv_bias,
+            kernel_size=config.conv_kernel,
+            groups=self.conv_dim,
+            padding=config.conv_kernel - 1,
+        )
+
+        # projection of the input hidden states
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size,
+            projection_size,
+            bias=config.use_bias,
+        )
+        # selective projection used to make dt, B and C input dependant
+
+        # time step projection (discretization)
+        # instantiate once and copy inv_dt in init_weights of PretrainedModel
+        self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
+
+        # S4D real initialization. These are not discretized!
+        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
+        A = torch.arange(1, self.num_heads + 1)
+        self.A_log = nn.Parameter(torch.log(A))
+        self.A_log._no_weight_decay = True
+        self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon, group_size=self.intermediate_size // self.n_groups)
+        self.D = nn.Parameter(torch.ones(self.num_heads))
+        self.D._no_weight_decay = True
+
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.use_bias = config.use_bias
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
+                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
+                " https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def cuda_kernels_forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        # 1. Gated MLP's linear projection
+        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+        projected_states = self.in_proj(hidden_states)
+
+        # Set up dimensions for reshapes later
+        batch_size, seq_len, _ = hidden_states.shape
+        groups_time_state_size = self.n_groups * self.ssm_state_size
+        d_mlp = (
+            projected_states.shape[-1]
+            - 2 * self.intermediate_size
+            - 2 * self.n_groups * self.ssm_state_size
+            - self.num_heads
+        ) // 2
+
+        # Single step calculations via cache
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
+                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+            )
+
+            # 2. Convolution sequence transformation
+            hidden_states_B_C = causal_conv1d_update(
+                hidden_states_B_C,
+                cache_params.conv_states[self.layer_idx],
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                self.activation,
+            )
+
+            hidden_states, B, C = torch.split(
+                hidden_states_B_C,
+                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                dim=-1,
+            )
+
+            # 3. SSM transformation
+            A = -torch.exp(self.A_log.float())  # (nheads,)
+            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
+            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
+            D = self.D[:, None, ...].expand(-1, self.head_dim)
+            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
+            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
+            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
+            hidden_states = selective_state_update(
+                cache_params.ssm_states[self.layer_idx],
+                hidden_states_reshaped,
+                dt,
+                A,
+                B,
+                C,
+                D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+            )
+            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
+            hidden_states = self.norm(hidden_states, gate)
+
+            # 4. Final linear projection
+            out = self.out_proj(hidden_states)[:, None, ...]
+
+        # Fused calculations or step by step if no initialized cache is found
+        else:
+            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
+            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
+
+            # 2-4. Fused kernel for conv1d, SSM, and the final projection
+            if self.training and cache_params is None:
+                out = mamba_split_conv1d_scan_combined(
+                    projected_states,
+                    self.conv1d.weight.squeeze(1),
+                    self.conv1d.bias,
+                    self.dt_bias,
+                    A,
+                    D=self.D,
+                    chunk_size=self.chunk_size,
+                    seq_idx=None,  # was seq_idx
+                    activation=self.activation,
+                    rmsnorm_weight=self.norm.weight,
+                    rmsnorm_eps=self.norm.variance_epsilon,
+                    outproj_weight=self.out_proj.weight,
+                    outproj_bias=self.out_proj.bias,
+                    headdim=self.head_dim,
+                    ngroups=self.n_groups,
+                    norm_before_gate=False,
+                    return_final_states=False,
+                    **dt_limit_kwargs,
+                )
+
+            else:
+                _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+                )
+
+                # 2. Convolution sequence transformation
+                # Init cache
+                if cache_params is not None:
+                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                    conv_states = nn.functional.pad(
+                        hidden_states_B_C_transposed,
+                        (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
+                    )
+                    cache_params.update_conv_state(
+                        layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
+                    )
+
+                if self.activation not in ["silu", "swish"]:
+                    hidden_states_B_C = self.act(
+                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
+                    )
+                else:
+                    hidden_states_B_C = causal_conv1d_fn(
+                        x=hidden_states_B_C.transpose(1, 2),
+                        weight=self.conv1d.weight.squeeze(1),
+                        bias=self.conv1d.bias,
+                        activation=self.activation,
+                    ).transpose(1, 2)
+                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+                hidden_states, B, C = torch.split(
+                    hidden_states_B_C,
+                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                    dim=-1,
+                )
+
+                # 3. SSM transformation
+                scan_output, ssm_state = mamba_chunk_scan_combined(
+                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
+                    dt,
+                    A,
+                    B.view(batch_size, seq_len, self.n_groups, -1),
+                    C.view(batch_size, seq_len, self.n_groups, -1),
+                    chunk_size=self.chunk_size,
+                    D=self.D,
+                    z=None,
+                    seq_idx=None,
+                    return_final_states=True,
+                    dt_bias=self.dt_bias,
+                    dt_softplus=True,
+                    **dt_limit_kwargs,
+                )
+
+                # Init cache
+                if ssm_state is not None and cache_params is not None:
+                    cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+                scan_output = scan_output.view(batch_size, seq_len, -1)
+
+                # Multiply "gate" branch and apply extra normalization layer
+                scan_output = self.norm(scan_output, gate)
+
+                # 4. Final linear projection
+                out = self.out_proj(scan_output)
+        return out
+
+    # fmt: off
+    def torch_forward(self, input_states, cache_params: Optional[HybridMambaAttentionDynamicCache]=None, cache_position:Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
+
+        # 1. Gated MLP's linear projection
+        input_states = apply_mask_to_padding_states(input_states, attention_mask)
+        projected_states = self.in_proj(input_states)
+        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
+        _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
+        )
+
+        # 2. Convolution sequence transformation
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
+
+            # We need to guarantee that anything regarding the cache is on the same device
+            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
+
+            hidden_states_B_C = torch.sum(
+                conv_states * self.conv1d.weight.squeeze(1), dim=-1
+            )
+            if self.use_conv_bias:
+                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
+            hidden_states_B_C = self.act(hidden_states_B_C)
+        else:
+            # Init cache
+            if cache_params is not None:
+                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                conv_states = nn.functional.pad(
+                    hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
+                )
+                cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
+
+            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))
+
+        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+        hidden_states, B, C = torch.split(
+            hidden_states_B_C,
+            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
+            dim=-1
+        )
+
+        # 3. SSM transformation
+        A = -torch.exp(self.A_log.float())                            # [num_heads]
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            # We need to guarantee that anything regarding the cache is on the same device
+            cache_device = cache_params.ssm_states.device
+
+            # Note: there is no need to pad parameter matrices here, as there is just one new token
+            # for batched generation
+            dt = dt[:, 0, :][:, None, ...]
+            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
+            # [num_heads] -> [num_heads, head_dim]
+            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
+
+            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            # [bsz, num_heads, head_dim, state_size]
+            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
+
+            # Discretize B
+            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
+            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
+            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
+            B = B.reshape(batch_size, -1, B.shape[-1])
+            # [bsz, num_heads, head_dim, state_size]
+            dB = dt[..., None] * B[..., None, :]
+
+            # Discretize x into dB
+            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
+            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
+            dBx = (dB * hidden_states[..., None]).to(device=cache_device)
+
+            # State calculation
+            cache_params.update_ssm_state(
+                layer_idx=self.layer_idx,
+                new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
+            )
+
+            # Subsequent output
+            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
+            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
+            C = C.reshape(batch_size, -1, C.shape[-1])
+            # [bsz, num_heads, head_dim]
+
+            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
+            # Reshape ssm_states to merge the first two dimensions
+            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
+            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
+            y = torch.bmm(ssm_states_reshaped, C_reshaped)
+            y = y.view(batch_size, self.num_heads, self.head_dim)
+
+            # D skip connection
+            # [num_heads] -> [num_heads, head_dim]
+            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
+            y = (y + hidden_states * D).to(y.dtype)
+
+            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
+            y = y.reshape(batch_size, -1)[:, None, ...]
+        else:
+            # begin ssd naive implementation without einsums
+            dt = nn.functional.softplus(dt + self.dt_bias)
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
+            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            B = B.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
+            C = C.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
+            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
+
+            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
+
+            # Discretize x and A
+            hidden_states = hidden_states * dt[..., None]
+            A = A.to(hidden_states.dtype) * dt
+
+            # Rearrange into blocks/chunks
+            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
+
+            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
+            A = A.permute(0, 3, 1, 2)
+            A_cumsum = torch.cumsum(A, dim=-1)
+
+            # 1. Compute the output for each intra-chunk (diagonal blocks)
+            # This is the analog of a causal mask
+            L = torch.exp(segment_sum(A))
+
+            # Contraction of C and B to get G (attention-weights like)
+            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]  # shape: (b, c, l, s, h, n)
+            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)
+
+            # Compute M, equivalent to applying attention mask to weights
+            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
+            M = M_intermediate.sum(dim=-1)
+
+            # Compute Y_diag (apply to values)
+            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
+
+            # 2. Compute the state for each intra-chunk
+            # (right term of low-rank factorization of off-diagonal blocks; B terms)
+            decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
+            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
+            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
+
+            # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
+            # (middle term of factorization of off-diag blocks; A terms)
+            if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
+            else:
+                previous_states = torch.zeros_like(states[:, :1])
+            states = torch.cat([previous_states, states], dim=1)
+            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
+            states, ssm_state = new_states[:, :-1], new_states[:, -1]
+
+            # 4. Compute state -> output conversion per chunk
+            # (left term of low-rank factorization of off-diagonal blocks; C terms)
+            state_decay_out = torch.exp(A_cumsum)
+            C_times_states = (C[..., None, :] * states[:, :, None, ...])
+            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
+            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
+
+            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
+            y = Y_diag + Y_off
+            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
+            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
+
+            y = y + D_residual
+            # Cutting off padded chunks
+            if pad_size > 0:
+                y = y[:, :seq_len, :, :]
+            y = y.reshape(batch_size, seq_len, -1)
+
+            # Init cache
+            if ssm_state is not None and cache_params is not None:
+                cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+        scan_output = self.norm(y, gate)
+
+        # end ssd naive
+
+        # 4. Final linear projection
+        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
+        return contextualized_states
+    # fmt: on
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
+            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
+        dtype = hidden_states.dtype
+        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
+            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)
+
+
+class NemotronHRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        NemotronHRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        # Weights are in float32
+        return (self.weight.to(torch.float32) * hidden_states).to(input_dtype)
+
+class NemotronHBlock(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.residual_in_fp32 = config.residual_in_fp32
+        self.norm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        # M: Mamba2, *: Attention, -: MLP
+        self.block_type = config.layers_block_type[layer_idx]
+        if self.block_type == "mamba":
+            self.mixer = NemotronHMamba2Mixer(config, layer_idx=layer_idx)
+        elif self.block_type == "attention":
+            self.mixer = NEMOTRONH_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+        elif self.block_type == "mlp":
+            self.mixer = NemotronHMLP(config, layer_idx=layer_idx)
+        elif self.block_type == "moe":
+            self.mixer = NemotronHMOE(config, layer_idx=layer_idx)
+        else:
+            raise ValueError(f"Invalid layer pattern {config.hybrid_override_pattern[layer_idx]}")
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        with torch.cuda.stream(torch.cuda.default_stream(hidden_states.device)):
+            # * Use torch.cuda.stream() to avoid NaN issues when using multiple GPUs
+            residual = hidden_states
+            hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
+            if self.residual_in_fp32:
+                residual = residual.to(torch.float32)
+
+            if self.block_type == "mamba":
+                hidden_states = self.mixer(
+                    hidden_states, cache_params=cache_params, cache_position=cache_position
+                )
+            elif self.block_type == "attention":
+                hidden_states = self.mixer(
+                    hidden_states, cache_position=cache_position
+                )
+                hidden_states = hidden_states[0]
+            elif self.block_type in ["mlp", "moe"]:
+                hidden_states = self.mixer(
+                    hidden_states
+                )
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            hidden_states = residual + hidden_states
+            return hidden_states
+
+
+# Copied from transformers.models.nemotron.modeling_nemotron Nemotron->NemotronH
+class NemotronHMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.intermediate_size
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.mlp_hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.up_proj(x)))
+
+
+class NemotronHMOE(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.experts = nn.ModuleList(
+            [
+                NemotronHMLP(config, intermediate_size=config.moe_intermediate_size, layer_idx=layer_idx)
+                for _ in range(config.n_routed_experts)
+            ]
+        )
+        self.gate = NemotronHTopkRouter(config)
+        self.shared_experts = NemotronHMLP(
+            config=config, intermediate_size=config.moe_shared_expert_intermediate_size, layer_idx=layer_idx
+        )
+
+    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
+        r"""
+        CALL FOR CONTRIBUTION! I don't have time to optimise this right now, but expert weights need to be fused
+        to not have to do a loop here (deepseek has 256 experts soooo yeah).
+        """
+        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)
+        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts))
+        expert_mask = expert_mask.permute(2, 0, 1)
+
+        for expert_idx in range(len(self.experts)):
+            expert = self.experts[expert_idx]
+            mask = expert_mask[expert_idx]
+            token_indices, weight_indices = torch.where(mask)
+
+            if token_indices.numel() > 0:
+                expert_weights = topk_weights[token_indices, weight_indices]
+                expert_input = hidden_states[token_indices]
+                expert_output = expert(expert_input)
+                weighted_output = expert_output * expert_weights.unsqueeze(-1)
+                final_hidden_states.index_add_(0, token_indices, weighted_output)
+            else:
+                # Local empty expert: no-op compute that still marks params as used.
+                expert_dtype = expert.down_proj.weight.dtype
+                dummy_out = expert(torch.zeros_like(hidden_states[0]).unsqueeze(0).to(expert_dtype))
+                final_hidden_states = final_hidden_states + dummy_out
+
+        # in original deepseek, the output of the experts are gathered once we leave this module
+        # thus the moe module is itelsf an IsolatedParallel module
+        # and all expert are "local" meaning we shard but we don't gather
+        return final_hidden_states.type(hidden_states.dtype)
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        topk_indices, topk_weights = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class NemotronHTopkRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size), dtype=torch.float32))
+        self.register_buffer("e_score_correction_bias", torch.zeros(self.n_routed_experts, dtype=torch.float32))
+
+    @torch.no_grad()
+    def get_topk_indices(self, scores):
+        scores_for_choice = scores.view(-1, self.n_routed_experts) + self.e_score_correction_bias.unsqueeze(0)
+        group_scores = (
+            scores_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        return topk_indices
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        scores = router_logits.sigmoid()
+        topk_indices = self.get_topk_indices(scores)
+        topk_weights = scores.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class NemotronHAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: NemotronHConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        if hasattr(config, "head_dim") and config.head_dim is not None:
+            self.head_dim = config.head_dim
+        else:
+            self.head_dim = config.hidden_size // self.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.head_dim * self.num_heads, self.hidden_size, bias=config.attention_bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        # position_embeddings: Tuple[torch.Tensor, torch.Tensor], #TODO
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        #attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.num_heads * self.head_dim)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
+#class JambaFlashAttention2(JambaAttention):
+class NemotronHFlashAttention2(NemotronHAttention):
+    """
+    Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        #attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
+#class JambaSdpaAttention(JambaAttention):
+class NemotronHSdpaAttention(NemotronHAttention):
+    """
+    Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from NemotronHAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "NemotronHModel is using NemotronHSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+        is_causal = True if self.is_causal and causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+NEMOTRONH_ATTENTION_CLASSES = {
+    "eager": NemotronHAttention,
+    "flash_attention_2": NemotronHFlashAttention2,
+    "sdpa": NemotronHSdpaAttention,
+}
+
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2PreTrainedModel
+class NemotronHPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = NemotronHConfig
+    base_model_prefix = "backbone"
+    _no_split_modules = ["NemotronHBlock"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, NemotronHMamba2Mixer):
+            if getattr(module.dt_bias, "_is_hf_initialized", False):
+                return
+            module.A_log._no_weight_decay = True
+            module.D._no_weight_decay = True
+
+            dt = torch.exp(
+                torch.rand(self.config.mamba_num_heads)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+
+            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                module.dt_bias.copy_(inv_dt)
+            module.dt_bias._no_reinit = True
+
+        if isinstance(module, nn.Linear):
+            if module.bias is not None:
+                if not getattr(module.bias, "_no_reinit", False):
+                    nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=self.config.initializer_range)
+
+        # TODO: Check
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            for name, p in module.named_parameters():
+                if getattr(p, "_is_hf_initialized", False):
+                    continue
+                if name in ["out_proj.weight"]:
+                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                    # We need to reinit p since this code could be called multiple times
+                    # Having just p *= scale would repeatedly scale it down
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(self.config.num_hidden_layers)
+
+
+@dataclass
+# Copied from transformers.models.mamba.modeling_mamba2.Mamba2Output with MAMBA2->NemotronH,Mamba2->NemotronH
+class NemotronHOutput(ModelOutput):
+    """
+    Class for the NemotronH model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.mamba2.modeling_mamba2.MambaCausalLMOutput with Mamba2->NemotronH
+class NemotronHCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`HybridMambaAttentionDynamicCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+NEMOTRONH_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`NemotronHConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+NEMOTRONH_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary.
+
+            If `cache_params.seqlen_offset>0`, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        position_ids (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        cache_params (`HybridMambaAttentionDynamicCache`, *optional*):
+            If passed along, the model uses the previous state in all the blocks (which will give the output for the
+            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
+        use_cache (`bool`, *optional*):
+            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            The position of the current input in the cache. This is used to ensure that the cache is correctly updated.
+            If `cache_params` is passed, `cache_position` should also be passed.
+        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+"""
+
+
+@add_start_docstrings(
+    "The bare NemotronH Model transformer outputting raw hidden-states without any specific head on top.",
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHModel(NemotronHPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([NemotronHBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        # Initialize weights and apply final processing
+        self._register_load_state_dict_pre_hook(self.load_hook)
+        self.post_init()
+
+    def load_hook(self, state_dict, prefix, *args):
+        for k in state_dict:
+            if "embedding." in k:
+                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
+                break
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, NemotronHOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        # use_cache = use_cache if use_cache is not None else self.config.use_cache
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # From zamba_modeling.py
+        if use_cache and cache_params is None:
+            logger.warning_once(
+                "NemotronH requires an initialized `NemotronHHybridDynamicCache` to return a cache. None was "
+                "provided, so no cache will be returned."
+            )
+
+        hidden_states = inputs_embeds
+
+        if cache_position is None:
+            cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
+        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        # Until HERE
+
+        for layer_idx, mixer_block in enumerate(self.layers):
+            # Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
+            if mixer_block.block_type == "mamba":
+                layer_mask = mamba_mask
+            elif mixer_block.block_type == "attention":
+                layer_mask = causal_mask
+            elif mixer_block.block_type in ["mlp", "moe"]:
+                layer_mask = None
+            else:
+                raise ValueError(f"Invalid block_type: {self.block_type}")
+
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__, hidden_states, cache_params, cache_position, layer_mask
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    cache_position=cache_position,
+                    attention_mask=layer_mask,
+                )
+
+            # TODO: Store attentions
+            # if output_attentions:
+            #     if layer_outputs[1] is not None:
+            #         # append attentions only of attention layers. Mamba layers return `None` as the attention weights
+            #         all_self_attns += (layer_outputs[1],)
+
+            # TODO (Check): should it happen before the forward pass?
+            # if output_hidden_states:
+            #     all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return NemotronHOutput(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    # Copied from transformers.models.jamba.modeling_jamba.JambaModel._update_causal_mask
+    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        target_length = cache_position[-1] + 1
+
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    def _update_mamba_mask(self, attention_mask, cache_position):
+        """
+        No need for zeroing states when
+            1. Cached forward
+            2. Attending to all inputs
+        """
+        mamba_mask = attention_mask
+        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
+            mamba_mask = None
+        return mamba_mask
+
+
+@add_start_docstrings(
+    """
+    The NEMOTRONH Model transformer with a language modeling head on top (linear layer with weights not tied to the input
+    embeddings).
+    """,
+    NEMOTRONH_START_DOCSTRING,
+)
+class NemotronHForCausalLM(NemotronHPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = NemotronHModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
+        # Overwitten -- uses `cache_params` as opposed to `past_key_values`
+        empty_past_kv = past_key_values is None
+
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
+        #              (we can't check exception 3 while compiling)
+        if not empty_past_kv:
+            if (
+                inputs_embeds is not None  # Exception 1
+                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
+            ):
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+        else:
+            past_key_values = HybridMambaAttentionDynamicCache(
+                self.config, input_ids.shape[0], self.dtype, device=self.device
+            )
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if not empty_past_kv:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and empty_past_kv:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+                "logits_to_keep": self.config.num_logits_to_keep,
+                "cache_position": cache_position,
+            }
+        )
+        return model_inputs
+
+    @add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=NemotronHCausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,  # for now we need this for generation
+    ) -> Union[Tuple, NemotronHCausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        nemotron_h_outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = nemotron_h_outputs[0]
+
+        # TODO: Check zamba_modeling.py: https://github.com/huggingface/transformers/blob/d7188ba600e36d3fd191b12e19f1b3bb81a8404f/src/transformers/models/zamba/modeling_zamba.py#L1284C1-L1286C2
+        #logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+        logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + nemotron_h_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return NemotronHCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=nemotron_h_outputs.cache_params,
+            hidden_states=nemotron_h_outputs.hidden_states,
+            attentions=nemotron_h_outputs.attentions,
+        )

nemo-evaluator-launcher-configs/local_nvidia-nemotron-3-nano-30b-a3b-base.yaml

@@ -0,0 +1,78 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+#
+# How to use:
+#
+# 1. copy this file locally or clone the repository
+# 2. (optional) uncomment limit_samples in the config file to run with 10 samples for quick testing
+# 3. export your HF token in the terminal (some benchmark datasets might be gated)
+# 4. run `nemo-evaluator-launcher run --config path/to/local_nvidia-nemotron-nano-3-30b-a3b-base.yaml`
+#
+# ⚠️  WARNING:
+#     Always run full evaluations (without limit_samples) for actual benchmark results.
+#     Using a subset of samples is solely for testing configuration and setup.
+#     Results from such test runs should NEVER be used to compare models or
+#     report benchmark performance.
+defaults:
+  - execution: local
+  - deployment: vllm
+  - _self_
+
+execution:
+  output_dir: NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
+  # mode: sequential  # enables sequential execution
+
+# specify deployment arguments
+deployment:
+  image: vllm/vllm-openai:v0.12.0
+  checkpoint_path: null
+  hf_model_handle: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
+  served_model_name: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
+  tensor_parallel_size: 1
+  data_parallel_size: 1
+  extra_args: "--max-model-len 262144 --mamba_ssm_cache_dtype float32 --no-enable-prefix-caching"
+
+# specify the benchmarks to evaluate
+evaluation:
+  env_vars:
+    HF_TOKEN: HF_TOKEN
+  nemo_evaluator_config:  # global config settings that apply to all tasks
+    config:
+      params:
+        max_retries: 5  # number of retries for API requests
+        request_timeout: 360  # timeout for API requests in seconds
+        parallelism: 4  # number of parallel requests
+        # limit_samples: 10 # uncomment to limit number of samples for quick testing
+        extra:
+          tokenizer: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
+          tokenizer_backend: huggingface
+  tasks:
+    - name: adlr_mmlu_pro_5_shot_base
+    - name: adlr_mmlu
+    - name: adlr_agieval_en_cot
+    - name: adlr_humaneval_greedy
+    - name: adlr_mbpp_sanitized_3_shot_greedy
+    - name: adlr_gsm8k_cot_8_shot
+    - name: adlr_minerva_math_nemo_4_shot
+    - name: adlr_math_500_4_shot_sampled
+    - name: adlr_arc_challenge_llama_25_shot
+    - name: hellaswag
+    - name: openbookqa
+    - name: piqa
+    - name: adlr_race
+    - name: adlr_winogrande_5_shot
+    - name: adlr_global_mmlu_lite_5_shot
+    - name: adlr_mgsm_native_cot_8_shot

nemo-evaluator-launcher-configs/local_qwen3-30b-a3b-base.yaml

@@ -0,0 +1,77 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+#
+# How to use:
+#
+# 1. copy this file locally
+# 2. (optional) uncomment limit_samples in the config file to run with 10 samples for quick testing
+# 3. export your HF token in the terminal (some benchmark datasets might be gated)
+# 4. run `nemo-evaluator-launcher run --config path/to/local_qwen3-30b-a3b-base.yaml`
+#
+# ⚠️  WARNING:
+#     Always run full evaluations (without limit_samples) for actual benchmark results.
+#     Using a subset of samples is solely for testing configuration and setup.
+#     Results from such test runs should NEVER be used to compare models or
+#     report benchmark performance.
+defaults:
+  - execution: local
+  - deployment: vllm
+  - _self_
+
+execution:
+  output_dir: Qwen3-30B-A3B-Base
+  # mode: sequential  # enables sequential execution
+
+# specify deployment arguments
+deployment:
+  image: vllm/vllm-openai:v0.11.0
+  checkpoint_path: null
+  hf_model_handle: Qwen/Qwen3-30B-A3B-Base
+  served_model_name: Qwen/Qwen3-30B-A3B-Base
+  tensor_parallel_size: 1
+  data_parallel_size: 1
+
+# specify the benchmarks to evaluate
+evaluation:
+  env_vars:
+    HF_TOKEN: HF_TOKEN
+  nemo_evaluator_config:  # global config settings that apply to all tasks
+    config:
+      params:
+        max_retries: 5  # number of retries for API requests
+        request_timeout: 360  # timeout for API requests in seconds
+        parallelism: 4  # number of parallel requests
+        # limit_samples: 10 # uncomment to limit number of samples for quick testing
+        extra:
+          tokenizer: Qwen/Qwen3-30B-A3B-Base
+          tokenizer_backend: huggingface
+  tasks:
+    - name: adlr_mmlu_pro_5_shot_base
+    - name: adlr_mmlu
+    - name: adlr_agieval_en_cot
+    - name: adlr_humaneval_greedy
+    - name: adlr_mbpp_sanitized_3_shot_greedy
+    - name: adlr_gsm8k_cot_8_shot
+    - name: adlr_minerva_math_nemo_4_shot
+    - name: adlr_math_500_4_shot_sampled
+    - name: adlr_arc_challenge_llama_25_shot
+    - name: hellaswag
+    - name: openbookqa
+    - name: piqa
+    - name: adlr_race
+    - name: adlr_winogrande_5_shot
+    - name: adlr_global_mmlu_lite_5_shot
+    - name: adlr_mgsm_native_cot_8_shot

privacy.md

@@ -0,0 +1,13 @@
+| Field | Response |
+| :---- | :---- |
+| Generatable or reverse engineerable personal data? | No |
+| Personal data used to create this model? | No |
+| Was consent obtained for any personal data used? | Not Applicable |
+| A description of any methods implemented in data acquisition or processing, if any, to address the prevalence of personal data in the training data, where relevant and applicable. | We used only prompts that do not contain any personal data for synthetic data generation. |
+| How often is the dataset reviewed? | Before Release |
+| Is there provenance for all datasets used in training? | Yes |
+| Does data labeling (annotation, metadata) comply with privacy laws? | Yes |
+| Is data compliant with data subject requests for data correction or removal, if such a request was made? | No, not possible with externally-sourced data. |
+| Applicable Privacy Policy | [NVIDIA Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) |
+| During AI model development, strict adherence to copyright policy ensured compliance through risk mitigation and legal reviews. Post-data collection, reserved rights content is identified and removed, with verified opt-out processes for rightsholders. Detailed records document due diligence and transparency. | True |
+| We employ automated tools and data processing techniques during pre-training to identify and filter certain categories of personal information. Scans of training datasets detected no PII. | True. We employ automated tools and data processing techniques to scan for Personally Identifiable Information (PII) during pre-training to identify and filter certain categories of personal information, including public-facing contact details such as email addresses and phone numbers. Scans of Common Crawl, CC-News, and Wikimedia datasets did not detect PII in the majority of samples. However, Microsoft Presidio indicated potential findings including business contact information embedded in natural language, such as email addresses and phone numbers. These were removed using verified instances of PII through a combination of automated filtering and human-in-the-loop validation. This evaluation used a 3,000-sample subset per dataset, identified as the optimal threshold for maximizing embedder accuracy. |

safety.md

@@ -0,0 +1,9 @@
+| Field | Response |
+| :---- | :---- |
+| Model Application Field(s): | Chat, Instruction Following, Chatbot Development, Code Generation, Reasoning, Customer Service |
+| Describe the life critical impact (if present). | Not Applicable |
+| Description of methods implemented in data acquisition or processing, if any, to address other types of potentially harmful data in the training, testing, and validation data: | We used a guard model for content safety to exclude potentially harmful data from training.  |
+| Description of any methods implemented in data acquisition or processing, if any, to address illegal or harmful content in the training data, including, but not limited to, child sexual abuse material (CSAM) and non-consensual intimate imagery (NCII) | We used a Gemma-3 4B-based guard model trained on [Nemotron Content Safety Dataset v2](https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0) for content safety to exclude potentially illegal or harmful content from the training. |
+| Use Case Restrictions: | Abide by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). |
+| Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development.  Restrictions enforce dataset access during training, and dataset license constraints adhered to. |
+| This AI model was developed based on our policies to ensure responsible data handling and risk mitigation. The datasets used for training have been scanned for harmful content and illegal content, consistent with our policies including scanning for Child Sexual Abuse Material (CSAM). Ongoing review and monitoring mechanisms are in place based on our policies and to maintain data integrity. | True. We use [Nemotron Content Safety Dataset V2](https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0) and an internal safety dataset specialized for minority sexuality for content safety evaluation to ensure the safety of this model. |

special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:623c34567aebb18582765289fbe23d901c62704d6518d71866e0e58db892b5b7
+size 17077484