Add Nemotron HF Support (#31699)

* Add nemotron support

* fix inference

* add unit test

* add layernorm1p as a class to avoid meta device mismatch

* test fixed

* Add copied_from statements

* remove pretraining_tp args

* remove nemotronlayernorm

* force LN computation done in FP32

* remove nemotrontokenizer and use llamatokenizer

* license update

* add option for kv_channels for minitron8b

* remove assert

* o_proj fixed

* o_proj reshape

* add gated_proj option

* typo

* remove todos

* fix broken test after merging latest main

* remove nezha/nat after meging main

* chnage default config to 15b model

* add nemo conversion script

* rename conversion script

* remove gate_proj option

* pr comment resolved

* fix unit test

* rename kv_channels to head_dim

* resolve PR issue

* add nemotron md

* fix broken tests

* refactor rope for nemotron

* test fix

* remove linearscaling

* whitespace and import

* fix some copied-from

* code style fix

* reformatted

* add position_embedding to nemotronattention

* rope refactor to only use config, copied-from fix

* format

* Run make fix-copies

* nemotron md with autodoc

* doc  fix

* fix order

* pass check_config_docstrings.py

* fix config_attributes

* remove all llama BC related code

* Use PreTrainedTokenizerFast

* ruff check examples

* conversion script update

* add nemotron to toctree

docs/source/en/_toctree.yml

@@ -468,6 +468,8 @@
         title: MT5
       - local: model_doc/mvp
         title: MVP
+      - local: model_doc/nemotron
+        title: Nemotron
       - local: model_doc/nezha
         title: NEZHA
       - local: model_doc/nllb

docs/source/en/index.md

@@ -222,6 +222,7 @@ Flax), PyTorch, and/or TensorFlow.
 |               [MusicGen Melody](model_doc/musicgen_melody)               |       ✅        |         ❌         |      ❌      |
 |                           [MVP](model_doc/mvp)                           |       ✅        |         ❌         |      ❌      |
 |                           [NAT](model_doc/nat)                           |       ✅        |         ❌         |      ❌      |
+|                      [Nemotron](model_doc/nemotron)                      |       ✅        |         ❌         |      ❌      |
 |                         [Nezha](model_doc/nezha)                         |       ✅        |         ❌         |      ❌      |
 |                          [NLLB](model_doc/nllb)                          |       ✅        |         ❌         |      ❌      |
 |                      [NLLB-MOE](model_doc/nllb-moe)                      |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/nemotron.md

+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+Copyright (c) 2024, 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.
+
+-->
+
+# Nemotron
+
+## Nemotron
+
+### License
+
+The use of this model is governed by the [NVIDIA AI Foundation Models Community License Agreement](https://developer.nvidia.com/downloads/nv-ai-foundation-models-license).
+
+### Description
+
+Nemotron-4 is a family of enterprise ready generative text models compatible with [NVIDIA NeMo Framework](https://www.nvidia.com/en-us/ai-data-science/generative-ai/nemo-framework/).
+
+NVIDIA NeMo is an end-to-end, cloud-native platform to build, customize, and deploy generative AI models anywhere. It includes training and inferencing frameworks, guardrailing toolkits, data curation tools, and pretrained models, offering enterprises an easy, cost-effective, and fast way to adopt generative AI. To get access to NeMo Framework, please sign up at [this link](https://developer.nvidia.com/nemo-framework/join).
+
+### References
+
+[Announcement Blog](https://developer.nvidia.com/blog/nvidia-ai-foundation-models-build-custom-enterprise-chatbots-and-co-pilots-with-production-ready-llms/)
+
+### Model Architecture
+
+**Architecture Type:** Transformer
+
+**Network Architecture:** Transformer Decoder (auto-regressive language model).
+
+## Minitron
+
+### Minitron 4B Base
+
+Minitron is a family of small language models (SLMs) obtained by pruning NVIDIA's [Nemotron-4 15B](https://arxiv.org/abs/2402.16819) model. We prune model embedding size, attention heads, and MLP intermediate dimension, following which, we perform continued training with distillation to arrive at the final models.
+
+Deriving the Minitron 8B and 4B models from the base 15B model using our approach requires up to **40x fewer training tokens** per model compared to training from scratch; this results in **compute cost savings of 1.8x** for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. Please refer to our [arXiv paper](https://arxiv.org/abs/2407.14679) for more details.
+
+Minitron models are for research and development only.
+
+### HuggingFace Quickstart
+
+The following code provides an example of how to load the Minitron-4B model and use it to perform text generation.
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# Load the tokenizer and model
+model_path = 'nvidia/Minitron-4B-Base'
+tokenizer  = AutoTokenizer.from_pretrained(model_path)
+
+device = 'cuda'
+dtype  = torch.bfloat16
+model  = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=dtype, device_map=device)
+
+# Prepare the input text
+prompt = 'Complete the paragraph: our solar system is'
+inputs = tokenizer.encode(prompt, return_tensors='pt').to(model.device)
+
+# Generate the output
+outputs = model.generate(inputs, max_length=20)
+
+# Decode and print the output
+output_text = tokenizer.decode(outputs[0])
+print(output_text)
+```
+
+### License
+
+Minitron is released under the [NVIDIA Open Model License Agreement](https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf).
+
+### Evaluation Results
+
+*5-shot performance.* Language Understanding evaluated using [Massive Multitask Language Understanding](https://arxiv.org/abs/2009.03300):
+
+| Average |
+| :---- |
+| 58.6 |
+
+*Zero-shot performance.* Evaluated using select datasets from the [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) with additions:
+
+| HellaSwag | Winogrande | GSM8K| ARC-C | XLSum |
+| :------------- | :------------- | :------------- | :------------- | :------------- |
+| 75.0 | 74.0 | 24.1  | 50.9 | 29.5
+
+
+*Code generation performance*. Evaluated using [HumanEval](https://github.com/openai/human-eval):
+
+| p@1, 0-Shot |
+| :------------- |
+| 23.3 |
+
+Please refer to our [paper](https://arxiv.org/abs/2407.14679) for the full set of results.
+
+### Citation
+
+If you find our work helpful, please consider citing our paper:
+```
+@article{minitron2024,
+      title={Compact Language Models via Pruning and Knowledge Distillation},
+      author={Saurav Muralidharan and Sharath Turuvekere Sreenivas and Raviraj Joshi and Marcin Chochowski and Mostofa Patwary and Mohammad Shoeybi and Bryan Catanzaro and Jan Kautz and Pavlo Molchanov},
+      journal={arXiv preprint arXiv:2407.14679},
+      year={2024},
+      url={https://arxiv.org/abs/2407.14679},
+}
+```
+
+## NemotronConfig
+
+[[autodoc]] NemotronConfig
+
+
+## NemotronModel
+
+[[autodoc]] NemotronModel
+    - forward
+
+
+## NemotronForCausalLM
+
+[[autodoc]] NemotronForCausalLM
+    - forward
+
+## NemotronForSequenceClassification
+
+[[autodoc]] NemotronForSequenceClassification
+    - forward
+
+
+## NemotronForQuestionAnswering
+
+[[autodoc]] NemotronForQuestionAnswering
+    - forward
+
+
+## NemotronForTokenClassification
+
+[[autodoc]] NemotronForTokenClassification
+    - forward
\ No newline at end of file

docs/source/en/perf_infer_gpu_one.md

@@ -67,6 +67,7 @@ FlashAttention-2 is currently supported for the following architectures:
 * [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [Nemotron](https://huggingface.co/docs/transformers/model_doc/nemotron)
 * [NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)
 * [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
 * [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel)
@@ -228,6 +229,7 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [Nemotron](https://huggingface.co/docs/transformers/model_doc/nemotron)
 * [ViT](https://huggingface.co/docs/transformers/model_doc/vit#transformers.ViTModel)
 * [ViTHybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid#transformers.ViTHybridModel)
 * [ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae#transformers.ViTMAEModel)

src/transformers/__init__.py

@@ -592,6 +592,7 @@ _import_structure = {
         "MusicgenMelodyDecoderConfig",
     ],
     "models.mvp": ["MvpConfig", "MvpTokenizer"],
+    "models.nemotron": ["NemotronConfig"],
     "models.nllb": [],
     "models.nllb_moe": ["NllbMoeConfig"],
     "models.nougat": ["NougatProcessor"],
@@ -2742,6 +2743,16 @@ else:
             "MvpPreTrainedModel",
         ]
     )
+    _import_structure["models.nemotron"].extend(
+        [
+            "NemotronForCausalLM",
+            "NemotronForQuestionAnswering",
+            "NemotronForSequenceClassification",
+            "NemotronForTokenClassification",
+            "NemotronModel",
+            "NemotronPreTrainedModel",
+        ]
+    )
     _import_structure["models.nllb_moe"].extend(
         [
             "NllbMoeForConditionalGeneration",
@@ -5286,6 +5297,7 @@ if TYPE_CHECKING:
         MusicgenMelodyDecoderConfig,
     )
     from .models.mvp import MvpConfig, MvpTokenizer
+    from .models.nemotron import NemotronConfig
     from .models.nllb_moe import NllbMoeConfig
     from .models.nougat import NougatProcessor
     from .models.nystromformer import (
@@ -7187,6 +7199,14 @@ if TYPE_CHECKING:
             MvpModel,
             MvpPreTrainedModel,
         )
+        from .models.nemotron import (
+            NemotronForCausalLM,
+            NemotronForQuestionAnswering,
+            NemotronForSequenceClassification,
+            NemotronForTokenClassification,
+            NemotronModel,
+            NemotronPreTrainedModel,
+        )
         from .models.nllb_moe import (
             NllbMoeForConditionalGeneration,
             NllbMoeModel,

src/transformers/models/__init__.py

@@ -159,6 +159,7 @@ from . import (
     musicgen,
     musicgen_melody,
     mvp,
+    nemotron,
     nllb,
     nllb_moe,
     nougat,

src/transformers/models/auto/configuration_auto.py

@@ -177,6 +177,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("musicgen_melody", "MusicgenMelodyConfig"),
         ("mvp", "MvpConfig"),
         ("nat", "NatConfig"),
+        ("nemotron", "NemotronConfig"),
         ("nezha", "NezhaConfig"),
         ("nllb-moe", "NllbMoeConfig"),
         ("nougat", "VisionEncoderDecoderConfig"),
@@ -469,6 +470,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("musicgen_melody", "MusicGen Melody"),
         ("mvp", "MVP"),
         ("nat", "NAT"),
+        ("nemotron", "Nemotron"),
         ("nezha", "Nezha"),
         ("nllb", "NLLB"),
         ("nllb-moe", "NLLB-MOE"),

src/transformers/models/auto/modeling_auto.py

@@ -169,6 +169,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("musicgen_melody", "MusicgenMelodyModel"),
         ("mvp", "MvpModel"),
         ("nat", "NatModel"),
+        ("nemotron", "NemotronModel"),
         ("nezha", "NezhaModel"),
         ("nllb-moe", "NllbMoeModel"),
         ("nystromformer", "NystromformerModel"),
@@ -481,6 +482,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("musicgen", "MusicgenForCausalLM"),
         ("musicgen_melody", "MusicgenMelodyForCausalLM"),
         ("mvp", "MvpForCausalLM"),
+        ("nemotron", "NemotronForCausalLM"),
         ("olmo", "OlmoForCausalLM"),
         ("open-llama", "OpenLlamaForCausalLM"),
         ("openai-gpt", "OpenAIGPTLMHeadModel"),
@@ -902,6 +904,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("mra", "MraForSequenceClassification"),
         ("mt5", "MT5ForSequenceClassification"),
         ("mvp", "MvpForSequenceClassification"),
+        ("nemotron", "NemotronForSequenceClassification"),
         ("nezha", "NezhaForSequenceClassification"),
         ("nystromformer", "NystromformerForSequenceClassification"),
         ("open-llama", "OpenLlamaForSequenceClassification"),
@@ -983,6 +986,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("mra", "MraForQuestionAnswering"),
         ("mt5", "MT5ForQuestionAnswering"),
         ("mvp", "MvpForQuestionAnswering"),
+        ("nemotron", "NemotronForQuestionAnswering"),
         ("nezha", "NezhaForQuestionAnswering"),
         ("nystromformer", "NystromformerForQuestionAnswering"),
         ("opt", "OPTForQuestionAnswering"),
@@ -1078,6 +1082,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("mpt", "MptForTokenClassification"),
         ("mra", "MraForTokenClassification"),
         ("mt5", "MT5ForTokenClassification"),
+        ("nemotron", "NemotronForTokenClassification"),
         ("nezha", "NezhaForTokenClassification"),
         ("nystromformer", "NystromformerForTokenClassification"),
         ("persimmon", "PersimmonForTokenClassification"),

src/transformers/models/nemotron/__init__.py

+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024, 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.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_nemotron": ["NemotronConfig"],
+}
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_nemotron"] = [
+        "NemotronForQuestionAnswering",
+        "NemotronForCausalLM",
+        "NemotronModel",
+        "NemotronPreTrainedModel",
+        "NemotronForSequenceClassification",
+        "NemotronForTokenClassification",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_nemotron import NemotronConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_nemotron import (
+            NemotronForCausalLM,
+            NemotronForQuestionAnswering,
+            NemotronForSequenceClassification,
+            NemotronForTokenClassification,
+            NemotronModel,
+            NemotronPreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/nemotron/configuration_nemotron.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024, 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.
+"""Nemotron model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...modeling_rope_utils import rope_config_validation
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`NemotronModel`]. It is used to instantiate an Nemotron
+    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 Nemotron-8B.
+    e.g. [nvidia/nemotron-3-8b-base-4k-hf](https://huggingface.co/nvidia/nemotron-3-8b-base-4k-hf).
+    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 256000):
+            Vocabulary size of the Nemotron model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`NemotronModel`]
+        hidden_size (`int`, *optional*, defaults to 6144):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 24576):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 48):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        head_dim (`int`, *optional*):
+            Projection weights dimension in multi-head attention. Set to hidden_size // num_attention_heads if None
+        num_key_value_heads (`int`, *optional*):
+            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. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"relu2"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.0134):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the normalization layers.
+        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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 2):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 3):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        partial_rotary_factor (`float`, *optional*, defaults to 0.5): Percentage of the query and keys which will have rotary embedding.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj and down_proj layers in the MLP layers.
+
+    ```python
+    >>> from transformers import NemotronModel, NemotronConfig
+
+    >>> # Initializing a Nemotron nemotron-15b style configuration
+    >>> configuration = NemotronConfig()
+
+    >>> # Initializing a model from the nemotron-15b style configuration
+    >>> model = NemotronModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "nemotron"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=256000,
+        hidden_size=6144,
+        intermediate_size=24576,
+        num_hidden_layers=32,
+        num_attention_heads=48,
+        head_dim=None,
+        num_key_value_heads=None,
+        hidden_act="relu2",
+        max_position_embeddings=4096,
+        initializer_range=0.0134,
+        norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=2,
+        eos_token_id=3,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        partial_rotary_factor=0.5,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.norm_eps = norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.partial_rotary_factor = partial_rotary_factor
+        rope_config_validation(self)
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+
+        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,
+        )

src/transformers/models/nemotron/convert_nemotron_nemo_to_hf.py

+# Copyright (c) 2024, 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.
+
+import json
+import os
+import shutil
+from argparse import ArgumentParser
+from collections import OrderedDict
+
+import torch
+from nemo.collections.common.tokenizers.huggingface.auto_tokenizer import AutoTokenizer
+from nemo.collections.nlp.models.language_modeling.megatron_gpt_model import MegatronGPTModel
+from nemo.collections.nlp.parts.nlp_overrides import NLPDDPStrategy
+from nemo.utils import logging
+from pytorch_lightning import Trainer
+
+from transformers import LlamaTokenizer, PreTrainedTokenizerFast
+from transformers.convert_slow_tokenizer import LlamaConverter
+
+
+"""
+Script to convert a nemotron checkpoint in nemo (mcore path) into a HuggingFace checkpoint.
+This script can be used to 1) generate only the HF weights, or 2) generate an entire HF model folder.
+
+1) Generate only HF weights from a nemo file:
+
+    python convert_nemotron_nemo_to_hf.py \
+    --input_name_or_path /path/to/file.nemo or /path/to/extracted_folder \
+    --output_path /path/to/pytorch_model.bin
+
+2) Generate the full HF model folder
+
+    python convert_nemotron_nemo_to_hf.py \
+    --input_name_or_path /path/to/file.nemo or /path/to/extracted_folder \
+    --hf_input_path /path/to/input_hf_folder \
+    --hf_output_path /path/to/output_hf_folder \
+
+    Use the --cpu-only flag if the model cannot fit in the GPU (e.g. Nemotron4 340b).
+    However this option makes the conversion script significantly slower.
+"""
+
+
+def get_args():
+    parser = ArgumentParser()
+    parser.add_argument(
+        "--input_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to .nemo file or extracted folder",
+    )
+    parser.add_argument("--output_path", type=str, default=None, required=False, help="Path to HF .bin file")
+    parser.add_argument(
+        "--hf_input_path",
+        type=str,
+        default=None,
+        help="A HF model path, " "e.g. a folder containing https://huggingface.co/nvidia/Minitron-8B-Base",
+    )
+    parser.add_argument(
+        "--hf_output_path",
+        type=str,
+        default=None,
+        help="Output HF model path, " "with the same format as above but user's own weights",
+    )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default=None,
+        help="Precision of output weights."
+        "Defaults to precision of the input nemo weights (model.cfg.trainer.precision)",
+    )
+    parser.add_argument(
+        "--cpu-only",
+        action="store_true",
+        help="Load model in cpu only. Useful if the model cannot fit in GPU memory, "
+        "but this option makes the conversion script significantly slower.",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def convert_hf_config(nemo_config, tokenizer, vocab_size, dtype, hf_output_path, hf_url="nvidia/Minitron-8B-Base"):
+    """
+    Convert NeMo config to HF config
+    """
+    NEMO_ACT2HF = {
+        "squared-relu": "relu2",
+        "fast-swiglu": "silu",
+    }
+    DTYPE2HF = {
+        torch.bfloat16: "bfloat16",
+        torch.float16: "float16",
+        torch.float32: "float32",
+    }
+    hf_config = {
+        "_name_or_path": hf_url,
+        "architectures": ["NemotronForCausalLM"],
+        "bos_token_id": tokenizer.bos_id,
+        "eos_token_id": tokenizer.eos_id,
+        "hidden_act": NEMO_ACT2HF[nemo_config.activation],
+        "hidden_size": nemo_config.hidden_size,
+        "initializer_range": nemo_config.init_method_std,
+        "intermediate_size": nemo_config.ffn_hidden_size,
+        "max_position_embeddings": nemo_config.max_position_embeddings,
+        "model_type": "nemotron",
+        "num_attention_heads": nemo_config.num_attention_heads,
+        "num_hidden_layers": nemo_config.num_layers,
+        "num_key_value_heads": nemo_config.get("num_query_groups", nemo_config.num_attention_heads),
+        "norm_eps": nemo_config.layernorm_epsilon,
+        "rope_theta": nemo_config.get("rotary_base", 10000),
+        "partial_rotary_factor": nemo_config.get("rotary_percentage", 1.0),
+        "tie_word_embeddings": False,
+        "torch_dtype": DTYPE2HF[dtype],
+        "transformers_version": "4.32.0.dev0",  # TODO
+        "use_cache": True,
+        "vocab_size": vocab_size,
+    }
+    if nemo_config.kv_channels is not None:
+        hf_config["kv_channels"] = nemo_config.kv_channels
+    json.dump(hf_config, open(f"{hf_output_path}/config.json", "w"), indent=2)
+
+
+def convert(input_nemo_file, output_hf_file, precision=None, cpu_only=False) -> None:
+    """
+    Convert NeMo weights to HF weights
+    """
+    dummy_trainer = Trainer(devices=1, accelerator="cpu", strategy=NLPDDPStrategy())
+    model_config = MegatronGPTModel.restore_from(input_nemo_file, trainer=dummy_trainer, return_config=True)
+    model_config.tensor_model_parallel_size = 1
+    model_config.pipeline_model_parallel_size = 1
+    model_config.sequence_parallel = False
+    model_config.transformer_engine = True
+    if cpu_only:
+        map_location = torch.device("cpu")
+        model_config.use_cpu_initialization = True
+        model_config.dist_ckpt_load_on_device = False
+    else:
+        map_location = None
+
+    if cpu_only:
+        logging.info("******** Loading model on CPU. This will take a significant amount of time.")
+
+    model = MegatronGPTModel.restore_from(
+        input_nemo_file, trainer=dummy_trainer, override_config_path=model_config, map_location=map_location
+    )
+
+    vocab_size = model.padded_vocab_size
+
+    if precision is None:
+        precision = model.cfg.precision
+    if precision in [32, "32"]:
+        dtype = torch.float32
+    elif precision in [16, "16", "16-mixed"]:
+        dtype = torch.float16
+    elif precision in ["bf16", "bf16-mixed"]:
+        dtype = torch.bfloat16
+    else:
+        logging.warning(f"Precision string {precision} is not recognized, falling back to fp32")
+        dtype = torch.float32  # fallback
+    logging.info(f"Using precision {dtype}")
+
+    def param_to_weights(param):
+        return param.to(dtype)
+
+    checkpoint = OrderedDict()
+
+    hidden_size = model.cfg.hidden_size
+    head_num = model.cfg.num_attention_heads
+    num_layers = model.cfg.num_layers
+    ffn_hidden_size = model.cfg.ffn_hidden_size
+    num_query_groups = model.cfg.get("num_query_groups", head_num)  # different num_query_groups for 70B
+    if num_query_groups is None:
+        num_query_groups = head_num
+    heads_per_group = head_num // num_query_groups
+    qkv_total_dim = head_num + 2 * num_query_groups
+
+    # Embedding
+    embed_weight = model.state_dict()["model.embedding.word_embeddings.weight"]
+    embed_weights_base_name = "model.embed_tokens.weight"
+    checkpoint[embed_weights_base_name] = param_to_weights(embed_weight)
+
+    for l in range(int(num_layers)):
+        print(f"converting layer {l}")
+
+        qkv_weights = model.state_dict()[f"model.decoder.layers.{l}.self_attention.linear_qkv.weight"]
+        qkv_weights = qkv_weights.reshape([qkv_total_dim, -1, hidden_size])
+
+        q_slice = torch.cat(
+            [
+                torch.arange((heads_per_group + 2) * i, (heads_per_group + 2) * i + heads_per_group)
+                for i in range(num_query_groups)
+            ]
+        )
+        k_slice = torch.arange(heads_per_group, qkv_total_dim, (heads_per_group + 2))
+        v_slice = torch.arange(heads_per_group + 1, qkv_total_dim, (heads_per_group + 2))
+        ## Example of slices
+        ## (without GQA): num_query_groups = head_num = 32,
+        ## q_slice = [0, 3, 6, 9 , ... 90, 93]
+        ## k_slice = [1, 4, 7, 10, ... 91, 94]
+        ## v_slice = [2, 5, 8, 11, ... 92, 95]
+        ## (with GQA): num_query_groups = 8, head_num = 64
+        ## q_slice = [0, 1, .. 6, 7, 10, 11, .. 16, 17, 20, 21, .. 67, 70, ... 76, 77]
+        ## k_slice = [8, 18, 28, ... 68, 78]
+        ## v_slice = [9, 19, 29, ... 69, 79]
+
+        q_weights_base_name = f"model.layers.{l}.self_attn.q_proj.weight"
+        k_weights_base_name = f"model.layers.{l}.self_attn.k_proj.weight"
+        v_weights_base_name = f"model.layers.{l}.self_attn.v_proj.weight"
+
+        checkpoint[q_weights_base_name] = param_to_weights(qkv_weights[q_slice].reshape(-1, hidden_size))
+        checkpoint[k_weights_base_name] = param_to_weights(qkv_weights[k_slice].reshape(-1, hidden_size))
+        checkpoint[v_weights_base_name] = param_to_weights(qkv_weights[v_slice].reshape(-1, hidden_size))
+
+        # attention dense
+        o_weight = model.state_dict()[f"model.decoder.layers.{l}.self_attention.linear_proj.weight"]
+        o_weight_base_name = f"model.layers.{l}.self_attn.o_proj.weight"
+        checkpoint[o_weight_base_name] = param_to_weights(o_weight)
+
+        # mlp
+        mlp_weights = model.state_dict()[f"model.decoder.layers.{l}.mlp.linear_fc1.weight"]
+        mlp_up_proj_weight = model.state_dict()[f"model.decoder.layers.{l}.mlp.linear_fc2.weight"]
+
+        if mlp_weights.shape[0] != mlp_up_proj_weight.shape[1]:
+            # Has projection (used for swi-glu)
+            logging.warning(
+                "Gated projection layers detected in NeMo checkpoint. Currently Nemotron HF does not support gated MLP."
+            )
+            assert mlp_weights.shape[0] == 2 * mlp_up_proj_weight.shape[1]
+
+            mlp_down_proj_weight = mlp_weights[:ffn_hidden_size, :]
+            mlp_gate_proj_weight = mlp_weights[ffn_hidden_size:, :]
+
+            mlp_down_proj_base_name = f"model.layers.{l}.mlp.gate_proj.weight"
+            mlp_gate_proj_base_name = f"model.layers.{l}.mlp.up_proj.weight"
+
+            checkpoint[mlp_down_proj_base_name] = param_to_weights(mlp_down_proj_weight)
+            checkpoint[mlp_gate_proj_base_name] = param_to_weights(mlp_gate_proj_weight)
+        else:
+            mlp_down_proj_weight = mlp_weights
+            mlp_down_proj_base_name = f"model.layers.{l}.mlp.up_proj.weight"
+            checkpoint[mlp_down_proj_base_name] = param_to_weights(mlp_down_proj_weight)
+
+        mlp_up_proj_base_name = f"model.layers.{l}.mlp.down_proj.weight"
+        checkpoint[mlp_up_proj_base_name] = param_to_weights(mlp_up_proj_weight)
+
+        # layernorm
+        input_ln_weight = model.state_dict()[f"model.decoder.layers.{l}.self_attention.linear_qkv.layer_norm_weight"]
+        input_ln_base_name = f"model.layers.{l}.input_layernorm.weight"
+        checkpoint[input_ln_base_name] = param_to_weights(input_ln_weight)
+        if (
+            model.state_dict().get(f"model.decoder.layers.{l}.self_attention.linear_qkv.layer_norm_bias", None)
+            is not None
+        ):
+            input_ln_bias = model.state_dict()[f"model.decoder.layers.{l}.self_attention.linear_qkv.layer_norm_bias"]
+            input_ln_bias_name = f"model.layers.{l}.input_layernorm.bias"
+            checkpoint[input_ln_bias_name] = param_to_weights(input_ln_bias)
+
+        post_attn_ln_weight = model.state_dict()[f"model.decoder.layers.{l}.mlp.linear_fc1.layer_norm_weight"]
+        post_attn_ln_base_name = f"model.layers.{l}.post_attention_layernorm.weight"
+        checkpoint[post_attn_ln_base_name] = param_to_weights(post_attn_ln_weight)
+        if model.state_dict().get(f"model.decoder.layers.{l}.mlp.linear_fc1.layer_norm_bias", None) is not None:
+            post_attn_ln_bias = model.state_dict()[f"model.decoder.layers.{l}.mlp.linear_fc1.layer_norm_bias"]
+            post_attn_ln_bias_name = f"model.layers.{l}.post_attention_layernorm.bias"
+            checkpoint[post_attn_ln_bias_name] = param_to_weights(post_attn_ln_bias)
+
+        print(f"done layer {l}")
+
+    final_ln_weight = model.state_dict()["model.decoder.final_layernorm.weight"]
+    final_ln_base_name = "model.norm.weight"
+    checkpoint[final_ln_base_name] = param_to_weights(final_ln_weight)
+    if model.state_dict().get("model.decoder.final_layernorm.bias", None) is not None:
+        final_ln_bias = model.state_dict()["model.decoder.final_layernorm.bias"]
+        final_ln_bias_name = "model.norm.bias"
+        checkpoint[final_ln_bias_name] = param_to_weights(final_ln_bias)
+
+    output_layer_weight = model.state_dict()["model.output_layer.weight"]
+    output_layer_base_name = "lm_head.weight"
+    checkpoint[output_layer_base_name] = param_to_weights(output_layer_weight)
+
+    os.makedirs(os.path.dirname(output_hf_file), exist_ok=True)
+    torch.save(checkpoint, output_hf_file)
+    logging.info(f"Weights saved to {output_hf_file}")
+
+    return model_config, model.tokenizer, dtype, vocab_size
+
+
+def extract_nemotron_tokenizer(nemo_file, model_config, output_hf_path, nemo_tokenizer):
+    tokenizer_cfg = model_config.tokenizer
+    if tokenizer_cfg.library == "sentencepiece":
+        # For sentencepiece tokenizer, we are wrapping with HF's LlamaTokenizer
+        # and convert it to a PreTrainedTokenizerFast
+        tokenizer_fn = tokenizer_cfg.model[5:]
+        output_tokenizer = f"{output_hf_path}/tokenizer.model"
+        if nemo_file.endswith(".nemo"):
+            import tarfile
+
+            archive = tarfile.open(nemo_file, "r")
+            tokenizer_filename = "./" + tokenizer_fn  # exclude 'nemo:' prefix
+            archive.extract(tokenizer_filename, output_hf_path)
+            archive.close()
+            os.rename(f"{output_hf_path}/{tokenizer_fn}", output_tokenizer)
+        elif os.path.isdir(nemo_file):
+            shutil.copy(f"{nemo_file}/{tokenizer_fn}", output_tokenizer)
+        # We use LlamaTokenizer for sentencepiece based tokenizer
+        tokenizer = LlamaTokenizer.from_pretrained(output_hf_path, legacy=False)
+        # Convert the LlamaTokenizer to a PreTrainedTokenizerFast instance
+        tokenizer = PreTrainedTokenizerFast(
+            tokenizer_object=LlamaConverter(tokenizer).converted(), model_input_names=["input_ids", "token_type_ids"]
+        )
+        tokenizer.save_pretrained(output_hf_path)
+        logging.info(f"Setencepiece tokenizer has been saved to {output_tokenizer}")
+    elif isinstance(nemo_tokenizer, AutoTokenizer):
+        nemo_tokenizer.tokenizer.save_pretrained(output_hf_path)
+        logging.info(f"HF AutoTokenizer has been saved to {output_hf_path}")
+    else:
+        raise ValueError(f"Unsupported tokenizer type: library: {tokenizer_cfg.library}, type: {tokenizer_cfg.type}")
+
+
+if __name__ == "__main__":
+    args = get_args()
+    if not args.hf_output_path:
+        assert args.output_path is not None, "Need to provide either output_path or hf_output_path"
+    else:
+        args.output_path = f"{args.hf_output_path}/pytorch_model.bin"
+        logging.info(f"weight will be saved to {args.output_path}")
+
+    nemo_config, nemo_tokenizer, dtype, vocab_size = convert(
+        args.input_name_or_path, args.output_path, precision=args.precision, cpu_only=args.cpu_only
+    )
+    if args.hf_input_path and args.hf_output_path:
+        convert_hf_config(nemo_config, nemo_tokenizer, vocab_size, dtype, args.hf_output_path, args.hf_input_path)
+        extract_nemotron_tokenizer(args.input_name_or_path, nemo_config, args.hf_output_path, nemo_tokenizer)
+    else:
+        logging.info("`hf_input_path` and/or `hf_output_path` not provided, not generating full HF model.")
+        logging.info(f".bin file is saved to {args.output_path}")

src/transformers/models/nemotron/modeling_nemotron.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024, 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 Nemotron model."""
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import Size, Tensor, nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache, StaticCache
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_flash_attention_utils import _flash_attention_forward
+from ...modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import ALL_LAYERNORM_LAYERS
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_nemotron import NemotronConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "NemotronConfig"
+
+
+def _cast_if_autocast_enabled(*args):
+    if not torch.is_autocast_enabled():
+        return args
+    else:
+        return torch.cuda.amp.autocast_mode._cast(args, torch.get_autocast_gpu_dtype())
+
+
+class NemotronLayerNorm1P(nn.LayerNorm):
+    def __init__(
+        self,
+        normalized_shape: Union[int, List[int], Size],
+        eps: float = 1e-5,
+        elementwise_affine: bool = True,
+        bias: bool = True,
+        device=None,
+        dtype=None,
+    ):
+        super().__init__(normalized_shape, eps, elementwise_affine, bias, device, dtype)
+
+    def forward(self, input: Tensor) -> Tensor:
+        args = _cast_if_autocast_enabled(input, self.normalized_shape, self.weight + 1, self.bias, self.eps)
+        with torch.cuda.amp.autocast(enabled=False):
+            return F.layer_norm(*args)
+
+
+ALL_LAYERNORM_LAYERS.append(NemotronLayerNorm1P)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronRotaryEmbedding(nn.Module):
+    # Ignore copy
+    def __init__(
+        self,
+        config: NemotronConfig,
+        device=None,
+    ):
+        super().__init__()
+
+        self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_kwargs = None
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len, **self.rope_kwargs
+            )
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    rot_dim = cos.shape[-1]
+    # If q_pass/k_pass is empty, rotary pos embedding is applied to all tensor q/k
+    q, q_pass = q[..., :rot_dim], q[..., rot_dim:]
+    k, k_pass = k[..., :rot_dim], k[..., rot_dim:]
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return torch.cat((q_embed, q_pass), dim=-1), torch.cat((k_embed, k_pass), dim=-1)
+
+
+class NemotronMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = 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.hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.up_proj(x)))
+
+
+# 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 NemotronAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: NemotronConfig, 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
+        self.head_dim = config.head_dim
+        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.rope_theta = config.rope_theta
+        self.partial_rotary_factor = config.partial_rotary_factor
+        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],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = 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 position_embeddings is not None:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronFlashAttention2(NemotronAttention):
+    """
+    Nemotron flash attention module. This module inherits from `NemotronAttention` 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()
+
+    # Ignore copy
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = 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 isinstance(past_key_value, StaticCache):
+            raise ValueError(
+                "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
+                "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
+            )
+
+        output_attentions = False
+
+        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).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 position_embeddings is not None:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # 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 the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (NemotronRMSNorm handles it correctly)
+
+        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)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            position_ids=position_ids,
+            dropout=dropout_rate,
+            sliding_window=getattr(self, "sliding_window", None),
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronSdpaAttention(NemotronAttention):
+    """
+    Nemotron attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `NemotronAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Ignore copy
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> 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(
+                "NemotronModel is using NemotronSdpaAttention, 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,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+            )
+
+        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 position_embeddings is not None:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        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 causal_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.
+        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, -1)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+NEMOTRON_ATTENTION_CLASSES = {
+    "eager": NemotronAttention,
+    "flash_attention_2": NemotronFlashAttention2,
+    "sdpa": NemotronSdpaAttention,
+}
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronDecoderLayer(nn.Module):
+    # Ignore copy
+    def __init__(self, config: NemotronConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = NEMOTRON_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+
+        self.mlp = NemotronMLP(config)
+        self.input_layernorm = NemotronLayerNorm1P(config.hidden_size, eps=config.norm_eps)
+        self.post_attention_layernorm = NemotronLayerNorm1P(config.hidden_size, eps=config.norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*):
+                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
+                query_sequence_length, key_sequence_length)` if default attention is used.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence
+            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            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,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+NEMOTRON_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 ([`NemotronConfig`]):
+            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.
+"""
+
+
+@add_start_docstrings(
+    "The bare Nemotron Model outputting raw hidden-states without any specific head on top.",
+    NEMOTRON_START_DOCSTRING,
+)
+class NemotronPreTrainedModel(PreTrainedModel):
+    config_class = NemotronConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["NemotronDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+NEMOTRON_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` 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)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        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.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        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 `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Nemotron Model outputting raw hidden-states without any specific head on top.",
+    NEMOTRON_START_DOCSTRING,
+)
+class NemotronModel(NemotronPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`NemotronDecoderLayer`]
+
+    Args:
+        config: NemotronConfig
+    """
+
+    def __init__(self, config: NemotronConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [NemotronDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = NemotronLayerNorm1P(config.hidden_size, eps=config.norm_eps)
+        self.rotary_emb = NemotronRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(NEMOTRON_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = 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,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        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
+        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):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        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
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if cache_position is None:
+            cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
+        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
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
+            if attention_mask.max() != 0:
+                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
+            causal_mask = attention_mask
+        else:
+            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
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 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"
+            and not output_attentions
+        ):
+            # 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
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronForCausalLM(NemotronPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = NemotronModel(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.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(NEMOTRON_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    # Ignore copy (doc string different)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = 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,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, NemotronForCausalLM
+
+        >>> model = NemotronForCausalLM.from_pretrained("nvidia/nemotron-3-8b-base-4k-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("nvidia/nemotron-3-8b-base-4k-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        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
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    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,
+    ):
+        # 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
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                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]
+
+        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 past_key_values:
+                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 cache_position[0] == 0:
+            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,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+
+@add_start_docstrings(
+    """
+    The Nemotron Model transformer with a sequence classification head on top (linear layer).
+
+    [`NemotronForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    NEMOTRON_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronForSequenceClassification(NemotronPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = NemotronModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(NEMOTRON_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+The Nemotron Model transformer with a span classification head on top for extractive question-answering tasks like
+SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    NEMOTRON_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForQuestionAnswering with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronForQuestionAnswering(NemotronPreTrainedModel):
+    base_model_prefix = "transformer"
+
+    # Copied from transformers.models.bloom.modeling_bloom.BloomForQuestionAnswering.__init__ with Bloom->Nemotron
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = NemotronModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.transformer.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.transformer.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(NEMOTRON_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1).to(start_logits.device)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1).to(end_logits.device)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Nemotron Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    NEMOTRON_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
+class NemotronForTokenClassification(NemotronPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = NemotronModel(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(NEMOTRON_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -6338,6 +6338,48 @@ class MvpPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class NemotronForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NemotronForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NemotronForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NemotronForTokenClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NemotronModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NemotronPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class NllbMoeForConditionalGeneration(metaclass=DummyObject):
     _backends = ["torch"]
 

tests/models/nemotron/test_modeling_nemotron.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024, 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.
+"""Testing suite for the PyTorch Nemotron model."""
+
+import tempfile
+import unittest
+
+import pytest
+from parameterized import parameterized
+
+from transformers import NemotronConfig, is_torch_available
+from transformers.testing_utils import (
+    is_flaky,
+    require_flash_attn,
+    require_read_token,
+    require_torch,
+    require_torch_gpu,
+    require_torch_sdpa,
+    slow,
+    torch_device,
+)
+
+from ...models.gemma.test_modeling_gemma import GemmaModelTest, GemmaModelTester
+from ...test_configuration_common import ConfigTester
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        AutoTokenizer,
+        NemotronForCausalLM,
+        NemotronForQuestionAnswering,
+        NemotronForSequenceClassification,
+        NemotronForTokenClassification,
+        NemotronModel,
+    )
+
+
+class NemotronModelTester(GemmaModelTester):
+    if is_torch_available():
+        config_class = NemotronConfig
+        model_class = NemotronModel
+        for_causal_lm_class = NemotronForCausalLM
+        for_sequence_class = NemotronForSequenceClassification
+        for_token_class = NemotronForTokenClassification
+
+
+@require_torch
+class NemotronModelTest(GemmaModelTest):
+    # Need to use `0.8` instead of `0.9` for `test_cpu_offload`
+    # This is because we are hitting edge cases with the causal_mask buffer
+    model_split_percents = [0.5, 0.7, 0.8]
+    all_model_classes = (
+        (
+            NemotronModel,
+            NemotronForCausalLM,
+            NemotronForSequenceClassification,
+            NemotronForQuestionAnswering,
+            NemotronForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    all_generative_model_classes = (NemotronForCausalLM,) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": NemotronModel,
+            "text-classification": NemotronForSequenceClassification,
+            "text-generation": NemotronForCausalLM,
+            "zero-shot": NemotronForSequenceClassification,
+            "question-answering": NemotronForQuestionAnswering,
+            "token-classification": NemotronForTokenClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+    fx_compatible = False
+
+    # used in `test_torch_compile`
+    _torch_compile_test_ckpt = "nvidia/nemotron-3-8b-base-4k-hf"
+
+    def setUp(self):
+        self.model_tester = NemotronModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=NemotronConfig, hidden_size=37)
+
+    @require_torch_sdpa
+    @slow
+    @unittest.skip(
+        reason="Due to custom causal mask, there is a slightly too big difference between eager and sdpa in bfloat16."
+    )
+    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
+    def test_eager_matches_sdpa_inference(self, torch_dtype: str):
+        pass
+
+    @unittest.skip("Eager and SDPA do not produce the same outputs, thus this test fails")
+    def test_model_outputs_equivalence(self, **kwargs):
+        pass
+
+    @require_torch_sdpa
+    @require_torch_gpu
+    @slow
+    def test_sdpa_equivalence(self):
+        for model_class in self.all_model_classes:
+            if not model_class._supports_sdpa:
+                self.skipTest(reason="Model does not support SDPA")
+
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                model_sdpa = model_class.from_pretrained(
+                    tmpdirname, torch_dtype=torch.float16, attn_implementation="sdpa"
+                )
+                model_sdpa.to(torch_device)
+
+                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, attn_implementation="eager")
+                model.to(torch_device)
+
+                dummy_input = inputs_dict[model_class.main_input_name]
+                dummy_input = dummy_input.to(torch_device)
+                outputs = model(dummy_input, output_hidden_states=True)
+                outputs_sdpa = model_sdpa(dummy_input, output_hidden_states=True)
+
+                logits = outputs.hidden_states[-1]
+                logits_sdpa = outputs_sdpa.hidden_states[-1]
+
+                # nemotron sdpa needs a high tolerance
+                assert torch.allclose(logits_sdpa, logits, atol=1e-2)
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @is_flaky()
+    @slow
+    def test_flash_attn_2_equivalence(self):
+        for model_class in self.all_model_classes:
+            if not model_class._supports_flash_attn_2:
+                self.skipTest(reason="Model does not support Flash Attention 2")
+
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                model_fa = model_class.from_pretrained(
+                    tmpdirname, torch_dtype=torch.float16, attn_implementation="flash_attention_2"
+                )
+                model_fa.to(torch_device)
+
+                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, attn_implementation="eager")
+                model.to(torch_device)
+
+                dummy_input = inputs_dict[model_class.main_input_name]
+                dummy_input = dummy_input.to(torch_device)
+                outputs = model(dummy_input, output_hidden_states=True)
+                outputs_fa = model_fa(dummy_input, output_hidden_states=True)
+
+                logits = outputs.hidden_states[-1]
+                logits_fa = outputs_fa.hidden_states[-1]
+
+                # nemotron flash attention 2 needs a high tolerance
+                assert torch.allclose(logits_fa, logits, atol=1e-2)
+
+
+@require_torch_gpu
+class NemotronIntegrationTest(unittest.TestCase):
+    # This variable is used to determine which CUDA device are we using for our runners (A10 or T4)
+    # Depending on the hardware we get different logits / generations
+    cuda_compute_capability_major_version = None
+
+    @classmethod
+    def setUpClass(cls):
+        if is_torch_available() and torch.cuda.is_available():
+            # 8 is for A100 / A10 and 7 for T4
+            cls.cuda_compute_capability_major_version = torch.cuda.get_device_capability()[0]
+
+    @slow
+    @require_read_token
+    def test_nemotron_8b_generation_sdpa(self):
+        text = ["What is the largest planet in solar system?"]
+        EXPECTED_TEXT = [
+            "What is the largest planet in solar system?\nAnswer: Jupiter\n\nWhat is the answer",
+        ]
+        model_id = "thhaus/nemotron3-8b"
+        model = NemotronForCausalLM.from_pretrained(
+            model_id, torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa"
+        )
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        inputs = tokenizer(text, return_tensors="pt").to(torch_device)
+
+        output = model.generate(**inputs, do_sample=False)
+        output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT, output_text)
+
+    @slow
+    @require_read_token
+    def test_nemotron_8b_generation_eager(self):
+        text = ["What is the largest planet in solar system?"]
+        EXPECTED_TEXT = [
+            "What is the largest planet in solar system?\nAnswer: Jupiter\n\nWhat is the answer",
+        ]
+        model_id = "thhaus/nemotron3-8b"
+        model = NemotronForCausalLM.from_pretrained(
+            model_id, torch_dtype=torch.float16, device_map="auto", attn_implementation="eager"
+        )
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        inputs = tokenizer(text, return_tensors="pt").to(torch_device)
+
+        output = model.generate(**inputs, do_sample=False)
+        output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT, output_text)
+
+    @slow
+    @require_read_token
+    def test_nemotron_8b_generation_fa2(self):
+        text = ["What is the largest planet in solar system?"]
+        EXPECTED_TEXT = [
+            "What is the largest planet in solar system?\nAnswer: Jupiter\n\nWhat is the answer",
+        ]
+        model_id = "thhaus/nemotron3-8b"
+        model = NemotronForCausalLM.from_pretrained(
+            model_id, torch_dtype=torch.float16, device_map="auto", attn_implementation="flash_attention_2"
+        )
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        inputs = tokenizer(text, return_tensors="pt").to(torch_device)
+
+        output = model.generate(**inputs, do_sample=False)
+        output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT, output_text)