add OpenELM-270M

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,47 @@
+Copyright (C) 2024 Apple Inc. All Rights Reserved.
+
+Disclaimer: IMPORTANT:  This Apple software is supplied to you by Apple
+Inc. ("Apple") in consideration of your agreement to the following
+terms, and your use, installation, modification or redistribution of
+this Apple software constitutes acceptance of these terms.  If you do
+not agree with these terms, please do not use, install, modify or
+redistribute this Apple software.
+
+In consideration of your agreement to abide by the following terms, and
+subject to these terms, Apple grants you a personal, non-exclusive
+license, under Apple's copyrights in this original Apple software (the
+"Apple Software"), to use, reproduce, modify and redistribute the Apple
+Software, with or without modifications, in source and/or binary forms;
+provided that if you redistribute the Apple Software in its entirety and
+without modifications, you must retain this notice and the following
+text and disclaimers in all such redistributions of the Apple Software.
+Neither the name, trademarks, service marks or logos of Apple Inc. may
+be used to endorse or promote products derived from the Apple Software
+without specific prior written permission from Apple.  Except as
+expressly stated in this notice, no other rights or licenses, express or
+implied, are granted by Apple herein, including but not limited to any
+patent rights that may be infringed by your derivative works or by other
+works in which the Apple Software may be incorporated.
+
+The Apple Software is provided by Apple on an "AS IS" basis.  APPLE
+MAKES NO WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION
+THE IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY AND FITNESS
+FOR A PARTICULAR PURPOSE, REGARDING THE APPLE SOFTWARE OR ITS USE AND
+OPERATION ALONE OR IN COMBINATION WITH YOUR PRODUCTS.
+
+IN NO EVENT SHALL APPLE BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL
+OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) ARISING IN ANY WAY OUT OF THE USE, REPRODUCTION,
+MODIFICATION AND/OR DISTRIBUTION OF THE APPLE SOFTWARE, HOWEVER CAUSED
+AND WHETHER UNDER THEORY OF CONTRACT, TORT (INCLUDING NEGLIGENCE),
+STRICT LIABILITY OR OTHERWISE, EVEN IF APPLE HAS BEEN ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+
+
+-------------------------------------------------------------------------------
+SOFTWARE DISTRIBUTED IN THIS REPOSITORY:
+
+This software includes a number of subcomponents with separate
+copyright notices and license terms - please see the file ACKNOWLEDGEMENTS.
+-------------------------------------------------------------------------------

README.md

@@ -0,0 +1,163 @@
+---
+license: other
+license_name: apple-sample-code-license
+license_link: LICENSE
+---
+
+# OpenELM
+
+*Sachin Mehta, Mohammad Hossein Sekhavat, Qingqing Cao, Maxwell Horton, Yanzi Jin, Chenfan Sun, Iman Mirzadeh, Mahyar Najibi, Dmitry Belenko, Peter Zatloukal, Mohammad Rastegari*
+
+We introduce **OpenELM**, a family of **Open**-source **E**fficient **L**anguage **M**odels. OpenELM uses a layer-wise scaling strategy to efficiently allocate parameters within each layer of the transformer model, leading to enhanced accuracy. We pretrained OpenELM models using the [CoreNet](https://github.com/apple/corenet) library. We release both pretrained and instruction tuned models with 270M, 450M, 1.1B and 3B parameters.
+
+Our pre-training dataset contains RefinedWeb, deduplicated PILE, a subset of RedPajama, and a subset of Dolma v1.6, totaling approximately 1.8 trillion tokens. Please check license agreements and terms of these datasets before using them.
+
+
+
+## Usage
+
+We have provided an example function to generate output from OpenELM models loaded via [HuggingFace Hub](https://huggingface.co/docs/hub/) in `generate_openelm.py`.
+
+You can try the model by running the following command:
+```
+python generate_openelm.py --model apple/OpenELM-270M --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2
+```
+Please refer to [this link](https://huggingface.co/docs/hub/security-tokens) to obtain your hugging face access token.
+
+Additional arguments to the hugging face generate function can be passed via `generate_kwargs`. As an example, to speedup the inference, you can try [lookup token speculative generation](https://huggingface.co/docs/transformers/generation_strategies) by passing the `prompt_lookup_num_tokens` argument as follows:
+```
+python generate_openelm.py --model apple/OpenELM-270M --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2 prompt_lookup_num_tokens=10
+```
+Alternatively, try model-wise speculative generation with an [assistive model](https://huggingface.co/blog/assisted-generation) by passing a smaller model through the `assistant_model` argument, for example:
+```
+python generate_openelm.py --model apple/OpenELM-270M --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2 --assistant_model [SMALLER_MODEL]
+```
+
+## Main Results
+
+### Zero-Shot
+
+| **Model Size**                                                              | **ARC-c** | **ARC-e** | **BoolQ** | **HellaSwag** | **PIQA**  | **SciQ**  | **WinoGrande** | **Average** |
+|-----------------------------------------------------------------------------|-----------|-----------|-----------|---------------|-----------|-----------|----------------|-------------|
+| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M)                   | 26.45     | 45.08     | **53.98** | 46.71         | 69.75     | **84.70** | **53.91**      | 54.37       |
+| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **30.55** | **46.68** | 48.56     | **52.07**     | **70.78** | 84.40     | 52.72          | **55.11**   |
+| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M)                   | 27.56     | 48.06     | 55.78     | 53.97         | 72.31     | 87.20     | 58.01          | 57.56       |
+| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **30.38** | **50.00** | **60.37** | **59.34**     | **72.63** | **88.00** | **58.96**      | **59.95**   |
+| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B)                   | 32.34     | **55.43** | 63.58     | 64.81         | **75.57** | **90.60** | 61.72          | 63.44       |
+| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **37.97** | 52.23     | **70.00** | **71.20**     | 75.03     | 89.30     | **62.75**      | **65.50**   |
+| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B)                       | 35.58     | 59.89     | 67.40     | 72.44         | 78.24     | **92.70** | 65.51          | 67.39       |
+| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct)     | **39.42** | **61.74** | **68.17** | **76.36**     | **79.00** | 92.50     | **66.85**      | **69.15**   |
+
+### LLM360
+
+| **Model Size**                                                              | **ARC-c** | **HellaSwag** | **MMLU**  | **TruthfulQA** | **WinoGrande** | **Average** |
+|-----------------------------------------------------------------------------|-----------|---------------|-----------|----------------|----------------|-------------|
+| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M)                   | 27.65     | 47.15         | 25.72     | **39.24**      | **53.83**      | 38.72       |
+| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **32.51** | **51.58**     | **26.70** | 38.72          | 53.20          | **40.54**   |
+| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M)                   | 30.20     | 53.86         | **26.01** | 40.18          | 57.22          | 41.50       |
+| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **33.53** | **59.31**     | 25.41     | **40.48**      | **58.33**      | **43.41**   |
+| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B)                   | 36.69     | 65.71         | **27.05** | 36.98          | 63.22          | 45.93       |
+| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **41.55** | **71.83**     | 25.65     | **45.95**      | **64.72**      | **49.94**   |
+| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B)                       | 42.24     | 73.28         | **26.76** | 34.98          | 67.25          | 48.90       |
+| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct)     | **47.70** | **76.87**     | 24.80     | **38.76**      | **67.96**      | **51.22**   |
+
+
+### OpenLLM Leaderboard
+
+| **Model Size**                                                              | **ARC-c** | **CrowS-Pairs** | **HellaSwag** | **MMLU**  | **PIQA**  | **RACE**  | **TruthfulQA** | **WinoGrande** | **Average** |
+|-----------------------------------------------------------------------------|-----------|-----------------|---------------|-----------|-----------|-----------|----------------|----------------|-------------|
+| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M)                   | 27.65     | **66.79**       | 47.15         | 25.72     | 69.75     | 30.91     | **39.24**      | **53.83**      | 45.13       |
+| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **32.51** | 66.01           | **51.58**     | **26.70** | **70.78** | 33.78     | 38.72          | 53.20          | **46.66**   |
+| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M)                   | 30.20     | **68.63**       | 53.86         | **26.01** | 72.31     | 33.11     | 40.18          | 57.22          | 47.69       |
+| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **33.53** | 67.44           | **59.31**     | 25.41     | **72.63** | **36.84** | **40.48**      | **58.33**      | **49.25**   |
+| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B)                   | 36.69     | **71.74**       | 65.71         | **27.05** | **75.57** | 36.46     | 36.98          | 63.22          | 51.68       |
+| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **41.55** | 71.02           | **71.83**     | 25.65     | 75.03     | **39.43** | **45.95**      | **64.72**      | **54.40**   |
+| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B)                       | 42.24     | **73.29**       | 73.28         | **26.76** | 78.24     | **38.76** | 34.98          | 67.25          | 54.35       |
+| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct)     | **47.70** | 72.33           | **76.87**     | 24.80     | **79.00** | 38.47     | **38.76**      | **67.96**      | **55.73**   |
+
+See the technical report for more results and comparison.
+
+## Evaluation
+
+### Setup
+
+Install the following dependencies:
+
+```bash
+
+# install public lm-eval-harness
+
+harness_repo="public-lm-eval-harness"
+git clone https://github.com/EleutherAI/lm-evaluation-harness ${harness_repo}
+cd ${harness_repo}
+# use main branch on 03-15-2024, SHA is dc90fec
+git checkout dc90fec
+pip install -e .
+cd ..
+
+# 66d6242 is the main branch on 2024-04-01 
+pip install datasets@git+https://github.com/huggingface/datasets.git@66d6242
+pip install tokenizers>=0.15.2 transformers>=4.38.2 sentencepiece>=0.2.0
+
+```
+
+### Evaluate OpenELM
+
+```bash
+
+# OpenELM-270M
+hf_model=OpenELM-270M
+
+# this flag is needed because lm-eval-harness set add_bos_token to False by default, but OpenELM uses LLaMA tokenizer which requires add_bos_token to be True
+tokenizer=meta-llama/Llama-2-7b-hf
+add_bos_token=True
+batch_size=1
+
+mkdir lm_eval_output
+
+shot=0
+task=arc_challenge,arc_easy,boolq,hellaswag,piqa,race,winogrande,sciq,truthfulqa_mc2
+lm_eval --model hf \
+        --model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
+        --tasks ${task} \
+        --device cuda:0 \
+        --num_fewshot ${shot} \
+        --output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
+        --batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
+
+shot=5
+task=mmlu,winogrande
+lm_eval --model hf \
+        --model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
+        --tasks ${task} \
+        --device cuda:0 \
+        --num_fewshot ${shot} \
+        --output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
+        --batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
+
+shot=25
+task=arc_challenge,crows_pairs_english
+lm_eval --model hf \
+        --model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
+        --tasks ${task} \
+        --device cuda:0 \
+        --num_fewshot ${shot} \
+        --output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
+        --batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
+
+shot=10
+task=hellaswag
+lm_eval --model hf \
+        --model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
+        --tasks ${task} \
+        --device cuda:0 \
+        --num_fewshot ${shot} \
+        --output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
+        --batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
+
+```
+
+
+## Bias, Risks, and Limitations
+
+The release of OpenELM models aims to empower and enrich the open research community by providing access to state-of-the-art language models. Trained on publicly available datasets, these models are made available without any safety guarantees. Consequently, there exists the possibility of these models producing outputs that are inaccurate, harmful, biased, or objectionable in response to user prompts. Thus, it is imperative for users and developers to undertake thorough safety testing and implement appropriate filtering mechanisms tailored to their specific requirements.

config.json

@@ -0,0 +1,88 @@
+{
+  "activation_fn_name": "swish",
+  "architectures": [
+    "OpenELMForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_openelm.OpenELMConfig",
+    "AutoModelForCausalLM": "modeling_openelm.OpenELMForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "ffn_dim_divisor": 256,
+  "ffn_multipliers": [
+    0.5,
+    0.73,
+    0.97,
+    1.2,
+    1.43,
+    1.67,
+    1.9,
+    2.13,
+    2.37,
+    2.6,
+    2.83,
+    3.07,
+    3.3,
+    3.53,
+    3.77,
+    4.0
+  ],
+  "ffn_with_glu": true,
+  "head_dim": 64,
+  "initializer_range": 0.02,
+  "max_context_length": 2048,
+  "model_dim": 1280,
+  "model_type": "openelm",
+  "normalization_layer_name": "rms_norm",
+  "normalize_qk_projections": true,
+  "num_gqa_groups": 4,
+  "num_kv_heads": [
+    3,
+    3,
+    3,
+    3,
+    3,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    5,
+    5,
+    5,
+    5
+  ],
+  "num_query_heads": [
+    12,
+    12,
+    12,
+    12,
+    12,
+    16,
+    16,
+    16,
+    16,
+    16,
+    16,
+    16,
+    20,
+    20,
+    20,
+    20
+  ],
+  "num_transformer_layers": 16,
+  "qkv_multipliers": [
+    0.5,
+    1.0
+  ],
+  "rope_freq_constant": 10000,
+  "rope_max_length": 4096,
+  "share_input_output_layers": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.39.3",
+  "use_cache": true,
+  "vocab_size": 32000
+}

configuration_openelm.py

@@ -0,0 +1,318 @@
+#
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+#
+
+"""Implements HF OpenELMConfig based on PretrainedConfig"""
+from numbers import Number
+from typing import List, Optional, Union
+
+import numpy as np
+from transformers import PretrainedConfig
+
+
+def make_divisible(
+    v: Union[float, int],
+    divisor: Optional[int] = 8,
+    min_value: Optional[Union[float, int]] = None,
+) -> Union[float, int]:
+    """
+    This function is taken from the original tf repo.
+    It ensures that all layers have a channel number that is divisible by the divisor
+    It can be seen at:
+    https://github.com/tensorflow/models/blob/2cfc99eff5e5eb729c6793d2f3d03aa1c9be2b15/research/slim/nets/mobilenet/mobilenet.py#L62
+
+    Args:
+        v: input value
+        divisor: default to 8
+        min_value: minimum divisor value
+    Returns:
+        new_v: new divisible value
+    """
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+def compute_heads(model_dim: int, head_dim: int) -> int:
+    """Compute the number of heads.
+
+    Args:
+        model_dim: Model dimension.
+        head_dim: Head dimension.
+
+    Returns:
+        An integer denoting number of heads in multi-head attention is returned.
+
+    Raises:
+        ValueError: if model dimension is not divisible by head dimension.
+    """
+    if model_dim % head_dim == 0:
+        return model_dim // head_dim
+    else:
+        raise ValueError(
+            f"Model dimension should be divisible by head dimension. Got: {model_dim} and {head_dim}."
+        )
+
+
+OpenELM_CONFIGS = {
+    "OpenELM-270M": dict(
+        num_transformer_layers=16,
+        model_dim=1280,
+        head_dim=64,
+        num_gqa_groups=4,
+        normalize_qk_projections=True,
+        share_input_output_layers=True,
+        # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
+        ffn_multipliers=(0.5, 4.0),
+        qkv_multipliers=(0.5, 1.0),
+    ),
+    "OpenELM-450M": dict(
+        num_transformer_layers=20,
+        model_dim=1536,
+        head_dim=64,
+        num_gqa_groups=4,
+        normalize_qk_projections=True,
+        share_input_output_layers=True,
+        # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
+        ffn_multipliers=(0.5, 4.0),
+        qkv_multipliers=(0.5, 1.0),
+    ),
+    "OpenELM-1_1B": dict(
+        num_transformer_layers=28,
+        model_dim=2048,
+        head_dim=64,
+        num_gqa_groups=4,
+        normalize_qk_projections=True,
+        share_input_output_layers=True,
+        # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
+        ffn_multipliers=(0.5, 4.0),
+        qkv_multipliers=(0.5, 1.0),
+    ),
+    "OpenELM-3B": dict(
+        num_transformer_layers=36,
+        model_dim=3072,
+        head_dim=128,
+        num_gqa_groups=4,
+        normalize_qk_projections=True,
+        share_input_output_layers=True,
+        # Vary the FFN and QKV multipliers to create variable FFN and attention layers respectively.
+        ffn_multipliers=(0.5, 4.0),
+        qkv_multipliers=(0.5, 1.0),
+    ),
+}
+
+
+class OpenELMConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`OpenELMModel`]. It is used to instantiate an OpenELM model according to the specified arguments, defining the model architecture.
+
+    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 32000):
+            Vocabulary size of the OpenELM model.
+        max_context_length (`int`, *optional*, defaults to 2048):
+            Maximum number of input tokens.
+        num_transformer_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer decoder.
+        model_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        head_dim (`int`, *optional*, defaults to 128):
+            The attention head dimension.
+        qkv_multipliers (`Union[Number, List[Number]]`, *optional*, defaults to 1.0):
+            If the qkv_multipliers is a Number, then all attention layers have the same latent dimensions,
+            resulting in uniform allocation of parameters.
+            If the qkv_multipliers is a List of Number, then each attention layer have different latent dimensions
+            assuming qkv_multipliers[0] != qkv_multipliers[1]. This results in variable allocation of parameters in attention layer.
+            This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
+        num_query_heads (`Union[int, None]`, *optional*, defaults to None):
+            The number of query heads, computed from `compute_heads(model_dim=model_dim, head_dim=head_dim)`.
+        num_gqa_groups (`int`, *optional*, defaults to 1):
+            This variable allows to switch between multi-head attention, group query attention, and multi-query attention.
+            When num_gqa_groups == 1, then it is multi-head attention.
+            When 1 < num_gqa_groups < num_heads and num_heads is divisible by num_gqa_groups, then it is group query attention
+            When num_gqa_groups == num_heads, then it is multi-query attention
+        ffn_multipliers (`Union[Number, List[Number]]`, *optional*, defaults to 4.0):
+            Feed-forward network (FFN) multipliers.
+            If the ffn_multipliers is a Number, then all FFN layers have the same latent dimensions,
+            resulting in uniform allocation of parameters.
+            If the ffn_multipliers is a List of Number, then each FFN layer have different latent dimensions
+            assuming ffn_multipliers[0] != ffn_multipliers[1]. This results in variable allocation of parameters in FFN layer.
+            This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
+        ffn_with_glu (`bool`, *optional*, defaults to True):
+            Whether to use FFN with Gated Linear Unit (GLU)
+        ffn_dim_divisor (`int`, *optional*, defaults to 256):
+            The ffn layer dimension divisor.
+        activation_fn_name (`str` or `function`, *optional*, defaults to `"swish"`):
+            The non-linear activation function (function or string) in the decoder.
+        normalization_layer_name (`str` or `function`, *optional*, defaults to `"rms_norm"`):
+            Type of normalization layer.
+        normalize_qk_projections (`bool`, *optional*, defaults to False):
+            Whether to normalize queries and keys after projections
+        share_input_output_layers (`bool`, *optional*, defaults to False):
+            Whether to share the embedding between input and output linear layer
+        rope_freq_constant (`int`, *optional*, defaults to 10000):
+            The base period of the RoPE embeddings.
+        rope_max_length (`int`, *optional*, defaults to 4096):
+            That rope_max_length is set to twice of max_context_length.
+            This allows flexibility in token lengths during training or fine-tuning.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        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`.
+        bos_token_id (`int`, *optional*, defaults to 2):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+    """
+
+    model_type = "openelm"
+
+    def __init__(
+        self,
+        vocab_size: int = 32000,
+        max_context_length: int = 2048,
+        num_transformer_layers: int = 12,
+        model_dim: int = 2048,
+        head_dim: int = 128,
+        qkv_multipliers: Union[Number, List[Number]] = 1.0,
+        num_query_heads: Union[int, None] = None,
+        num_gqa_groups: int = 1,
+        ffn_multipliers: Union[Number, List[Number]] = 4.0,
+        ffn_with_glu: bool = True,
+        ffn_dim_divisor: int = 256,
+        activation_fn_name: str = "swish",
+        normalization_layer_name: str = "rms_norm",
+        normalize_qk_projections: bool = False,
+        share_input_output_layers: bool = False,
+        rope_freq_constant: int = 10000,
+        rope_max_length: int = 4096,
+        initializer_range: float = 0.02,
+        use_cache: bool = True,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ) -> None:
+        self.vocab_size = vocab_size
+        self.max_context_length = max_context_length
+        self.num_transformer_layers = num_transformer_layers
+        self.model_dim = model_dim
+        self.head_dim = head_dim
+        self.qkv_multipliers = qkv_multipliers
+        self.num_query_heads = num_query_heads
+        self.num_gqa_groups = num_gqa_groups
+        self.ffn_multipliers = ffn_multipliers
+        self.ffn_with_glu = ffn_with_glu
+        self.ffn_dim_divisor = ffn_dim_divisor
+        self.activation_fn_name = activation_fn_name
+        self.normalization_layer_name = normalization_layer_name
+        self.normalize_qk_projections = normalize_qk_projections
+        self.share_input_output_layers = share_input_output_layers
+        self.rope_freq_constant = rope_freq_constant
+        self.rope_max_length = rope_max_length
+        self.num_query_heads = (
+            compute_heads(model_dim=model_dim, head_dim=head_dim)
+            if num_query_heads is None
+            else num_query_heads
+        )
+        self.initializer_range = initializer_range
+
+        self.__post_init__()
+        super().__init__(
+            use_cache=use_cache,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+
+    def __post_init__(self) -> None:
+        if self.num_gqa_groups is not None:
+            head_multiple_of = self.num_gqa_groups
+        else:
+            head_multiple_of = 2
+
+        if isinstance(self.qkv_multipliers, Number):
+            # All attention layers have the same latent dimensions, resulting in uniform allocation of parameters.
+            qkv_dim = make_divisible(
+                self.model_dim * self.qkv_multipliers,
+                divisor=self.head_dim * head_multiple_of,
+            )
+            query_dims = [int(qkv_dim)] * self.num_transformer_layers
+
+        elif (
+            isinstance(self.qkv_multipliers, (tuple, list))
+            and len(self.qkv_multipliers) == 2
+        ):
+            # Each attention layer have different latent dimensions assuming qkv_multipliers[0] != qkv_multipliers[1].
+            # This results in variable allocation of parameters in attention layer.
+            # This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
+            qkv_multipliers = [
+                round(v, 2)
+                for v in np.linspace(
+                    self.qkv_multipliers[0],
+                    self.qkv_multipliers[1],
+                    num=self.num_transformer_layers,
+                    dtype=float,
+                )
+            ]
+            # Make sure that scaled model dimension is divisible by scaled head dimension.
+            query_dims = [
+                int(
+                    make_divisible(
+                        self.model_dim * m, divisor=self.head_dim * head_multiple_of
+                    )
+                )
+                for m in qkv_multipliers
+            ]
+        else:
+            raise NotImplementedError(
+                f"QKV multipliers should be a single number or a list containing exactly two numbers. Got: {qkv_multipliers}."
+            )
+
+        # compute the number of query, key, and value heads
+        # For multi-head and multi-query attention, the number of heads for query, key, and value are the same.
+        # For group query attention, the number of key and value heads are the same.
+        self.num_query_heads = [
+            int(compute_heads(q_dim, self.head_dim)) for q_dim in query_dims
+        ]
+        self.num_kv_heads = [
+            q_heads // self.num_gqa_groups for q_heads in self.num_query_heads
+        ]
+
+        # Feed-forward network (FFN) multipliers
+        if isinstance(self.ffn_multipliers, Number):
+            # All FFN layers have the same latent dimensions, resulting in uniform allocation of parameters.
+            self.ffn_multipliers = [self.ffn_multipliers] * self.num_transformer_layers
+        elif isinstance(self.ffn_multipliers, (tuple, list)):
+            # Each FFN layer have different latent dimensions assuming ffn_multipliers[0] != ffn_multipliers[1].
+            # This results in variable allocation of parameters in FFN layer.
+            # This scaling is known as layer-wise or block-wise scaling: https://arxiv.org/abs/2008.00623
+            if len(self.ffn_multipliers) == 2:
+                self.ffn_multipliers = [
+                    round(v, 2)
+                    for v in np.linspace(
+                        self.ffn_multipliers[0],
+                        self.ffn_multipliers[1],
+                        num=self.num_transformer_layers,
+                        dtype=float,
+                    )
+                ]
+            else:
+                assert (
+                    len(self.ffn_multipliers) == self.num_transformer_layers
+                ), f"{len(self.ffn_multipliers)=}!={self.num_transformer_layers=}"
+        else:
+            raise NotImplementedError(
+                f"FFN multipliers should be a single number or a list containing exactly two numbers. Got: {qkv_multipliers}."
+            )
+
+        # check num_query_heads divisible by num_kv_heads for every layer
+        for layer_idx in range(len(query_dims)):
+            assert self.num_query_heads[layer_idx] % self.num_kv_heads[layer_idx] == 0

generate_openelm.py

@@ -0,0 +1,240 @@
+#
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+#
+
+"""Module to generate OpenELM output given a model and an input prompt."""
+import os
+import logging
+import time
+import argparse
+from typing import Optional, Union
+import torch
+
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+
+def generate(
+    prompt: str,
+    model: Union[str, AutoModelForCausalLM],
+    hf_access_token: str = None,
+    tokenizer: Union[str, AutoTokenizer] = 'meta-llama/Llama-2-7b-hf',
+    device: Optional[str] = None,
+    max_length: int = 1024,
+    assistant_model: Optional[Union[str, AutoModelForCausalLM]] = None,
+    generate_kwargs: Optional[dict] = None,
+) -> str:
+    """ Generates output given a prompt.
+
+    Args:
+        prompt: The string prompt.
+        model: The LLM Model. If a string is passed, it should be the path to
+            the hf converted checkpoint.
+        hf_access_token: Hugging face access token.
+        tokenizer: Tokenizer instance. If model is set as a string path,
+            the tokenizer will be loaded from the checkpoint.
+        device: String representation of device to run the model on. If None
+            and cuda available it would be set to cuda:0 else cpu.
+        max_length: Maximum length of tokens, input prompt + generated tokens.
+        assistant_model: If set, this model will be used for
+            speculative generation. If a string is passed, it should be the
+            path to the hf converted checkpoint.
+        generate_kwargs: Extra kwargs passed to the hf generate function.
+
+    Returns:
+        output_text: output generated as a string.
+        generation_time: generation time in seconds.
+
+    Raises:
+        ValueError: If device is set to CUDA but no CUDA device is detected.
+        ValueError: If tokenizer is not set.
+        ValueError: If hf_access_token is not specified.
+    """
+    if not device:
+        if torch.cuda.is_available() and torch.cuda.device_count():
+            device = "cuda:0"
+            logging.warning(
+                'inference device is not set, using cuda:0, %s',
+                torch.cuda.get_device_name(0)
+            )
+        else:
+            device = 'cpu'
+            logging.warning(
+                (
+                    'No CUDA device detected, using cpu, '
+                    'expect slower speeds.'
+                )
+            )
+
+    if 'cuda' in device and not torch.cuda.is_available():
+        raise ValueError('CUDA device requested but no CUDA device detected.')
+
+    if not tokenizer:
+        raise ValueError('Tokenizer is not set in the generate function.')
+
+    if not hf_access_token:
+        raise ValueError((
+            'Hugging face access token needs to be specified. '
+            'Please refer to https://huggingface.co/docs/hub/security-tokens'
+            ' to obtain one.'
+            )
+        )
+
+    if isinstance(model, str):
+        checkpoint_path = model
+        model = AutoModelForCausalLM.from_pretrained(
+            checkpoint_path,
+            trust_remote_code=True
+        )
+    model.to(device).eval()
+    if isinstance(tokenizer, str):
+        tokenizer = AutoTokenizer.from_pretrained(
+            tokenizer,
+            token=hf_access_token,
+        )
+
+    # Speculative mode
+    draft_model = None
+    if assistant_model:
+        draft_model = assistant_model
+        if isinstance(assistant_model, str):
+            draft_model = AutoModelForCausalLM.from_pretrained(
+                assistant_model,
+                trust_remote_code=True
+            )
+        draft_model.to(device).eval()
+
+    # Prepare the prompt
+    tokenized_prompt = tokenizer(prompt)
+    tokenized_prompt = torch.tensor(
+        tokenized_prompt['input_ids'],
+        device=device
+    )
+
+    tokenized_prompt = tokenized_prompt.unsqueeze(0)
+
+    # Generate
+    stime = time.time()
+    output_ids = model.generate(
+        tokenized_prompt,
+        max_length=max_length,
+        pad_token_id=0,
+        assistant_model=draft_model,
+        **(generate_kwargs if generate_kwargs else {}),
+    )
+    generation_time = time.time() - stime
+
+    output_text = tokenizer.decode(
+        output_ids[0].tolist(),
+        skip_special_tokens=True
+    )
+
+    return output_text, generation_time
+
+
+def openelm_generate_parser():
+    """Argument Parser"""
+
+    class KwargsParser(argparse.Action):
+        """Parser action class to parse kwargs of form key=value"""
+        def __call__(self, parser, namespace, values, option_string=None):
+            setattr(namespace, self.dest, dict())
+            for val in values:
+                if '=' not in val:
+                    raise ValueError(
+                        (
+                            'Argument parsing error, kwargs are expected in'
+                            ' the form of key=value.'
+                        )
+                    )
+                kwarg_k, kwarg_v = val.split('=')
+                try:
+                    converted_v = int(kwarg_v)
+                except ValueError:
+                    try:
+                        converted_v = float(kwarg_v)
+                    except ValueError:
+                        converted_v = kwarg_v            
+                getattr(namespace, self.dest)[kwarg_k] = converted_v
+
+    parser = argparse.ArgumentParser('OpenELM Generate Module')
+    parser.add_argument(
+        '--model',
+        dest='model',
+        help='Path to the hf converted model.',
+        required=True,
+        type=str,
+    )
+    parser.add_argument(
+        '--hf_access_token',
+        dest='hf_access_token',
+        help='Hugging face access token, starting with "hf_".',
+        type=str,
+    )
+    parser.add_argument(
+      '--prompt',
+      dest='prompt',
+      help='Prompt for LLM call.',
+      default='',
+      type=str,
+    )
+    parser.add_argument(
+        '--device',
+        dest='device',
+        help='Device used for inference.',
+        type=str,
+    )
+    parser.add_argument(
+        '--max_length',
+        dest='max_length',
+        help='Maximum length of tokens.',
+        default=256,
+        type=int,
+    )
+    parser.add_argument(
+        '--assistant_model',
+        dest='assistant_model',
+        help=(
+            (
+                'If set, this is used as a draft model '
+                'for assisted speculative generation.'
+            )
+        ),
+        type=str,
+    )
+    parser.add_argument(
+        '--generate_kwargs',
+        dest='generate_kwargs',
+        help='Additional kwargs passed to the HF generate function.',
+        type=str,
+        nargs='*',
+        action=KwargsParser,
+    )
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = openelm_generate_parser()
+    prompt = args.prompt
+
+    output_text, genertaion_time = generate(
+        prompt=prompt,
+        model=args.model,
+        device=args.device,
+        max_length=args.max_length,
+        assistant_model=args.assistant_model,
+        generate_kwargs=args.generate_kwargs,
+        hf_access_token=args.hf_access_token,
+    )
+
+    print_txt = (
+        f'\r\n{"=" * os.get_terminal_size().columns}\r\n'
+        '\033[1m Prompt + Generated Output\033[0m\r\n'
+        f'{"-" * os.get_terminal_size().columns}\r\n'
+        f'{output_text}\r\n'
+        f'{"-" * os.get_terminal_size().columns}\r\n'
+        '\r\nGeneration took'
+        f'\033[1m\033[92m {round(genertaion_time, 2)} \033[0m'
+        'seconds.\r\n'
+    )
+    print(print_txt)

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.39.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afbeade8abad5a1ed4df7dabdf1773aa0215c5bb7563537ac62d33fff9a2ecfa
+size 1086123184

modeling_openelm.py

@@ -0,0 +1,1008 @@
+#
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+#
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import Tensor, nn
+from torch.nn import CrossEntropyLoss
+from torch.nn import functional as F
+from transformers import PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+# this import has to be relative, otherwise, when setting trust_remote_code=True
+# huggingface transformers won't be able to load the module correctly
+from .configuration_openelm import OpenELMConfig, make_divisible
+
+
+class OpenELMRMSNorm(nn.Module):
+    def __init__(self, num_features: int, eps: float = 1e-6):
+        """
+        Initialize the OpenELMRMSNorm normalization layer.
+
+        Args:
+            dim (int): The dimension of the input tensor.
+            eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
+
+        Attributes:
+            eps (float): A small value added to the denominator for numerical stability.
+            weight (nn.Parameter): Learnable scaling parameter.
+
+        """
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(num_features))
+        self.num_features = num_features
+
+    def _norm(self, x: Tensor) -> Tensor:
+        """
+        Apply the OpenELMRMSNorm normalization to the input tensor.
+
+        Args:
+            x (torch.Tensor): The input tensor.
+
+        Returns:
+            torch.Tensor: The normalized tensor.
+
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Forward pass through the OpenELMRMSNorm layer.
+
+        Args:
+            x (torch.Tensor): The input tensor.
+
+        Returns:
+            torch.Tensor: The output tensor after applying OpenELMRMSNorm.
+
+        """
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return (
+            super().extra_repr() + f"num_features={self.num_features}, eps={self.eps}"
+        )
+
+
+class OpenELMPreTrainedModel(PreTrainedModel):
+    config_class = OpenELMConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["OpenELMDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def __init__(self, *inputs, **kwargs) -> None:
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module) -> None:
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, OpenELMRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+def _rotate_half(x: Tensor) -> Tensor:
+    x1, x2 = x.chunk(2, dim=-1)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def _apply_rotary_pos_emb(x: Tensor, pos_sin: Tensor, pos_cos: Tensor) -> Tensor:
+    return (x * pos_cos) + (_rotate_half(x) * pos_sin)
+
+
+class OpenELMRotaryEmbedding(torch.nn.Module):
+    """
+    The rotary position embeddings (aka RoPE) from `RoFormer <https://arxiv.org/abs/2104.09864>`_.
+
+    RoPE encodes the position information of tokens using a rotation matrix, and is able to capture
+    explicit relative positional dependencies.
+
+    Args:
+        model_dim: The dimensionality of the model's hidden state.
+        max_seq_length: Maximum sequence length.
+        freq_constant: A constant used for computing frequencies.
+    """
+
+    def __init__(
+        self, model_dim: int, max_seq_length: int, freq_constant: int = 10000
+    ) -> None:
+        inv_freq = 1.0 / (
+            freq_constant
+            ** (torch.arange(0, model_dim, 2, dtype=torch.float32) / model_dim)
+        )
+        super().__init__()
+
+        self.model_dim = model_dim
+        self.freq_constant = freq_constant
+        self.max_seq_length = max_seq_length
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._cached_cos = None
+        self._cached_sin = None
+        self._cached_seq_length = max_seq_length
+        self._compute_sin_cos_embeddings(max_seq_length)
+
+    def extra_repr(self) -> str:
+        return f"\tmodel_dim={self.model_dim}, max_seq_length={self.max_seq_length}, freq_constant={self.freq_constant}"
+
+    def _compute_sin_cos_embeddings(
+        self,
+        key_len: int,
+        key_device: torch.device = torch.device("cpu"),
+        key_dtype: torch.dtype = torch.float32,
+    ) -> None:
+        """
+        Compute sine and cos embeddings.
+
+        Args:
+            key_len: Number of tokens in the key embeddings in the transformer model.
+            device: Device where the key embeddings are stored.
+            key_dtype: Data type of the key embeddings.
+
+        Returns:
+            None
+
+        ...note:
+            We recalculate the sine and cosine embeddings if any of the following conditions are met:
+                1. The number of tokens in key embeddings are greater than the cached sequence length.
+                2. Sine and cosine caches are empty.
+                3. The device and data type of sine and cosine embeddings does not match with the key embeddings.
+        """
+        if (
+            key_len > self._cached_seq_length
+            or self._cached_cos is None
+            or (self._cached_cos is not None and self._cached_cos.device != key_device)
+            or (self._cached_cos is not None and self._cached_cos.dtype != key_dtype)
+            or self._cached_sin is None
+            or (self._cached_sin is not None and self._cached_sin.device != key_device)
+            or (self._cached_sin is not None and self._cached_sin.dtype != key_dtype)
+        ):
+            self._cached_seq_length = max(key_len, self._cached_seq_length)
+
+            # The shape of 'pos_index' is [number of key tokens]
+            pos_index = torch.arange(
+                self._cached_seq_length,
+                dtype=torch.float32,
+                device=self.inv_freq.device,
+            )
+            # The shape of 'pos_index_theta' is [number of key tokens, model dimension]
+            pos_index_theta = torch.einsum("i,j->ij", pos_index, self.inv_freq)
+            # The shape of 'emb' is [number of key tokens, model dimension]
+            emb = torch.cat((pos_index_theta, pos_index_theta), dim=-1)
+
+            # the shape of cos and sin embeddings is [number of key tokens, model_dim]
+            cos_emb = emb.cos().to(dtype=key_dtype, device=key_device)
+            sin_emb = emb.sin().to(dtype=key_dtype, device=key_device)
+
+            # the shape of cached cos and sin embeddings is [1, 1, number of key tokens, model_dim]
+            self._cached_cos = cos_emb[None, None, :, :]
+            self._cached_sin = sin_emb[None, None, :, :]
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        The forward function of RoPE embeddings.
+
+        Args:
+            query: Query embeddings in the transformer model. The shape of query embeddings is
+                [Batch, number of query heads, number of query tokens, model dimension].
+            key: Key embeddings in the transformer model. The shape of key embeddings is
+                [Batch, number of key heads, number of key tokens, model dimension].
+
+        Returns:
+            A tuple containing the query and key embeddings with positional information. The shape of the returned query
+            and key embeddings is the same as the input query and key embeddings respectively.
+
+        ...note:
+            The RoPE embedding computation is done in full-precision. After the computation, input query and key tensors
+            are casted to original input datatype.
+        """
+        dim = key.shape[-1]
+        key_len = key.shape[2]
+        query_len = query.shape[2]
+
+        assert dim == self.model_dim
+        assert key.device == query.device
+        assert key.dtype == query.dtype
+
+        # In the context of self-attention, the lengths of keys and queries are equal.
+        # However, in generation tasks, such as predicting the next token in a sequence, the lengths of keys and queries
+        # can differ. For instance, when employing key-value (KV) caching for sequence prediction, the keys
+        # represent embeddings of previous tokens and the current token, while the query corresponds
+        # to the embedding of the current token only.
+        assert (
+            key_len >= query_len
+        ), "Number of keys has to be greater than or equal to number of queries."
+
+        query_float = query.float()
+        key_float = key.float()
+
+        self._compute_sin_cos_embeddings(
+            key_len, key_device=key_float.device, key_dtype=key_float.dtype
+        )
+        query_float = _apply_rotary_pos_emb(
+            x=query_float,
+            pos_sin=self._cached_sin[..., key_len - query_len : key_len, :],
+            pos_cos=self._cached_cos[..., key_len - query_len : key_len, :],
+        )
+        key_float = _apply_rotary_pos_emb(
+            x=key_float,
+            pos_sin=self._cached_sin[..., :key_len, :],
+            pos_cos=self._cached_cos[..., :key_len, :],
+        )
+
+        return query_float.type_as(query), key_float.type_as(key)
+
+
+class OpenELMMultiHeadCausalAttention(nn.Module):
+    def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
+        super().__init__()
+        self.layer_idx = layer_idx
+        head_dim = config.head_dim
+        q_heads = config.num_query_heads[layer_idx]
+        k_heads = config.num_kv_heads[layer_idx]
+        v_heads = config.num_kv_heads[layer_idx]
+
+        self.qkv_proj = nn.Linear(
+            in_features=config.model_dim,
+            out_features=(q_heads + k_heads + v_heads) * head_dim,
+            bias=False,
+        )
+
+        self.pos_embedding = OpenELMRotaryEmbedding(
+            model_dim=config.head_dim,
+            max_seq_length=config.rope_max_length,
+            freq_constant=config.rope_freq_constant,
+        )
+
+        if config.normalize_qk_projections:
+            self.q_norm = OpenELMRMSNorm(
+                num_features=config.head_dim,
+            )
+            self.k_norm = OpenELMRMSNorm(
+                num_features=config.head_dim,
+            )
+        else:
+            self.q_norm = None
+            self.k_norm = None
+
+        self.out_proj = nn.Linear(
+            in_features=q_heads * head_dim,
+            out_features=config.model_dim,
+            bias=False,
+        )
+
+        self.head_dim = config.head_dim
+        self.num_q_heads = q_heads
+        self.num_k_heads = k_heads
+        self.num_v_heads = v_heads
+        self.transformer_dim = config.model_dim
+        self.num_groups = self.num_q_heads // self.num_k_heads
+
+    def extra_repr(self) -> str:
+        return (
+            super().extra_repr()
+            + f"query_heads={self.num_q_heads}, key_heads={self.num_k_heads}, value_heads={self.num_v_heads}"
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = 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]]]:
+        """
+        Forward pass of multi-head self-attention.
+
+        Args:
+            hidden_states: Input tensor of the shape [batch size, sequence length, model dimension].
+            past_key_value: Tensor storing the cached keys and values.
+            output_attentions: output attention weights.
+            use_cache: Specifies whether to use kv-cache for generation.
+            cache_position: used for updating the kv-cache.
+
+        Returns:
+            The output of the same shape as the input, optionally with a tensor containing cached keys and values.
+        """
+
+        # scaled_dot_product_attention does not return attention weights, set output_attentions to False
+        output_attentions = False
+        batch_size, seq_length, d_model = hidden_states.size()
+
+        # [B, S, d] --> [B, S, (q_h + k_h + v_h) * h]
+        qkv = self.qkv_proj(hidden_states)
+        # [B, S, (q_h + k_h + v_h) * h] --> [B, S, (q_h + k_h + v_h), h]
+        qkv = qkv.reshape(
+            batch_size,
+            seq_length,
+            self.num_q_heads + self.num_k_heads + self.num_v_heads,
+            self.head_dim,
+        )
+        # [B, S, (q_h + k_h + v_h), h] --> [B, (q_h + k_h + v_h), S, h]
+        qkv = qkv.transpose(1, 2)
+        # [B, (q_h + k_h + v_h), S, h] --> [B, q_h, S h], [B, k_h, S, h], [B, v_h, S, h]
+        queries, keys, values = qkv.split(
+            [self.num_q_heads, self.num_k_heads, self.num_v_heads], dim=1
+        )
+
+        if self.q_norm is not None:
+            queries = self.q_norm(queries)
+
+        if self.k_norm is not None:
+            keys = self.k_norm(keys)
+
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            # cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            cache_kwargs = {"cache_position": cache_position}
+            keys, values = past_key_value.update(
+                keys, values, self.layer_idx, cache_kwargs
+            )
+
+        # Add positional embedding
+        queries, keys = self.pos_embedding(queries, keys)
+
+        if self.num_groups != 1:
+            # GQA
+            # [B, k_h, S, h] --> [B, q_h, S, h]
+            keys = keys.repeat_interleave(self.num_groups, dim=1)
+            # [B, v_h, S, h] --> [B, q_h, S, h]
+            values = values.repeat_interleave(self.num_groups, dim=1)
+
+        causal_mask = attention_mask
+        if attention_mask is not None and cache_position is not None:
+            causal_mask = causal_mask[:, :, cache_position, : keys.shape[-2]]
+
+        attn_output = F.scaled_dot_product_attention(
+            queries,
+            keys,
+            values,
+            attn_mask=causal_mask,
+            dropout_p=0,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(
+            batch_size, seq_length, self.num_q_heads * self.head_dim
+        )
+        attn_output = self.out_proj(attn_output)
+        if not output_attentions:
+            attn_weights = None
+        return attn_output, attn_weights, past_key_value
+
+
+class OpenELMFeedForwardNetwork(nn.Module):
+    def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
+        super().__init__()
+        ffn_multiplier = config.ffn_multipliers[layer_idx]
+        intermediate_dim = int(
+            make_divisible(
+                ffn_multiplier * config.model_dim,
+                divisor=config.ffn_dim_divisor,
+            )
+        )
+        if config.ffn_with_glu:
+            # FFN with Gated linear unit, as described in https://arxiv.org/abs/2002.05202v1.
+            self.proj_1 = nn.Linear(
+                in_features=config.model_dim,
+                out_features=2 * intermediate_dim,
+                bias=False,
+            )
+            self.proj_2 = nn.Linear(
+                in_features=intermediate_dim,
+                out_features=config.model_dim,
+                bias=False,
+            )
+            self.ffn_with_glu = True
+        else:
+            # Standard FFN, as described in https://arxiv.org/abs/1706.03762
+            self.proj_1 = nn.Linear(
+                in_features=config.model_dim,
+                out_features=intermediate_dim,
+                bias=False,
+            )
+            self.proj_2 = nn.Linear(
+                in_features=intermediate_dim,
+                out_features=config.model_dim,
+                bias=False,
+            )
+            self.ffn_with_glu = False
+
+        self.act = ACT2FN[config.activation_fn_name]
+
+    def extra_repr(self) -> str:
+        return super().extra_repr() + f"(ffn_with_glu) : {self.ffn_with_glu}"
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Forward function of FFN layer.
+
+        Args:
+            x: Input tensor of the shape [batch size, sequence length, model dimension].
+
+        Returns:
+            A tensor of the same shape as the input.
+        """
+        if self.ffn_with_glu:
+            y_12 = self.proj_1(x)
+            y_1, y_2 = y_12.chunk(2, dim=-1)
+            y = self.act(y_1) * y_2
+            return self.proj_2(y)
+        else:
+            return self.proj_2(self.act(self.proj_1(x)))
+
+
+class OpenELMDecoderLayer(nn.Module):
+    def __init__(self, config: OpenELMConfig, layer_idx: int) -> None:
+        super().__init__()
+        self.attn = OpenELMMultiHeadCausalAttention(config=config, layer_idx=layer_idx)
+        self.ffn = OpenELMFeedForwardNetwork(config=config, layer_idx=layer_idx)
+        self.ffn_norm = OpenELMRMSNorm(
+            num_features=config.model_dim,
+        )
+        self.attn_norm = OpenELMRMSNorm(
+            num_features=config.model_dim,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **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
+        """
+        residual = hidden_states
+        hidden_states = self.attn_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.ffn_norm(hidden_states)
+        hidden_states = self.ffn(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
+
+
+class OpenELMModel(OpenELMPreTrainedModel):
+    config_class = OpenELMConfig
+
+    def __init__(self, config: OpenELMConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.token_embeddings = nn.Embedding(
+            embedding_dim=config.model_dim,
+            num_embeddings=config.vocab_size,
+        )
+
+        self.layers = nn.ModuleList(
+            OpenELMDecoderLayer(config=config, layer_idx=layer_idx)
+            for layer_idx in range(config.num_transformer_layers)
+        )
+        self.norm = OpenELMRMSNorm(num_features=config.model_dim)
+        if config.share_input_output_layers:
+            self.classifier = None
+        else:
+            self.classifier = nn.Linear(
+                in_features=config.model_dim,
+                out_features=config.vocab_size,
+                bias=False,
+            )
+        self.num_transformer_layers = config.num_transformer_layers
+        self.gradient_checkpointing = False
+
+        # Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
+        # NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_context_length`.
+        causal_mask = torch.full(
+            (config.max_context_length, config.max_context_length),
+            fill_value=True,
+            dtype=torch.bool,
+        )
+        self.register_buffer(
+            "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+        self.reset_parameters(config=config)
+
+    def get_input_embeddings(self):
+        return self.token_embeddings
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.token_embeddings = new_embeddings
+
+    def reset_parameters(self, config: OpenELMConfig) -> None:
+        """Initialize the layers in Language Model
+
+        The initialization scheme is followed, following `OPT <https://arxiv.org/pdf/2205.01068.pdf>`_.
+
+        Args:
+            use_megatron_std: Use standard deviation as described in Megatron-LM.
+
+        Returns:
+            None
+        """
+        for module in self.modules():
+            if isinstance(module, nn.Linear):
+                std = module.in_features**-0.5
+                torch.nn.init.normal_(module.weight, mean=0.0, std=std)
+                if module.bias is not None:
+                    torch.nn.init.zeros_(module.bias)
+            elif isinstance(module, nn.Embedding):
+                std = module.embedding_dim**-0.5
+                torch.nn.init.normal_(module.weight, mean=0.0, std=std)
+            elif isinstance(module, OpenELMRMSNorm):
+                if module.weight is not None:
+                    torch.nn.init.ones_(module.weight)
+                if hasattr(module, "bias") and module.bias is not None:
+                    torch.nn.init.zeros_(module.bias)
+
+        model_dim = config.model_dim
+        n_layers = config.num_transformer_layers
+        std = (model_dim**-0.5) * ((2 * n_layers) ** -0.5)
+        for param_name, param in self.named_parameters():
+            if param_name.endswith("out_proj.weight") or param_name.endswith(
+                "ffn.proj_2.weight"
+            ):
+                torch.nn.init.normal_(param, mean=0.0, std=std)
+
+    def forward(
+        self,
+        input_ids: 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,
+        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.token_embeddings(input_ids)
+
+        past_seen_tokens = 0
+        if use_cache:  # kept for BC (cache positions)
+            if not isinstance(past_key_values, StaticCache):
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_seen_tokens = past_key_values.get_seq_length()
+
+        if cache_position is None:
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + 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)
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # 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,
+                )
+            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,
+                )
+
+            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 = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache()
+                if isinstance(next_decoder_cache, Cache)
+                else next_decoder_cache
+            )
+        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,
+        )
+
+    def _update_causal_mask(self, attention_mask, input_tensor):
+        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
+
+        batch_size, seq_length = input_tensor.shape[:2]
+        dtype = input_tensor.dtype
+        device = input_tensor.device
+
+        # support going beyond cached `max_position_embedding`
+        if seq_length > self.causal_mask.shape[-1]:
+            causal_mask = torch.full(
+                (2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]),
+                fill_value=1,
+            )
+            self.register_buffer(
+                "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False
+            )
+
+        # We use the current dtype to avoid any overflows
+        min_dtype = torch.finfo(dtype).min
+        causal_mask = (
+            self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype)
+            * min_dtype
+        )
+
+        causal_mask = causal_mask.to(dtype=dtype, device=device)
+        if attention_mask is not None and attention_mask.dim() == 2:
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[
+                :, None, None, :
+            ].eq(0.0)
+            causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+        if self.config._attn_implementation == "sdpa" and attention_mask is not None:
+            # For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
+            is_tracing = (
+                torch.jit.is_tracing()
+                or isinstance(input_tensor, torch.fx.Proxy)
+                or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
+            )
+            if not is_tracing and torch.any(attention_mask != 1):
+                # Attend to all tokens in masked rows from 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 = causal_mask.mul(
+                    ~torch.all(causal_mask == min_dtype, dim=-1, keepdim=True)
+                ).to(dtype)
+
+        return causal_mask
+
+
+class OpenELMForCausalLM(OpenELMPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: OpenELMConfig):
+        super().__init__(config)
+        self.transformer = OpenELMModel(config)
+        self.vocab_size = config.vocab_size
+        if config.share_input_output_layers:
+            self.lm_head = None
+        else:
+            self.lm_head = nn.Linear(config.model_dim, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.transformer.token_embeddings
+
+    def set_input_embeddings(self, value):
+        self.transformer.token_embeddings = 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.transformer = decoder
+
+    def get_decoder(self):
+        return self.transformer
+
+    def forward(
+        self,
+        input_ids: 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,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        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.transformer(
+            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]
+        if self.lm_head is None:
+            # shared
+            logits = F.linear(
+                hidden_states, weight=self.transformer.token_embeddings.weight
+            )
+        else:
+            logits = self.lm_head(hidden_states)
+        logits = logits[:, : self.config.vocab_size]
+        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,
+        **kwargs,
+    ):
+        past_length = 0
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if (
+                attention_mask is not None
+                and attention_mask.shape[1] > input_ids.shape[1]
+            ):
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        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 self.generation_config.cache_implementation == "static":
+            # generation with static cache
+            cache_position = kwargs.get("cache_position", None)
+            if cache_position is None:
+                past_length = 0
+            else:
+                past_length = cache_position[-1] + 1
+            input_ids = input_ids[:, past_length:]
+            position_ids = position_ids[:, past_length:]
+
+        # we should only keep a `cache_position` in generate, and do +=1.
+        # same goes for position ids. Could also help with continued generation.
+        cache_position = torch.arange(
+            past_length,
+            past_length + position_ids.shape[-1],
+            device=position_ids.device,
+        )
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
+            # We could use `next_tokens` directly instead.
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids.contiguous(),
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx.to(past_state.device))
+                    for past_state in layer_past
+                ),
+            )
+        return reordered_past