Model save

README.md

@@ -0,0 +1,100 @@
+---
+library_name: transformers
+tags:
+- trl
+- cpo
+- generated_from_trainer
+model-index:
+- name: OpenELM-1_1B-SimPO
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# OpenELM-1_1B-SimPO
+
+This model was trained from scratch on an unknown dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.8496
+- Rewards/chosen: -1.1328
+- Rewards/rejected: -1.7031
+- Rewards/accuracies: 0.6680
+- Rewards/margins: 0.5742
+- Logps/rejected: -171.0
+- Logps/chosen: -113.0
+- Logits/rejected: 1.2422
+- Logits/chosen: -0.5781
+- Nll Loss: 0.0
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 5e-05
+- train_batch_size: 8
+- eval_batch_size: 16
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 64
+- total_eval_batch_size: 64
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 3
+
+### Training results
+
+| Training Loss | Epoch  | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen | Nll Loss |
+|:-------------:|:------:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:--------------:|:------------:|:---------------:|:-------------:|:--------:|
+| 0.9346        | 0.1047 | 100  | 0.9349          | -0.3320        | -0.4180          | 0.6133             | 0.0864          | -41.75         | -33.25       | -7.9688         | -8.5625       | 0.0      |
+| 0.9139        | 0.2093 | 200  | 0.9069          | -0.4844        | -0.6367          | 0.6270             | 0.1504          | -63.5          | -48.5        | -2.4375         | -3.4531       | 0.0      |
+| 0.907         | 0.3140 | 300  | 0.9099          | -0.6914        | -0.8359          | 0.6055             | 0.1416          | -83.5          | -69.5        | -4.0            | -5.1875       | 0.0      |
+| 0.901         | 0.4186 | 400  | 0.8957          | -0.8359        | -1.0156          | 0.6328             | 0.1748          | -101.0         | -84.0        | 0.0164          | -1.7422       | 0.0      |
+| 0.8752        | 0.5233 | 500  | 0.8768          | -0.7266        | -0.9570          | 0.6582             | 0.2324          | -95.5          | -72.5        | 0.8555          | -0.5625       | 0.0      |
+| 0.8808        | 0.6279 | 600  | 0.8742          | -0.8633        | -1.0938          | 0.6445             | 0.2334          | -109.5         | -86.0        | 3.2344          | 2.1562        | 0.0      |
+| 0.8277        | 0.7326 | 700  | 0.8679          | -0.5195        | -0.7734          | 0.6445             | 0.2520          | -77.5          | -52.0        | 0.3496          | -0.7930       | 0.0      |
+| 0.8341        | 0.8373 | 800  | 0.8503          | -0.8047        | -1.0859          | 0.6602             | 0.2773          | -108.5         | -80.5        | 1.3047          | 0.2188        | 0.0      |
+| 0.8333        | 0.9419 | 900  | 0.8454          | -0.8984        | -1.2188          | 0.6660             | 0.3184          | -121.5         | -90.0        | 1.8438          | 0.6406        | 0.0      |
+| 0.8071        | 1.0466 | 1000 | 0.8441          | -1.0           | -1.3359          | 0.6699             | 0.3340          | -133.0         | -100.0       | 1.3516          | 0.1504        | 0.0      |
+| 0.7845        | 1.1512 | 1100 | 0.8307          | -0.8477        | -1.2266          | 0.6660             | 0.3809          | -122.5         | -84.5        | 0.3301          | -1.5078       | 0.0      |
+| 0.7483        | 1.2559 | 1200 | 0.8353          | -0.9453        | -1.3281          | 0.6758             | 0.3809          | -133.0         | -94.5        | 0.9805          | -0.4160       | 0.0      |
+| 0.7802        | 1.3605 | 1300 | 0.8363          | -0.6211        | -1.0             | 0.7051             | 0.3828          | -100.5         | -62.0        | 0.3418          | -1.5859       | 0.0      |
+| 0.7499        | 1.4652 | 1400 | 0.8228          | -0.9727        | -1.4141          | 0.7012             | 0.4414          | -141.0         | -97.0        | 1.4531          | -0.1719       | 0.0      |
+| 0.6966        | 1.5699 | 1500 | 0.8231          | -1.0625        | -1.5234          | 0.6836             | 0.4609          | -152.0         | -106.0       | 1.5             | -0.3301       | 0.0      |
+| 0.6921        | 1.6745 | 1600 | 0.8222          | -1.0703        | -1.5469          | 0.6875             | 0.4766          | -155.0         | -107.0       | 2.25            | 0.6133        | 0.0      |
+| 0.7162        | 1.7792 | 1700 | 0.8106          | -1.0312        | -1.5391          | 0.6953             | 0.5078          | -154.0         | -103.0       | 2.4688          | 0.6992        | 0.0      |
+| 0.714         | 1.8838 | 1800 | 0.8183          | -1.0938        | -1.625           | 0.6855             | 0.5312          | -162.0         | -109.5       | 2.1875          | 0.0579        | 0.0      |
+| 0.7068        | 1.9885 | 1900 | 0.8164          | -0.9727        | -1.5078          | 0.7031             | 0.5352          | -151.0         | -97.5        | 1.9922          | 0.3184        | 0.0      |
+| 0.4781        | 2.0931 | 2000 | 0.8475          | -1.1875        | -1.7109          | 0.6797             | 0.5273          | -171.0         | -119.0       | 1.7344          | 0.0977        | 0.0      |
+| 0.4964        | 2.1978 | 2100 | 0.8455          | -1.0           | -1.5547          | 0.6875             | 0.5547          | -155.0         | -100.0       | 0.9219          | -0.9258       | 0.0      |
+| 0.4723        | 2.3025 | 2200 | 0.8475          | -1.1016        | -1.6562          | 0.6934             | 0.5586          | -166.0         | -110.0       | 1.2969          | -0.4648       | 0.0      |
+| 0.5051        | 2.4071 | 2300 | 0.8480          | -1.1328        | -1.6953          | 0.6895             | 0.5664          | -170.0         | -113.0       | 1.4141          | -0.2891       | 0.0      |
+| 0.4647        | 2.5118 | 2400 | 0.8463          | -1.1406        | -1.7188          | 0.6758             | 0.5742          | -171.0         | -114.0       | 1.4531          | -0.3496       | 0.0      |
+| 0.4442        | 2.6164 | 2500 | 0.8527          | -1.2344        | -1.7969          | 0.6680             | 0.5664          | -180.0         | -123.5       | 1.5859          | -0.1436       | 0.0      |
+| 0.4349        | 2.7211 | 2600 | 0.8505          | -1.1172        | -1.6953          | 0.6699             | 0.5742          | -169.0         | -112.0       | 1.2422          | -0.5898       | 0.0      |
+| 0.4514        | 2.8257 | 2700 | 0.8493          | -1.1172        | -1.6953          | 0.6738             | 0.5781          | -169.0         | -112.0       | 1.1953          | -0.6406       | 0.0      |
+| 0.459         | 2.9304 | 2800 | 0.8496          | -1.1328        | -1.7031          | 0.6680             | 0.5742          | -171.0         | -113.0       | 1.2422          | -0.5781       | 0.0      |
+
+
+### Framework versions
+
+- Transformers 4.44.2
+- Pytorch 2.3.0
+- Datasets 3.0.0
+- Tokenizers 0.19.1

all_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 2.998430141287284,
+    "total_flos": 0.0,
+    "train_loss": 0.6978365646815009,
+    "train_runtime": 9531.6254,
+    "train_samples": 61134,
+    "train_samples_per_second": 19.241,
+    "train_steps_per_second": 0.301
+}

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

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.44.2"
+}

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

train_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 2.998430141287284,
+    "total_flos": 0.0,
+    "train_loss": 0.6978365646815009,
+    "train_runtime": 9531.6254,
+    "train_samples": 61134,
+    "train_samples_per_second": 19.241,
+    "train_steps_per_second": 0.301
+}