From d0ed91720f46ef9d5b47e6110dddb2a3e71482b7 Mon Sep 17 00:00:00 2001
From: Eustache Le Bihan <eulebihan@gmail.com>
Date: Fri, 13 Dec 2024 18:42:07 +0100
Subject: [PATCH] fix cache handling and causality for cross attention

---
 .../moonshine/configuration_moonshine.py      |  236 ++
 .../moonshine/convert_usefulsensors_to_hf.py  |  145 ++
 .../models/moonshine/modeling_moonshine.py    | 2078 +++++++++++++++++
 .../models/moonshine/modular_moonshine.py     |   26 +-
 4 files changed, 2469 insertions(+), 16 deletions(-)
 create mode 100644 src/transformers/models/moonshine/configuration_moonshine.py
 create mode 100644 src/transformers/models/moonshine/convert_usefulsensors_to_hf.py
 create mode 100644 src/transformers/models/moonshine/modeling_moonshine.py

diff --git a/src/transformers/models/moonshine/configuration_moonshine.py b/src/transformers/models/moonshine/configuration_moonshine.py
new file mode 100644
index 00000000000000..44f043d6d3c3dd
--- /dev/null
+++ b/src/transformers/models/moonshine/configuration_moonshine.py
@@ -0,0 +1,236 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/moonshine/modular_moonshine.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_moonshine.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+
+from ...configuration_utils import PretrainedConfig
+
+
+class MoonshineConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MoonshineModel`]. It is used to instantiate a Moonshine
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Moonshine
+    [UsefulSensors/moonshine](https://huggingface.co/UsefulSensors/moonshine).
+
+    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 32768):
+            Vocabulary size of the Moonshine model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`MoonshineModel`].
+        hidden_size (`int`, *optional*, defaults to 288):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 6):
+            Number of hidden layers in the Transformer encoder and decoder.
+        num_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder and decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        encoder_hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder.
+        decoder_hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. TODO: check this
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        decoder_start_token_id (`int`, *optional*, defaults to 1):
+            Corresponds to the "<|startoftranscript|>" token, which is automatically used when no `decoder_input_ids`
+            are provided to the `generate` function. It is used to guide the model`s generation process depending on
+            the task.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
+            Whether the model is used as an encoder/decoder or not.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings. TODO: check this
+        partial_rotary_factor (`float`, *optional*, defaults to 0.5):
+            Percentage of the query and keys which will have rotary embedding. TODO: check this
+        ff_mult (`int`, *optional*, defaults to 4):
+            Factor by which to scale the intermediate size.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        qk_layernorm (`bool`, *optional*, defaults to `False`):
+            Whether or not to normalize the Queries and Keys after projecting the hidden states.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Denotes beginning of sequences token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            Denotes end of sequences token id.
+        apply_spec_augment (`bool`, *optional*, defaults to `False`):
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
+            [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
+            Recognition](https://arxiv.org/abs/1904.08779).
+        mask_time_prob (`float`, *optional*, defaults to 0.05):
+            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking
+            procecure generates `mask_time_prob*len(time_axis)/mask_time_length` independent masks over the axis. If
+            reasoning from the propability of each feature vector to be chosen as the start of the vector span to be
+            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the
+            actual percentage of masked vectors. This is only relevant if `apply_spec_augment == True`.
+        mask_time_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the time axis.
+        mask_time_min_masks (`int`, *optional*, defaults to 2),:
+            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,
+            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <
+            mask_time_min_masks''
+        mask_feature_prob (`float`, *optional*, defaults to 0.0):
+            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The
+            masking procecure generates `mask_feature_prob*len(feature_axis)/mask_time_length` independent masks over
+            the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector
+            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap
+            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is
+            True`.
+        mask_feature_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the feature axis.
+        mask_feature_min_masks (`int`, *optional*, defaults to 0),:
+            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time
+            step, irrespectively of `mask_feature_prob`. Only relevant if
+            `mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks`.
+
+    Example:
+
+    ```python
+    >>> from transformers import MoonshineModel, MoonshineConfig
+
+    >>> # Initializing a Moonshine style configuration
+    >>> configuration = MoonshineConfig().from_pretrained("UsefulSensors/moonshine")
+
+    >>> # Initializing a model from the configuration
+    >>> model = MoonshineModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "moonshine"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32768,
+        hidden_size=288,
+        intermediate_size=None,
+        num_hidden_layers=6,
+        num_attention_heads=8,
+        num_key_value_heads=None,
+        encoder_hidden_act="gelu",
+        decoder_hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        decoder_start_token_id=1,
+        use_cache=True,
+        is_encoder_decoder=True,
+        rope_theta=10000.0,
+        partial_rotary_factor=0.5,
+        attention_bias=False,
+        attention_dropout=0.0,
+        qk_layernorm=False,
+        rope_scaling=None,
+        ff_mult=4,
+        bos_token_id=1,
+        eos_token_id=2,
+        apply_spec_augment=False,
+        mask_time_prob=0.05,
+        mask_time_length=10,
+        mask_time_min_masks=2,
+        mask_feature_prob=0.0,
+        mask_feature_length=10,
+        mask_feature_min_masks=0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.encoder_hidden_act = encoder_hidden_act
+        self.decoder_hidden_act = decoder_hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.is_encoder_decoder = is_encoder_decoder
+        self.rope_theta = rope_theta
+        self.partial_rotary_factor = partial_rotary_factor
+
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.qk_layernorm = qk_layernorm
+        self.rope_scaling = rope_scaling
+        self.ff_mult = ff_mult
+
+        # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
+        self.apply_spec_augment = apply_spec_augment
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
+        self.mask_time_min_masks = mask_time_min_masks
+        self.mask_feature_prob = mask_feature_prob
+        self.mask_feature_length = mask_feature_length
+        self.mask_feature_min_masks = mask_feature_min_masks
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )
diff --git a/src/transformers/models/moonshine/convert_usefulsensors_to_hf.py b/src/transformers/models/moonshine/convert_usefulsensors_to_hf.py
new file mode 100644
index 00000000000000..cf0010fd552ddd
--- /dev/null
+++ b/src/transformers/models/moonshine/convert_usefulsensors_to_hf.py
@@ -0,0 +1,145 @@
+#!/usr/bin/env python
+"""Converts a Moonshine model in Useful Sensors format to Hugging Face format."""
+# Copyright 2022 The HuggingFace Inc. team and the OpenAI team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from huggingface_hub import hf_hub_download
+
+import h5py
+import torch
+import numpy as np
+import re
+
+from transformers.models.moonshine.modeling_moonshine import MoonshineConfig
+from transformers.models.moonshine.modeling_moonshine import MoonshineDecoder
+
+
+# Copied from https://github.com/usefulsensors/moonshine/blob/a1d77cc573b0471ac4602b86f67b3f48d67df1a9/moonshine/model.py
+def _get_weights(model_name):
+    repo = "UsefulSensors/moonshine"
+
+    return (
+        hf_hub_download(repo, f"{x}.weights.h5", subfolder=model_name)
+        for x in ("preprocessor", "encoder", "decoder")
+    )
+
+
+def _read_h5_weights(group, current_key="", weights={}):
+    for key in group.keys():
+        full_key = f"{current_key}.{key}" if current_key else key
+        if isinstance(group[key], h5py.Dataset):
+            w = np.array(group[key])
+            w = torch.from_numpy(w)
+            if len(w.shape) > 1:
+                if len(w.shape) == 3:
+                    hidden_size = max(list(w.shape))
+                    try:
+                        w = w.reshape(hidden_size, hidden_size)
+                    except RuntimeError:
+                        # meaning its a conv layers
+                        pass
+                w = w.transpose(0, -1)
+            weights[full_key] = w
+        else:
+            _read_h5_weights(group[key], full_key, weights)
+    return weights
+
+
+def _convert_layer_names(name, gated_mlp=False):
+    name = re.sub(r'layers\.functional(?:_(\d+))?\.layers', lambda m: f'layers.{m.group(1) if m.group(1) else "0"}', name, count=1)
+    if gated_mlp:
+        name = re.sub(r"functional\.layers\.dense\.", "mlp.up_proj.", name)
+        name = re.sub(r"functional\.layers\.dense_1\.", "mlp.down_proj.", name)
+    else:
+        name = re.sub(r"functional\.layers\.sequential\.layers\.dense\.", "mlp.fc1.", name)
+        name = re.sub(r"functional\.layers\.sequential\.layers\.dense_1\.", "mlp.fc2.", name)
+    name = re.sub(r"layers\.sequential\.layers\.conv1d\.", "conv1.", name)
+    name = re.sub(r"layers\.sequential\.layers\.conv1d_1\.", "conv2.", name)
+    name = re.sub(r"layers\.sequential\.layers\.conv1d_2\.", "conv3.", name)
+    name = re.sub(r"layers\.sequential\.layers\.group_normalization\.", "groupnorm.", name)
+    name = re.sub(r"mha_with_rope\.key_dense", "self_attn.k_proj", name)
+    name = re.sub(r"mha_with_rope\.query_dense", "self_attn.q_proj", name)
+    name = re.sub(r"mha_with_rope\.value_dense", "self_attn.v_proj", name)
+    name = re.sub(r"mha_with_rope\.output_dense", "self_attn.dense", name)
+    name = re.sub(r"mha_precomputed_kv\.key_dense", "encoder_attn.k_proj", name)
+    name = re.sub(r"mha_precomputed_kv\.query_dense", "encoder_attn.q_proj", name)
+    name = re.sub(r"mha_precomputed_kv\.value_dense", "encoder_attn.v_proj", name)
+    name = re.sub(r"mha_precomputed_kv\.output_dense", "encoder_attn.dense", name)
+    name = re.sub(r"mha_causal_with_rope\.key_dense", "self_attn.k_proj", name)
+    name = re.sub(r"mha_causal_with_rope\.query_dense", "self_attn.q_proj", name)
+    name = re.sub(r"mha_causal_with_rope\.value_dense", "self_attn.v_proj", name)
+    name = re.sub(r"mha_causal_with_rope\.output_dense", "self_attn.dense", name)
+    name = re.sub(r"layer_normalization\.", "input_layernorm.", name)
+    name = re.sub(r"layer_normalization_1\.", "post_attention_layernorm.", name)
+    name = re.sub(r"layer_normalization_2\.", "final_layernorm.", name)
+    name = re.sub(r"vars\.0", "weight", name)
+    name = re.sub(r"vars\.1", "bias", name)
+    name = re.sub(r"layers\.reversible_embedding", "embed_tokens", name)
+
+    return name
+
+
+def _convert_weights(weights, encoder=True):
+    if "layers.rotary_embedding.vars.0" in weights:
+        weights.pop("layers.rotary_embedding.vars.0")
+
+    converted_weights = {}
+    if encoder:
+        converted_weights["layer_norm.weight"] = weights.pop("layers.layer_normalization.vars.0")
+    else:
+        converted_weights["norm.weight"] = weights.pop("layers.layer_normalization.vars.0")
+
+    for name, w in weights.items():
+        if encoder:
+            new_name = _convert_layer_names(name)
+        else:
+            new_name = _convert_layer_names(name, gated_mlp=True)
+        converted_weights[new_name] = w
+
+    return converted_weights
+
+
+def convert_usefulsensors_moonshine_to_hf(model_name, pytorch_dump_folder_path):
+    preprocessor_weights_path, encoder_weights_path, decoder_weights_path = _get_weights(model_name)
+    
+    with h5py.File(preprocessor_weights_path, 'r') as f:
+        loaded_preprocessor_weights = _read_h5_weights(f, weights={})
+
+    with h5py.File(encoder_weights_path, 'r') as f:
+        loaded_encoder_weights = _read_h5_weights(f, weights={})
+
+    with h5py.File(decoder_weights_path, 'r') as f:
+        loaded_decoder_weights = _read_h5_weights(f, weights={})
+
+    encoder_state_dict = {**loaded_encoder_weights, **loaded_preprocessor_weights}
+    encoder_state_dict = _convert_weights(encoder_state_dict)
+
+    converted_decoder_weights = _convert_weights(loaded_decoder_weights, encoder=False)
+    converted_decoder_weights['embed_tokens.weight'] = converted_decoder_weights['embed_tokens.weight'].T
+
+
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # # Required parameters
+    parser.add_argument("--model_name", type=str, help="Path to the downloaded checkpoints")
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    args = parser.parse_args()
+
+    convert_usefulsensors_moonshine_to_hf(
+        args.model_name, args.pytorch_dump_folder_path
+    )
+
diff --git a/src/transformers/models/moonshine/modeling_moonshine.py b/src/transformers/models/moonshine/modeling_moonshine.py
new file mode 100644
index 00000000000000..9f46520f40d86d
--- /dev/null
+++ b/src/transformers/models/moonshine/modeling_moonshine.py
@@ -0,0 +1,2078 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/moonshine/modular_moonshine.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_moonshine.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+
+import copy
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+from packaging import version
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache
+from ...generation import GenerationMixin
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_flash_attention_utils import FlashAttentionKwargs, _flash_attention_forward
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPast,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from ...modeling_utils import PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    get_torch_version,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_moonshine import MoonshineConfig
+
+
+if is_flash_attn_2_available():
+    from ...modeling_flash_attention_utils import _flash_attention_forward
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "MoonshineConfig"
+
+
+class MoonshineRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim=None,
+        max_position_embeddings=2048,
+        base=10000,
+        device=None,
+        scaling_factor=1.0,
+        rope_type="default",
+        config: Optional[MoonshineConfig] = None,
+    ):
+        super().__init__()
+        # TODO (joao): remove the `if` below, only used for BC
+        self.rope_kwargs = {}
+        if config is None:
+            logger.warning_once(
+                "`MoonshineRotaryEmbedding` can now be fully parameterized by passing the model config through the "
+                "`config` argument. All other arguments will be removed in v4.46"
+            )
+            self.rope_kwargs = {
+                "rope_type": rope_type,
+                "factor": scaling_factor,
+                "dim": dim,
+                "base": base,
+                "max_position_embeddings": max_position_embeddings,
+            }
+            self.rope_type = rope_type
+            self.max_seq_len_cached = max_position_embeddings
+            self.original_max_seq_len = max_position_embeddings
+        else:
+            # BC: "rope_type" was originally "type"
+            if config.rope_scaling is not None:
+                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+            else:
+                self.rope_type = "default"
+            self.max_seq_len_cached = config.max_position_embeddings
+            self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len, **self.rope_kwargs
+            )
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.stack((freqs, freqs), dim=-1)
+            emb = emb.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class MoonshineNonGatedMLP(nn.Module):
+    def __init__(self, config: MoonshineConfig, hidden_act: str):
+        super().__init__()
+        config = copy.deepcopy(config)
+        config.hidden_act = hidden_act
+        if config.intermediate_size is None:
+            config.intermediate_size = config.hidden_size * config.ff_mult
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class MoonshineGatedMLP(nn.Module):
+    def __init__(self, config: MoonshineConfig, hidden_act: str):
+        super().__init__()
+        config = copy.deepcopy(config)
+        config.hidden_act = hidden_act
+        if config.intermediate_size is None:
+            config.intermediate_size = config.hidden_size * config.ff_mult * 2
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
+        self.down_proj = nn.Linear(self.intermediate_size // 2, self.hidden_size, bias=True)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_state):
+        hidden_state = self.up_proj(hidden_state)
+        hidden_state, gate = hidden_state.chunk(2, dim=-1)
+        hidden_state = self.act_fn(gate) * hidden_state
+        return self.down_proj(hidden_state)
+
+
+class MoonshineMLP:
+    def __new__(cls, config: MoonshineConfig, hidden_act: str):
+        if hidden_act == "gelu":
+            return MoonshineNonGatedMLP(config, hidden_act)
+        elif hidden_act == "silu":
+            return MoonshineGatedMLP(config, hidden_act)
+        else:
+            raise ValueError(f"Unsupported activation function: {hidden_act}, please use 'gelu' or 'silu'")
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def rotate_every_two(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_embed = (q * cos) + (rotate_every_two(q) * sin)
+    k_embed = (k * cos) + (rotate_every_two(k) * sin)
+    return q_embed, k_embed
+
+
+class MoonshineAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: MoonshineConfig, layer_idx: Optional[int] = None, is_causal: bool = False):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.rope_theta = config.rope_theta
+        self.rotary_ndims = max(config.hidden_size // config.num_attention_heads // 2, 32)
+
+        self.is_causal = is_causal
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.dense = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
+
+        self.qk_layernorm = config.qk_layernorm
+        if self.qk_layernorm:
+            self.q_layernorm = nn.LayerNorm(
+                config.hidden_size // self.num_heads, eps=config.layer_norm_eps, elementwise_affine=True
+            )
+            self.k_layernorm = nn.LayerNorm(
+                config.hidden_size // self.num_heads, eps=config.layer_norm_eps, elementwise_affine=True
+            )
+
+        self.rotary_emb = MoonshineRotaryEmbedding(
+            dim=self.rotary_ndims,
+            max_position_embeddings=config.max_position_embeddings,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        key_value_states: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        if past_key_value is not None:
+            is_updated = past_key_value.is_updated.get(self.layer_idx)
+            if is_cross_attention:
+                # after the first generated id, we can subsequently re-use all key/value_states from cache
+                past_key_value.is_updated[self.layer_idx] = True
+                past_key_value = past_key_value.cross_attention_cache
+            else:
+                past_key_value = past_key_value.self_attention_cache
+
+        # use key_value_states if cross attention
+        current_states = key_value_states if key_value_states is not None else hidden_states
+        if is_cross_attention and past_key_value and is_updated:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value.key_cache[self.layer_idx]
+            value_states = past_key_value.value_cache[self.layer_idx]
+        else:
+            key_states = self.k_proj(current_states)
+            value_states = self.v_proj(current_states)
+            key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
+                )
+
+        if self.qk_layernorm:
+            query_states = self.q_layernorm(query_states)
+            key_states = self.k_layernorm(key_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        if not is_cross_attention:
+            if position_embeddings is None:
+                logger.warning_once(
+                    "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                    "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                    "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                    "removed and `position_embeddings` will be mandatory."
+                )
+                cos, sin = self.rotary_emb(value_states, position_ids)
+            else:
+                cos, sin = position_embeddings
+
+            # Partial rotary embedding
+            query_rot, query_pass = (
+                query_states[..., : self.rotary_ndims],
+                query_states[..., self.rotary_ndims :],
+            )
+            key_rot, key_pass = (
+                key_states[..., : self.rotary_ndims],
+                key_states[..., self.rotary_ndims :],
+            )
+            # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+            query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
+
+            # [batch_size, seq_length, num_heads, head_dim]
+            query_states = torch.cat((query_rot, query_pass), dim=-1)
+            key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+        if past_key_value is not None:
+            if not is_cross_attention:
+                cache_kwargs = {
+                    "sin": sin,
+                    "cos": cos,
+                    "partial_rotation_size": self.rotary_ndims,
+                    "cache_position": cache_position,
+                }
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, cache_kwargs
+                )
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        # Queries and keys upcast to fp32 is required by Moonshine-2 to avoid overflow
+        attn_weights = torch.matmul(
+            query_states.to(torch.float32), key_states.to(torch.float32).transpose(2, 3)
+        ) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights += causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class MoonshineFlashAttention2(MoonshineAttention):
+    """
+    Moonshine flash attention module. This module inherits from `MoonshineAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        key_value_states: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        if past_key_value is not None:
+            is_updated = past_key_value.is_updated.get(self.layer_idx)
+            if is_cross_attention:
+                # after the first generated id, we can subsequently re-use all key/value_states from cache
+                past_key_value.is_updated[self.layer_idx] = True
+                past_key_value = past_key_value.cross_attention_cache
+            else:
+                past_key_value = past_key_value.self_attention_cache
+
+        # use key_value_states if cross attention
+        current_states = key_value_states if key_value_states is not None else hidden_states
+        if is_cross_attention and past_key_value and is_updated:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value.key_cache[self.layer_idx]
+            value_states = past_key_value.value_cache[self.layer_idx]
+        else:
+            key_states = self.k_proj(current_states)
+            value_states = self.v_proj(current_states)
+            key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
+                )
+
+        if self.qk_layernorm:
+            query_states = self.q_layernorm(query_states)
+            key_states = self.k_layernorm(key_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        if not is_cross_attention:
+            if position_embeddings is None:
+                logger.warning_once(
+                    "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                    "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                    "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                    "removed and `position_embeddings` will be mandatory."
+                )
+                cos, sin = self.rotary_emb(value_states, position_ids)
+            else:
+                cos, sin = position_embeddings
+
+            # Partial rotary embedding
+            query_rot, query_pass = (
+                query_states[..., : self.rotary_ndims],
+                query_states[..., self.rotary_ndims :],
+            )
+            key_rot, key_pass = (
+                key_states[..., : self.rotary_ndims],
+                key_states[..., self.rotary_ndims :],
+            )
+            # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+            query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
+
+            # [batch_size, seq_length, num_heads, head_dim]
+            query_states = torch.cat((query_rot, query_pass), dim=-1)
+            key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+        if past_key_value is not None:
+            if not is_cross_attention:
+                cache_kwargs = {
+                    "sin": sin,
+                    "cos": cos,
+                    "partial_rotation_size": self.rotary_ndims,
+                    "cache_position": cache_position,
+                }
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, cache_kwargs
+                )
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_dropout = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32.
+
+        if query_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            position_ids=position_ids,
+            dropout=attn_dropout,
+            softmax_scale=None,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class MoonshineSdpaAttention(MoonshineAttention):
+    """
+    SDPA attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `MoonshineAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.require_contiguous_qkv = version.parse(get_torch_version()) < version.parse("2.2.0")
+
+    # Adapted from MoonshineAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        key_value_states: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        if past_key_value is not None:
+            is_updated = past_key_value.is_updated.get(self.layer_idx)
+            if is_cross_attention:
+                # after the first generated id, we can subsequently re-use all key/value_states from cache
+                past_key_value.is_updated[self.layer_idx] = True
+                past_key_value = past_key_value.cross_attention_cache
+            else:
+                past_key_value = past_key_value.self_attention_cache
+
+        # use key_value_states if cross attention
+        current_states = key_value_states if key_value_states is not None else hidden_states
+        if is_cross_attention and past_key_value and is_updated:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value.key_cache[self.layer_idx]
+            value_states = past_key_value.value_cache[self.layer_idx]
+        else:
+            key_states = self.k_proj(current_states)
+            value_states = self.v_proj(current_states)
+            key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
+                )
+
+        if self.qk_layernorm:
+            query_states = self.q_layernorm(query_states)
+            key_states = self.k_layernorm(key_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        if not is_cross_attention:
+            if position_embeddings is None:
+                logger.warning_once(
+                    "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                    "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                    "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                    "removed and `position_embeddings` will be mandatory."
+                )
+                cos, sin = self.rotary_emb(value_states, position_ids)
+            else:
+                cos, sin = position_embeddings
+
+            # Partial rotary embedding
+            query_rot, query_pass = (
+                query_states[..., : self.rotary_ndims],
+                query_states[..., self.rotary_ndims :],
+            )
+            key_rot, key_pass = (
+                key_states[..., : self.rotary_ndims],
+                key_states[..., self.rotary_ndims :],
+            )
+            # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+            query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
+
+            # [batch_size, seq_length, num_heads, head_dim]
+            query_states = torch.cat((query_rot, query_pass), dim=-1)
+            key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+        if past_key_value is not None:
+            if not is_cross_attention:
+                cache_kwargs = {
+                    "sin": sin,
+                    "cos": cos,
+                    "partial_rotation_size": self.rotary_ndims,
+                    "cache_position": cache_position,
+                }
+                key_states, value_states = past_key_value.update(
+                    key_states, value_states, self.layer_idx, cache_kwargs
+                )
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is broken in torch==2.1.2 when using non-contiguous inputs and a custom
+        # attn_mask, so we need to call `.contiguous()` here. This was fixed in torch==2.2.0.
+        # Reference: https://github.com/pytorch/pytorch/issues/112577
+        if self.require_contiguous_qkv and query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        is_causal = True if self.is_causal and causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.dense(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+MOONSHINE_ATTENTION_CLASSES = {
+    "eager": MoonshineAttention,
+    "flash_attention_2": MoonshineFlashAttention2,
+    "sdpa": MoonshineSdpaAttention,
+}
+
+
+class MoonshineEncoderLayer(nn.Module):
+    def __init__(self, config: MoonshineConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+
+        self.mlp = MoonshineMLP(config, config.encoder_hidden_act)
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*):
+                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
+                query_sequence_length, key_sequence_length)` if default attention is used.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence
+            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class MoonshineDecoderLayer(nn.Module):
+    def __init__(self, config: MoonshineConfig, layer_idx: int = None):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](
+            config=config, layer_idx=layer_idx, is_causal=True
+        )
+        self.encoder_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](
+            config=config, layer_idx=layer_idx, is_causal=False
+        )
+
+        self.mlp = MoonshineMLP(config, config.decoder_hidden_act)
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+        self.final_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        encoder_position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        encoder_position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = 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.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence
+            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            encoder_position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, encoder_seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.post_attention_layernorm(hidden_states)
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                position_ids=encoder_position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                position_embeddings=encoder_position_embeddings,
+            )
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 1 of present_key_value tuple
+            present_key_value = (present_key_value, cross_attn_present_key_value)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+MOONSHINE_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`MoonshineConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Moonshine Model outputting raw hidden-states without any specific head on top.",
+    MOONSHINE_START_DOCSTRING,
+)
+class MoonshinePreTrainedModel(PreTrainedModel):
+    config_class = MoonshineConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MoonshineDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+MOONSHINE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Moonshine Model outputting raw hidden-states without any specific head on top.",
+    MOONSHINE_START_DOCSTRING,
+)
+class MoonshineEncoder(MoonshinePreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MoonshineDecoderLayer`]
+
+    Args:
+        config: MoonshineConfig
+    """
+
+    main_input_name = "input_features"
+
+    def __init__(self, config: MoonshineConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.conv1 = nn.Conv1d(1, embed_dim, kernel_size=127, stride=64, bias=False)
+        self.conv2 = nn.Conv1d(embed_dim, 2 * embed_dim, kernel_size=7, stride=3)
+        self.conv3 = nn.Conv1d(2 * embed_dim, embed_dim, kernel_size=3, stride=2)
+        self.groupnorm = nn.GroupNorm(num_groups=1, num_channels=embed_dim, eps=1e-5)
+
+        self.rotary_emb = MoonshineRotaryEmbedding(
+            dim=max(config.hidden_size // config.num_attention_heads // 2, 32),
+            max_position_embeddings=config.max_position_embeddings,
+        )
+
+        self.layers = nn.ModuleList([MoonshineEncoderLayer(config, idx) for idx in range(config.num_hidden_layers)])
+        self.layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps, bias=False)
+
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.conv1
+
+    def set_input_embeddings(self, value: nn.Module):
+        self.conv1 = value
+
+    @add_start_docstrings_to_model_forward(MOONSHINE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> 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_features is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        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.preprocess(input_features)
+
+        # kept for BC (non `Cache` `past_key_values` inputs)
+        return_legacy_cache = False
+        if use_cache or past_key_values is not None:
+            if isinstance(past_key_values, Cache) and not isinstance(past_key_values, EncoderDecoderCache):
+                past_key_values = EncoderDecoderCache(past_key_values, DynamicCache())
+            elif not isinstance(past_key_values, EncoderDecoderCache):
+                return_legacy_cache = True
+                logger.warning_once(
+                    "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.43.0. "
+                    "You should pass an instance of `EncoderDecoderCache` instead, e.g. "
+                    "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`."
+                )
+                past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            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, cache_position, past_key_values, output_attentions
+        )
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            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.layer_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_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: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+    def _freeze_parameters(self):
+        for param in self.parameters():
+            param.requires_grad = False
+        self._requires_grad = False
+
+    def preprocess(self, input_features: torch.FloatTensor):
+        input_features = input_features.unsqueeze(1)
+        inputs_embeds = nn.functional.tanh(self.conv1(input_features))
+        inputs_embeds = self.groupnorm(inputs_embeds)
+        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
+        inputs_embeds = nn.functional.gelu(self.conv3(inputs_embeds))
+        inputs_embeds = inputs_embeds.permute(0, 2, 1)
+        return inputs_embeds
+
+
+@add_start_docstrings(
+    "The bare Moonshine Model outputting raw hidden-states without any specific head on top.",
+    MOONSHINE_START_DOCSTRING,
+)
+class MoonshineDecoder(MoonshinePreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MoonshineDecoderLayer`]
+
+    Args:
+        config: MoonshineConfig
+    """
+
+    def __init__(self, config: MoonshineConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [MoonshineDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
+        self.rotary_emb = MoonshineRotaryEmbedding(
+            dim=max(config.hidden_size // config.num_attention_heads // 2, 32),
+            max_position_embeddings=config.max_position_embeddings,
+        )
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(MOONSHINE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        encoder_position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> 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 must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        # kept for BC (non `Cache` `past_key_values` inputs)
+        return_legacy_cache = False
+        return_self_attention_cache = False
+        if use_cache or past_key_values is not None:
+            if isinstance(past_key_values, Cache) and not isinstance(past_key_values, EncoderDecoderCache):
+                return_self_attention_cache = True
+                past_key_values = EncoderDecoderCache(past_key_values, DynamicCache())
+            elif not isinstance(past_key_values, EncoderDecoderCache):
+                return_legacy_cache = True
+                logger.warning_once(
+                    "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.43.0. "
+                    "You should pass an instance of `EncoderDecoderCache` instead, e.g. "
+                    "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`."
+                )
+                past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            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)
+
+        if encoder_position_ids is None:
+            encoder_position_ids = torch.arange(
+                encoder_hidden_states.shape[1], device=encoder_hidden_states.device
+            ).unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        encoder_position_embeddings = self.rotary_emb(encoder_hidden_states, encoder_position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) 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,
+                    encoder_hidden_states,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    position_ids=position_ids,
+                    encoder_position_ids=encoder_position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    encoder_position_embeddings=encoder_position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        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 = past_key_values if use_cache else None
+        if return_self_attention_cache:
+            next_cache = past_key_values.self_attention_cache
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_cache()
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+def _compute_mask_indices(
+    shape: Tuple[int, int],
+    mask_prob: float,
+    mask_length: int,
+    attention_mask: Optional[torch.LongTensor] = None,
+    min_masks: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
+    ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
+    CPU as part of the preprocessing during training.
+
+    Args:
+        shape: The shape for which to compute masks. This should be of a tuple of size 2 where
+               the first element is the batch size and the second element is the length of the axis to span.
+        mask_prob:  The percentage of the whole axis (between 0 and 1) which will be masked. The number of
+                    independently generated mask spans of length `mask_length` is computed by
+                    `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
+                    actual percentage will be smaller.
+        mask_length: size of the mask
+        min_masks: minimum number of masked spans
+        attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
+                        each batch dimension.
+    """
+    batch_size, sequence_length = shape
+
+    if mask_length < 1:
+        raise ValueError("`mask_length` has to be bigger than 0.")
+
+    if mask_length > sequence_length:
+        raise ValueError(
+            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
+            f" and `sequence_length`: {sequence_length}`"
+        )
+
+    # epsilon is used for probabilistic rounding
+    epsilon = np.random.rand(1).item()
+
+    def compute_num_masked_span(input_length):
+        """Given input length, compute how many spans should be masked"""
+        num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
+        num_masked_span = max(num_masked_span, min_masks)
+
+        # make sure num masked span <= sequence_length
+        if num_masked_span * mask_length > sequence_length:
+            num_masked_span = sequence_length // mask_length
+
+        # make sure num_masked span is also <= input_length - (mask_length - 1)
+        if input_length - (mask_length - 1) < num_masked_span:
+            num_masked_span = max(input_length - (mask_length - 1), 0)
+
+        return num_masked_span
+
+    # compute number of masked spans in batch
+    input_lengths = (
+        attention_mask.sum(-1).detach().tolist()
+        if attention_mask is not None
+        else [sequence_length for _ in range(batch_size)]
+    )
+
+    # SpecAugment mask to fill
+    spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
+    spec_aug_mask_idxs = []
+
+    max_num_masked_span = compute_num_masked_span(sequence_length)
+
+    if max_num_masked_span == 0:
+        return spec_aug_mask
+
+    for input_length in input_lengths:
+        # compute num of masked spans for this input
+        num_masked_span = compute_num_masked_span(input_length)
+
+        # get random indices to mask
+        spec_aug_mask_idx = np.random.choice(
+            np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
+        )
+
+        # pick first sampled index that will serve as a dummy index to pad vector
+        # to ensure same dimension for all batches due to probabilistic rounding
+        # Picking first sample just pads those vectors twice.
+        if len(spec_aug_mask_idx) == 0:
+            # this case can only happen if `input_length` is strictly smaller then
+            # `sequence_length` in which case the last token has to be a padding
+            # token which we can use as a dummy mask id
+            dummy_mask_idx = sequence_length - 1
+        else:
+            dummy_mask_idx = spec_aug_mask_idx[0]
+
+        spec_aug_mask_idx = np.concatenate(
+            [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
+        )
+        spec_aug_mask_idxs.append(spec_aug_mask_idx)
+
+    spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)
+
+    # expand masked indices to masked spans
+    spec_aug_mask_idxs = np.broadcast_to(
+        spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)
+
+    # add offset to the starting indexes so that indexes now create a span
+    offsets = np.arange(mask_length)[None, None, :]
+    offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
+        batch_size, max_num_masked_span * mask_length
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
+
+    # ensure that we cannot have indices larger than sequence_length
+    if spec_aug_mask_idxs.max() > sequence_length - 1:
+        spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1
+
+    # scatter indices to mask
+    np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
+
+    return spec_aug_mask
+
+
+@add_start_docstrings(
+    "The bare Moonshine Model outputting raw hidden-states without any specific head on top.",
+    MOONSHINE_START_DOCSTRING,
+)
+class MoonshineModel(MoonshinePreTrainedModel):
+    def __init__(self, config: MoonshineConfig):
+        super().__init__(config)
+        self.encoder = MoonshineEncoder(config)
+        self.decoder = MoonshineDecoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Moonshine encoder so that its parameters will
+        not be updated during training.
+        """
+        self.encoder._freeze_parameters()
+
+    def _mask_input_features(
+        self,
+        input_features: torch.FloatTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        """
+        Masks extracted features along time axis and/or along feature axis according to
+        [SpecAugment](https://arxiv.org/abs/1904.08779).
+        """
+
+        # `config.apply_spec_augment` can set masking to False
+        if not getattr(self.config, "apply_spec_augment", True):
+            return input_features
+
+        # generate indices & apply SpecAugment along time axis
+        batch_size, hidden_size, sequence_length = input_features.size()
+
+        if self.config.mask_time_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along time axis
+            mask_time_indices = _compute_mask_indices(
+                (batch_size, sequence_length),
+                mask_prob=self.config.mask_time_prob,
+                mask_length=self.config.mask_time_length,
+                attention_mask=attention_mask,
+                min_masks=self.config.mask_time_min_masks,
+            )
+            mask_time_indices = torch.tensor(mask_time_indices, device=input_features.device, dtype=torch.bool)
+            mask_time_indices = mask_time_indices[:, None].expand(-1, hidden_size, -1)
+            input_features[mask_time_indices] = 0
+
+        if self.config.mask_feature_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along feature axis
+            mask_feature_indices = _compute_mask_indices(
+                (batch_size, hidden_size),
+                mask_prob=self.config.mask_feature_prob,
+                mask_length=self.config.mask_feature_length,
+                min_masks=self.config.mask_feature_min_masks,
+            )
+            mask_feature_indices = torch.tensor(mask_feature_indices, device=input_features.device, dtype=torch.bool)
+            input_features[mask_feature_indices] = 0
+
+        return input_features
+
+    @add_start_docstrings_to_model_forward(MOONSHINE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Union[EncoderDecoderCache, Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        decoder_position_ids: Optional[Tuple[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[torch.Tensor], Seq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+        Returns:
+
+        Example:
+
+        ```python
+         >>> import torch
+         >>> from transformers import AutoFeatureExtractor, WhisperModel
+         >>> from datasets import load_dataset
+
+         >>> model = WhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        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 encoder_outputs is None:
+            input_features = self._mask_input_features(input_features, attention_mask=attention_mask)
+
+            encoder_outputs = self.encoder(
+                input_features,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            position_ids=decoder_position_ids,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("self.model.config.pad_token_id has to be defined.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+class MoonshineForConditionalGeneration(MoonshinePreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["proj_out.weight"]
+
+    def __init__(self, config: MoonshineConfig):
+        super().__init__(config)
+        self.model = MoonshineModel(config)
+        self.proj_out = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def get_output_embeddings(self):
+        return self.proj_out
+
+    def set_output_embeddings(self, new_embeddings):
+        self.proj_out = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.model.get_input_embeddings()
+
+    @property
+    def encoder(self):
+        return self.get_encoder()
+
+    @add_start_docstrings_to_model_forward(MOONSHINE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Union[EncoderDecoderCache, Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        decoder_position_ids: Optional[Tuple[torch.LongTensor]] = 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[torch.Tensor], Seq2SeqLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`. `sequence_length` should be smaller than or equal to `config.max_target_positions`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if labels.shape[1] > self.max_target_positions:
+                raise ValueError(
+                    f"Labels' sequence length {labels.shape[1]} cannot exceed the maximum allowed length of {self.max_target_positions} tokens."
+                )
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            decoder_position_ids=decoder_position_ids,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+        lm_logits = self.proj_out(outputs[0])
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # move labels to correct device to enable PP
+            labels = labels.to(lm_logits.device)
+            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
diff --git a/src/transformers/models/moonshine/modular_moonshine.py b/src/transformers/models/moonshine/modular_moonshine.py
index be24c7837e8360..bea8370b7313ea 100644
--- a/src/transformers/models/moonshine/modular_moonshine.py
+++ b/src/transformers/models/moonshine/modular_moonshine.py
@@ -462,6 +462,8 @@ def forward(
             value_states = self.v_proj(current_states)
             key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
             value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, {"cache_position": cache_position})
 
         if self.qk_layernorm:
             query_states = self.q_layernorm(query_states)
@@ -505,11 +507,7 @@ def forward(
                     "partial_rotation_size": self.rotary_ndims,
                     "cache_position": cache_position,
                 }
-            else:
-                cache_kwargs = {
-                    "cache_position": cache_position,
-                }
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         key_states = repeat_kv(key_states, self.num_key_value_groups)
         value_states = repeat_kv(value_states, self.num_key_value_groups)
@@ -587,6 +585,8 @@ def forward(
             value_states = self.v_proj(current_states)
             key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
             value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, {"cache_position": cache_position})
 
         if self.qk_layernorm:
             query_states = self.q_layernorm(query_states)
@@ -630,11 +630,7 @@ def forward(
                     "partial_rotation_size": self.rotary_ndims,
                     "cache_position": cache_position,
                 }
-            else:
-                cache_kwargs = {
-                    "cache_position": cache_position,
-                }
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
         # to be able to avoid many of these transpose/reshape/view.
@@ -732,6 +728,8 @@ def forward(
             value_states = self.v_proj(current_states)
             key_states = key_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
             value_states = value_states.view(bsz, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            if is_cross_attention and past_key_value is not None:
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, {"cache_position": cache_position})
 
         if self.qk_layernorm:
             query_states = self.q_layernorm(query_states)
@@ -775,11 +773,7 @@ def forward(
                     "partial_rotation_size": self.rotary_ndims,
                     "cache_position": cache_position,
                 }
-            else:
-                cache_kwargs = {
-                    "cache_position": cache_position,
-                }
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         key_states = repeat_kv(key_states, self.num_key_value_groups)
         value_states = repeat_kv(value_states, self.num_key_value_groups)
@@ -839,7 +833,7 @@ def __init__(self, config: MoonshineConfig, layer_idx: int = None):
         self.hidden_size = config.hidden_size
 
         self.self_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx, is_causal=True)
-        self.encoder_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx, is_causal=True)
+        self.encoder_attn = MOONSHINE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx, is_causal=False)
 
         self.mlp = MoonshineMLP(config, config.decoder_hidden_act) 
         self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)