diff --git a/src/transformers/modeling_rope_utils.py b/src/transformers/modeling_rope_utils.py
index c617420a5896de..55d63fa4514a88 100644
--- a/src/transformers/modeling_rope_utils.py
+++ b/src/transformers/modeling_rope_utils.py
@@ -279,15 +279,10 @@ def _compute_longrope_parameters(
     # `original_max_position_embeddings` field containing the pretrained value. They use the ratio between these two
     # values to compute the default attention scaling factor, instead of using `factor`.
     if hasattr(config, "original_max_position_embeddings"):
-        if seq_len and seq_len < config.original_max_position_embeddings:
-            expanded_max_position_embeddings = config.original_max_position_embeddings
-        else:
-            expanded_max_position_embeddings = config.max_position_embeddings
         max_position_embeddings = config.original_max_position_embeddings
-        factor = expanded_max_position_embeddings / max_position_embeddings
+        factor = config.max_position_embeddings / config.original_max_position_embeddings
     else:
         max_position_embeddings = config.max_position_embeddings
-        expanded_max_position_embeddings = max_position_embeddings * factor
 
     # Sets the attention factor as suggested in the paper
     if attention_factor is None:
@@ -297,7 +292,7 @@ def _compute_longrope_parameters(
             attention_factor = math.sqrt(1 + math.log(factor) / math.log(max_position_embeddings))
 
     # Compute the inverse frequencies -- scaled based on the target sequence length
-    if expanded_max_position_embeddings > max_position_embeddings:
+    if seq_len and seq_len > max_position_embeddings:
         ext_factors = torch.tensor(long_factor, dtype=torch.float32, device=device)
     else:
         ext_factors = torch.tensor(short_factor, dtype=torch.float32, device=device)
diff --git a/src/transformers/models/aria/modeling_aria.py b/src/transformers/models/aria/modeling_aria.py
index 6481d6f3c434c7..3547a1a0b49d50 100644
--- a/src/transformers/models/aria/modeling_aria.py
+++ b/src/transformers/models/aria/modeling_aria.py
@@ -723,11 +723,7 @@ def _init_weights(self, module):
 
 
 class AriaTextRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: AriaTextConfig,
-        device=None,
-    ):
+    def __init__(self, config: AriaTextConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/bamba/modeling_bamba.py b/src/transformers/models/bamba/modeling_bamba.py
index c89d8d7853008d..e17e5ef712cf7a 100644
--- a/src/transformers/models/bamba/modeling_bamba.py
+++ b/src/transformers/models/bamba/modeling_bamba.py
@@ -120,11 +120,7 @@ def __init__(self, config: BambaConfig, batch_size, dtype=torch.float16, device=
 
 
 class BambaRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: BambaConfig,
-        device=None,
-    ):
+    def __init__(self, config: BambaConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/cohere/modeling_cohere.py b/src/transformers/models/cohere/modeling_cohere.py
index 7b8b9547ac1c33..58a1891f00ebc2 100644
--- a/src/transformers/models/cohere/modeling_cohere.py
+++ b/src/transformers/models/cohere/modeling_cohere.py
@@ -1,3 +1,9 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/cohere/modular_cohere.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_cohere.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 # coding=utf-8
 # Copyright 2024 Cohere team. All rights reserved.
 #
@@ -20,13 +26,10 @@
 
 # This file is based on the LLama model definition file in transformers
 
-"""PyTorch Cohere model."""
 
-import math
-from typing import List, Optional, Tuple, Union
+from typing import Callable, List, Optional, Tuple, Union
 
 import torch
-import torch.utils.checkpoint
 from torch import nn
 
 from ...activations import ACT2FN
@@ -34,31 +37,21 @@
 from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
-from ...modeling_outputs import (
-    BaseModelOutputWithPast,
-    CausalLMOutputWithPast,
-)
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
-from ...modeling_utils import PreTrainedModel
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...pytorch_utils import ALL_LAYERNORM_LAYERS
 from ...utils import (
+    LossKwargs,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
-    is_flash_attn_2_available,
-    is_flash_attn_greater_or_equal_2_10,
     logging,
     replace_return_docstrings,
 )
 from .configuration_cohere import CohereConfig
 
 
-if is_flash_attn_2_available():
-    from ...modeling_flash_attention_utils import _flash_attention_forward
-
-
 logger = logging.get_logger(__name__)
-
 _CONFIG_FOR_DOC = "CohereConfig"
 
 
@@ -79,49 +72,17 @@ def forward(self, hidden_states):
         return hidden_states.to(input_dtype)
 
 
-ALL_LAYERNORM_LAYERS.append(CohereLayerNorm)
-
-
 class CohereRotaryEmbedding(nn.Module):
-    # Note: the forward pass of this RoPE is slightly different from Llama's, resulting in different `sin`/`cos` for
-    # the same parameterization. The differences are highlighted with a comment.
-
-    def __init__(
-        self,
-        dim=None,
-        max_position_embeddings=2048,
-        base=10000,
-        device=None,
-        scaling_factor=1.0,
-        rope_type="default",
-        config: Optional[CohereConfig] = None,
-    ):
+    def __init__(self, config: CohereConfig, device=None):
         super().__init__()
-        # TODO (joao): remove the `if` below, only used for BC
         self.rope_kwargs = {}
-        if config is None:
-            logger.warning_once(
-                "`CohereRotaryEmbedding` 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
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
         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.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]
@@ -161,7 +122,7 @@ def forward(self, x, position_ids):
         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.repeat_interleave(freqs, 2, dim=-1)  # This line differs from Llama's implementation
+            emb = torch.repeat_interleave(freqs, 2, dim=-1)  # diff from Llama: we interleave() instead of cat()
             cos = emb.cos()
             sin = emb.sin()
 
@@ -172,6 +133,60 @@ def forward(self, x, position_ids):
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
 
+class CohereMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+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 eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
 def rotate_half(x):
     # Split and rotate. Note that this function is different from e.g. Llama.
     x1 = x[..., ::2]
@@ -210,36 +225,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed.to(dtype=dtype), k_embed.to(dtype=dtype)
 
 
-class CohereMLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
-        self.act_fn = ACT2FN[config.hidden_act]
-
-    # Ignore copy
-    def forward(self, x):
-        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
-        return down_proj
-
-
-# Copied from transformers.models.llama.modeling_llama.repeat_kv
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
 class CohereAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -247,162 +232,57 @@ def __init__(self, config: CohereConfig, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
-        if layer_idx is None:
-            logger.warning_once(
-                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
-                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
-                "when creating this class."
-            )
-
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
         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.max_position_embeddings = config.max_position_embeddings
-        self.rope_theta = config.rope_theta
         self.is_causal = True
-        self.use_qk_norm = config.use_qk_norm
-
-        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(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.use_qk_norm = config.use_qk_norm
         if self.use_qk_norm:
             # When sharding the model using Tensor Parallelism, need to be careful to use n_local_heads
-            self.q_norm = CohereLayerNorm(hidden_size=(self.num_heads, self.head_dim), eps=config.layer_norm_eps)
-            self.k_norm = CohereLayerNorm(
-                hidden_size=(self.num_key_value_heads, self.head_dim), eps=config.layer_norm_eps
+            self.q_norm = CohereLayerNorm(
+                hidden_size=(config.num_attention_heads, self.head_dim), eps=config.layer_norm_eps
             )
-
-        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
-        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
-        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
-        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
-
-    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: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-        **kwargs,
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        bsz, q_len, _ = hidden_states.size()
-
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
-        if self.use_qk_norm:
-            query_states = self.q_norm(query_states)
-            key_states = self.k_norm(key_states)
-
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_value is not None:
-            # sin and cos are specific to RoPE models; position_ids needed for the static cache
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
-
-        if attention_mask is not None:  # no matter the length, we just slice it
-            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
-            attn_weights = attn_weights + causal_mask
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
-        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        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()}"
+            self.k_norm = CohereLayerNorm(
+                hidden_size=(config.num_key_value_heads, self.head_dim), eps=config.layer_norm_eps
             )
 
-        attn_output = attn_output.transpose(1, 2).contiguous()
-
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-
-        attn_output = self.o_proj(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-# NO LONGER EXIST Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->Cohere
-# TODO cyril: modular
-class CohereFlashAttention2(CohereAttention):
-    """
-    Cohere flash attention module. This module inherits from `CohereAttention` as the weights of the module stays
-    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
-    flash attention and deal with padding tokens in case the input contains any of them.
-    """
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
-        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
-        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
-        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
-
-    # Ignore copy
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.LongTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
         past_key_value: Optional[Cache] = None,
-        output_attentions: bool = False,
-        use_cache: bool = False,
         cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-        **kwargs,
+        **kwargs: Unpack[FlashAttentionKwargs],
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if isinstance(past_key_value, StaticCache):
-            raise ValueError(
-                "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
-                "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
-            )
-        output_attentions = False
-
-        bsz, q_len, _ = hidden_states.size()
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
 
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
 
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
-        if self.use_qk_norm:
+        if self.use_qk_norm:  # main diff from Llama
             query_states = self.q_norm(query_states)
             key_states = self.k_norm(key_states)
 
         query_states = query_states.transpose(1, 2)
         key_states = key_states.transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
 
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
@@ -412,169 +292,37 @@ def forward(
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
-        # to be able to avoid many of these transpose/reshape/view.
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-
-        dropout_rate = self.attention_dropout if self.training else 0.0
-
-        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
-        # therefore the input hidden states gets silently casted in float32. Hence, we need
-        # cast them back in the correct dtype just to be sure everything works as expected.
-        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
-        # in fp32. (CohereLayerNorm handles it correctly)
-
-        input_dtype = query_states.dtype
-        if input_dtype == torch.float32:
-            if torch.is_autocast_enabled():
-                target_dtype = torch.get_autocast_gpu_dtype()
-            # Handle the case where the model is quantized
-            elif hasattr(self.config, "_pre_quantization_dtype"):
-                target_dtype = self.config._pre_quantization_dtype
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
             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)
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        attn_output = _flash_attention_forward(
+        attn_output, attn_weights = attention_interface(
+            self,
             query_states,
             key_states,
             value_states,
             attention_mask,
-            q_len,
-            dropout=dropout_rate,
-            use_top_left_mask=self._flash_attn_uses_top_left_mask,
-            is_causal=self.is_causal,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
         )
 
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
         attn_output = self.o_proj(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-class CohereSdpaAttention(CohereAttention):
-    """
-    Cohere attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
-    `CohereAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
-    SDPA API.
-    """
-
-    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: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if output_attentions:
-            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
-            logger.warning_once(
-                "CohereModel is using CohereSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
-                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-            )
-            return super().forward(
-                hidden_states=hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_value,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-                cache_position=cache_position,
-                position_embeddings=position_embeddings,
-            )
-
-        bsz, q_len, _ = hidden_states.size()
-
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
-        if self.use_qk_norm:
-            query_states = self.q_norm(query_states)
-            key_states = self.k_norm(key_states)
-
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_value is not None:
-            # sin and cos are specific to RoPE models; cache_position needed for the static cache
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        causal_mask = attention_mask
-        # if attention_mask is not None and cache_position is not None:
-        if attention_mask is not None:
-            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
-
-        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
-        # Reference: https://github.com/pytorch/pytorch/issues/112577.
-        if query_states.device.type == "cuda" and causal_mask is not None:
-            query_states = query_states.contiguous()
-            key_states = key_states.contiguous()
-            value_states = value_states.contiguous()
-
-        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
-        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
-        is_causal = True if causal_mask is None and q_len > 1 else False
-
-        attn_output = torch.nn.functional.scaled_dot_product_attention(
-            query_states,
-            key_states,
-            value_states,
-            attn_mask=causal_mask,
-            dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=is_causal,
-        )
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
-
-        attn_output = self.o_proj(attn_output)
-
-        return attn_output, None, past_key_value
-
-
-COHERE_ATTENTION_CLASSES = {
-    "eager": CohereAttention,
-    "flash_attention_2": CohereFlashAttention2,
-    "sdpa": CohereSdpaAttention,
-}
+        return attn_output, attn_weights
 
 
 class CohereDecoderLayer(nn.Module):
     def __init__(self, config: CohereConfig, layer_idx: int):
         super().__init__()
         self.hidden_size = config.hidden_size
-
-        self.self_attn = COHERE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
-
+        self.self_attn = CohereAttention(config=config, layer_idx=layer_idx)
         self.mlp = CohereMLP(config)
         self.input_layernorm = CohereLayerNorm(hidden_size=(config.hidden_size), eps=config.layer_norm_eps)
 
@@ -583,11 +331,12 @@ def forward(
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Tuple[torch.Tensor]] = 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,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -595,13 +344,13 @@ def forward(
             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.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
             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*):
@@ -613,7 +362,7 @@ def forward(
         hidden_states = self.input_layernorm(hidden_states)
 
         # Self Attention
-        hidden_states_attention, self_attn_weights, present_key_value = self.self_attn(
+        hidden_states_attention, self_attn_weights = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
@@ -622,6 +371,7 @@ def forward(
             use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
+            **kwargs,
         )
 
         # Fully Connected
@@ -631,19 +381,16 @@ def forward(
         hidden_states = residual + hidden_states_attention + hidden_states_mlp
 
         outputs = (hidden_states,)
-
         if output_attentions:
             outputs += (self_attn_weights,)
 
-        if use_cache:
-            outputs += (present_key_value,)
-
         return outputs
 
 
 COHERE_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 etc.).
+    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
@@ -661,7 +408,6 @@ def forward(
     "The bare Cohere Model outputting raw hidden-states without any specific head on top.",
     COHERE_START_DOCSTRING,
 )
-# Copied from transformers.models.llama.modeling_llama.LlamaPreTrainedModel with Llama->Cohere
 class CoherePreTrainedModel(PreTrainedModel):
     config_class = CohereConfig
     base_model_prefix = "model"
@@ -754,6 +500,10 @@ def _init_weights(self, module):
             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.
 """
 
 
@@ -761,8 +511,6 @@ def _init_weights(self, module):
     "The bare Cohere Model outputting raw hidden-states without any specific head on top.",
     COHERE_START_DOCSTRING,
 )
-# copied from transformers.models.llama.modeling_llama.LlamaModel with Llama->Cohere, LLAMA->COHERE
-# TODO cyril: modular
 class CohereModel(CoherePreTrainedModel):
     """
     Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`CohereDecoderLayer`]
@@ -771,7 +519,6 @@ class CohereModel(CoherePreTrainedModel):
         config: CohereConfig
     """
 
-    # Ignore copy
     def __init__(self, config: CohereConfig):
         super().__init__(config)
         self.padding_idx = config.pad_token_id
@@ -800,7 +547,7 @@ def forward(
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Cache] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
@@ -1023,11 +770,13 @@ def _prepare_4d_causal_attention_mask_with_cache_position(
         return causal_mask
 
 
-# TODO: re-enable check: Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with Llama->Cohere
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
 class CohereForCausalLM(CoherePreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
 
-    # Ignore copy
     def __init__(self, config):
         super().__init__(config)
         self.model = CohereModel(config)
@@ -1035,6 +784,7 @@ def __init__(self, config):
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
         self.logit_scale = config.logit_scale
         self.tie_word_embeddings = config.tie_word_embeddings
+
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -1056,7 +806,6 @@ def set_decoder(self, decoder):
     def get_decoder(self):
         return self.model
 
-    # Ignore copy
     @add_start_docstrings_to_model_forward(COHERE_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
@@ -1064,7 +813,7 @@ def forward(
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
@@ -1073,7 +822,7 @@ def forward(
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         num_logits_to_keep: int = 0,
-        **loss_kwargs,
+        **kwargs: Unpack[KwargsForCausalLM],
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         r"""
         Args:
@@ -1123,16 +872,17 @@ def forward(
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
             cache_position=cache_position,
+            **kwargs,
         )
 
         hidden_states = outputs[0]
         # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        logits = logits * self.logit_scale
+        logits = logits * self.logit_scale  # main diff from Llama
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
 
         if not return_dict:
             output = (logits,) + outputs[1:]
diff --git a/src/transformers/models/cohere/modular_cohere.py b/src/transformers/models/cohere/modular_cohere.py
new file mode 100644
index 00000000000000..b3483103461e9a
--- /dev/null
+++ b/src/transformers/models/cohere/modular_cohere.py
@@ -0,0 +1,393 @@
+# coding=utf-8
+# Copyright 2024 Cohere team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+
+# This file is based on the LLama model definition file in transformers
+
+"""PyTorch Cohere model."""
+
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...cache_utils import Cache
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_outputs import CausalLMOutputWithPast
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...processing_utils import Unpack
+from ...pytorch_utils import ALL_LAYERNORM_LAYERS
+from ...utils import LossKwargs, logging
+from ..llama.modeling_llama import (
+    LlamaAttention,
+    LlamaForCausalLM,
+    LlamaMLP,
+    LlamaModel,
+    LlamaRotaryEmbedding,
+    eager_attention_forward,
+)
+from .configuration_cohere import CohereConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "CohereConfig"
+
+
+class CohereLayerNorm(nn.Module):
+    def __init__(self, hidden_size=None, eps=1e-5, bias=False):
+        """The hidden size can be a tuple or an int. The tuple is used for QKNorm to normalize across head_dim"""
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        mean = hidden_states.mean(-1, keepdim=True)
+        variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
+        hidden_states = (hidden_states - mean) * torch.rsqrt(variance + self.variance_epsilon)
+        hidden_states = self.weight.to(torch.float32) * hidden_states
+        return hidden_states.to(input_dtype)
+
+
+ALL_LAYERNORM_LAYERS.append(CohereLayerNorm)
+
+
+class CohereRotaryEmbedding(LlamaRotaryEmbedding):
+    @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.repeat_interleave(freqs, 2, dim=-1)  # diff from Llama: we interleave() instead of cat()
+            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)
+
+
+def rotate_half(x):
+    # Split and rotate. Note that this function is different from e.g. Llama.
+    x1 = x[..., ::2]
+    x2 = x[..., 1::2]
+    rot_x = torch.stack([-x2, x1], dim=-1).flatten(-2)
+    return rot_x
+
+
+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.
+    """
+    dtype = q.dtype
+    q = q.float()
+    k = k.float()
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed.to(dtype=dtype), k_embed.to(dtype=dtype)
+
+
+class CohereMLP(LlamaMLP):
+    def __init__(self, config):
+        super().__init__(config)
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+
+class CohereAttention(LlamaAttention):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: CohereConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)
+        self.use_qk_norm = config.use_qk_norm
+        if self.use_qk_norm:
+            # When sharding the model using Tensor Parallelism, need to be careful to use n_local_heads
+            self.q_norm = CohereLayerNorm(
+                hidden_size=(config.num_attention_heads, self.head_dim), eps=config.layer_norm_eps
+            )
+            self.k_norm = CohereLayerNorm(
+                hidden_size=(config.num_key_value_heads, self.head_dim), eps=config.layer_norm_eps
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
+
+        if self.use_qk_norm:  # main diff from Llama
+            query_states = self.q_norm(query_states)
+            key_states = self.k_norm(key_states)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class CohereDecoderLayer(nn.Module):
+    def __init__(self, config: CohereConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = CohereAttention(config=config, layer_idx=layer_idx)
+        self.mlp = CohereMLP(config)
+        self.input_layernorm = CohereLayerNorm(hidden_size=(config.hidden_size), eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> 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.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            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`).
+            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.
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states_attention, self_attn_weights = 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,
+        )
+
+        # Fully Connected
+        hidden_states_mlp = self.mlp(hidden_states)
+
+        # Add everything together
+        hidden_states = residual + hidden_states_attention + hidden_states_mlp
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class CohereModel(LlamaModel):
+    def __init__(self, config: CohereConfig):
+        super().__init__(config)
+        self.layers = nn.ModuleList(
+            [CohereDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.rotary_emb = CohereRotaryEmbedding(config=config)
+        self.norm = CohereLayerNorm(hidden_size=(config.hidden_size), eps=config.layer_norm_eps)
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class CohereForCausalLM(LlamaForCausalLM):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = CohereModel(config)
+        self.logit_scale = config.logit_scale
+        self.tie_word_embeddings = config.tie_word_embeddings
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >> from transformers import AutoTokenizer, CohereForCausalLM
+
+        >> model = CohereForCausalLM.from_pretrained("CohereForAI/c4ai-command-r-v01")
+        >> tokenizer = AutoTokenizer.from_pretrained("CohereForAI/c4ai-command-r-v01")
+
+        >> prompt = "Hey, are you conscious? Can you talk to me?"
+        >> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >> # Generate
+        >> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        logits = logits * self.logit_scale  # main diff from Llama
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
diff --git a/src/transformers/models/cohere2/modeling_cohere2.py b/src/transformers/models/cohere2/modeling_cohere2.py
index 1ffa4bffddc3df..d0463573d2b1d8 100644
--- a/src/transformers/models/cohere2/modeling_cohere2.py
+++ b/src/transformers/models/cohere2/modeling_cohere2.py
@@ -28,10 +28,13 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, HybridCache
 from ...generation import GenerationMixin
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from ...modeling_utils import PreTrainedModel
+from ...processing_utils import Unpack
 from ...utils import (
+    LossKwargs,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
     is_flash_attn_2_available,
@@ -46,50 +49,20 @@
 
 
 logger = logging.get_logger(__name__)
-
 _CONFIG_FOR_DOC = "Cohere2Config"
 
 
 class Cohere2RotaryEmbedding(nn.Module):
-    # Note: the forward pass of this RoPE is slightly different from Llama's, resulting in different `sin`/`cos` for
-    # the same parameterization. The differences are highlighted with a comment.
-
-    def __init__(
-        self,
-        dim=None,
-        max_position_embeddings=2048,
-        base=10000,
-        device=None,
-        scaling_factor=1.0,
-        rope_type="default",
-        config: Optional[Cohere2Config] = None,
-    ):
+    def __init__(self, config: Cohere2Config, device=None):
         super().__init__()
-        # TODO (joao): remove the `if` below, only used for BC
         self.rope_kwargs = {}
-        if config is None:
-            logger.warning_once(
-                "`Cohere2RotaryEmbedding` 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
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
         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.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]
@@ -129,7 +102,7 @@ def forward(self, x, position_ids):
         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.repeat_interleave(freqs, 2, dim=-1)  # This line differs from Llama's implementation
+            emb = torch.repeat_interleave(freqs, 2, dim=-1)  # diff from Llama: we interleave() instead of cat()
             cos = emb.cos()
             sin = emb.sin()
 
@@ -157,6 +130,18 @@ def forward(self, hidden_states):
         return hidden_states.to(input_dtype)
 
 
+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_half(x):
     # Split and rotate. Note that this function is different from e.g. Llama.
     x1 = x[..., ::2]
@@ -195,18 +180,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed.to(dtype=dtype), k_embed.to(dtype=dtype)
 
 
-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 eager_attention_forward(
     config: Cohere2Config,
     query: torch.Tensor,
@@ -425,7 +398,6 @@ def __init__(self, config):
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
         self.act_fn = ACT2FN[config.hidden_act]
 
-    # Ignore copy
     def forward(self, x):
         down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
         return down_proj
@@ -436,7 +408,6 @@ def __init__(self, config: Cohere2Config, layer_idx: int):
         super().__init__()
         self.hidden_size = config.hidden_size
         self.self_attn = Cohere2Attention(config, layer_idx)
-
         self.mlp = Cohere2MLP(config)
         self.input_layernorm = Cohere2LayerNorm(hidden_size=(config.hidden_size), eps=config.layer_norm_eps)
         self.config = config
@@ -521,7 +492,8 @@ def forward(
 
 COHERE2_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 etc.).
+    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
@@ -874,11 +846,13 @@ def _prepare_4d_causal_attention_mask_with_cache_position(
         return causal_mask
 
 
-# TODO: re-enable check: Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with Llama->Cohere2
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
 class Cohere2ForCausalLM(Cohere2PreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
 
-    # Ignore copy
     def __init__(self, config: Cohere2Config):
         super().__init__(config)
         self.model = Cohere2Model(config)
@@ -886,6 +860,7 @@ def __init__(self, config: Cohere2Config):
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
         self.logit_scale = config.logit_scale
         self.tie_word_embeddings = config.tie_word_embeddings
+
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -907,7 +882,6 @@ def set_decoder(self, decoder):
     def get_decoder(self):
         return self.model
 
-    # Ignore copy
     @add_start_docstrings_to_model_forward(COHERE2_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
@@ -915,7 +889,7 @@ def forward(
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
@@ -924,7 +898,7 @@ def forward(
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         num_logits_to_keep: int = 0,
-        **loss_kwargs,
+        **kwargs: Unpack[KwargsForCausalLM],
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         r"""
         Args:
@@ -974,16 +948,17 @@ def forward(
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
             cache_position=cache_position,
+            **kwargs,
         )
 
         hidden_states = outputs[0]
         # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        logits = logits * self.logit_scale
+        logits = logits * self.logit_scale  # main diff from Llama
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
 
         if not return_dict:
             output = (logits,) + outputs[1:]
diff --git a/src/transformers/models/gemma/modeling_gemma.py b/src/transformers/models/gemma/modeling_gemma.py
index e2ea12b03fe434..f2611b16c34558 100644
--- a/src/transformers/models/gemma/modeling_gemma.py
+++ b/src/transformers/models/gemma/modeling_gemma.py
@@ -92,11 +92,7 @@ def forward(self, x):
 
 
 class GemmaRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: GemmaConfig,
-        device=None,
-    ):
+    def __init__(self, config: GemmaConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/gemma2/modeling_gemma2.py b/src/transformers/models/gemma2/modeling_gemma2.py
index 67fc6c86a3bac6..0bf2b154f9c221 100644
--- a/src/transformers/models/gemma2/modeling_gemma2.py
+++ b/src/transformers/models/gemma2/modeling_gemma2.py
@@ -324,11 +324,7 @@ def forward(
 
 
 class Gemma2RotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: Gemma2Config,
-        device=None,
-    ):
+    def __init__(self, config: Gemma2Config, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/glm/modeling_glm.py b/src/transformers/models/glm/modeling_glm.py
index 95ad0d9719951d..87a1c75e5084ed 100644
--- a/src/transformers/models/glm/modeling_glm.py
+++ b/src/transformers/models/glm/modeling_glm.py
@@ -62,7 +62,6 @@ def __init__(self, config):
         self.config = config
         self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
         self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
-
         self.activation_fn = ACT2FN[config.hidden_act]
 
     def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
@@ -256,11 +255,7 @@ def extra_repr(self):
 
 
 class GlmRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: GlmConfig,
-        device=None,
-    ):
+    def __init__(self, config: GlmConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/granite/modeling_granite.py b/src/transformers/models/granite/modeling_granite.py
index 2e045e149d95de..49b6b7ed2df237 100644
--- a/src/transformers/models/granite/modeling_granite.py
+++ b/src/transformers/models/granite/modeling_granite.py
@@ -309,11 +309,7 @@ def forward(
 
 
 class GraniteRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: GraniteConfig,
-        device=None,
-    ):
+    def __init__(self, config: GraniteConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py
index 5be33c26414cd7..61e7fa0f6ca429 100644
--- a/src/transformers/models/llama/modeling_llama.py
+++ b/src/transformers/models/llama/modeling_llama.py
@@ -80,11 +80,7 @@ def extra_repr(self):
 
 
 class LlamaRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: LlamaConfig,
-        device=None,
-    ):
+    def __init__(self, config: LlamaConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/mistral/modeling_mistral.py b/src/transformers/models/mistral/modeling_mistral.py
index 90c38895b4280b..8c0d8af3ece737 100644
--- a/src/transformers/models/mistral/modeling_mistral.py
+++ b/src/transformers/models/mistral/modeling_mistral.py
@@ -270,11 +270,7 @@ def forward(
 
 
 class MistralRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: MistralConfig,
-        device=None,
-    ):
+    def __init__(self, config: MistralConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/mixtral/modeling_mixtral.py b/src/transformers/models/mixtral/modeling_mixtral.py
index 84ed327d9be920..a726b69fb6688f 100644
--- a/src/transformers/models/mixtral/modeling_mixtral.py
+++ b/src/transformers/models/mixtral/modeling_mixtral.py
@@ -392,11 +392,7 @@ def forward(
 
 
 class MixtralRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: MixtralConfig,
-        device=None,
-    ):
+    def __init__(self, config: MixtralConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/olmo/modeling_olmo.py b/src/transformers/models/olmo/modeling_olmo.py
index 11d3d99f4f72c9..8f43f303dc8869 100644
--- a/src/transformers/models/olmo/modeling_olmo.py
+++ b/src/transformers/models/olmo/modeling_olmo.py
@@ -274,11 +274,7 @@ def forward(
 
 
 class OlmoRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: OlmoConfig,
-        device=None,
-    ):
+    def __init__(self, config: OlmoConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/olmo2/modeling_olmo2.py b/src/transformers/models/olmo2/modeling_olmo2.py
index 49ae798e7f1101..dc5893cad55473 100644
--- a/src/transformers/models/olmo2/modeling_olmo2.py
+++ b/src/transformers/models/olmo2/modeling_olmo2.py
@@ -275,11 +275,7 @@ def forward(
 
 
 class Olmo2RotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: Olmo2Config,
-        device=None,
-    ):
+    def __init__(self, config: Olmo2Config, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/phi/modeling_phi.py b/src/transformers/models/phi/modeling_phi.py
index 477896decd5318..b4079c3ef7d417 100644
--- a/src/transformers/models/phi/modeling_phi.py
+++ b/src/transformers/models/phi/modeling_phi.py
@@ -270,11 +270,7 @@ def forward(
 
 
 class PhiRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: PhiConfig,
-        device=None,
-    ):
+    def __init__(self, config: PhiConfig, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/phi3/modeling_phi3.py b/src/transformers/models/phi3/modeling_phi3.py
index 908fd982b9c73c..551f610908bb5d 100644
--- a/src/transformers/models/phi3/modeling_phi3.py
+++ b/src/transformers/models/phi3/modeling_phi3.py
@@ -1,3 +1,9 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/phi3/modular_phi3.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_phi3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 # coding=utf-8
 # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
 #
@@ -13,34 +19,31 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""PyTorch Phi-3 model."""
 
-import math
-import warnings
-from typing import List, Optional, Tuple, Union
+from typing import Callable, List, Optional, Tuple, Union
 
 import torch
-import torch.utils.checkpoint
 from torch import nn
-from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
 from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
-from ...modeling_flash_attention_utils import _flash_attention_forward
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
     SequenceClassifierOutputWithPast,
     TokenClassifierOutput,
 )
-from ...modeling_utils import PreTrainedModel
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
 from ...utils import (
+    LossKwargs,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
-    is_flash_attn_greater_or_equal_2_10,
     logging,
     replace_return_docstrings,
 )
@@ -53,186 +56,24 @@
 _CONFIG_FOR_DOC = "Phi3Config"
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
-class Phi3RMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        """
-        Phi3RMSNorm is equivalent to T5LayerNorm
-        """
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-    def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
-
-
-# copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
-# TODO cyril: modular
-class Phi3RotaryEmbedding(nn.Module):
-    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
         super().__init__()
 
-        self.dim = dim
-        self.max_position_embeddings = max_position_embeddings
-        self.base = base
-
-        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
-        self.register_buffer("inv_freq", tensor=inv_freq, persistent=False)
-
-    @torch.no_grad()
-    def forward(self, x, position_ids, seq_len=None):
-        # x: [bs, num_attention_heads, seq_len, head_size]
-        self.inv_freq.to(x.device)
-        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 since bfloat16 loses precision on long contexts
-        # See https://github.com/huggingface/transformers/pull/29285
-        device_type = x.device.type
-        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
-            cos = emb.cos()
-            sin = emb.sin()
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
-class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
-    def __init__(self, dim, config, device=None):
-        warnings.warn(
-            "The class Phi3SuScaledRotaryEmbedding is deprecated and will be removed in version 5 of Transformers. Please"
-            " use Phi3LongRoPEScaledRotaryEmbedding instead.",
-            FutureWarning,
-        )
-        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
-
-        self.short_factor = config.rope_scaling["short_factor"]
-        self.long_factor = config.rope_scaling["long_factor"]
-        self.original_max_position_embeddings = config.original_max_position_embeddings
-
-    @torch.no_grad()
-    def forward(self, x, position_ids, seq_len=None):
-        seq_len = torch.max(position_ids) + 1
-        if seq_len > self.original_max_position_embeddings:
-            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
-        else:
-            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
-        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
-        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
-        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 since bfloat16 loses precision on long contexts
-        # See https://github.com/huggingface/transformers/pull/29285
-        device_type = x.device.type
-        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
-            scale = self.max_position_embeddings / self.original_max_position_embeddings
-            if scale <= 1.0:
-                scaling_factor = 1.0
-            else:
-                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
-            cos = emb.cos() * scaling_factor
-            sin = emb.sin() * scaling_factor
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
-class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
-    def __init__(self, dim, config, device=None):
-        warnings.warn(
-            "The class Phi3YarnScaledRotaryEmbedding is deprecated and will be removed in version 5 of Transformers",
-            FutureWarning,
-        )
-        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
-
-        self.short_factor = config.rope_scaling["short_factor"]
-        self.long_factor = config.rope_scaling["long_factor"]
-        self.original_max_position_embeddings = config.original_max_position_embeddings
-
-    @torch.no_grad()
-    def forward(self, x, position_ids, seq_len=None):
-        seq_len = torch.max(position_ids) + 1
-        if seq_len > self.original_max_position_embeddings:
-            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
-        else:
-            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
-
-        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
-        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
-
-        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 since bfloat16 loses precision on long contexts
-        # See https://github.com/huggingface/transformers/pull/29285
-        device_type = x.device.type
-        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
-
-            scale = self.max_position_embeddings / self.original_max_position_embeddings
-            if scale <= 1.0:
-                scaling_factor = 1.0
-            else:
-                scaling_factor = 0.1 * math.log(scale) + 1.0
-
-            cos = emb.cos() * scaling_factor
-            sin = emb.sin() * scaling_factor
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
-class Phi3LongRoPEScaledRotaryEmbedding(Phi3RotaryEmbedding):
-    def __init__(self, dim, config, device=None):
-        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
-
-        self.short_factor = config.rope_scaling["short_factor"]
-        self.long_factor = config.rope_scaling["long_factor"]
-        self.original_max_position_embeddings = config.original_max_position_embeddings
-
-    @torch.no_grad()
-    def forward(self, x, position_ids, seq_len=None):
-        seq_len = seq_len or torch.max(position_ids) + 1
-        if seq_len > self.original_max_position_embeddings:
-            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
-        else:
-            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
-
-        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
-        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
-
-        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
-        position_ids_expanded = position_ids[:, None, :].float()
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+        self.activation_fn = ACT2FN[config.hidden_act]
 
-        # Force float32 since bfloat16 loses precision on long contexts
-        # See https://github.com/huggingface/transformers/pull/29285
-        device_type = x.device.type
-        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
 
-            scale = self.max_position_embeddings / self.original_max_position_embeddings
-            if scale <= 1.0:
-                scaling_factor = 1.0
-            else:
-                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
 
-            cos = emb.cos() * scaling_factor
-            sin = emb.sin() * scaling_factor
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+        return self.down_proj(up_states)
 
 
-# Copied from transformers.models.llama.modeling_llama.rotate_half
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., : x.shape[-1] // 2]
@@ -240,7 +81,6 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
 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.
 
@@ -268,26 +108,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
-class Phi3MLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-
-        self.config = config
-        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
-        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
-
-        self.activation_fn = ACT2FN[config.hidden_act]
-
-    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
-        up_states = self.gate_up_proj(hidden_states)
-
-        gate, up_states = up_states.chunk(2, dim=-1)
-        up_states = up_states * self.activation_fn(gate)
-
-        return self.down_proj(up_states)
-
-
-# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
 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,
@@ -300,6 +120,32 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
 class Phi3Attention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -307,374 +153,117 @@ def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
-        if layer_idx is None:
-            logger.warning_once(
-                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
-                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
-                "when creating this class."
-            )
-
-        self.attention_dropout = config.attention_dropout
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_dim = self.hidden_size // self.num_heads
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
         self.num_key_value_heads = config.num_key_value_heads
-        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
-        self.max_position_embeddings = config.max_position_embeddings
-        self.original_max_position_embeddings = config.original_max_position_embeddings
-        self.rope_theta = config.rope_theta
-        self.rope_scaling = config.rope_scaling
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
         self.is_causal = True
 
-        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})."
-            )
-
-        op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
-        self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
-        self._init_rope()
-
-    def _init_rope(self):
-        if self.rope_scaling is None:
-            self.rotary_emb = Phi3RotaryEmbedding(
-                self.head_dim,
-                max_position_embeddings=self.max_position_embeddings,
-                base=self.rope_theta,
-            )
-        else:
-            scaling_type = self.config.rope_scaling["type"]
-            if scaling_type == "longrope":
-                self.rotary_emb = Phi3LongRoPEScaledRotaryEmbedding(self.head_dim, self.config)
-            else:
-                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
+        op_size = config.num_attention_heads * self.head_dim + 2 * (config.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(config.hidden_size, op_size, bias=False)
 
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
         past_key_value: Optional[Cache] = None,
-        output_attentions: bool = False,
-        use_cache: bool = False,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
-
-        bsz, q_len, _ = hidden_states.size()
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
 
         qkv = self.qkv_proj(hidden_states)
-        query_pos = self.num_heads * self.head_dim
+        query_pos = self.config.num_attention_heads * self.head_dim
         query_states = qkv[..., :query_pos]
         key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
         value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
 
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
 
-        kv_seq_len = key_states.shape[-2]
-        if past_key_value is not None:
-            if self.layer_idx is None:
-                raise ValueError(
-                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
-                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
-                    "with a layer index."
-                )
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
-
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
         if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        # repeat k/v heads if n_kv_heads < n_heads
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        attn_weights = torch.matmul(query_states, key_states.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.o_proj(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-class Phi3FlashAttention2(Phi3Attention):
-    """
-    Phi-3 flash attention module. This module inherits from `Phi3Attention` 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.LongTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Cache] = None,
-        output_attentions: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        # Phi3FlashAttention2 attention does not support output_attentions
-
-        output_attentions = False
-
-        bsz, q_len, _ = hidden_states.size()
-
-        qkv = self.qkv_proj(hidden_states)
-        query_pos = self.num_heads * self.head_dim
-        query_states = qkv[..., :query_pos]
-        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
-        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
-
-        # Flash attention requires the input to have the shape
-        # batch_size x seq_length x head_dim x hidden_dim
-        # therefore we just need to keep the original shape
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        kv_seq_len = key_states.shape[-2]
-        if past_key_value is not None:
-            if self.layer_idx is None:
-                raise ValueError(
-                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
-                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
-                    "with a layer index."
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                 )
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-
-        # Because the input can be padded, the absolute sequence length depends on the max position id.
-        rotary_seq_len = (
-            max(kv_seq_len, position_ids[:, -1].max().item() + 1) if position_ids is not None else kv_seq_len
-        )
-
-        cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len, position_ids=position_ids)
-
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
-
-        if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        # repeat k/v heads if n_kv_heads < n_heads
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        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.qkv_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)
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        # Reashape to the expected shape for Flash Attention
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-
-        attn_output = _flash_attention_forward(
+        attn_output, attn_weights = attention_interface(
+            self,
             query_states,
             key_states,
             value_states,
             attention_mask,
-            q_len,
-            position_ids=position_ids,
-            dropout=attn_dropout,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
             sliding_window=getattr(self.config, "sliding_window", None),
-            use_top_left_mask=self._flash_attn_uses_top_left_mask,
-            is_causal=self.is_causal,
+            **kwargs,
         )
 
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
         attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
 
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-# NO LONGER EXIST copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
-# TODO cyril: modular
-class Phi3SdpaAttention(Phi3Attention):
-    """
-    Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
-    `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
-    SDPA API.
-    """
-
-    # Adapted from Phi3Attention.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,
-        output_attentions: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if output_attentions:
-            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
-            logger.warning_once(
-                "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
-                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-            )
-            return super().forward(
-                hidden_states=hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_value,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-            )
-
-        bsz, q_len, _ = hidden_states.size()
-
-        qkv = self.qkv_proj(hidden_states)
-        query_pos = self.num_heads * self.head_dim
-        query_states = qkv[..., :query_pos]
-        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
-        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
-
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        kv_seq_len = key_states.shape[-2]
-        if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
-
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
-
-        if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
-            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 = attention_mask[:, :, :, : key_states.shape[-2]]
-
-        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
-        # Reference: https://github.com/pytorch/pytorch/issues/112577.
-        if query_states.device.type == "cuda" and attention_mask is not None:
-            query_states = query_states.contiguous()
-            key_states = key_states.contiguous()
-            value_states = value_states.contiguous()
 
-        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
-        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
-        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
-        is_causal = True if causal_mask is None and q_len > 1 else False
-
-        attn_output = torch.nn.functional.scaled_dot_product_attention(
-            query_states,
-            key_states,
-            value_states,
-            attn_mask=causal_mask,
-            dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=is_causal,
-        )
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
-
-        attn_output = self.o_proj(attn_output)
-
-        return attn_output, None, past_key_value
+class Phi3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Phi3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
 
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
 
-PHI3_ATTENTION_CLASSES = {
-    "eager": Phi3Attention,
-    "flash_attention_2": Phi3FlashAttention2,
-    "sdpa": Phi3SdpaAttention,
-}
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
 
 
 class Phi3DecoderLayer(nn.Module):
     def __init__(self, config: Phi3Config, layer_idx: int):
         super().__init__()
-
-        self.config = config
-        self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
-
+        self.hidden_size = config.hidden_size
+        self.self_attn = Phi3Attention(config=config, layer_idx=layer_idx)
         self.mlp = Phi3MLP(config)
         self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
+        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.config = config
         self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
         self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
-        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
     def forward(
         self,
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        past_key_value: Optional[Cache] = None,
         output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -685,26 +274,25 @@ def forward(
             position_ids (`torch.LongTensor` of shape `({0})`, *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_value (`Cache`, *optional*): cached past key and value projection states
             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
             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
-        attn_outputs, self_attn_weights, present_key_value = self.self_attn(
+        hidden_states, self_attn_weights = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
@@ -712,26 +300,95 @@ def forward(
             output_attentions=output_attentions,
             use_cache=use_cache,
             cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
         )
-
-        hidden_states = residual + self.resid_attn_dropout(attn_outputs)
+        hidden_states = residual + self.resid_attn_dropout(hidden_states)  # main diff with Llama
 
         residual = hidden_states
         hidden_states = self.post_attention_layernorm(hidden_states)
         hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + self.resid_mlp_dropout(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)  # main diff with Llama
 
         outputs = (hidden_states,)
-
         if output_attentions:
             outputs += (self_attn_weights,)
 
-        if use_cache:
-            outputs += (present_key_value,)
-
         return outputs
 
 
+class Phi3RotaryEmbedding(nn.Module):
+    def __init__(self, config: Phi3Config, device=None):
+        super().__init__()
+        self.rope_kwargs = {}
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and 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
+        if self.rope_type == "longrope":
+            long_inv_freq, _ = self.rope_init_fn(
+                self.config, device, seq_len=config.original_max_position_embeddings + 1, **self.rope_kwargs
+            )
+            self.register_buffer("long_inv_freq", long_inv_freq, persistent=False)
+
+    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):
+        inv_freq = self.inv_freq
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+            inv_freq = self.inv_freq
+        elif self.rope_type == "longrope":
+            seq_len = torch.max(position_ids) + 1
+            if seq_len > self.config.original_max_position_embeddings:
+                inv_freq = self.long_inv_freq
+
+        # Core RoPE block
+        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
 PHI3_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
@@ -750,7 +407,7 @@ def forward(
 
 
 @add_start_docstrings(
-    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    "The bare Phi3 Model outputting raw hidden-states without any specific head on top.",
     PHI3_START_DOCSTRING,
 )
 class Phi3PreTrainedModel(PreTrainedModel):
@@ -758,11 +415,12 @@ class Phi3PreTrainedModel(PreTrainedModel):
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _no_split_modules = ["Phi3DecoderLayer"]
-    _skip_keys_device_placement = "past_key_values"
+    _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
     _version = "0.0.5"
 
     def _init_weights(self, module):
@@ -853,7 +511,7 @@ def _init_weights(self, module):
 
 
 @add_start_docstrings(
-    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    "The bare Phi3 Model outputting raw hidden-states without any specific head on top.",
     PHI3_START_DOCSTRING,
 )
 class Phi3Model(Phi3PreTrainedModel):
@@ -870,14 +528,13 @@ def __init__(self, config: Phi3Config):
         self.vocab_size = config.vocab_size
 
         self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
-        self.embed_dropout = nn.Dropout(config.embd_pdrop)
         self.layers = nn.ModuleList(
             [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
         )
-        self._attn_implementation = config._attn_implementation
         self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
+        self.rotary_emb = Phi3RotaryEmbedding(config=config)
         self.gradient_checkpointing = False
+
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -893,54 +550,43 @@ def forward(
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        past_key_values: Optional[Cache] = 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:
-            if use_cache:
-                logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                )
-                use_cache = False
-
-        # kept for BC (non `Cache` `past_key_values` inputs)
-        return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
-            return_legacy_cache = True
-            if past_key_values is None:
-                past_key_values = DynamicCache()
-            else:
-                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-                logger.warning_once(
-                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
-                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
-                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
-                )
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
         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)
 
@@ -950,12 +596,14 @@ def forward(
 
         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:
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
             if output_hidden_states:
                 all_hidden_states += (hidden_states,)
 
@@ -969,6 +617,7 @@ def forward(
                     output_attentions,
                     use_cache,
                     cache_position,
+                    position_embeddings,
                 )
             else:
                 layer_outputs = decoder_layer(
@@ -979,13 +628,12 @@ def forward(
                     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],)
 
@@ -995,18 +643,13 @@ def forward(
         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(
+        output = BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
-            past_key_values=next_cache,
+            past_key_values=past_key_values if use_cache else None,
             hidden_states=all_hidden_states,
             attentions=all_self_attns,
         )
+        return output if return_dict else output.to_tuple()
 
     def _update_causal_mask(
         self,
@@ -1017,6 +660,14 @@ def _update_causal_mask(
         output_attentions: bool,
     ):
         if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and past_key_values is not None:
+                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
+                if is_padding_right:
+                    raise ValueError(
+                        "You are attempting to perform batched generation with padding_side='right'"
+                        " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
+                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                    )
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
@@ -1084,7 +735,6 @@ def _update_causal_mask(
         return causal_mask
 
     @staticmethod
-    # Copied from transformers.models.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->Phi3
     def _prepare_4d_causal_attention_mask_with_cache_position(
         attention_mask: torch.Tensor,
         sequence_length: int,
@@ -1152,10 +802,13 @@ def _prepare_4d_causal_attention_mask_with_cache_position(
         return causal_mask
 
 
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
 class Phi3ForCausalLM(Phi3PreTrainedModel, GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
     def __init__(self, config):
         super().__init__(config)
         self.model = Phi3Model(config)
@@ -1165,31 +818,24 @@ def __init__(self, config):
         # Initialize weights and apply final processing
         self.post_init()
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
     def get_input_embeddings(self):
         return self.model.embed_tokens
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
     def set_input_embeddings(self, value):
         self.model.embed_tokens = value
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
     def get_output_embeddings(self):
         return self.lm_head
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.lm_head = new_embeddings
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
     def set_decoder(self, decoder):
         self.model = decoder
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
     def get_decoder(self):
         return self.model
 
-    # Ignore copy
     @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
@@ -1197,7 +843,7 @@ def forward(
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
@@ -1206,7 +852,7 @@ def forward(
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         num_logits_to_keep: int = 0,
-        **loss_kwargs,
+        **kwargs: Unpack[KwargsForCausalLM],
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         r"""
         Args:
@@ -1227,27 +873,17 @@ def forward(
         ```python
         >>> from transformers import AutoTokenizer, Phi3ForCausalLM
 
-        >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+        >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf")
 
-        >>> prompt = "This is an example script ."
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
         >>> inputs = tokenizer(prompt, return_tensors="pt")
 
         >>> # Generate
         >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
         >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
         ```"""
-        if (
-            use_cache
-            and self.config.rope_scaling
-            and cache_position is not None
-            and cache_position[0] == self.config.original_max_position_embeddings
-        ):
-            logger.warning(
-                f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed."
-            )
-
         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
@@ -1265,6 +901,8 @@ def forward(
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
         )
 
         hidden_states = outputs[0]
@@ -1273,7 +911,7 @@ def forward(
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
 
         if not return_dict:
             output = (logits,) + outputs[1:]
@@ -1329,7 +967,7 @@ def prepare_inputs_for_generation(
 
 @add_start_docstrings(
     """
-    The [`Phi3Model`] with a sequence classification head on top (linear layer).
+    The Phi3 Model transformer with a sequence classification head on top (linear layer).
 
     [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
     (e.g. GPT-2) do.
@@ -1342,7 +980,6 @@ def prepare_inputs_for_generation(
     """,
     PHI3_START_DOCSTRING,
 )
-# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
 class Phi3ForSequenceClassification(Phi3PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
@@ -1381,7 +1018,7 @@ def forward(
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        model_outputs = self.model(
+        transformer_outputs = self.model(
             input_ids,
             attention_mask=attention_mask,
             position_ids=position_ids,
@@ -1392,7 +1029,7 @@ def forward(
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        hidden_states = model_outputs[0]
+        hidden_states = transformer_outputs[0]
         logits = self.score(hidden_states)
 
         if input_ids is not None:
@@ -1420,44 +1057,48 @@ def forward(
             loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
 
         if not return_dict:
-            output = (pooled_logits,) + model_outputs[1:]
+            output = (pooled_logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output
 
         return SequenceClassifierOutputWithPast(
             loss=loss,
             logits=pooled_logits,
-            past_key_values=model_outputs.past_key_values,
-            hidden_states=model_outputs.hidden_states,
-            attentions=model_outputs.attentions,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
         )
 
 
 @add_start_docstrings(
     """
-    [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
-    Named-Entity-Recognition (NER) tasks.
+    The Phi3 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
     """,
     PHI3_START_DOCSTRING,
 )
-# Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
 class Phi3ForTokenClassification(Phi3PreTrainedModel):
-    def __init__(self, config: Phi3Config):
+    def __init__(self, config):
         super().__init__(config)
         self.num_labels = config.num_labels
-
         self.model = Phi3Model(config)
-        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+        if getattr(config, "classifier_dropout", None) is not None:
             classifier_dropout = config.classifier_dropout
-        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+        elif getattr(config, "hidden_dropout", None) is not None:
             classifier_dropout = config.hidden_dropout
         else:
             classifier_dropout = 0.1
         self.dropout = nn.Dropout(classifier_dropout)
-        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
 
         # Initialize weights and apply final processing
         self.post_init()
 
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
     @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         checkpoint=_CHECKPOINT_FOR_DOC,
@@ -1467,16 +1108,16 @@ def __init__(self, config: Phi3Config):
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
         attention_mask: Optional[torch.Tensor] = None,
-        inputs_embeds: Optional[torch.Tensor] = None,
-        labels: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-        **deprecated_arguments,
-    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+    ) -> Union[Tuple, TokenClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
             Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
@@ -1485,38 +1126,32 @@ def forward(
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        model_outputs = self.model(
+        outputs = self.model(
             input_ids,
-            past_key_values=past_key_values,
             attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-
-        hidden_states = model_outputs[0]
-        hidden_states = self.dropout(hidden_states)
-        logits = self.classifier(hidden_states)
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
 
         loss = None
         if labels is not None:
-            # move labels to correct device to enable model parallelism
-            labels = labels.to(logits.device)
-            batch_size, seq_length = labels.shape
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(
-                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
-            )
+            loss = self.loss_function(logits, labels, self.config)
 
         if not return_dict:
-            output = (logits,) + model_outputs[2:]
+            output = (logits,) + outputs[2:]
             return ((loss,) + output) if loss is not None else output
 
         return TokenClassifierOutput(
             loss=loss,
             logits=logits,
-            hidden_states=model_outputs.hidden_states,
-            attentions=model_outputs.attentions,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
         )
diff --git a/src/transformers/models/phi3/modular_phi3.py b/src/transformers/models/phi3/modular_phi3.py
new file mode 100644
index 00000000000000..af44efadd341cf
--- /dev/null
+++ b/src/transformers/models/phi3/modular_phi3.py
@@ -0,0 +1,304 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. 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.
+
+"""PyTorch Phi-3 model."""
+
+from typing import Callable, Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...processing_utils import Unpack
+from ...utils import logging
+from ..mistral.modeling_mistral import (
+    MistralDecoderLayer,
+    MistralForCausalLM,
+    MistralForSequenceClassification,
+    MistralForTokenClassification,
+    MistralPreTrainedModel,
+    MistralRotaryEmbedding,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+)
+from .configuration_phi3 import Phi3Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
+_CONFIG_FOR_DOC = "Phi3Config"
+
+
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+        self.activation_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
+
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
+
+        return self.down_proj(up_states)
+
+
+class Phi3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        op_size = config.num_attention_heads * self.head_dim + 2 * (config.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(config.hidden_size, op_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.config.num_attention_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Phi3DecoderLayer(MistralDecoderLayer):
+    def __init__(self, config: Phi3Config, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.config = config
+        self.self_attn = Phi3Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Phi3MLP(config)
+        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
+
+    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,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> 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, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            position_ids (`torch.LongTensor` of shape `({0})`, *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_value (`Cache`, *optional*): cached past key and value projection states
+            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`).
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence
+            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 = 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 + self.resid_attn_dropout(hidden_states)  # main diff with Llama
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)  # main diff with Llama
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class Phi3RotaryEmbedding(MistralRotaryEmbedding):
+    def __init__(self, config: Phi3Config, device=None):
+        super().__init__(config, device)
+        if self.rope_type == "longrope":
+            long_inv_freq, _ = self.rope_init_fn(
+                self.config, device, seq_len=config.original_max_position_embeddings + 1, **self.rope_kwargs
+            )
+            self.register_buffer("long_inv_freq", long_inv_freq, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq = self.inv_freq
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+            inv_freq = self.inv_freq
+        elif self.rope_type == "longrope":
+            seq_len = torch.max(position_ids) + 1
+            if seq_len > self.config.original_max_position_embeddings:
+                inv_freq = self.long_inv_freq
+
+        # Core RoPE block
+        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3PreTrainedModel(MistralPreTrainedModel):
+    _version = "0.0.5"
+
+
+class Phi3ForCausalLM(MistralForCausalLM, Phi3PreTrainedModel):
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        num_logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the
+        # process
+
+        # When the first time input length reached long and short factor switching point, enforce re-compute cache
+        # It will cause downside of slower at this single token position, however, better than current failure.
+        if (
+            past_key_values
+            and self.config.rope_scaling
+            and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1
+        ):
+            past_length = cache_position[0]
+            if past_length <= self.config.original_max_position_embeddings:
+                past_key_values = None
+
+        model_inputs = Phi3PreTrainedModel().prepare_inputs_for_generation(
+            input_ids=input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            use_cache=use_cache,
+            num_logits_to_keep=num_logits_to_keep,
+            **kwargs,
+        )
+        return model_inputs
+
+
+class Phi3ForSequenceClassification(MistralForSequenceClassification):
+    pass
+
+
+class Phi3ForTokenClassification(MistralForTokenClassification):
+    pass
diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py
index 36fb1ddf1390ac..c0e0d88431c268 100644
--- a/src/transformers/models/qwen2/modeling_qwen2.py
+++ b/src/transformers/models/qwen2/modeling_qwen2.py
@@ -283,11 +283,7 @@ def forward(
 
 
 class Qwen2RotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: Qwen2Config,
-        device=None,
-    ):
+    def __init__(self, config: Qwen2Config, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/src/transformers/models/starcoder2/modeling_starcoder2.py b/src/transformers/models/starcoder2/modeling_starcoder2.py
index 3b4fdbcb81ccc4..a510ca1e1ca26a 100644
--- a/src/transformers/models/starcoder2/modeling_starcoder2.py
+++ b/src/transformers/models/starcoder2/modeling_starcoder2.py
@@ -274,11 +274,7 @@ def forward(
 
 
 class Starcoder2RotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        config: Starcoder2Config,
-        device=None,
-    ):
+    def __init__(self, config: Starcoder2Config, device=None):
         super().__init__()
         self.rope_kwargs = {}
         # BC: "rope_type" was originally "type"
diff --git a/tests/models/phi3/test_modeling_phi3.py b/tests/models/phi3/test_modeling_phi3.py
index 2c5557dfd67aae..6ec663c6636fa6 100644
--- a/tests/models/phi3/test_modeling_phi3.py
+++ b/tests/models/phi3/test_modeling_phi3.py
@@ -459,6 +459,9 @@ def test_model_rope_scaling_short_long_factor(self, scaling_type):
             "long_factor": [5.0 for _ in range(n_factors)],
         }
         input_tensor = ids_tensor([1, 4090], config.vocab_size)
+        # Make sure we don't have padding tokens. If this is the case, then the actual number of "true" tokens may be shorter
+        # than `config.original_max_position_embeddings + 5`, invalidating this test
+        input_tensor[input_tensor == config.pad_token_id] += 1
         model = Phi3ForCausalLM(config)
         model.to(torch_device)
         model.eval()