From 92abe6033491dcaa958235e551f40f6b417d3771 Mon Sep 17 00:00:00 2001
From: fxmarty <9808326+fxmarty@users.noreply.github.com>
Date: Wed, 31 Jul 2024 20:03:07 +0200
Subject: [PATCH] >3-5x faster torch.compile forward compilation for
 autoregressive decoder models (#32227)

* draft

* apply changes to all relevant archs

* rerun ci - check_docstrings.py failing?

* fix docstring

* move 2D->4D mask creation to modeling file

* repo consistency

* fix the batch size = 1 case - calling contiguous is not enough

* nit

* style

* propagate to gemma/gemma-2

* prepare inputs for gemma generation

* implement test and tiny fix in gemma2

* Update src/transformers/models/bloom/modeling_bloom.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix copies

* ci pass

* fix gemma's test_compile_static_cache tests

* flacky

* retrigger ci

---------

Co-authored-by: sanchit-gandhi <sanchit@huggingface.co>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
---
 src/transformers/generation/utils.py          |   1 -
 .../models/bloom/modeling_bloom.py            |  87 +++++++++----
 .../models/chameleon/modeling_chameleon.py    |  87 +++++++++----
 .../models/cohere/modeling_cohere.py          | 114 +++++++++++++----
 src/transformers/models/dbrx/modeling_dbrx.py | 114 +++++++++++++----
 .../models/gemma/modeling_gemma.py            | 112 +++++++++++++----
 .../models/gemma2/modeling_gemma2.py          | 117 +++++++++++++----
 .../models/jetmoe/modeling_jetmoe.py          |  91 ++++++++++----
 .../models/llama/modeling_llama.py            | 113 +++++++++++++----
 .../models/mistral/modeling_mistral.py        |   4 +-
 .../models/mixtral/modeling_mixtral.py        |  92 ++++++++++----
 src/transformers/models/olmo/modeling_olmo.py | 114 +++++++++++++----
 .../models/persimmon/modeling_persimmon.py    | 118 ++++++++++++++----
 src/transformers/models/phi/modeling_phi.py   | 118 ++++++++++++++----
 src/transformers/models/phi3/modeling_phi3.py | 118 ++++++++++++++----
 .../models/qwen2/modeling_qwen2.py            | 118 ++++++++++++++----
 .../models/qwen2_moe/modeling_qwen2_moe.py    | 118 ++++++++++++++----
 .../models/stablelm/modeling_stablelm.py      | 118 ++++++++++++++----
 .../models/starcoder2/modeling_starcoder2.py  | 118 ++++++++++++++----
 .../models/whisper/modeling_whisper.py        |  87 +++++++++----
 tests/models/gemma/test_modeling_gemma.py     |   2 +-
 tests/test_modeling_common.py                 |  63 +++++++++-
 22 files changed, 1601 insertions(+), 423 deletions(-)

diff --git a/src/transformers/generation/utils.py b/src/transformers/generation/utils.py
index a5e53a294b..c9683791cb 100644
--- a/src/transformers/generation/utils.py
+++ b/src/transformers/generation/utils.py
@@ -13,7 +13,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 import copy
 import inspect
 import warnings
diff --git a/src/transformers/models/bloom/modeling_bloom.py b/src/transformers/models/bloom/modeling_bloom.py
index 2cf9bf57c1..c1caae6c68 100644
--- a/src/transformers/models/bloom/modeling_bloom.py
+++ b/src/transformers/models/bloom/modeling_bloom.py
@@ -45,6 +45,60 @@ _CHECKPOINT_FOR_DOC = "bigscience/bloom-560m"
 _CONFIG_FOR_DOC = "BloomConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
     """
     Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
@@ -774,27 +828,18 @@ class BloomModel(BloomPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
diff --git a/src/transformers/models/chameleon/modeling_chameleon.py b/src/transformers/models/chameleon/modeling_chameleon.py
index f7237ebb0c..955c5ed683 100644
--- a/src/transformers/models/chameleon/modeling_chameleon.py
+++ b/src/transformers/models/chameleon/modeling_chameleon.py
@@ -50,6 +50,60 @@ if is_flash_attn_2_available():
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "ChameleonConfig"
@@ -1417,27 +1471,18 @@ class ChameleonModel(ChameleonPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
diff --git a/src/transformers/models/cohere/modeling_cohere.py b/src/transformers/models/cohere/modeling_cohere.py
index 6257eeb995..4bc03ade89 100644
--- a/src/transformers/models/cohere/modeling_cohere.py
+++ b/src/transformers/models/cohere/modeling_cohere.py
@@ -59,6 +59,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "CohereConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 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"""
@@ -895,27 +949,18 @@ class CohereModel(CoherePreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1082,11 +1127,36 @@ class CohereForCausalLM(CoherePreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/dbrx/modeling_dbrx.py b/src/transformers/models/dbrx/modeling_dbrx.py
index 9e590de9be..f07b910fcf 100644
--- a/src/transformers/models/dbrx/modeling_dbrx.py
+++ b/src/transformers/models/dbrx/modeling_dbrx.py
@@ -45,6 +45,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "DbrxConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with Gemma->Dbrx
 class DbrxRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -1147,27 +1201,18 @@ class DbrxModel(DbrxPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1353,11 +1398,36 @@ class DbrxForCausalLM(DbrxPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/gemma/modeling_gemma.py b/src/transformers/models/gemma/modeling_gemma.py
index bfedbea288..2710234717 100644
--- a/src/transformers/models/gemma/modeling_gemma.py
+++ b/src/transformers/models/gemma/modeling_gemma.py
@@ -52,6 +52,60 @@ from .configuration_gemma import GemmaConfig
 logger = logging.get_logger(__name__)
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 class GemmaRMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-6):
         super().__init__()
@@ -914,27 +968,17 @@ class GemmaModel(GemmaPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1094,12 +1138,36 @@ class GemmaForCausalLM(GemmaPreTrainedModel):
             position_ids.masked_fill_(attention_mask == 0, 1)
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
 
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/gemma2/modeling_gemma2.py b/src/transformers/models/gemma2/modeling_gemma2.py
index ce2ee1ef1a..2fda315bb5 100644
--- a/src/transformers/models/gemma2/modeling_gemma2.py
+++ b/src/transformers/models/gemma2/modeling_gemma2.py
@@ -54,6 +54,60 @@ if is_flash_attn_2_available():
 logger = logging.get_logger(__name__)
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 class Gemma2RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-6):
         super().__init__()
@@ -845,27 +899,17 @@ class Gemma2Model(Gemma2PreTrainedModel):
         else:
             target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
         return causal_mask
 
 
@@ -1024,12 +1068,41 @@ class Gemma2ForCausalLM(Gemma2PreTrainedModel):
             position_ids.masked_fill_(attention_mask == 0, 1)
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s
+                # `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride
+                # during the decoding. Here, simply using `.contiguous()` is not sufficient as in the
+                # batch size = 1 case, `position_ids` is already contiguous but with varying stride
+                # which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
 
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            # The clone here is for the same reason as for `position_ids`.
+            model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format)}
+
+        if isinstance(past_key_values, HybridCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/jetmoe/modeling_jetmoe.py b/src/transformers/models/jetmoe/modeling_jetmoe.py
index 14f56a66cf..a3fc645e5a 100644
--- a/src/transformers/models/jetmoe/modeling_jetmoe.py
+++ b/src/transformers/models/jetmoe/modeling_jetmoe.py
@@ -25,9 +25,7 @@ from torch.nn import functional as F
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
@@ -54,6 +52,60 @@ _CHECKPOINT_FOR_DOC = "jetmoe"
 _CONFIG_FOR_DOC = "JetMoeConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
 def load_balancing_loss_func(
     gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
@@ -1121,27 +1173,18 @@ class JetMoeModel(JetMoePreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py
index 0db0ed60fc..dd053c805f 100644
--- a/src/transformers/models/llama/modeling_llama.py
+++ b/src/transformers/models/llama/modeling_llama.py
@@ -55,6 +55,59 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "LlamaConfig"
 
 
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 class LlamaRMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         """
@@ -1029,27 +1082,18 @@ class LlamaModel(LlamaPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1216,11 +1260,36 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/mistral/modeling_mistral.py b/src/transformers/models/mistral/modeling_mistral.py
index 01bc0f67de..1a2b732e85 100644
--- a/src/transformers/models/mistral/modeling_mistral.py
+++ b/src/transformers/models/mistral/modeling_mistral.py
@@ -1072,7 +1072,6 @@ class MistralForCausalLM(MistralPreTrainedModel):
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.prepare_inputs_for_generation
     def prepare_inputs_for_generation(
         self,
         input_ids,
@@ -1100,6 +1099,9 @@ class MistralForCausalLM(MistralPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
diff --git a/src/transformers/models/mixtral/modeling_mixtral.py b/src/transformers/models/mixtral/modeling_mixtral.py
index 78608a7973..1ff2a6c250 100644
--- a/src/transformers/models/mixtral/modeling_mixtral.py
+++ b/src/transformers/models/mixtral/modeling_mixtral.py
@@ -30,10 +30,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-    _prepare_4d_causal_attention_mask,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
@@ -70,6 +67,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "MixtralConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 def load_balancing_loss_func(
     gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
 ) -> float:
@@ -1104,27 +1155,18 @@ class MixtralModel(MixtralPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
diff --git a/src/transformers/models/olmo/modeling_olmo.py b/src/transformers/models/olmo/modeling_olmo.py
index 16e8711188..61a8a2bf6b 100644
--- a/src/transformers/models/olmo/modeling_olmo.py
+++ b/src/transformers/models/olmo/modeling_olmo.py
@@ -57,6 +57,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "OlmoConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 class OlmoLayerNorm(nn.Module):
     """LayerNorm but with no learnable weight or bias."""
 
@@ -941,27 +995,18 @@ class OlmoModel(OlmoPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1126,11 +1171,36 @@ class OlmoForCausalLM(OlmoPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/persimmon/modeling_persimmon.py b/src/transformers/models/persimmon/modeling_persimmon.py
index 8b4e02a256..885d744266 100644
--- a/src/transformers/models/persimmon/modeling_persimmon.py
+++ b/src/transformers/models/persimmon/modeling_persimmon.py
@@ -29,9 +29,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -48,6 +46,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "PersimmonConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 #  Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Persimmon
 class PersimmonRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -755,27 +807,18 @@ class PersimmonModel(PersimmonPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -945,11 +988,36 @@ class PersimmonForCausalLM(PersimmonPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/phi/modeling_phi.py b/src/transformers/models/phi/modeling_phi.py
index 1289910381..c445459763 100644
--- a/src/transformers/models/phi/modeling_phi.py
+++ b/src/transformers/models/phi/modeling_phi.py
@@ -26,9 +26,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -59,6 +57,60 @@ _CHECKPOINT_FOR_DOC = "microsoft/phi-1"
 _CONFIG_FOR_DOC = "PhiConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Phi
 class PhiRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -1039,27 +1091,18 @@ class PhiModel(PhiPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1230,11 +1273,36 @@ class PhiForCausalLM(PhiPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/phi3/modeling_phi3.py b/src/transformers/models/phi3/modeling_phi3.py
index 6bbfae9845..000c6f4c0a 100644
--- a/src/transformers/models/phi3/modeling_phi3.py
+++ b/src/transformers/models/phi3/modeling_phi3.py
@@ -26,9 +26,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -57,6 +55,60 @@ _CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
 _CONFIG_FOR_DOC = "Phi3Config"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
 class Phi3RMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
@@ -1082,27 +1134,18 @@ class Phi3Model(Phi3PreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1273,11 +1316,36 @@ class Phi3ForCausalLM(Phi3PreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py
index 4b51aa7bc8..c66380ca84 100644
--- a/src/transformers/models/qwen2/modeling_qwen2.py
+++ b/src/transformers/models/qwen2/modeling_qwen2.py
@@ -29,9 +29,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -61,6 +59,60 @@ _CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
 _CONFIG_FOR_DOC = "Qwen2Config"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
 class Qwen2RMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
@@ -941,27 +993,18 @@ class Qwen2Model(Qwen2PreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1125,11 +1168,36 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
index 4573ccb380..2e0c75ed84 100644
--- a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
+++ b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
@@ -30,9 +30,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
@@ -60,6 +58,60 @@ _CHECKPOINT_FOR_DOC = "Qwen/Qwen1.5-MoE-A2.7B"
 _CONFIG_FOR_DOC = "Qwen2MoeConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
 def load_balancing_loss_func(
     gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
@@ -1114,27 +1166,18 @@ class Qwen2MoeModel(Qwen2MoePreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1321,11 +1364,36 @@ class Qwen2MoeForCausalLM(Qwen2MoePreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/stablelm/modeling_stablelm.py b/src/transformers/models/stablelm/modeling_stablelm.py
index 086f525644..2f326184e1 100755
--- a/src/transformers/models/stablelm/modeling_stablelm.py
+++ b/src/transformers/models/stablelm/modeling_stablelm.py
@@ -29,9 +29,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -59,6 +57,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "StableLmConfig"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->StableLm
 class StableLmRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -1032,27 +1084,18 @@ class StableLmModel(StableLmPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1223,11 +1266,36 @@ class StableLmForCausalLM(StableLmPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/starcoder2/modeling_starcoder2.py b/src/transformers/models/starcoder2/modeling_starcoder2.py
index 18dbe51045..7cf767b70e 100644
--- a/src/transformers/models/starcoder2/modeling_starcoder2.py
+++ b/src/transformers/models/starcoder2/modeling_starcoder2.py
@@ -29,9 +29,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-)
+from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -59,6 +57,60 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "Starcoder2Config"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Starcoder2
 class Starcoder2RotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
@@ -915,27 +967,18 @@ class Starcoder2Model(Starcoder2PreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1101,11 +1144,36 @@ class Starcoder2ForCausalLM(Starcoder2PreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and cache_position[0] == 0:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
 
         model_inputs.update(
             {
diff --git a/src/transformers/models/whisper/modeling_whisper.py b/src/transformers/models/whisper/modeling_whisper.py
index c4d59a3604..8785d5681f 100644
--- a/src/transformers/models/whisper/modeling_whisper.py
+++ b/src/transformers/models/whisper/modeling_whisper.py
@@ -59,6 +59,60 @@ _CONFIG_FOR_DOC = "WhisperConfig"
 _CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
 
 
+# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
 def sinusoids(length: int, channels: int, max_timescale: float = 10000) -> torch.Tensor:
     """Returns sinusoids for positional embedding"""
     if channels % 2 != 0:
@@ -1412,27 +1466,18 @@ class WhisperDecoder(WhisperPreTrainedModel):
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            if attention_mask.max() != 0:
-                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
diff --git a/tests/models/gemma/test_modeling_gemma.py b/tests/models/gemma/test_modeling_gemma.py
index 36f5966948..46a40c6f66 100644
--- a/tests/models/gemma/test_modeling_gemma.py
+++ b/tests/models/gemma/test_modeling_gemma.py
@@ -816,7 +816,7 @@ class GemmaIntegrationTest(unittest.TestCase):
         # Dynamic Cache
         generated_ids = model.generate(**inputs, max_new_tokens=NUM_TOKENS_TO_GENERATE, do_sample=False)
         dynamic_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
-        self.assertEqual(EXPECTED_TEXT_COMPLETION[8], dynamic_text)  # Both GPU architectures have the same output
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, dynamic_text)  # Both GPU architectures have the same output
 
         # Static Cache
         generated_ids = model.generate(
diff --git a/tests/test_modeling_common.py b/tests/test_modeling_common.py
index 4e7f9bdf14..148f33e048 100755
--- a/tests/test_modeling_common.py
+++ b/tests/test_modeling_common.py
@@ -22,6 +22,7 @@ import os.path
 import random
 import re
 import tempfile
+import time
 import warnings
 from collections import defaultdict
 from typing import Dict, List, Tuple
@@ -37,6 +38,7 @@ from transformers import (
     AutoModelForCausalLM,
     AutoModelForSequenceClassification,
     AutoTokenizer,
+    GenerationConfig,
     PretrainedConfig,
     PreTrainedModel,
     is_torch_available,
@@ -4605,7 +4607,6 @@ class ModelTesterMixin:
         tokenizer = AutoTokenizer.from_pretrained(ckpt)
         model = AutoModelForCausalLM.from_pretrained(ckpt, torch_dtype=torch.float16).to(torch_device)
 
-        model.generation_config.max_new_tokens = 4
         model.generation_config.max_new_tokens = 4
 
         model.generation_config.cache_implementation = "static"
@@ -4617,6 +4618,66 @@ class ModelTesterMixin:
         for i in range(n_iter):
             _ = model.generate(**input_ids, do_sample=False)
 
+    @slow
+    @require_torch_gpu  # Testing cuda graphs.
+    @require_read_token
+    def test_compile_cuda_graph_time(self):
+        if version.parse(torch.__version__) < version.parse("2.3"):
+            self.skipTest(reason="This test requires torch >= 2.3 to run.")
+
+        # TODO felix: All models supporting `StaticCache` or `torch.compile` should be tested.
+        # At the moment, only llama, gemma and gemma2 are tested here!
+        if not hasattr(self, "_torch_compile_test_ckpt"):
+            self.skipTest(f"{self.__class__.__name__} doesn't have the attribute `_torch_compile_test_ckpt`.")
+        ckpt = self._torch_compile_test_ckpt
+
+        os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+        tokenizer = AutoTokenizer.from_pretrained(ckpt)
+        model = AutoModelForCausalLM.from_pretrained(ckpt, torch_dtype=torch.float16).to(torch_device)
+
+        cache_implementation = "static"
+        if model.config.model_type == "gemma2":
+            cache_implementation = "hybrid"
+
+        new_tokens = 50
+        gen_config = GenerationConfig(
+            max_new_tokens=new_tokens,
+            min_new_tokens=new_tokens,
+            use_cache=True,
+            pad_token_id=tokenizer.pad_token_id,
+            num_beams=1,
+            do_sample=False,
+            eos_token_id=None,  # This is required for min_new_tokens to actually have an effect.
+        )
+        model.generation_config.eos_token_id = None  # greedy_search falls back on this eos_token_id that we need to set to None as well for min_new_tokens to have an effect.
+
+        model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+
+        inp = tokenizer("Why cats are cute?", return_tensors="pt").to(torch_device)
+
+        # First run: the first run warms up each graph, which does things like CuBlas or Triton benchmarking
+        start = time.perf_counter()
+        _ = model.generate(**inp, generation_config=gen_config, cache_implementation=cache_implementation)
+        end = time.perf_counter()
+        graph_warmup_time = end - start
+
+        # Second run: CUDA Graph recording, and replays it
+        start = time.perf_counter()
+        _ = model.generate(**inp, generation_config=gen_config, cache_implementation=cache_implementation)
+        end = time.perf_counter()
+        record_time = end - start
+
+        # Finally: we hit the optimized, CUDA Graph replay path
+        start = time.perf_counter()
+        _ = model.generate(**inp, generation_config=gen_config, cache_implementation=cache_implementation)
+        end = time.perf_counter()
+        opt_time = end - start
+
+        # For the recording step, we expect only two cuda graphs and this step should be much faster than the first.
+        self.assertTrue(record_time < 0.15 * graph_warmup_time)
+        self.assertTrue(opt_time < record_time)
+
 
 global_rng = random.Random()