modeling_gelinear.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/gemma2/modular_gemma2.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_gemma2.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2024 Google Inc. HuggingFace Inc. team. All rights reserved.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Callable, List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.activations import ACT2FN 
from transformers.cache_utils import Cache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from transformers.utils.deprecation import deprecate_kwarg
from transformers.models.gemma2.configuration_gemma2 import Gemma2Config
from einops import rearrange
from .chunk_cache import HybridCache
import gc
logger = logging.get_logger(__name__)


_CHECKPOINT_FOR_DOC = "google/gemma2-7b"
_CONFIG_FOR_DOC = "Gemma2Config"

class HedgehogFeatureMap(nn.Module):
    def __init__(
        self,
        num_heads: int,
        head_dim: int,  # input dim
        feature_dim: int,  # output dim
        # dtype: torch.dtype,
        # device: torch.device,
        bias: bool = False,
        eps: float = 1e-12,
    ):
        super().__init__()

        self.layer = nn.Parameter(
            torch.zeros(
                (num_heads, head_dim, feature_dim),
                # dtype=dtype,
                # device=device,
            )
        )
        nn.init.kaiming_uniform_(self.layer)
        if bias:
            self.bias = nn.Parameter(
                torch.zeros(
                    (1, num_heads, 1, 1),
                    # dtype=dtype,
                    # device=device,
                )
            )
            nn.init.kaiming_uniform_(self.bias)
        else:
            self.bias = 0.0  # hack
        self.eps = eps

    def forward(self, x: torch.Tensor):
        """
        x = (batch_size, num_heads, seq_len, head_dim)
        """
        output = torch.einsum("hdf,bhld->bhlf", self.layer, x) + self.bias
        output = torch.cat(
            [torch.softmax(output, dim=-1), torch.softmax(-output, dim=-1)], dim=-1
        ).clamp(min=self.eps)
        return output

def pad(x, chunk_size=64):
    T = x.shape[-2]
    padded_seq_len = ceildiv(T, chunk_size) * chunk_size
    if x.shape[-2] % chunk_size != 0:
        x = F.pad(x, (0, 0, 0, padded_seq_len - T))

    return x


def ceildiv(a, b):
    return -(a // -b)


# def chunk_linear_attn(q, k, v, chunk_size=64, cached_kv=None):
#     q, k, v = map(lambda x: pad(x), [q, k, v])
#     q = rearrange(q, "b h (n c) d -> b h n c d", c=chunk_size) * (q.shape[-1] ** -0.5)
#     k = rearrange(k, "b h (n c) d -> b h n c d", c=chunk_size)
#     v = rearrange(v, "b h (n c) d -> b h n c d", c=chunk_size)
#     kv = k.transpose(-1, -2) @ v
#     kv = kv.cumsum(2)
#     if cached_kv is not None:
#         kv += cached_kv
#     kv = torch.cat([torch.zeros_like(kv[:, :, :1]), kv[:, :, :-1]], dim=2)
#     inter = q @ kv  # (b, h, n, c, d) @ (b, h, n, d, d) -> (b, h, n, c, d)
#     intra = (
#         (q @ k.transpose(-1, -2)).masked_fill_(
#             torch.triu(
#                 torch.ones(chunk_size, chunk_size, dtype=bool, device=q.device),
#                 diagonal=1,
#             ),
#             0,
#         )
#     ) @ v
#     o = inter + intra
#     return rearrange(o, "b h n c d -> b h (n c) d")


def chunk_linear_attn(q, k, v, chunk_size=64, cached_kv=None):
    q, k, v = map(lambda x: pad(x,chunk_size=chunk_size), [q, k, v])
    q = rearrange(q, "b h (n c) d -> b h n c d", c=chunk_size) * (q.shape[-1] ** -0.5)
    k = rearrange(k, "b h (n c) d -> b h n c d", c=chunk_size)
    v = rearrange(v, "b h (n c) d -> b h n c d", c=chunk_size)
    kv = k.transpose(-1, -2) @ v
    if cached_kv is None:
        kv = kv.cumsum(2) 
        cached_kv = kv[:,:,-1:,...]
        kv = torch.cat([torch.zeros_like(kv[:, :, :1]), kv[:, :, :-1]], dim=2)
        inter = q @ kv  # (b, h, n, c, d) @ (b, h, n, d, d) -> (b, h, n, c, d)
        intra = (
            (q @ k.transpose(-1, -2)).masked_fill_(
                torch.triu(
                    torch.ones(chunk_size, chunk_size, dtype=bool, device=q.device),
                    diagonal=1,
                ),
                0,
            )
        ) @ v
        o = inter + intra
    elif cached_kv is not None:
        kv += cached_kv
        cached_kv = kv
        o = q @ kv
    return rearrange(o, "b h n c d -> b h (n c) d"), cached_kv


class Gemma2LinearAttention(nn.Module):
    def __init__(self, config: Gemma2Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = config.query_pre_attn_scalar**-0.5
        self.attention_dropout = self.config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.attn_logit_softcapping = self.config.attn_logit_softcapping
        self.feature_dim = config.feature_dim
        self.sliding_window = config.sliding_window if not bool(layer_idx % 2) else None
        self.feature_map_q = HedgehogFeatureMap(
            num_heads=config.num_attention_heads,
            head_dim=self.head_dim,
            feature_dim=self.feature_dim,
        )
        self.feature_map_k = HedgehogFeatureMap(
            num_heads=config.num_attention_heads,
            head_dim=self.head_dim,
            feature_dim=self.feature_dim,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)
        batch_size = hidden_states.shape[0]
        seq_len = hidden_states.shape[1]
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)


        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        query_states = self.feature_map_q(query_states)
        key_states = self.feature_map_k(key_states)
        chunk_size = 64
        cache = None
        if past_key_value is not None:
            if past_key_value._get_chunk_cache(self.layer_idx) is not None:
                cache = past_key_value._get_chunk_cache(self.layer_idx)
                chunk_size = 1
        output, cache = chunk_linear_attn(query_states, key_states, value_states,chunk_size=chunk_size, cached_kv=cache)
        past_key_value._static_update(self.layer_idx,cache)
        
        output = (
            output.transpose(1, 2)
            .contiguous()[:, :seq_len, ...]
            .view(batch_size, seq_len, -1)
        )
        attn_output = self.o_proj(output)
        return attn_output, output


class Gemma2RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float())
        # Llama does x.to(float16) * w whilst Gemma2 is (x * w).to(float16)
        # See https://github.com/huggingface/transformers/pull/29402
        output = output * (1.0 + self.weight.float())
        return output.type_as(x)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.eps}"


class Gemma2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_activation]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    dropout: float = 0.0,
    scaling: Optional[float] = None,
    softcap: Optional[float] = None,
    **kwargs,
) -> Tuple[torch.Tensor, torch.Tensor]:
    if scaling is None:
        scaling = module.head_dim**-0.5

    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling

    if softcap is not None:
        attn_weights = attn_weights / softcap
        attn_weights = torch.tanh(attn_weights)
        attn_weights = attn_weights * softcap
    if attention_mask is not None:  # no matter the length, we just slice it
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


class Gemma2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Gemma2Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = config.query_pre_attn_scalar**-0.5
        self.attention_dropout = self.config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.attn_logit_softcapping = self.config.attn_logit_softcapping
        self.sliding_window = config.sliding_window if not bool(layer_idx % 2) else None

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {
                "sin": sin,
                "cos": cos,
                "cache_position": cache_position,
                "sliding_window": self.sliding_window,
            }
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

            # Here we need to slice as we use a static cache by default, but FA2 does not support it
            if attention_mask is not None and self.config._attn_implementation == "flash_attention_2":
                seq_len = attention_mask.shape[-1]
                key_states, value_states = key_states[:, :, :seq_len, :], value_states[:, :, :seq_len, :]

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=self.attention_dropout if self.training else 0.0,
            scaling=self.scaling,
            sliding_window=self.sliding_window,
            softcap=self.attn_logit_softcapping,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Gemma2DecoderLayer(nn.Module):
    def __init__(self, config: Gemma2Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.config = config
        self.is_sliding = not bool(layer_idx % 2)
        if self.is_sliding:
            self.self_attn = Gemma2Attention(config=config, layer_idx=layer_idx)
        else:
            self.self_attn = Gemma2LinearAttention(config=config, layer_idx=layer_idx)
        self.mlp = Gemma2MLP(config)
        self.input_layernorm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.pre_feedforward_layernorm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.sliding_window = config.sliding_window

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        last_cache_position: int = 0,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if self.is_sliding and attention_mask is not None:  # efficient SDPA and no padding
            # In prefill, we may be larger than sliding window
            effective_seq_len = max(cache_position.shape[0], self.sliding_window)
            # For FA2, the mask is 2D and is of shape [bs, processed_tokens] (not [bs, max_cache_len]),
            # thus we must slice from the right (at most `effective_seq_len` elements)
            if self.config._attn_implementation == "flash_attention_2":
                attention_mask = attention_mask[:, -effective_seq_len:]
            # Otherwise, the mask is 4D of shape [bs, 1, query_len, max_cache_len] thus we must slice
            # from the left, with an offset if we are beyond the sliding window
            else:
                min_dtype = torch.finfo(attention_mask.dtype).min
                sliding_window_mask = torch.tril(
                    torch.ones_like(attention_mask, dtype=torch.bool), diagonal=-self.sliding_window
                )
                attention_mask = torch.where(sliding_window_mask, min_dtype, attention_mask)
                # In case we are beyond the sliding window, we need to correctly offset the mask slicing
                # `last_cache_position` is equivalent to `cache_position[-1]` but without breaking dynamo
                offset = last_cache_position - effective_seq_len
                # Should only be used when beyond the sliding window (i.e. offset > 0)
                offset = max(0, offset)
                attention_mask = attention_mask[:, :, :, offset : offset + effective_seq_len]

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.pre_feedforward_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


class Gemma2RotaryEmbedding(nn.Module):
    def __init__(self, config: Gemma2Config, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            # This .to() is needed if the model has been moved to a device after being initialized (because
            # the buffer is automatically moved, but not the original copy)
            self.original_inv_freq = self.original_inv_freq.to(device)
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


GEMMA2_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`Gemma2Config`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare Gemma2 Model outputting raw hidden-states without any specific head on top.",
    GEMMA2_START_DOCSTRING,
)
class Gemma2PreTrainedModel(PreTrainedModel):
    config_class = Gemma2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Gemma2DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True
    _supports_attention_backend = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()


GEMMA2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            Two formats are allowed:
            - a [`~cache_utils.Cache`] instance, see our
            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
            cache format.

            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
            legacy cache format will be returned.

            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
            of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings(
    "The bare Gemma2 Model outputting raw hidden-states without any specific head on top.",
    GEMMA2_START_DOCSTRING,
)
class Gemma2Model(Gemma2PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Gemma2DecoderLayer`]

    Args:
        config: Gemma2Config
    """

    def __init__(self, config: Gemma2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Gemma2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Gemma2RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.past_key_values = None
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[HybridCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        last_cache_position: Optional[int] = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
            )
            use_cache = False

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        # if use_cache and past_key_values is None and not self.training:
        #     batch_size, seq_len, _ = inputs_embeds.shape
        #     # NOTE: ideally, `HybridCache` should be initialized outside the model with `layer_device_map`
        #     past_key_values = HybridCache(
        #         self.config,
        #         max_batch_size=batch_size,
        #         max_cache_len=seq_len,
        #         dtype=inputs_embeds.dtype,
        #         device=self.device,
        #     )
        old_key_cache = past_key_values.key_cache
        old_value_cache = past_key_values.value_cache
        if self.past_key_values is None:
            self.past_key_values = HybridCache(
                    self.config,
                    max_batch_size=past_key_values.max_batch_size,
                    max_cache_len=past_key_values.max_cache_len,
                    dtype=inputs_embeds.dtype,
                    device=self.device,
                ) 
        self.past_key_values.key_cache = old_key_cache
        self.past_key_values.value_cache = old_value_cache
        if inputs_embeds.shape[1] > 1:
            self.past_key_values.chunk_cache = {}
        del past_key_values
        del old_key_cache
        del old_value_cache
        gc.collect()
        torch.cuda.empty_cache()
        if cache_position is None:
            past_seen_tokens = self.past_key_values.get_seq_length() if self.past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        # This is needed to correctly slice the mask without data-dependent slicing later on if using dynamo tracing
        # (retrieving the same value from `cache_position` later on would crash dynamo)
        if last_cache_position is None:
            last_cache_position = 0
            if attention_mask is not None:
                # In case a 4d mask is passed directly without using `generate`, we have to rely on cache_position
                # It will break dynamo tracing but there are no way around it (and it should never happen in practice)
                last_cache_position = (
                    attention_mask.shape[-1] if attention_mask.dim() == 2 else cache_position[-1].item()
                )
        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, self.past_key_values, output_attentions
        )

        # embed positions
        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # normalized
        # Gemma2 downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
        # See https://github.com/huggingface/transformers/pull/29402
        normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
        hidden_states = hidden_states * normalizer

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    self.past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    last_cache_position,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    position_embeddings=position_embeddings,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_value=self.past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    last_cache_position=last_cache_position,
                    **flash_attn_kwargs,
                )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        output = BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=self.past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )
        return output if return_dict else output.to_tuple()

    @torch.no_grad()
    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: HybridCache,
        output_attentions: bool,
    ):
        # Flash Attention currently doesn't support static cache but Gemma2 work only with static cache.
        # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape
        # to cut out keys/values trailing 0 used in static cache. This workaround should be compile compatible
        # as it doesn't cause dynamic control issues.
        if self.config._attn_implementation == "flash_attention_2":
            return attention_mask

        dtype, device = input_tensor.dtype, input_tensor.device
        sequence_length = input_tensor.shape[1]
        if isinstance(past_key_values, (HybridCache, StaticCache)):
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )
        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to place the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class Gemma2ForCausalLM(Gemma2PreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = Gemma2Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[HybridCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **loss_kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

            logits_to_keep (`int` or `torch.Tensor`, *optional*):
                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
                This is useful when using packed tensor format (single dimension for batch and sequence length).

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, Gemma2ForCausalLM

        >>> model = Gemma2ForCausalLM.from_pretrained("google/gemma-2-9b")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")

        >>> prompt = "What is your favorite condiment?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "What is your favorite condiment?"
        ```"""

        if self.training and self.config._attn_implementation != "eager":
            logger.warning_once(
                "It is strongly recommended to train Gemma2 models with the `eager` attention implementation "
                f"instead of `{self.config._attn_implementation}`. Use `eager` with `AutoModelForCausalLM.from_pretrained('<path-to-checkpoint>', attn_implementation='eager')`."
            )
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
            **loss_kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        if self.config.final_logit_softcapping is not None:
            logits = logits / self.config.final_logit_softcapping
            logits = torch.tanh(logits)
            logits = logits * self.config.final_logit_softcapping

        loss = None
        if labels is not None:
            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten: has a special cache type, `HybridCache`
        
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            position_ids=position_ids,
            use_cache=use_cache,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )

        # This is needed to correctly slice the mask without data-dependent slicing later on if using dynamo tracing
        # (retrieving the same value from `cache_position` later on would crash dynamo)
        model_inputs["last_cache_position"] = attention_mask.shape[-1] if attention_mask is not None else 0
        if logits_to_keep is None:
            _ = model_inputs.pop("logits_to_keep", None)

        if (
            isinstance(past_key_values, HybridCache)
            and attention_mask.ndim == 2
            and not self.config._attn_implementation == "flash_attention_2"
        ):
            if model_inputs["inputs_embeds"] is not None:
                batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape
                device = model_inputs["inputs_embeds"].device
            else:
                batch_size, sequence_length = model_inputs["input_ids"].shape
                device = model_inputs["input_ids"].device

            attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position(
                attention_mask,
                sequence_length=sequence_length,
                target_length=past_key_values.get_max_cache_shape(),
                dtype=self.lm_head.weight.dtype,
                device=device,
                cache_position=cache_position,
                batch_size=batch_size,
            )
            model_inputs["attention_mask"] = attention_mask

        return model_inputs



__all__ = [
    "Gemma2ForCausalLM",
]