image-match/python/py/Lib/site-packages/transformers/models/mamba/modeling_mamba.py

# Copyright 2024 state-spaces/mamba org and HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch MAMBA model."""

import math
from dataclasses import dataclass

import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import lazy_load_kernel
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_utils import PreTrainedModel
from ...utils import (
    ModelOutput,
    auto_docstring,
    logging,
)
from ...utils.import_utils import (
    is_mambapy_available,
    is_torch_greater_or_equal,
    is_tracing,
    resolve_internal_import,
)
from .configuration_mamba import MambaConfig


logger = logging.get_logger(__name__)

if is_torch_greater_or_equal("2.9.0"):
    from torch._higher_order_ops.associative_scan import associative_scan
else:
    associative_scan = None


if is_mambapy_available():
    from mambapy.pscan import pscan
else:
    pscan = None


class MambaMixer(nn.Module):
    """
    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
    and is why Mamba is called **selective** state spaces)
    """

    def __init__(self, config: MambaConfig, layer_idx: int, initialize_mixer_weights: bool = True):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.state_size
        self.conv_kernel_size = config.conv_kernel
        self.intermediate_size = config.intermediate_size
        self.time_step_rank = int(config.time_step_rank)
        self.layer_idx = layer_idx
        self.use_conv_bias = config.use_conv_bias
        self.conv1d = nn.Conv1d(
            in_channels=self.intermediate_size,
            out_channels=self.intermediate_size,
            bias=config.use_conv_bias,
            kernel_size=config.conv_kernel,
            groups=self.intermediate_size,
            padding=config.conv_kernel - 1,
        )

        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]

        self.use_mambapy = config.use_mambapy
        self.use_associative_scan = config.use_associative_scan

        # projection of the input hidden states
        self.in_proj = nn.Linear(self.hidden_size, self.intermediate_size * 2, bias=config.use_bias)
        # selective projection used to make dt, B and C input dependent
        self.x_proj = nn.Linear(self.intermediate_size, self.time_step_rank + self.ssm_state_size * 2, bias=False)
        # time step projection (discretization)
        self.dt_proj = nn.Linear(self.time_step_rank, self.intermediate_size, bias=True)

        # S4D real initialization. These are not discretized!
        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
        self.A_log = nn.Parameter(torch.empty(self.intermediate_size, self.ssm_state_size))
        self.D = nn.Parameter(torch.empty(self.intermediate_size))
        if initialize_mixer_weights and self.dt_proj.weight.device.type != "meta":
            self.init_mamba_weights()
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
        self.use_bias = config.use_bias

        global causal_conv1d, causal_conv1d_update, causal_conv1d_fn
        causal_conv1d = lazy_load_kernel("causal-conv1d")
        causal_conv1d_update = getattr(causal_conv1d, "causal_conv1d_update", None)
        causal_conv1d_fn = getattr(causal_conv1d, "causal_conv1d_fn", None)

        global mamba_ssm, selective_state_update, selective_scan_fn, mamba_inner_fn
        mamba_ssm = lazy_load_kernel("mamba-ssm")
        selective_state_update = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.selective_state_update.selective_state_update"
        )
        selective_scan_fn = getattr(mamba_ssm, "selective_scan_fn", None)
        mamba_inner_fn = getattr(mamba_ssm, "mamba_inner_fn", None)

        self.warn_slow_implementation()

    @torch.no_grad()
    def init_mamba_weights(self):
        A = torch.arange(1, self.ssm_state_size + 1, dtype=torch.float32, device=self.A_log.device)[None, :]
        A = A.expand(self.intermediate_size, -1).contiguous()
        init.copy_(self.A_log, torch.log(A))
        init.ones_(self.D)

        dt_init_std = self.config.time_step_rank**-0.5 * self.config.time_step_scale
        if self.config.time_step_init_scheme == "constant":
            init.constant_(self.dt_proj.weight, dt_init_std)
        elif self.config.time_step_init_scheme == "random":
            init.uniform_(self.dt_proj.weight, -dt_init_std, dt_init_std)

        dt = torch.exp(
            torch.rand(self.intermediate_size, device=self.dt_proj.bias.device, dtype=torch.float32)
            * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
            + math.log(self.config.time_step_min)
        ).clamp(min=self.config.time_step_floor)
        # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
        inv_dt = dt + torch.log(-torch.expm1(-dt))
        init.copy_(self.dt_proj.bias, inv_dt)

    def warn_slow_implementation(self):
        is_fast_path_available = all(
            (selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)
        )
        if not is_fast_path_available:
            if self.use_mambapy:
                if is_mambapy_available():
                    logger.warning_once(
                        "The fast path is not available because one of `(selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)`"
                        " is None. Falling back to the mamba.py backend. To install follow https://github.com/state-spaces/mamba/#installation for mamba-ssm and"
                        " install the kernels library using `pip install kernels` or https://github.com/Dao-AILab/causal-conv1d for causal-conv1d"
                    )
                else:
                    raise ImportError(
                        "use_mambapy is set to True but the mambapy package is not installed. To install it follow https://github.com/alxndrTL/mamba.py."
                    )
            else:
                logger.warning_once(
                    "The fast path is not available because one of `(selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)`"
                    " is None. Falling back to the sequential implementation of Mamba, as use_mambapy is set to False. To install follow https://github.com/state-spaces/mamba/#installation for mamba-ssm and"
                    " install the kernels library using `pip install kernels` or https://github.com/Dao-AILab/causal-conv1d for causal-conv1d. For the mamba.py backend, follow https://github.com/alxndrTL/mamba.py."
                )

    def cuda_kernels_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.LongTensor | None = None,
    ):
        # 1. Gated MLP's linear projection
        projected_states = self.in_proj(hidden_states).transpose(1, 2)

        if self.training and cache_params is None:  # Doesn't support outputting the states -> used for training
            contextualized_states = mamba_inner_fn(
                projected_states,
                self.conv1d.weight,
                self.conv1d.bias if self.use_conv_bias else None,
                self.x_proj.weight,
                self.dt_proj.weight,
                self.out_proj.weight,
                self.out_proj.bias.float() if self.use_bias else None,
                -torch.exp(self.A_log.float()),
                None,  # input-dependent B
                None,  # input-dependent C
                self.D.float(),
                delta_bias=self.dt_proj.bias.float(),
                delta_softplus=True,
            )

        else:
            hidden_states, gate = projected_states.chunk(2, dim=1)

            if attention_mask is not None:
                hidden_states = hidden_states * attention_mask.unsqueeze(1)

            is_decoding = cache_params is not None and cache_params.has_previous_state(self.layer_idx)

            # 2. Convolution sequence transformation
            conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0), self.conv1d.weight.size(2))
            if is_decoding:
                hidden_states = causal_conv1d_update(
                    hidden_states.squeeze(-1),
                    cache_params.layers[self.layer_idx].conv_states,
                    conv_weights,
                    self.conv1d.bias,
                    self.activation,
                )
                hidden_states = hidden_states.unsqueeze(-1)
            else:
                if cache_params is not None:
                    conv_states = nn.functional.pad(
                        hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0)
                    )
                    cache_params.update_conv_state(conv_states, self.layer_idx)
                hidden_states = causal_conv1d_fn(
                    hidden_states, conv_weights, self.conv1d.bias, activation=self.activation
                )

            if attention_mask is not None:
                hidden_states = hidden_states * attention_mask.unsqueeze(1)

            # 3. State Space Model sequence transformation
            # 3.a. input varying initialization of time_step, B and C
            ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
            time_step, B, C = torch.split(
                ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
            )
            discrete_time_step = self.dt_proj.weight @ time_step.transpose(1, 2)

            A = -torch.exp(self.A_log.float())
            # 3.c perform the recurrence y ← SSM(A, B, C)(x)
            time_proj_bias = self.dt_proj.bias.float() if hasattr(self.dt_proj, "bias") else None
            if is_decoding:
                scan_outputs = selective_state_update(
                    cache_params.layers[self.layer_idx].recurrent_states,
                    hidden_states[..., 0],
                    discrete_time_step[..., 0],
                    A,
                    B[:, 0],
                    C[:, 0],
                    self.D,
                    gate[..., 0],
                    time_proj_bias,
                    dt_softplus=True,
                ).unsqueeze(-1)
            else:
                scan_outputs, ssm_state = selective_scan_fn(
                    hidden_states,
                    discrete_time_step,
                    A,
                    B.transpose(1, 2),
                    C.transpose(1, 2),
                    self.D.float(),
                    gate,
                    time_proj_bias,
                    delta_softplus=True,
                    return_last_state=True,
                )
                if ssm_state is not None and cache_params is not None:
                    cache_params.update_recurrent_state(ssm_state, self.layer_idx)

            # 4. Final linear projection
            contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))
        return contextualized_states

    # fmt: off
    def slow_forward(self, input_states, cache_params: Cache | None=None, attention_mask: torch.LongTensor | None = None):
        batch_size, seq_len, _ = input_states.shape
        dtype = input_states.dtype
        # 1. Gated MLP's linear projection
        projected_states = self.in_proj(input_states).transpose(1, 2)                   # [batch, 2 * intermediate_size, seq_len]
        hidden_states, gate = projected_states.chunk(2, dim=1)

        if attention_mask is not None:
            hidden_states = hidden_states * attention_mask.unsqueeze(1)

        if cache_params is not None and cache_params.has_previous_state(self.layer_idx):
            ssm_state = cache_params.layers[self.layer_idx].recurrent_states.clone()
        else:
            ssm_state = torch.zeros(
                (batch_size, self.intermediate_size, self.ssm_state_size),
                device=hidden_states.device, dtype=dtype
            )

        # 2. Convolution sequence transformation
        if cache_params is not None:
            if not cache_params.has_previous_state(self.layer_idx):
                conv_state = nn.functional.pad(
                    hidden_states,
                    (self.conv_kernel_size - hidden_states.shape[-1], 0)
                )

                cache_params.update_conv_state(conv_state, self.layer_idx)
                hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])     # [batch, intermediate_size, seq_len]
            else:
                conv_state = cache_params.update_conv_state(hidden_states, self.layer_idx)
                conv_state = conv_state.to(self.conv1d.weight.device)
                hidden_states = torch.sum(conv_state * self.conv1d.weight[:, 0, :], dim=-1)
                if self.use_conv_bias:
                    hidden_states += self.conv1d.bias
                hidden_states = self.act(hidden_states).to(dtype).unsqueeze(-1)         # [batch, intermediate_size, 1] : decoding
        else:
            hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])         # [batch, intermediate_size, seq_len]

        if attention_mask is not None:
            hidden_states = hidden_states * attention_mask.unsqueeze(1)

        # 3. State Space Model sequence transformation
        # 3.a. Selection:  [batch, seq_len, self.time_step_rank + self.ssm_state_size * 2]
        ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
        time_step, B, C = torch.split(
            ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
        )
        discrete_time_step = self.dt_proj(time_step)                                    # [batch, seq_len, intermediate_size]
        discrete_time_step = nn.functional.softplus(discrete_time_step).transpose(1, 2) # [batch, intermediate_size, seq_len]

        # 3.b. Discretization: B and C to [batch, seq_len, intermediate_size, ssm_state_size] (SRAM)
        A = -torch.exp(self.A_log.float())                                              # [intermediate_size, ssm_state_size]
        discrete_A = torch.exp(A[None, :, None, :] * discrete_time_step[:, :, :, None]) # [batch, intermediate_size, seq_len, ssm_state_size]
        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediate_size, seq_len, ssm_state_size]
        deltaB_u = discrete_B * hidden_states[:, :, :, None].float()

        # 3.c perform the recurrence y ← SSM(A, B, C)(x)
        if self.use_mambapy and self.training and cache_params is None:
            hs = pscan(discrete_A.transpose(1, 2), deltaB_u.transpose(1, 2)) # [batch, seq_len, intermediate_size, ssm_state_size]

            scan_output = (hs @ C.unsqueeze(-1)).squeeze(3).transpose(1, 2) # [batch, intermediate_size, seq_len]
            scan_output = scan_output + hidden_states * self.D[None, :, None]
            scan_output = scan_output * self.act(gate)
        else:
            # Use associative_scan for parallel computation when available
            if self.use_associative_scan and associative_scan is not None and is_tracing(hidden_states) and cache_params is None:
                def combine_fn(left, right):
                    a_left, b_left = left
                    a_right, b_right = right
                    return (a_left * a_right, a_right * b_left + b_right)

                combine_mode = "pointwise" if discrete_A.device.type in ("cuda", "xpu") else "generic"
                _, all_h = associative_scan(combine_fn, (discrete_A, deltaB_u), dim=2, combine_mode=combine_mode)
                # all_h: [B, D, S, N] -> output: [B, D, S]
                scan_output = torch.matmul(all_h.permute(0, 2, 1, 3).to(dtype), C.unsqueeze(-1)).squeeze(-1).permute(0, 2, 1)
                ssm_state = all_h[:, :, -1, :]
            else:
                # Sequential loop for decoding or when associative_scan unavailable
                scan_outputs = []
                for i in range(seq_len):
                    ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediate_size, ssm_state]
                    scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediate_size, 1]
                    scan_outputs.append(scan_output[:, :, 0])
                scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, intermediate_size, seq_len]

            scan_output = scan_output + (hidden_states * self.D[None, :, None])
            scan_output = (scan_output * self.act(gate))

            if cache_params is not None:
                cache_params.update_recurrent_state(ssm_state, self.layer_idx)

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.transpose(1, 2))  # [batch, seq_len, hidden_size]
        return contextualized_states
    # fmt: on

    def forward(
        self,
        hidden_states,
        cache_params: Cache | None = None,
        attention_mask: torch.LongTensor | None = None,
        **kwargs,
    ):
        is_fast_path_available = all(
            (selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)
        )
        if is_fast_path_available and "cuda" in self.x_proj.weight.device.type and not is_tracing(hidden_states):
            return self.cuda_kernels_forward(hidden_states, cache_params, attention_mask)
        return self.slow_forward(hidden_states, cache_params, attention_mask)


class MambaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        MambaRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{self.weight.shape[0]}, eps={self.variance_epsilon}"


class MambaBlock(GradientCheckpointingLayer):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.residual_in_fp32 = config.residual_in_fp32
        self.norm = MambaRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.mixer = MambaMixer(config, layer_idx=layer_idx, initialize_mixer_weights=False)

    def forward(
        self,
        hidden_states,
        cache_params: Cache | None = None,
        attention_mask: torch.LongTensor | None = None,
        **kwargs,
    ):
        residual = hidden_states
        hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
        if self.residual_in_fp32:
            residual = residual.to(torch.float32)

        hidden_states = self.mixer(hidden_states, cache_params=cache_params, attention_mask=attention_mask)
        hidden_states = residual + hidden_states
        return hidden_states


@auto_docstring
class MambaPreTrainedModel(PreTrainedModel):
    config: MambaConfig
    base_model_prefix = "backbone"
    _no_split_modules = ["MambaBlock", "MambaMixer"]
    supports_gradient_checkpointing = True
    _is_stateful = True

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights."""
        std = self.config.initializer_range
        if isinstance(module, MambaMixer):
            # S4D real initialization. These are not discretized!
            # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
            module.init_mamba_weights()

            init.kaiming_uniform_(module.conv1d.weight, a=math.sqrt(5))
            if module.conv1d.bias is not None:
                init.zeros_(module.conv1d.bias)
            init.kaiming_uniform_(module.out_proj.weight, a=math.sqrt(5))

            if self.config.rescale_prenorm_residual:
                # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
                #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
                #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
                #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
                #
                # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                # We need to reinit p since this code could be called multiple times
                # Having just p *= scale would repeatedly scale it down
                p = module.out_proj.weight
                p /= math.sqrt(self.config.num_hidden_layers)

        if isinstance(module, nn.Linear):
            init.normal_(module.weight, std=std)
            if module.bias is not None:
                init.zeros_(module.bias)
        elif isinstance(module, MambaRMSNorm):
            init.ones_(module.weight)
        elif isinstance(module, nn.Embedding):
            init.normal_(module.weight, std=std)


@dataclass
@auto_docstring(
    custom_intro="""
    Class for the MAMBA model outputs.
    """
)
class MambaOutput(ModelOutput):
    r"""
    cache_params (`Cache`):
        The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
        avoid providing the old `input_ids`.

        Includes both the State space model state matrices after the selective scan, and the Convolutional states
    """

    last_hidden_state: torch.FloatTensor | None = None
    cache_params: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for causal language model (or autoregressive) outputs.
    """
)
class MambaCausalLMOutput(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    cache_params (`Cache`):
        The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
        avoid providing the old `input_ids`.

        Includes both the State space model state matrices after the selective scan, and the Convolutional states
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    cache_params: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None


@auto_docstring
class MambaModel(MambaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([MambaBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])

        self.gradient_checkpointing = False
        self.norm_f = MambaRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        # Initialize weights and apply final processing
        self._register_load_state_dict_pre_hook(self.load_hook)
        self.post_init()

    def load_hook(self, state_dict, prefix, *args):
        for k in state_dict:
            if "embedding." in k:
                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
                break

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, new_embeddings):
        self.embeddings = new_embeddings

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.LongTensor | None = None,
        cache_params: Cache | None = None,
        use_cache: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        attention_mask: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple | MambaOutput:
        r"""
        cache_params (`Cache`, *optional*):
            If passed along, the model uses the previous state in all the blocks (which will give the output for the
            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
        use_cache (`bool`, *optional*):
            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
        """
        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 if not self.training else False)
        return_dict = return_dict if return_dict is not None else self.config.return_dict

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

        if inputs_embeds is None:
            inputs_embeds = self.embeddings(input_ids)

        if self.gradient_checkpointing and self.training and use_cache:
            use_cache = False

        if use_cache and cache_params is None:
            cache_params = DynamicCache(config=self.config)

        hidden_states = inputs_embeds
        all_hidden_states = () if output_hidden_states else None
        for mixer_block in self.layers:
            hidden_states = mixer_block(
                hidden_states,
                cache_params=cache_params,
                attention_mask=attention_mask,
            )

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

        hidden_states = self.norm_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)

        return MambaOutput(
            last_hidden_state=hidden_states,
            cache_params=cache_params if use_cache else None,
            hidden_states=all_hidden_states,
        )


@auto_docstring(
    custom_intro="""
    The MAMBA Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """
)
class MambaForCausalLM(MambaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "backbone.embeddings.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.backbone = MambaModel(config)
        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.backbone.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.backbone.set_input_embeddings(new_embeddings)

    def prepare_inputs_for_generation(
        self,
        input_ids,
        inputs_embeds=None,
        use_cache=None,
        cache_params: Cache | None = None,
        attention_mask: torch.LongTensor | None = None,
        is_first_iteration: bool | None = False,
        **kwargs,
    ):
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            cache_params=cache_params,
            attention_mask=attention_mask,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if use_cache and not is_first_iteration:
            model_inputs["attention_mask"] = None

        return model_inputs

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        cache_params: Cache | None = None,
        labels: torch.LongTensor | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,  # for now we need this for generation
    ) -> tuple | MambaCausalLMOutput:
        r"""
        cache_params (`Cache`, *optional*):
            If passed along, the model uses the previous state in all the blocks (which will give the output for the
            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        use_cache (`bool`, *optional*):
            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
        """
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        mamba_outputs = self.backbone(
            input_ids,
            cache_params=cache_params,
            inputs_embeds=inputs_embeds,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            use_cache=use_cache,
            attention_mask=attention_mask,
        )

        hidden_states = mamba_outputs[0]
        # Only compute necessary logits
        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, :].to(self.lm_head.weight.dtype)).float()

        loss = None
        if labels is not None:
            # move labels to correct device
            labels = labels.to(logits.device)
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

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

        return MambaCausalLMOutput(
            loss=loss,
            logits=logits,
            cache_params=mamba_outputs.cache_params,
            hidden_states=mamba_outputs.hidden_states,
        )


__all__ = ["MambaForCausalLM", "MambaModel", "MambaPreTrainedModel"]