image-match/python/py/Lib/site-packages/transformers/models/mamba2/modeling_mamba2.py

# Copyright 2024 state-spaces/mamba2 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 MAMBA2 model."""

import math
from dataclasses import dataclass

import torch
from torch import nn

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, is_torchdynamo_compiling, logging
from ...utils.import_utils import resolve_internal_import
from .configuration_mamba2 import Mamba2Config


logger = logging.get_logger(__name__)


# Helper methods for segment sum computation


def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
    """
    Padding x tensor with `pad_size` on the seq_len dim (dim=1)

    Assumes that we only have tensors of either size 4 or 3
    """
    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)

    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)


def reshape_into_chunks(input_tensor, pad_size, chunk_size):
    """
    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
    simultaneously splitting it into chunk sequences.

    Assumes that we only have tensors of either size 4 or 3
    """
    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
    input_tensor = pad_tensor_by_size(input_tensor, pad_size)

    if len(input_tensor.shape) == 3:
        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
    else:
        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
        return input_tensor.reshape(
            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
        )


def segment_sum(input_tensor):
    """
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    """
    chunk_size = input_tensor.size(-1)
    # 1. expand input tensor to have an additional dimension and repeat along that dimension
    # [..., chunk_size] -> [..., chunk_size, chunk_size]
    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
    input_tensor = input_tensor.masked_fill(~mask, 0)
    # 3. compute actual cumsum
    tensor_segsum = torch.cumsum(input_tensor, dim=-2)

    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
    return tensor_segsum


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    # NOTE: attention mask is a 2D boolean tensor
    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


class MambaRMSNormGated(torch.nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states, gate=None):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)

        if gate is not None:
            hidden_states = hidden_states * nn.functional.silu(gate.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)


class Mamba2Mixer(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: Mamba2Config, layer_idx: int, initialize_mixer_weights: bool = True):
        super().__init__()
        self.num_heads = config.num_heads
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.state_size
        self.conv_kernel_size = config.conv_kernel
        self.intermediate_size = int(config.expand * self.hidden_size)
        self.time_step_rank = int(config.time_step_rank)
        self.layer_idx = layer_idx
        self.use_conv_bias = config.use_conv_bias
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]

        self.layer_norm_epsilon = config.layer_norm_epsilon
        self.rms_norm = config.rms_norm

        self.n_groups = config.n_groups
        self.head_dim = config.head_dim
        self.chunk_size = config.chunk_size

        self.time_step_limit = config.time_step_limit
        self.time_step_min = config.time_step_min
        self.time_step_max = config.time_step_max
        self.time_step_floor = config.time_step_floor

        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=config.use_conv_bias,
            kernel_size=config.conv_kernel,
            groups=self.conv_dim,
            padding=config.conv_kernel - 1,
        )

        # projection of the input hidden states
        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
        self.in_proj = nn.Linear(
            self.hidden_size,
            projection_size,
            bias=config.use_bias,
        )
        # selective projection used to make dt, B and C input dependent

        # time step projection (discretization)
        self.dt_bias = nn.Parameter(torch.empty(self.num_heads))

        # 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.num_heads))
        self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon)
        self.D = nn.Parameter(torch.empty(self.num_heads))
        if initialize_mixer_weights and self.dt_bias.device.type != "meta":
            self.init_mamba2_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_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 selective_state_update, mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
        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"
        )
        mamba_chunk_scan_combined = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.ssd_combined.mamba_chunk_scan_combined"
        )
        mamba_split_conv1d_scan_combined = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.ssd_combined.mamba_split_conv1d_scan_combined"
        )

        global is_fast_path_available
        is_fast_path_available = all(
            (
                selective_state_update,
                mamba_chunk_scan_combined,
                mamba_split_conv1d_scan_combined,
                causal_conv1d_fn,
                causal_conv1d_update,
            )
        )

        if not is_fast_path_available:
            logger.warning_once(
                "The fast path is not available because one of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
                " https://github.com/Dao-AILab/causal-conv1d"
            )

    @torch.no_grad()
    def init_mamba2_weights(self):
        A = torch.arange(1, self.num_heads + 1, device=self.A_log.device, dtype=torch.float32)
        init.copy_(self.A_log, torch.log(A))
        init.ones_(self.D)

        dt = torch.exp(
            torch.rand(self.num_heads, device=self.dt_bias.device, dtype=torch.float32)
            * (math.log(self.time_step_max) - math.log(self.time_step_min))
            + math.log(self.time_step_min)
        ).clamp(min=self.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_bias, inv_dt)

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

        # Set up dimensions for reshapes later
        batch_size, seq_len, _ = hidden_states.shape
        groups_time_state_size = self.n_groups * self.ssm_state_size
        d_mlp = (
            projected_states.shape[-1]
            - 2 * self.intermediate_size
            - 2 * self.n_groups * self.ssm_state_size
            - self.num_heads
        ) // 2

        # Single step calculations via cache
        if cache_params is not None and cache_params.has_previous_state(self.layer_idx):
            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
            )

            # 2. Convolution sequence transformation
            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                cache_params.layers[self.layer_idx].conv_states,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.activation,
            )

            hidden_states, B, C = torch.split(
                hidden_states_B_C,
                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                dim=-1,
            )

            # 3. SSM transformation
            A = -torch.exp(self.A_log.float())  # (nheads,)
            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
            D = self.D[:, None, ...].expand(-1, self.head_dim)
            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            hidden_states = selective_state_update(
                cache_params.layers[self.layer_idx].recurrent_states,
                hidden_states_reshaped,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
            )
            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
            hidden_states = self.norm(hidden_states, gate)

            # 4. Final linear projection
            out = self.out_proj(hidden_states)[:, None, ...]

        # Fused calculations or step by step if no initialized cache is found
        else:
            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}

            # 2-4. Fused kernel for conv1d, SSM, and the final projection
            if self.training and cache_params is None:
                out = mamba_split_conv1d_scan_combined(
                    projected_states,
                    self.conv1d.weight.squeeze(1),
                    self.conv1d.bias,
                    self.dt_bias,
                    A,
                    D=self.D,
                    chunk_size=self.chunk_size,
                    seq_idx=None,  # was seq_idx
                    activation=self.activation,
                    rmsnorm_weight=self.norm.weight,
                    rmsnorm_eps=self.norm.variance_epsilon,
                    outproj_weight=self.out_proj.weight,
                    outproj_bias=self.out_proj.bias,
                    headdim=self.head_dim,
                    ngroups=self.n_groups,
                    norm_before_gate=False,
                    return_final_states=False,
                    **dt_limit_kwargs,
                )

            else:
                _, _, gate, hidden_states_B_C, dt = projected_states.split(
                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
                )

                # 2. Convolution sequence transformation
                # Init cache
                if cache_params is not None:
                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
                    conv_states = nn.functional.pad(
                        hidden_states_B_C_transposed,
                        (self.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
                    )
                    conv_states = cache_params.update_conv_state(conv_states, layer_idx=self.layer_idx)

                if self.activation not in ["silu", "swish"]:
                    hidden_states_B_C = self.act(
                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
                    )
                else:
                    hidden_states_B_C = causal_conv1d_fn(
                        x=hidden_states_B_C.transpose(1, 2),
                        weight=self.conv1d.weight.squeeze(1),
                        bias=self.conv1d.bias,
                        activation=self.activation,
                    ).transpose(1, 2)

                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
                hidden_states, B, C = torch.split(
                    hidden_states_B_C,
                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                    dim=-1,
                )

                # 3. SSM transformation
                scan_output, ssm_state = mamba_chunk_scan_combined(
                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
                    dt,
                    A,
                    B.view(batch_size, seq_len, self.n_groups, -1),
                    C.view(batch_size, seq_len, self.n_groups, -1),
                    chunk_size=self.chunk_size,
                    D=self.D,
                    z=None,
                    seq_idx=None,
                    return_final_states=True,
                    dt_bias=self.dt_bias,
                    dt_softplus=True,
                    **dt_limit_kwargs,
                )

                # Init cache
                if ssm_state is not None and cache_params is not None:
                    cache_params.update_recurrent_state(ssm_state, layer_idx=self.layer_idx)

                scan_output = scan_output.view(batch_size, seq_len, -1)
                # Multiply "gate" branch and apply extra normalization layer
                scan_output = self.norm(scan_output, gate)

                # 4. Final linear projection
                out = self.out_proj(scan_output)
        return out

    # fmt: off
    def torch_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None
    ):
        batch_size, seq_len, _ = hidden_states.shape
        dtype = hidden_states.dtype

        # 1. Gated MLP's linear projection
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states = self.in_proj(hidden_states)
        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
        _, _, gate, hidden_states_B_C, dt = projected_states.split(
                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
        )
        hidden_states_B_C = hidden_states_B_C.transpose(1,2)

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

        # 2. Convolution sequence transformation
        if is_decoding:
            conv_states = cache_params.update_conv_state(hidden_states_B_C, layer_idx=self.layer_idx)

            hidden_states_B_C = torch.sum(
                conv_states * self.conv1d.weight.squeeze(1), dim=-1
            )
            if self.use_conv_bias:
                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
            hidden_states_B_C = self.act(hidden_states_B_C)
        else:
            # Init cache
            if cache_params is not None:
                conv_states = nn.functional.pad(
                    hidden_states_B_C, (self.conv_kernel_size - hidden_states_B_C.shape[-1], 0)
                )
                cache_params.update_conv_state(conv_states, layer_idx=self.layer_idx)

            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C)[..., :seq_len].transpose(1, 2))

        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
        hidden_states, B, C = torch.split(
            hidden_states_B_C,
            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
            dim=-1
        )

        # 3. SSM transformation
        A = -torch.exp(self.A_log.float())                            # [num_heads]
        if is_decoding:
            # We need to guarantee that anything regarding the cache is on the same device
            cache_device = cache_params.layers[self.layer_idx].device

            # Note: there is no need to pad parameter matrices here, as there is just one new token
            # for batched generation
            dt = dt[:, 0, :][:, None, ...]
            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
            # [num_heads] -> [num_heads, head_dim]
            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)

            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            # [bsz, num_heads, head_dim, state_size]
            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)

            # Discretize B
            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
            B = B.reshape(batch_size, -1, B.shape[-1])
            # [bsz, num_heads, head_dim, state_size]
            dB = dt[..., None] * B[..., None, :]

            # Discretize x into dB
            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
            dBx = (dB * hidden_states[..., None]).to(device=cache_device)

            # State calculation
            ssm_states = cache_params.layers[self.layer_idx].recurrent_states * dA + dBx
            ssm_states = cache_params.update_recurrent_state(ssm_states, layer_idx=self.layer_idx)

            # Subsequent output
            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
            C = C.reshape(batch_size, -1, C.shape[-1])
            # [bsz, num_heads, head_dim]

            # Reshape ssm_states to merge the first two dimensions
            ssm_states = ssm_states.to(device=C.device, dtype=C.dtype)
            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
            y = torch.bmm(ssm_states_reshaped, C_reshaped)
            y = y.view(batch_size, self.num_heads, self.head_dim)

            # D skip connection
            # [num_heads] -> [num_heads, head_dim]
            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
            y = (y + hidden_states * D).to(y.dtype)

            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
            y = y.reshape(batch_size, -1)[:, None, ...]
        else:
            # begin ssd naive implementation without einsums
            dt = nn.functional.softplus(dt + self.dt_bias)
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            B = B.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            C = C.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size

            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)

            # Discretize x and A
            hidden_states = hidden_states * dt[..., None]
            A = A.to(hidden_states.dtype) * dt

            # Rearrange into blocks/chunks
            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]

            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
            A = A.permute(0, 3, 1, 2)
            A_cumsum = torch.cumsum(A, dim=-1)

            # 1. Compute the output for each intra-chunk (diagonal blocks)
            # This is the analog of a causal mask
            L = torch.exp(segment_sum(A))

            # Contraction of C and B to get G (attention-weights like)
            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]  # shape: (b, c, l, s, h, n)
            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)

            # Compute M, equivalent to applying attention mask to weights
            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
            M = M_intermediate.sum(dim=-1)

            # Compute Y_diag (apply to values)
            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)

            # 2. Compute the state for each intra-chunk
            # (right term of low-rank factorization of off-diagonal blocks; B terms)
            decay_states = torch.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)

            # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
            # (middle term of factorization of off-diag blocks; A terms)
            previous_states = torch.zeros_like(states[:, :1])
            states = torch.cat([previous_states, states], dim=1)
            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
            decay_chunk = decay_chunk.transpose(1, 3)
            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
            states, ssm_state = new_states[:, :-1], new_states[:, -1]

            # 4. Compute state -> output conversion per chunk
            # (left term of low-rank factorization of off-diagonal blocks; C terms)
            state_decay_out = torch.exp(A_cumsum)
            C_times_states = (C[..., None, :] * states[:, :, None, ...])
            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])

            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
            y = Y_diag + Y_off
            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)

            y = y + D_residual
            # Cutting off padded chunks
            if pad_size > 0:
                y = y[:, :seq_len, :, :]
            y = y.reshape(batch_size, seq_len, -1)

            # Init cache
            if ssm_state is not None and cache_params is not None:
                cache_params.update_recurrent_state(ssm_state, layer_idx=self.layer_idx)

        scan_output = self.norm(y, gate)

        # end ssd naive

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

    def forward(
        self,
        hidden_states,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ):
        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type and not is_torchdynamo_compiling():
            return self.cuda_kernels_forward(hidden_states, cache_params, attention_mask)
        return self.torch_forward(hidden_states, cache_params, attention_mask)


class Mamba2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Mamba2RMSNorm 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)


class Mamba2Block(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 = Mamba2RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.mixer = Mamba2Mixer(config, layer_idx=layer_idx, initialize_mixer_weights=False)

    def forward(
        self,
        hidden_states,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | 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 Mamba2PreTrainedModel(PreTrainedModel):
    config: Mamba2Config
    base_model_prefix = "backbone"
    _no_split_modules = ["Mamba2Block"]
    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, Mamba2Mixer):
            # 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_mamba2_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, (Mamba2RMSNorm, MambaRMSNormGated)):
            init.ones_(module.weight)
        elif isinstance(module, nn.Embedding):
            init.normal_(module.weight, std=std)


@dataclass
@auto_docstring(
    custom_intro="""
    Class for the MAMBA2 model outputs.
    """
)
# Copied from transformers.models.mamba.modeling_mamba.MambaOutput with MAMBA->MAMBA2,Mamba->Mamba2
class Mamba2Output(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.
    """
)
# Copied from transformers.models.mamba.modeling_mamba.MambaCausalLMOutput with Mamba->Mamba2
class Mamba2CausalLMOutput(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 Mamba2Model(Mamba2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([Mamba2Block(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])

        self.gradient_checkpointing = False
        self.norm_f = Mamba2RMSNorm(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.Tensor | None = None,
        **kwargs,
    ) -> tuple | Mamba2Output:
        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 Mamba2Output(
            last_hidden_state=hidden_states,
            cache_params=cache_params if use_cache else None,
            hidden_states=all_hidden_states,
        )


@auto_docstring(
    custom_intro="""
    The MAMBA2 Model transformer with a language modeling head on top (linear layer with weights not tied to the input
    embeddings).
    """
)
class Mamba2ForCausalLM(Mamba2PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "backbone.embeddings.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.backbone = Mamba2Model(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.Tensor | 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,
        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,
        attention_mask: torch.Tensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,  # for now we need this for generation and loss_function
    ) -> tuple | Mamba2CausalLMOutput:
        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

        mamba2_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 = mamba2_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:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

        if not return_dict:
            output = (logits,) + mamba2_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return Mamba2CausalLMOutput(
            loss=loss,
            logits=logits,
            cache_params=mamba2_outputs.cache_params,
            hidden_states=mamba2_outputs.hidden_states,
        )


__all__ = ["Mamba2ForCausalLM", "Mamba2Model", "Mamba2PreTrainedModel"]