image-match/python/py/Lib/site-packages/transformers/models/bamba/modular_bamba.py

# Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch Bamba model."""

from typing import TypedDict

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...integrations.hub_kernels import lazy_load_kernel
from ...masking_utils import create_causal_mask
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from ...modeling_utils import PreTrainedModel
from ...processing_utils import Unpack
from ...utils import auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
from ...utils.generic import merge_with_config_defaults
from ...utils.import_utils import resolve_internal_import
from ...utils.output_capturing import capture_outputs
from ..jamba.modeling_jamba import JambaAttentionDecoderLayer
from ..llama.modeling_llama import (
    LlamaAttention,
    LlamaForCausalLM,
    LlamaMLP,
    LlamaRMSNorm,
    LlamaRotaryEmbedding,
    rotate_half,
)
from ..mamba2.modeling_mamba2 import (
    MambaRMSNormGated,
    apply_mask_to_padding_states,
    pad_tensor_by_size,
    reshape_into_chunks,
    segment_sum,
)
from .configuration_bamba import BambaConfig


logger = logging.get_logger(__name__)


class BambaFlashAttentionKwargs(TypedDict, total=False):
    """
    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
    Use cases include padding-free training and fewer `torch.compile` graph breaks.

    cu_seq_lens_q (`torch.LongTensor`):
        Gets cumulative sequence length for query state.
    cu_seq_lens_k (`torch.LongTensor`):
        Gets cumulative sequence length for key state.
    max_length_q (`int`):
        Maximum sequence length for query state.
    max_length_k (`int`):
        Maximum sequence length for key state.
    seq_idx (`torch.IntTensor`):
        Index of each packed sequence.
    """

    cu_seq_lens_q: torch.LongTensor
    cu_seq_lens_k: torch.LongTensor
    max_length_q: int
    max_length_k: int
    seq_idx: torch.IntTensor


class BambaRotaryEmbedding(LlamaRotaryEmbedding):
    pass


# Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Removes the interleaving of cos and sin from GLM

    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.
        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)

    # Keep half or full tensor for later concatenation
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


class BambaAttention(LlamaAttention):
    pass


class BambaRMSNormGated(MambaRMSNormGated):
    pass


# Adapted from transformers.models.mamba2.modeling_mamba2.Mamba2Mixer
class BambaMixer(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)

    The are a few differences between this and Mamba2Mixer:
    - The variable use_precomputed_states is slightly different due to the hybrid cache structure
    - There's a few non-obvious bugs fixed with batching in the slow path that exist in main
    - Some extra variables that our layer doesn't need have been removed
    - We ported most of the refactors in https://github.com/huggingface/transformers/pull/35154, which is (as of Dec 18, 2024) unmerged
    """

    def __init__(self, config: BambaConfig, layer_idx: int):
        super().__init__()
        self.num_heads = config.mamba_n_heads
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.mamba_d_state
        self.conv_kernel_size = config.mamba_d_conv
        self.intermediate_size = int(config.mamba_expand * self.hidden_size)
        self.layer_idx = layer_idx
        self.use_conv_bias = config.mamba_conv_bias
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.use_bias = config.mamba_proj_bias

        self.layer_norm_epsilon = config.rms_norm_eps

        self.n_groups = config.mamba_n_groups
        self.head_dim = config.mamba_d_head
        self.chunk_size = config.mamba_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.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.mamba_conv_bias,
            kernel_size=self.conv_kernel_size,
            groups=self.conv_dim,
            padding=self.conv_kernel_size - 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=self.use_bias,
        )
        # selective projection used to make dt, B and C input dependent

        # time step projection (discretization)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(torch.ones(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
        A = torch.arange(1, self.num_heads + 1)
        self.A_log = nn.Parameter(torch.log(A))
        self.norm = BambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon)
        self.D = nn.Parameter(torch.ones(self.num_heads))

        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.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"
            )
        else:
            logger.warning_once("The fast path for Bamba will be used when running the model on a GPU")

    def cuda_kernels_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        seq_idx: torch.IntTensor | 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

        use_precomputed_states = (
            cache_params is not None and cache_params.has_previous_state(self.layer_idx) and seq_len == 1
        )

        # getting projected states from cache if it exists
        if use_precomputed_states:
            gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
                [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=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(
                    [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
                )

                # 2. Convolution sequence transformation
                # Init cache
                if cache_params is not None:
                    # storing the states
                    # If we just take xBC[:, :, -self.d_conv :], it will error if seqlen < self.d_conv
                    # Instead F.pad will pad with zeros if seqlen < self.d_conv, and truncate otherwise.
                    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, 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,
                        seq_idx=seq_idx,
                    ).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=seq_idx,
                    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:
                    ssm_state = cache_params.update_recurrent_state(ssm_state, 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,
        input_states,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        batch_size, seq_len, _ = input_states.shape
        dtype = input_states.dtype

        # 1. Gated MLP's linear projection
        input_states = apply_mask_to_padding_states(input_states, attention_mask)
        projected_states = self.in_proj(input_states)
        gate, hidden_states_B_C, dt = projected_states.split(
                [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
        )
        hidden_states_B_C = hidden_states_B_C.transpose(1,2)

        use_precomputed_states = cache_params is not None and cache_params.has_previous_state(self.layer_idx) and seq_len == 1

        # 2. Convolution sequence transformation
        if use_precomputed_states:
            conv_states = cache_params.update_conv_state(hidden_states_B_C, 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)
                )
                conv_states = cache_params.update_conv_state(conv_states, 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 use_precomputed_states:
            # We need to guarantee that anything regarding the cache is on the same device
            cache_device = cache_params.layers[self.layer_idx].recurrent_states.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, 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]

            ssm_states = ssm_states.to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
            # Reshape ssm_states to merge the first two dimensions
            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:
                ssm_state = cache_params.update_recurrent_state(ssm_state, 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,
        seq_idx: torch.IntTensor | 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, seq_idx)
        if seq_idx is not None:
            raise NotImplementedError(
                "`seq_idx` support requires fast path support. Please install `mamba_ssm` and `causal_conv1d`"
            )
        dtype = hidden_states.dtype
        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

        return self.torch_forward(hidden_states, cache_params, attention_mask)


class BambaMLP(LlamaMLP):
    pass


class BambaRMSNorm(LlamaRMSNorm):
    pass


class BambaDecoderLayer(JambaAttentionDecoderLayer):
    def __init__(self, config: BambaConfig, layer_idx: int, layer_type: str = "mamba"):
        super().__init__(config, layer_idx)

        del self.self_attn

        num_experts = 1
        ffn_layer_class = BambaMLP if num_experts == 1 else None
        self.feed_forward = ffn_layer_class(config)

        self.layer_type = layer_type
        if layer_type == "mamba":
            self.mamba = BambaMixer(config=config, layer_idx=layer_idx)
        elif layer_type == "attention":
            self.self_attn = BambaAttention(config, layer_idx)
        else:
            raise ValueError("Invalid layer_type")

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[BambaFlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        if self.layer_type == "mamba":
            hidden_states = self.mamba(
                hidden_states=hidden_states,
                cache_params=past_key_values,
                attention_mask=attention_mask,
                **kwargs,
            )
            self_attn_weights = None
        elif self.layer_type == "attention":
            hidden_states, self_attn_weights = self.self_attn(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.pre_ff_layernorm(hidden_states)
        hidden_states = self.feed_forward(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states, self_attn_weights


@auto_docstring
class BambaPreTrainedModel(PreTrainedModel):
    config: BambaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BambaDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True
    _is_stateful = True
    _can_record_outputs = {
        "hidden_states": BambaDecoderLayer,
        "attentions": BambaAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, BambaMixer):
            init.ones_(module.dt_bias)
            init.copy_(module.A_log, torch.log(torch.arange(1, module.num_heads + 1)))
            init.ones_(module.D)


@auto_docstring
class BambaModel(BambaPreTrainedModel):
    def __init__(self, config: BambaConfig):
        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)
        decoder_layers = []
        for i in range(config.num_hidden_layers):
            decoder_layers.append(BambaDecoderLayer(config, layer_idx=i, layer_type=config.layers_block_type[i]))
        self.layers = nn.ModuleList(decoder_layers)

        self._attn_implementation = config._attn_implementation
        self.final_layernorm = BambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = BambaRotaryEmbedding(config=config)

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[BambaFlashAttentionKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        hidden_states = inputs_embeds

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

        if position_ids is None:
            position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device).unsqueeze(0)

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )
        mamba_mask = self._update_mamba_mask(attention_mask, past_key_values)
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        for i, decoder_layer in enumerate(self.layers):
            layer_mask = mamba_mask if self.config.layers_block_type[i] == "mamba" else causal_mask

            hidden_states, attn_weights = decoder_layer(
                hidden_states,
                attention_mask=layer_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.final_layernorm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )

    def _update_mamba_mask(self, attention_mask, past_key_values):
        """
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        """
        mamba_mask = attention_mask
        if (past_key_values is not None and past_key_values.has_previous_state()) or (
            attention_mask is not None and torch.all(attention_mask == 1)
        ):
            mamba_mask = None
        return mamba_mask


class BambaForCausalLM(LlamaForCausalLM):
    def __init__(self, config):
        super().__init__(config)
        self.z_loss_coefficient = config.z_loss_coefficient

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> 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]`.

        Example:

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

        >>> model = BambaForCausalLM.from_pretrained("...")
        >>> tokenizer = AutoTokenizer.from_pretrained("...")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        outputs: BaseModelOutputWithPast = 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,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        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, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
            if self.z_loss_coefficient > 0:
                z_loss = logits.logsumexp(dim=-1).to(dtype=loss.dtype).pow(2).mean()
                loss = loss + self.z_loss_coefficient * z_loss

        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,
        position_ids=None,
        use_cache=True,
        is_first_iteration=False,
        **kwargs,
    ):
        kwargs["logits_to_keep"] = self.config.num_logits_to_keep
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            position_ids=position_ids,
            use_cache=use_cache,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        return model_inputs


__all__ = ["BambaModel", "BambaForCausalLM", "BambaPreTrainedModel"]