modeling_rwkv.py

# coding=utf-8
# Copyright 2023 Bo Peng and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch RWKV model."""

import math
from dataclasses import dataclass
from pathlib import Path
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss

from ...modeling_utils import PreTrainedModel
from ...utils import (
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_bitsandbytes_available,
    is_ninja_available,
    is_torch_cuda_available,
    logging,
)
from .configuration_rwkv import RwkvConfig


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "RWKV/rwkv-4-169m-pile"
_CONFIG_FOR_DOC = "RwkvConfig"

RWKV_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "RWKV/rwkv-4-169m-pile",
    "RWKV/rwkv-4-430m-pile",
    "RWKV/rwkv-4-1b5-pile",
    "RWKV/rwkv-4-3b-pile",
    "RWKV/rwkv-4-7b-pile",
    "RWKV/rwkv-4-14b-pile",
    "RWKV/rwkv-raven-1b5",
    "RWKV/rwkv-raven-3b",
    "RWKV/rwkv-raven-7b",
    "RWKV/rwkv-raven-14b",
    # See all RWKV models at https://huggingface.co/models?filter=rwkv
]


rwkv_cuda_kernel = None


def load_wkv_cuda_kernel(context_length):
    from torch.utils.cpp_extension import load as load_kernel

    global rwkv_cuda_kernel

    kernel_folder = Path(__file__).resolve().parent.parent.parent / "kernels" / "rwkv"
    cuda_kernel_files = [kernel_folder / f for f in ["wkv_op.cpp", "wkv_cuda.cu", "wkv_cuda_bf16.cu"]]

    # Only load the kernel if it's not been loaded yet or if we changed the context length
    if rwkv_cuda_kernel is not None and rwkv_cuda_kernel.max_seq_length == context_length:
        return

    logger.info(f"Loading CUDA kernel for RWKV at context length of {context_length}.")

    flags = [
        "-res-usage",
        "--maxrregcount 60",
        "--use_fast_math",
        "-O3",
        "-Xptxas -O3",
        "--extra-device-vectorization",
        f"-DTmax={context_length}",
    ]
    rwkv_cuda_kernel = load_kernel(
        name=f"wkv_{context_length}",
        sources=cuda_kernel_files,
        verbose=(logging.get_verbosity() == logging.DEBUG),
        extra_cuda_cflags=flags,
    )
    rwkv_cuda_kernel.max_seq_length = context_length


class RwkvLinearAttention(torch.autograd.Function):
    @staticmethod
    def forward(ctx, time_decay, time_first, key, value, state=None, return_state=False):
        batch_size, seq_len, hidden_size = key.size()
        if seq_len > rwkv_cuda_kernel.max_seq_length:
            raise ValueError(
                f"Cannot process a batch with {seq_len} tokens at the same time, use a maximum of "
                f"{rwkv_cuda_kernel.max_seq_length} with this model."
            )
        if batch_size * hidden_size % min(hidden_size, 32) != 0:
            raise ValueError(
                f"The product of batch size ({batch_size}) and hidden size ({hidden_size}) needs to be a round "
                f"multiple of {min(hidden_size, 32)}."
            )

        ctx.input_dtype = key.dtype

        if (
            time_decay.device.type != "cuda"
            or time_first.device.type != "cuda"
            or key.device.type != "cuda"
            or value.device.type != "cuda"
        ):
            raise ValueError("Calling the CUDA kernel for wkv attention requires all tensors to be on CUDA devices.")

        time_decay = -torch.exp(time_decay.float().contiguous())
        if key.dtype == torch.float16:
            time_first = time_first.float()
            key = key.float()
            value = value.float()
        time_first = time_first.contiguous()
        key = key.contiguous()
        value = value.contiguous()
        # The CUDA kernel will fill this tensor.
        output = torch.empty_like(key, memory_format=torch.contiguous_format)
        if return_state or state is not None:
            if state is None:
                state = torch.zeros(
                    batch_size,
                    hidden_size,
                    3,
                    dtype=torch.float32,
                    device=key.device,
                    memory_format=torch.contiguous_format,
                )
                state[:, :, 2] -= 1e38
            else:
                state = torch.cat([s.unsqueeze(2) for s in state], dim=2).contiguous()
            if key.dtype == torch.bfloat16:
                forward_func = rwkv_cuda_kernel.forward_with_state_bf16
            else:
                forward_func = rwkv_cuda_kernel.forward_with_state
            forward_func(time_decay, time_first, key, value, output, state)
        else:
            forward_func = rwkv_cuda_kernel.forward_bf16 if key.dtype == torch.bfloat16 else rwkv_cuda_kernel.forward
            forward_func(time_decay, time_first, key, value, output)

        ctx.save_for_backward(time_decay, time_first, key, value, output)

        if state is not None:
            state = [s.squeeze(2) for s in torch.chunk(state, 3, dim=2)]

        return output.to(ctx.input_dtype), state

    @staticmethod
    # g stands for grad
    def backward(ctx, g_output, g_state=None):
        input_dtype = ctx.input_dtype

        time_decay, time_first, key, value, output = ctx.saved_tensors
        # The CUDA kernel will fill those tensors.
        g_time_decay = torch.empty_like(
            time_decay,
            memory_format=torch.contiguous_format,
            dtype=torch.bfloat16 if input_dtype == torch.bfloat16 else torch.float32,
        )
        g_time_first = torch.empty_like(time_first, memory_format=torch.contiguous_format)
        g_key = torch.empty_like(key, memory_format=torch.contiguous_format)
        g_value = torch.empty_like(value, memory_format=torch.contiguous_format)

        if input_dtype == torch.float16:
            g_output = g_output.float()
        backward_func = rwkv_cuda_kernel.backward_bf16 if input_dtype == torch.bfloat16 else rwkv_cuda_kernel.backward
        backward_func(
            time_decay,
            time_first,
            key,
            value,
            output,
            g_output.contiguous(),
            g_time_decay,
            g_time_first,
            g_key,
            g_value,
        )

        return (
            g_time_decay.to(input_dtype),
            g_time_first.to(input_dtype),
            g_key.to(input_dtype),
            g_value.to(input_dtype),
            None,
            None,
        )


def rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=None, return_state=False):
    # For CPU fallback. Will be slower and probably take more memory than the custom CUDA kernel if not executed
    # within a torch.no_grad.
    _, seq_length, _ = key.size()
    output = torch.zeros_like(key)

    if state is None:
        num_state = torch.zeros_like(key[:, 0], dtype=torch.float32)
        den_state = torch.zeros_like(key[:, 0], dtype=torch.float32)
        max_state = torch.zeros_like(key[:, 0], dtype=torch.float32) - 1e38
    else:
        num_state, den_state, max_state = state
    # For numerical stability
    #    real_numerator_state = num_state * torch.exp(max_state)
    #    real_denominator_state = den_state * torch.exp(max_state)

    time_decay = -torch.exp(time_decay)

    for current_index in range(seq_length):
        current_key = key[:, current_index].float()
        current_value = value[:, current_index]

        # wkv computation at time t
        max_for_output = torch.maximum(max_state, current_key + time_first)
        e1 = torch.exp(max_state - max_for_output)
        e2 = torch.exp(current_key + time_first - max_for_output)
        numerator = e1 * num_state + e2 * current_value
        denominator = e1 * den_state + e2
        output[:, current_index] = (numerator / denominator).to(output.dtype)

        # Update state for next iteration
        max_for_state = torch.maximum(max_state + time_decay, current_key)
        e1 = torch.exp(max_state + time_decay - max_for_state)
        e2 = torch.exp(current_key - max_for_state)
        num_state = e1 * num_state + e2 * current_value
        den_state = e1 * den_state + e2
        max_state = max_for_state

    if return_state or state is not None:
        state = [num_state, den_state, max_state]

    return output, state


def rwkv_linear_attention(time_decay, time_first, key, value, state=None, return_state=False):
    no_cuda = any(t.device.type != "cuda" for t in [time_decay, time_first, key, value])
    # Launching the CUDA kernel for just one token will actually be slower (there is no for loop in the CPU version
    # in this case).
    one_token = key.size(1) == 1
    if rwkv_cuda_kernel is None or no_cuda or one_token:
        return rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=state, return_state=return_state)
    else:
        return RwkvLinearAttention.apply(time_decay, time_first, key, value, state, return_state)


class RwkvSelfAttention(nn.Module):
    def __init__(self, config, layer_id=0):
        super().__init__()
        self.config = config
        kernel_loaded = rwkv_cuda_kernel is not None and rwkv_cuda_kernel.max_seq_length == config.context_length
        if is_ninja_available() and is_torch_cuda_available() and not kernel_loaded:
            try:
                load_wkv_cuda_kernel(config.context_length)
            except Exception:
                logger.info("Could not load the custom CUDA kernel for RWKV attention.")
        self.layer_id = layer_id
        hidden_size = config.hidden_size
        attention_hidden_size = (
            config.attention_hidden_size if config.attention_hidden_size is not None else hidden_size
        )
        self.attention_hidden_size = attention_hidden_size

        self.time_decay = nn.Parameter(torch.empty(attention_hidden_size))
        self.time_first = nn.Parameter(torch.empty(attention_hidden_size))

        self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
        self.time_mix_value = nn.Parameter(torch.empty(1, 1, hidden_size))
        self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))

        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
        self.key = nn.Linear(hidden_size, attention_hidden_size, bias=False)
        self.value = nn.Linear(hidden_size, attention_hidden_size, bias=False)
        self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
        self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)

    # TODO: maybe jit, otherwise move inside forward
    def extract_key_value(self, hidden, state=None):
        # Mix hidden with the previous timestep to produce key, value, receptance
        if hidden.size(1) == 1 and state is not None:
            shifted = state[1][:, :, self.layer_id]
        else:
            shifted = self.time_shift(hidden)
            if state is not None:
                shifted[:, 0] = state[1][:, :, self.layer_id]
        key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
        value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
        receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)

        key = self.key(key)
        value = self.value(value)
        receptance = torch.sigmoid(self.receptance(receptance))
        if state is not None:
            state[1][:, :, self.layer_id] = hidden[:, -1]
        return receptance, key, value, state

    def forward(self, hidden, state=None, use_cache=False):
        receptance, key, value, state = self.extract_key_value(hidden, state=state)
        layer_state = tuple(s[:, :, self.layer_id] for s in state[2:]) if state is not None else None
        rwkv, layer_state = rwkv_linear_attention(
            self.time_decay,
            self.time_first,
            key,
            value,
            state=layer_state,
            return_state=use_cache,
        )

        if layer_state is not None:
            state[2][:, :, self.layer_id] = layer_state[0]
            state[3][:, :, self.layer_id] = layer_state[1]
            state[4][:, :, self.layer_id] = layer_state[2]

        return self.output(receptance * rwkv), state


class RwkvFeedForward(nn.Module):
    def __init__(self, config, layer_id=0):
        super().__init__()
        self.config = config
        self.layer_id = layer_id
        hidden_size = config.hidden_size
        intermediate_size = (
            config.intermediate_size if config.intermediate_size is not None else 4 * config.hidden_size
        )

        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
        self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
        self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))

        self.key = nn.Linear(hidden_size, intermediate_size, bias=False)
        self.receptance = nn.Linear(hidden_size, hidden_size, bias=False)
        self.value = nn.Linear(intermediate_size, hidden_size, bias=False)

    def forward(self, hidden, state=None):
        if hidden.size(1) == 1 and state is not None:
            shifted = state[0][:, :, self.layer_id]
        else:
            shifted = self.time_shift(hidden)
            if state is not None:
                shifted[:, 0] = state[0][:, :, self.layer_id]
        key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
        receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)

        key = torch.square(torch.relu(self.key(key)))
        value = self.value(key)
        receptance = torch.sigmoid(self.receptance(receptance))

        if state is not None:
            state[0][:, :, self.layer_id] = hidden[:, -1]

        return receptance * value, state


class RwkvBlock(nn.Module):
    def __init__(self, config, layer_id):
        super().__init__()
        self.config = config
        self.layer_id = layer_id

        if layer_id == 0:
            self.pre_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        self.attention = RwkvSelfAttention(config, layer_id)
        self.feed_forward = RwkvFeedForward(config, layer_id)

    def forward(self, hidden, state=None, use_cache=False, output_attentions=False):
        if self.layer_id == 0:
            hidden = self.pre_ln(hidden)

        attention, state = self.attention(self.ln1(hidden), state=state, use_cache=use_cache)
        hidden = hidden + attention

        feed_forward, state = self.feed_forward(self.ln2(hidden), state=state)
        hidden = hidden + feed_forward

        outputs = (hidden, state)
        if output_attentions:
            outputs += (attention,)
        else:
            outputs += (None,)

        return outputs


class RwkvPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = RwkvConfig
    base_model_prefix = "rwkv"
    _no_split_modules = ["RwkvBlock"]
    _keep_in_fp32_modules = ["time_decay", "time_first"]
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, RwkvSelfAttention):
            layer_id = module.layer_id
            num_hidden_layers = module.config.num_hidden_layers
            hidden_size = module.config.hidden_size
            attention_hidden_size = module.attention_hidden_size

            ratio_0_to_1 = layer_id / (num_hidden_layers - 1)  # 0 to 1
            ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers)  # 1 to ~0

            time_weight = torch.tensor(
                [i / hidden_size for i in range(hidden_size)],
                dtype=module.time_mix_key.dtype,
                device=module.time_mix_key.device,
            )
            time_weight = time_weight[None, None, :]

            decay_speed = [
                -5 + 8 * (h / (attention_hidden_size - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
                for h in range(attention_hidden_size)
            ]
            decay_speed = torch.tensor(decay_speed, dtype=module.time_decay.dtype, device=module.time_decay.device)
            zigzag = (
                torch.tensor(
                    [(i + 1) % 3 - 1 for i in range(attention_hidden_size)],
                    dtype=module.time_first.dtype,
                    device=module.time_first.device,
                )
                * 0.5
            )

            with torch.no_grad():
                module.time_decay.data = decay_speed
                module.time_first.data = torch.ones_like(module.time_first * math.log(0.3) + zigzag)

                module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
                module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
                module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
        elif isinstance(module, RwkvFeedForward):
            layer_id = module.layer_id
            num_hidden_layers = module.config.num_hidden_layers
            hidden_size = module.config.hidden_size

            ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers)  # 1 to ~0

            time_weight = torch.tensor(
                [i / hidden_size for i in range(hidden_size)],
                dtype=module.time_mix_key.dtype,
                device=module.time_mix_key.device,
            )
            time_weight = time_weight[None, None, :]

            with torch.no_grad():
                module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
                module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0)


@dataclass
class RwkvOutput(ModelOutput):
    """
    Class for the RWKV model outputs.

    Args:
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`):
            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`.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    last_hidden_state: torch.FloatTensor = None
    state: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
class RwkvCausalLMOutput(ModelOutput):
    """
    Base class for causal language model (or autoregressive) outputs.

    Args:
        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).
        state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`):
            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`.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    state: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


RWKV_START_DOCSTRING = r"""

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

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

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

RWKV_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
            `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
            sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

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

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

            This is currently not used by `RwkvModel`, but will be supported in the future.

            [What are attention masks?](../glossary#attention-mask)
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *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 last state is returned and can be used to quickly generate the next logits.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare RWKV Model transformer outputting raw hidden-states without any specific head on top.",
    RWKV_START_DOCSTRING,
)
class RwkvModel(RwkvPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.blocks = nn.ModuleList([RwkvBlock(config, layer_id=idx) for idx in range(config.num_hidden_layers)])
        self.ln_out = nn.LayerNorm(config.hidden_size)

        self.layers_are_rescaled = False

        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.embeddings

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

    @add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=RwkvOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,  # noqa
        inputs_embeds: Optional[torch.FloatTensor] = None,
        state: Optional[List[torch.FloatTensor]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, RwkvOutput]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if self.training == self.layers_are_rescaled:
            self._rescale_layers()

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is None and inputs_embeds is None:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

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

        if use_cache and state is None:
            shape = (inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers)
            state = [
                torch.zeros(
                    *shape, dtype=inputs_embeds.dtype if i <= 1 else torch.float32, device=inputs_embeds.device
                )
                for i in range(5)
            ]
            state[4] -= 1e30

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

        hidden_states = inputs_embeds

        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for idx, block in enumerate(self.blocks):
            if self.gradient_checkpointing and self.training:
                hidden_states, state, attentions = self._gradient_checkpointing_func(
                    block.__call__, hidden_states, state, use_cache, output_attentions
                )
            else:
                hidden_states, state, attentions = block(
                    hidden_states, state=state, use_cache=use_cache, output_attentions=output_attentions
                )

            if (
                self.layers_are_rescaled
                and self.config.rescale_every > 0
                and (idx + 1) % self.config.rescale_every == 0
            ):
                hidden_states = hidden_states / 2

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if output_attentions:
                all_self_attentions = all_self_attentions + (attentions,)

        hidden_states = self.ln_out(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(x for x in [hidden_states, state, all_hidden_states, all_self_attentions] if x is not None)

        return RwkvOutput(
            last_hidden_state=hidden_states,
            state=state,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )

    def _rescale_layers(self):
        # Layers should be rescaled for inference only.
        if self.layers_are_rescaled == (not self.training):
            return
        if self.config.rescale_every > 0:
            with torch.no_grad():
                for block_id, block in enumerate(self.blocks):
                    if self.training:
                        block.attention.output.weight.mul_(2 ** int(block_id // self.config.rescale_every))
                        block.feed_forward.value.weight.mul_(2 ** int(block_id // self.config.rescale_every))
                    else:
                        # Deal with quantization statistics
                        if hasattr(block.attention.output.weight, "SCB"):
                            block.attention.output.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
                            block.feed_forward.value.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
                        elif hasattr(block.attention.output.weight, "quant_state"):
                            self._bnb_4bit_dequantize_and_rescale(block.attention.output, block_id)
                            self._bnb_4bit_dequantize_and_rescale(block.feed_forward.value, block_id)
                        else:
                            block.attention.output.weight.div_(2 ** int(block_id // self.config.rescale_every))
                            block.feed_forward.value.weight.div_(2 ** int(block_id // self.config.rescale_every))

        self.layers_are_rescaled = not self.training

    def _bnb_4bit_dequantize_and_rescale(self, target_layer, block_id):
        r"""
        Perform the dequantization and rescaling of the weights of a given layer. After that operation the layer will
        be quantized again.
        """
        if not is_bitsandbytes_available():
            raise ImportError("Please install bitsandbytes to use this method.")
        import bitsandbytes as bnb

        dequant_weights = bnb.functional.dequantize_4bit(target_layer.weight.data, target_layer.weight.quant_state)

        dequant_weights.div_(2 ** int(block_id // self.config.rescale_every))

        # re-quantize the model:
        # we need to put it first on CPU then back to the device
        # this will create an overhead :/
        # We set requires_grad=False as we cannot compute gradients on top of 4bit parameters anyway and to avoid
        # bugs with bnb
        quant_weight = bnb.nn.Params4bit(dequant_weights.to("cpu"), requires_grad=False).to(dequant_weights.device)
        setattr(target_layer, "weight", quant_weight)


@add_start_docstrings(
    """
    The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    RWKV_START_DOCSTRING,
)
class RwkvForCausalLM(RwkvPreTrainedModel):
    _tied_weights_keys = ["head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.rwkv = RwkvModel(config)
        self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    def get_output_embeddings(self):
        return self.head

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

    def generate(self, *args, **kwargs):
        # Thin wrapper to raise exceptions when trying to generate with methods that manipulate `past_key_values`.
        # RWKV is one of the few models that don't have it (it has `state` instead, which has different properties and
        # usage).
        try:
            gen_output = super().generate(*args, **kwargs)
        except AttributeError as exc:
            # Expected exception: "AttributeError: '(object name)' object has no attribute 'past_key_values'"
            if "past_key_values" in str(exc):
                raise AttributeError(
                    "You tried to call `generate` with a decoding strategy that manipulates `past_key_values`. RWKV "
                    "doesn't have that attribute, try another generation strategy instead. For the available "
                    "generation strategies, check this doc: https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
                )
            else:
                raise exc
        return gen_output

    def prepare_inputs_for_generation(self, input_ids, state=None, inputs_embeds=None, **kwargs):
        # only last token for inputs_ids if the state is passed along.
        if state is not None:
            input_ids = input_ids[:, -1].unsqueeze(-1)

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and state is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs["state"] = state
        return model_inputs

    @add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=RwkvCausalLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,  # noqa
        inputs_embeds: Optional[torch.FloatTensor] = None,
        state: Optional[List[torch.FloatTensor]] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, RwkvCausalLMOutput]:
        r"""
        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]`
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        rwkv_outputs = self.rwkv(
            input_ids,
            inputs_embeds=inputs_embeds,
            state=state,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = rwkv_outputs[0]

        logits = self.head(hidden_states)

        loss = None
        if labels is not None:
            # move labels to correct device to enable model parallelism
            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,) + rwkv_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return RwkvCausalLMOutput(
            loss=loss,
            logits=logits,
            state=rwkv_outputs.state,
            hidden_states=rwkv_outputs.hidden_states,
            attentions=rwkv_outputs.attentions,
        )