add rwkv encoder

aky15

2023-10-20 ab653d3871f72f7f6cd1ac3126b3df722f4c7943

add rwkv encoder

 funasr/build_utils/build_asr_model.py

@@ -42,6 +42,7 @@
from funasr.models.encoder.branchformer_encoder import BranchformerEncoder
from funasr.models.encoder.e_branchformer_encoder import EBranchformerEncoder
from funasr.models.encoder.transformer_encoder import TransformerEncoder
from funasr.models.encoder.rwkv_encoder import RWKVEncoder
from funasr.models.frontend.default import DefaultFrontend
from funasr.models.frontend.default import MultiChannelFrontend
from funasr.models.frontend.fused import FusedFrontends
@@ -119,6 +120,7 @@
        e_branchformer=EBranchformerEncoder,
        mfcca_enc=MFCCAEncoder,
        chunk_conformer=ConformerChunkEncoder,
        rwkv=RWKVEncoder,
    ),
    default="rnn",
)

 funasr/models/encoder/rwkv_encoder.py

New file
@@ -0,0 +1,158 @@
"""RWKV encoder definition for Transducer models."""

import math
from typing import Dict, List, Optional, Tuple

import torch
from typeguard import check_argument_types

from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.modules.rwkv import RWKV
from funasr.modules.layer_norm import LayerNorm
from funasr.modules.rwkv_subsampling import RWKVConvInput
from funasr.modules.nets_utils import make_source_mask

class RWKVEncoder(AbsEncoder):
    """RWKV encoder module.

    Based on https://arxiv.org/pdf/2305.13048.pdf.

    Args:
        vocab_size: Vocabulary size.
        output_size: Input/Output size.
        context_size: Context size for WKV computation.
        linear_size: FeedForward hidden size.
        attention_size: SelfAttention hidden size.
        normalization_type: Normalization layer type.
        normalization_args: Normalization layer arguments.
        num_blocks: Number of RWKV blocks.
        embed_dropout_rate: Dropout rate for embedding layer.
        att_dropout_rate: Dropout rate for the attention module.
        ffn_dropout_rate: Dropout rate for the feed-forward module.
    """

    def __init__(
        self,
        input_size: int,
        output_size: int = 512,
        context_size: int = 1024,
        linear_size: Optional[int] = None,
        attention_size: Optional[int] = None,
        num_blocks: int = 4,
        att_dropout_rate: float = 0.0,
        ffn_dropout_rate: float = 0.0,
        dropout_rate: float = 0.0,
        subsampling_factor: int =4,
        time_reduction_factor: int = 1,
        kernel: int = 3,
    ) -> None:
        """Construct a RWKVEncoder object."""
        super().__init__()

        assert check_argument_types()

        self.embed = RWKVConvInput(
            input_size,
            [output_size//4, output_size//2, output_size],
            subsampling_factor,
            conv_kernel_size=kernel,
            output_size=output_size,
        )

        self.subsampling_factor = subsampling_factor

        linear_size = output_size * 4 if linear_size is None else linear_size
        attention_size = output_size if attention_size is None else attention_size
        
        self.rwkv_blocks = torch.nn.ModuleList(
            [
                RWKV(
                    output_size,
                    linear_size,
                    attention_size,
                    context_size,
                    block_id,
                    num_blocks,
                    att_dropout_rate=att_dropout_rate,
                    ffn_dropout_rate=ffn_dropout_rate,
                    dropout_rate=dropout_rate,
                )
                for block_id in range(num_blocks)
            ]
        )

        self.embed_norm = LayerNorm(output_size)
        self.final_norm = LayerNorm(output_size)

        self._output_size = output_size
        self.context_size = context_size

        self.num_blocks = num_blocks
        self.time_reduction_factor = time_reduction_factor

    def output_size(self) -> int:
        return self._output_size

    def forward(self, x: torch.Tensor, x_len) -> torch.Tensor:
        """Encode source label sequences.

        Args:
            x: Encoder input sequences. (B, L)

        Returns:
            out: Encoder output sequences. (B, U, D)

        """
        _, length, _ = x.size()

        assert (
            length <= self.context_size * self.subsampling_factor
        ), "Context size is too short for current length: %d versus %d" % (
            length,
            self.context_size * self.subsampling_factor,
        )
        mask = make_source_mask(x_len).to(x.device)
        x, mask = self.embed(x, mask, None)
        x = self.embed_norm(x)
        olens = mask.eq(0).sum(1)

        for block in self.rwkv_blocks:
            x, _ = block(x)
        # for streaming inference
        # xs_pad = self.rwkv_infer(xs_pad)

        x = self.final_norm(x)

        if self.time_reduction_factor > 1:
            x = x[:,::self.time_reduction_factor,:]
            olens = torch.floor_divide(olens-1, self.time_reduction_factor) + 1

        return x, olens, None

    def rwkv_infer(self, xs_pad):

        batch_size = xs_pad.shape[0]

        hidden_sizes = [
            self._output_size for i in range(5)
        ]

        state = [
            torch.zeros(
                (batch_size, 1, hidden_sizes[i], self.num_rwkv_blocks),
                dtype=torch.float32,
                device=self.device,
            )
            for i in range(5)
        ]

        state[4] -= 1e-30

        xs_out = []
        for t in range(xs_pad.shape[1]):
            x_t = xs_pad[:,t,:]
            for idx, block in enumerate(self.rwkv_blocks):
                x_t, state = block(x_t, state=state)
            xs_out.append(x_t)
        xs_out = torch.stack(xs_out, dim=1)
        return xs_out

 funasr/modules/rwkv.py

New file
@@ -0,0 +1,145 @@
"""Receptance Weighted Key Value (RWKV) block definition.

Based/modified from https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v4/src/model.py

"""

from typing import Dict, Optional, Tuple

import torch

from funasr.modules.rwkv_attention import EncoderSelfAttention, DecoderSelfAttention
from funasr.modules.rwkv_feed_forward import FeedForward
from funasr.modules.layer_norm import LayerNorm

class RWKV(torch.nn.Module):
    """RWKV module.

    Args:
        size: Input/Output size.
        linear_size: Feed-forward hidden size.
        attention_size: SelfAttention hidden size.
        context_size: Context size for WKV computation.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.
        normalization_class: Normalization layer class.
        normalization_args: Normalization layer arguments.
        att_dropout_rate: Dropout rate for the attention module.
        ffn_dropout_rate: Dropout rate for the feed-forward module.

    """

    def __init__(
        self,
        size: int,
        linear_size: int,
        attention_size: int,
        context_size: int,
        block_id: int,
        num_blocks: int,
        att_dropout_rate: float = 0.0,
        ffn_dropout_rate: float = 0.0,
        dropout_rate: float = 0.0,
    ) -> None:
        """Construct a RWKV object."""
        super().__init__()

        self.layer_norm_att = LayerNorm(size)
        self.layer_norm_ffn = LayerNorm(size)

        self.att = EncoderSelfAttention(
            size, attention_size, context_size, block_id, att_dropout_rate, num_blocks
        )
        self.dropout_att = torch.nn.Dropout(p=dropout_rate)

        self.ffn = FeedForward(size, linear_size, block_id, ffn_dropout_rate, num_blocks)
        self.dropout_ffn = torch.nn.Dropout(p=dropout_rate)

    def forward(
        self,
        x: torch.Tensor,
        state: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """Compute receptance weighted key value.

        Args:
            x: RWKV input sequences. (B, L, size)
            state: Decoder hidden states. [5 x (B, D_att/size, N)]

        Returns:
            x: RWKV output sequences. (B, L, size)
            x: Decoder hidden states. [5 x (B, D_att/size, N)]

        """
        att, state = self.att(self.layer_norm_att(x), state=state)
        x = x + self.dropout_att(att)
        ffn, state = self.ffn(self.layer_norm_ffn(x), state=state)
        x = x + self.dropout_ffn(ffn)
        return x, state

class RWKVDecoderLayer(torch.nn.Module):
    """RWKV module.

    Args:
        size: Input/Output size.
        linear_size: Feed-forward hidden size.
        attention_size: SelfAttention hidden size.
        context_size: Context size for WKV computation.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.
        normalization_class: Normalization layer class.
        normalization_args: Normalization layer arguments.
        att_dropout_rate: Dropout rate for the attention module.
        ffn_dropout_rate: Dropout rate for the feed-forward module.

    """

    def __init__(
        self,
        size: int,
        linear_size: int,
        attention_size: int,
        context_size: int,
        block_id: int,
        num_blocks: int,
        att_dropout_rate: float = 0.0,
        ffn_dropout_rate: float = 0.0,
        dropout_rate: float = 0.0,
    ) -> None:
        """Construct a RWKV object."""
        super().__init__()

        self.layer_norm_att = LayerNorm(size)
        self.layer_norm_ffn = LayerNorm(size)

        self.att = DecoderSelfAttention(
            size, attention_size, context_size, block_id, att_dropout_rate, num_blocks
        )
        self.dropout_att = torch.nn.Dropout(p=dropout_rate)

        self.ffn = FeedForward(size, linear_size, block_id, ffn_dropout_rate, num_blocks)
        self.dropout_ffn = torch.nn.Dropout(p=dropout_rate)

    def forward(
        self,
        x: torch.Tensor,
        state: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """Compute receptance weighted key value.

        Args:
            x: RWKV input sequences. (B, L, size)
            state: Decoder hidden states. [5 x (B, D_att/size, N)]

        Returns:
            x: RWKV output sequences. (B, L, size)
            x: Decoder hidden states. [5 x (B, D_att/size, N)]

        """
        att, state = self.att(self.layer_norm_att(x), state=state)
        x = x + self.dropout_att(att)

        ffn, state = self.ffn(self.layer_norm_ffn(x), state=state)
        x = x + self.dropout_ffn(ffn)

        return x, state

 funasr/modules/rwkv_attention.py

New file
@@ -0,0 +1,632 @@
"""Attention (time mixing) modules for RWKV block.

Based/Modified from https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v4/src/model.py.

Some variables are renamed according to https://github.com/huggingface/transformers/blob/main/src/transformers/models/rwkv/modeling_rwkv.py.

"""  # noqa

import math
from importlib.util import find_spec
from pathlib import Path
from typing import List, Optional, Tuple, Union

import torch

wkv_kernel_encoder = None
wkv_kernel_decoder = None

class WKVLinearAttentionEncoder(torch.autograd.Function):
    """WKVLinearAttention function definition."""

    @staticmethod
    def forward(
        ctx,
        time_decay: torch.Tensor,
        time_first: torch.Tensor,
        key: torch.Tensor,
        value: torch.tensor,
    ) -> torch.Tensor:
        """WKVLinearAttention function forward pass.

        Args:
            time_decay: Channel-wise time decay vector. (D_att)
            time_first: Channel-wise time first vector. (D_att)
            key: Key tensor. (B, U, D_att)
            value: Value tensor. (B, U, D_att)

        Returns:
            out: Weighted Key-Value tensor. (B, U, D_att)

        """
        batch, length, dim = key.size()

        assert length <= wkv_kernel_encoder.context_size, (
            f"Cannot process key of length {length} while context_size "
            f"is ({wkv_kernel_encoder.context_size}). Limit should be increased."
        )

        assert batch * dim % min(dim, 32) == 0, (
            f"batch size ({batch}) by dimension ({dim}) should be a multiple of "
            f"{min(dim, 32)}"
        )

        ctx.input_dtype = key.dtype

        time_decay = -torch.exp(time_decay.float().contiguous())
        time_first = time_first.float().contiguous()

        key = key.float().contiguous()
        value = value.float().contiguous()

        out = torch.empty_like(key, memory_format=torch.contiguous_format)

        wkv_kernel_encoder.forward(time_decay, time_first, key, value, out)
        ctx.save_for_backward(time_decay, time_first, key, value, out)

        return out

    @staticmethod
    def backward(
        ctx, grad_output: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """WKVLinearAttention function backward pass.

        Args:
            grad_output: Output gradient. (B, U, D_att)

        Returns:
            grad_time_decay: Gradient for channel-wise time decay vector. (D_att)
            grad_time_first: Gradient for channel-wise time first vector. (D_att)
            grad_key: Gradient for key tensor. (B, U, D_att)
            grad_value: Gradient for value tensor. (B, U, D_att)

        """
        time_decay, time_first, key, value, output = ctx.saved_tensors
        grad_dtype = ctx.input_dtype

        batch, _, dim = key.size()

        grad_time_decay = torch.empty(
            (batch, dim),
            memory_format=torch.contiguous_format,
            dtype=time_decay.dtype,
            device=time_decay.device,
        )

        grad_time_first = torch.empty(
            (batch, dim),
            memory_format=torch.contiguous_format,
            dtype=time_decay.dtype,
            device=time_decay.device,
        )

        grad_key = torch.empty_like(key, memory_format=torch.contiguous_format)
        grad_value = torch.empty_like(value, memory_format=torch.contiguous_format)

        wkv_kernel_encoder.backward(
            time_decay,
            time_first,
            key,
            value,
            output,
            grad_output.contiguous(),
            grad_time_decay,
            grad_time_first,
            grad_key,
            grad_value,
        )

        grad_time_decay = torch.sum(grad_time_decay, dim=0)
        grad_time_first = torch.sum(grad_time_first, dim=0)

        return (
            grad_time_decay,
            grad_time_first,
            grad_key,
            grad_value,
        )

class WKVLinearAttentionDecoder(torch.autograd.Function):
    """WKVLinearAttention function definition."""

    @staticmethod
    def forward(
        ctx,
        time_decay: torch.Tensor,
        time_first: torch.Tensor,
        key: torch.Tensor,
        value: torch.tensor,
    ) -> torch.Tensor:
        """WKVLinearAttention function forward pass.

        Args:
            time_decay: Channel-wise time decay vector. (D_att)
            time_first: Channel-wise time first vector. (D_att)
            key: Key tensor. (B, U, D_att)
            value: Value tensor. (B, U, D_att)

        Returns:
            out: Weighted Key-Value tensor. (B, U, D_att)

        """
        batch, length, dim = key.size()

        assert length <= wkv_kernel_decoder.context_size, (
            f"Cannot process key of length {length} while context_size "
            f"is ({wkv_kernel.context_size}). Limit should be increased."
        )

        assert batch * dim % min(dim, 32) == 0, (
            f"batch size ({batch}) by dimension ({dim}) should be a multiple of "
            f"{min(dim, 32)}"
        )

        ctx.input_dtype = key.dtype

        time_decay = -torch.exp(time_decay.float().contiguous())
        time_first = time_first.float().contiguous()

        key = key.float().contiguous()
        value = value.float().contiguous()

        out = torch.empty_like(key, memory_format=torch.contiguous_format)

        wkv_kernel_decoder.forward(time_decay, time_first, key, value, out)
        ctx.save_for_backward(time_decay, time_first, key, value, out)

        return out

    @staticmethod
    def backward(
        ctx, grad_output: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """WKVLinearAttention function backward pass.

        Args:
            grad_output: Output gradient. (B, U, D_att)

        Returns:
            grad_time_decay: Gradient for channel-wise time decay vector. (D_att)
            grad_time_first: Gradient for channel-wise time first vector. (D_att)
            grad_key: Gradient for key tensor. (B, U, D_att)
            grad_value: Gradient for value tensor. (B, U, D_att)

        """
        time_decay, time_first, key, value, output = ctx.saved_tensors
        grad_dtype = ctx.input_dtype

        batch, _, dim = key.size()

        grad_time_decay = torch.empty(
            (batch, dim),
            memory_format=torch.contiguous_format,
            dtype=time_decay.dtype,
            device=time_decay.device,
        )

        grad_time_first = torch.empty(
            (batch, dim),
            memory_format=torch.contiguous_format,
            dtype=time_decay.dtype,
            device=time_decay.device,
        )

        grad_key = torch.empty_like(key, memory_format=torch.contiguous_format)
        grad_value = torch.empty_like(value, memory_format=torch.contiguous_format)

        wkv_kernel_decoder.backward(
            time_decay,
            time_first,
            key,
            value,
            output,
            grad_output.contiguous(),
            grad_time_decay,
            grad_time_first,
            grad_key,
            grad_value,
        )

        grad_time_decay = torch.sum(grad_time_decay, dim=0)
        grad_time_first = torch.sum(grad_time_first, dim=0)

        return (
            grad_time_decay,
            grad_time_first,
            grad_key,
            grad_value,
        )

def load_encoder_wkv_kernel(context_size: int) -> None:
    """Load WKV CUDA kernel.

    Args:
        context_size: Context size.

    """
    from torch.utils.cpp_extension import load

    global wkv_kernel_encoder

    if wkv_kernel_encoder is not None and wkv_kernel_encoder.context_size == context_size:
        return

    if find_spec("ninja") is None:
        raise ImportError(
            "Ninja package was not found. WKV kernel module can't be loaded "
            "for training. Please, 'pip install ninja' in your environment."
        )

    if not torch.cuda.is_available():
        raise ImportError(
            "CUDA is currently a requirement for WKV kernel loading. "
            "Please set your devices properly and launch again."
        )

    kernel_folder = Path(__file__).resolve().parent / "cuda_encoder"
    kernel_files = [kernel_folder / f for f in ["wkv_op.cpp", "wkv_cuda.cu"]]

    kernel_cflags = [
        "-res-usage",
        "--maxrregcount 60",
        "--use_fast_math",
        "-O3",
        "-Xptxas -O3",
        f"-DTmax={context_size}",
    ]
    wkv_kernel_encoder = load(
        name=f"encoder_wkv_{context_size}",
        sources=kernel_files,
        verbose=True,
        extra_cuda_cflags=kernel_cflags,
    )
    wkv_kernel_encoder.context_size = context_size

def load_decoder_wkv_kernel(context_size: int) -> None:
    """Load WKV CUDA kernel.

    Args:
        context_size: Context size.

    """
    from torch.utils.cpp_extension import load

    global wkv_kernel_decoder

    if wkv_kernel_decoder is not None and wkv_kernel_decoder.context_size == context_size:
        return

    if find_spec("ninja") is None:
        raise ImportError(
            "Ninja package was not found. WKV kernel module can't be loaded "
            "for training. Please, 'pip install ninja' in your environment."
        )

    if not torch.cuda.is_available():
        raise ImportError(
            "CUDA is currently a requirement for WKV kernel loading. "
            "Please set your devices properly and launch again."
        )

    kernel_folder = Path(__file__).resolve().parent / "cuda_decoder"
    kernel_files = [kernel_folder / f for f in ["wkv_op.cpp", "wkv_cuda.cu"]]

    kernel_cflags = [
        "-res-usage",
        "--maxrregcount 60",
        "--use_fast_math",
        "-O3",
        "-Xptxas -O3",
        f"-DTmax={context_size}",
    ]
    wkv_kernel_decoder = load(
        name=f"decoder_wkv_{context_size}",
        sources=kernel_files,
        verbose=True,
        extra_cuda_cflags=kernel_cflags,
    )
    wkv_kernel_decoder.context_size = context_size

class SelfAttention(torch.nn.Module):
    """SelfAttention module definition.

    Args:
        size: Input/Output size.
        attention_size: Attention hidden size.
        context_size: Context size for WKV kernel.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.

    """

    def __init__(
        self,
        size: int,
        attention_size: int,
        block_id: int,
        dropout_rate: float,
        num_blocks: int,
    ) -> None:
        """Construct a SelfAttention object."""
        super().__init__()
        self.time_shift = torch.nn.ZeroPad2d((0, 0, 1, -1))

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

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

        self.proj_key = torch.nn.Linear(size, attention_size, bias=True)
        self.proj_value = torch.nn.Linear(size, attention_size, bias=True)
        self.proj_receptance = torch.nn.Linear(size, attention_size, bias=True)

        self.proj_output = torch.nn.Linear(attention_size, size, bias=True)

        self.block_id = block_id

        self.reset_parameters(size, attention_size, block_id, num_blocks)
        self.dropout = torch.nn.Dropout(p=dropout_rate)

    def reset_parameters(
        self, size: int, attention_size: int, block_id: int, num_blocks: int
    ) -> None:
        """Reset module parameters.

        Args:
            size: Block size.
            attention_size: Attention hidden size.
            block_id: Block index.
            num_blocks: Number of blocks in the architecture.

        """
        ratio_0_to_1 = block_id / (num_blocks - 1)
        ratio_1_to_almost0 = 1.0 - (block_id / num_blocks)

        time_weight = torch.ones(1, 1, size)

        for i in range(size):
            time_weight[0, 0, i] = i / size

        decay_speed = [
            -5 + 8 * (h / (attention_size - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
            for h in range(attention_size)
        ]
        decay_speed = torch.tensor(
            decay_speed, dtype=self.time_decay.dtype, device=self.time_decay.device
        )

        zigzag = (
            torch.tensor(
                [(i + 1) % 3 - 1 for i in range(attention_size)],
                dtype=self.time_first.dtype,
                device=self.time_first.device,
            )
            * 0.5
        )

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

            self.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
            self.time_mix_value.data = (
                torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
            )
            self.time_mix_receptance.data = torch.pow(
                time_weight, 0.5 * ratio_1_to_almost0
            )

    @torch.no_grad()
    def wkv_linear_attention(
        self,
        time_decay: torch.Tensor,
        time_first: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        state: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
        """Compute WKV with state (i.e.: for inference).

        Args:
            time_decay: Channel-wise time decay vector. (D_att)
            time_first: Channel-wise time first vector. (D_att)
            key: Key tensor. (B, 1, D_att)
            value: Value tensor. (B, 1, D_att)
            state: Decoder hidden states. [3 x (B, D_att)]

        Returns:
            output: Weighted Key-Value. (B, 1, D_att)
            state: Decoder hidden states. [3 x (B, 1, D_att)]

        """
        num_state, den_state, max_state = state

        max_for_output = torch.maximum(max_state, (time_first + key))

        e1 = torch.exp(max_state - max_for_output)
        e2 = torch.exp((time_first + key) - max_for_output)

        numerator = e1 * num_state + e2 * value
        denominator = e1 * den_state + e2

        max_for_state = torch.maximum(key, (max_state + time_decay))

        e1 = torch.exp((max_state + time_decay) - max_for_state)
        e2 = torch.exp(key - max_for_state)

        wkv = numerator / denominator

        state = [e1 * num_state + e2 * value, e1 * den_state + e2, max_for_state]

        return wkv, state


class DecoderSelfAttention(SelfAttention):
    """SelfAttention module definition.

    Args:
        size: Input/Output size.
        attention_size: Attention hidden size.
        context_size: Context size for WKV kernel.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.

    """

    def __init__(
        self,
        size: int,
        attention_size: int,
        context_size: int,
        block_id: int,
        dropout_rate: float,
        num_blocks: int,
    ) -> None:
        """Construct a SelfAttention object."""
        super().__init__(
            size,
            attention_size,
            block_id,
            dropout_rate,
            num_blocks
        )
        load_decoder_wkv_kernel(context_size)

    def forward(
        self,
        x: torch.Tensor,
        state: Optional[List[torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, Optional[List[torch.Tensor]]]:
        """Compute time mixing.

        Args:
            x: SelfAttention input sequences. (B, U, size)
            state: Decoder hidden states. [5 x (B, 1, D_att, N)]

        Returns:
            x: SelfAttention output sequences. (B, U, size)

        """
        shifted_x = (
            self.time_shift(x) if state is None else state[1][..., self.block_id]
        )

        key = x * self.time_mix_key + shifted_x * (1 - self.time_mix_key)
        value = x * self.time_mix_value + shifted_x * (1 - self.time_mix_value)
        receptance = x * self.time_mix_receptance + shifted_x * (
            1 - self.time_mix_receptance
        )

        key = self.proj_key(key)
        value = self.proj_value(value)
        receptance = torch.sigmoid(self.proj_receptance(receptance))

        if state is not None:
            state[1][..., self.block_id] = x

            wkv, att_state = self.wkv_linear_attention(
                self.time_decay,
                self.time_first,
                key,
                value,
                tuple(s[..., self.block_id] for s in state[2:]),
            )

            state[2][..., self.block_id] = att_state[0]
            state[3][..., self.block_id] = att_state[1]
            state[4][..., self.block_id] = att_state[2]
        else:
            wkv = WKVLinearAttentionDecoder.apply(self.time_decay, self.time_first, key, value)

        wkv = self.dropout(wkv)
        x = self.proj_output(receptance * wkv)

        return x, state

class EncoderSelfAttention(SelfAttention):
    """SelfAttention module definition.

    Args:
        size: Input/Output size.
        attention_size: Attention hidden size.
        context_size: Context size for WKV kernel.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.

    """

    def __init__(
        self,
        size: int,
        attention_size: int,
        context_size: int,
        block_id: int,
        dropout_rate: float,
        num_blocks: int,
    ) -> None:
        """Construct a SelfAttention object."""
        super().__init__(
            size,
            attention_size,
            block_id,
            dropout_rate,
            num_blocks
        )
        load_encoder_wkv_kernel(context_size)

    def forward(
        self,
        x: torch.Tensor,
        state: Optional[List[torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, Optional[List[torch.Tensor]]]:
        """Compute time mixing.

        Args:
            x: SelfAttention input sequences. (B, U, size)
            state: Decoder hidden states. [5 x (B, 1, D_att, N)]

        Returns:
            x: SelfAttention output sequences. (B, U, size)

        """
        shifted_x = (
            self.time_shift(x) if state is None else state[1][..., self.block_id]
        )

        key = x * self.time_mix_key + shifted_x * (1 - self.time_mix_key)
        value = x * self.time_mix_value + shifted_x * (1 - self.time_mix_value)
        receptance = x * self.time_mix_receptance + shifted_x * (
            1 - self.time_mix_receptance
        )

        key = self.proj_key(key)
        value = self.proj_value(value)
        receptance = torch.sigmoid(self.proj_receptance(receptance))

        if state is not None:
            state[1][..., self.block_id] = x

            wkv, att_state = self.wkv_linear_attention(
                self.time_decay,
                self.time_first,
                key,
                value,
                tuple(s[..., self.block_id] for s in state[2:]),
            )

            state[2][..., self.block_id] = att_state[0]
            state[3][..., self.block_id] = att_state[1]
            state[4][..., self.block_id] = att_state[2]
        else:
            wkv = WKVLinearAttentionEncoder.apply(self.time_decay, self.time_first, key, value)

        wkv = self.dropout(wkv)
        x = self.proj_output(receptance * wkv)

        return x, state


 funasr/modules/rwkv_feed_forward.py

New file
@@ -0,0 +1,97 @@
"""Feed-forward (channel mixing) module for RWKV block.

Based/Modified from https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v4/src/model.py

Some variables are renamed according to https://github.com/huggingface/transformers/blob/main/src/transformers/models/rwkv/modeling_rwkv.py.

"""  # noqa

from typing import List, Optional, Tuple

import torch


class FeedForward(torch.nn.Module):
    """FeedForward module definition.

    Args:
        size: Input/Output size.
        hidden_size: Hidden size.
        block_id: Block index.
        num_blocks: Number of blocks in the architecture.

    """

    def __init__(
        self, size: int, hidden_size: int, block_id: int, dropout_rate: float, num_blocks: int
    ) -> None:
        """Construct a FeedForward object."""
        super().__init__()

        self.time_shift = torch.nn.ZeroPad2d((0, 0, 1, -1))

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

        self.proj_key = torch.nn.Linear(size, hidden_size, bias=True)
        self.proj_value = torch.nn.Linear(hidden_size, size, bias=True)
        self.proj_receptance = torch.nn.Linear(size, size, bias=True)

        self.block_id = block_id

        self.reset_parameters(size, block_id, num_blocks)
        self.dropout = torch.nn.Dropout(p=dropout_rate)

    def reset_parameters(self, size: int, block_id: int, num_blocks: int) -> None:
        """Reset module parameters.

        Args:
            size: Block size.
            block_id: Block index.
            num_blocks: Number of blocks in the architecture.

        """
        ratio_1_to_almost0 = 1.0 - (block_id / num_blocks)

        time_weight = torch.ones(1, 1, size)

        for i in range(size):
            time_weight[0, 0, i] = i / size

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

    def forward(
        self, x: torch.Tensor, state: Optional[List[torch.Tensor]] = None
    ) -> Tuple[torch.Tensor, Optional[List[torch.Tensor]]]:
        """Compute channel mixing.

        Args:
            x: FeedForward input sequences. (B, U, size)
            state: Decoder hidden state. [5 x (B, 1, size, N)]

        Returns:
            x: FeedForward output sequences. (B, U, size)
            state: Decoder hidden state. [5 x (B, 1, size, N)]

        """
        shifted_x = (
            self.time_shift(x) if state is None else state[0][..., self.block_id]
        )

        key = x * self.time_mix_key + shifted_x * (1 - self.time_mix_key)
        receptance = x * self.time_mix_receptance + shifted_x * (
            1 - self.time_mix_receptance
        )

        key = torch.square(torch.relu(self.proj_key(key)))
        value = self.proj_value(self.dropout(key))
        receptance = torch.sigmoid(self.proj_receptance(receptance))

        if state is not None:
            state[0][..., self.block_id] = x

        x = receptance * value

        return x, state

 funasr/modules/rwkv_subsampling.py

New file
@@ -0,0 +1,190 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

# Copyright 2019 Shigeki Karita
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Subsampling layer definition."""
import numpy as np
import torch
import torch.nn.functional as F
from funasr.modules.embedding import PositionalEncoding
import logging
from funasr.modules.streaming_utils.utils import sequence_mask
from funasr.modules.nets_utils import sub_factor_to_params, pad_to_len
from typing import Optional, Tuple, Union
import math

class TooShortUttError(Exception):
    """Raised when the utt is too short for subsampling.

    Args:
        message (str): Message for error catch
        actual_size (int): the short size that cannot pass the subsampling
        limit (int): the limit size for subsampling

    """

    def __init__(self, message, actual_size, limit):
        """Construct a TooShortUttError for error handler."""
        super().__init__(message)
        self.actual_size = actual_size
        self.limit = limit


def check_short_utt(ins, size):
    """Check if the utterance is too short for subsampling."""
    if isinstance(ins, Conv2dSubsampling2) and size < 3:
        return True, 3
    if isinstance(ins, Conv2dSubsampling) and size < 7:
        return True, 7
    if isinstance(ins, Conv2dSubsampling6) and size < 11:
        return True, 11
    if isinstance(ins, Conv2dSubsampling8) and size < 15:
        return True, 15
    return False, -1


class RWKVConvInput(torch.nn.Module):
    """Streaming ConvInput module definition.
    Args:
        input_size: Input size.
        conv_size: Convolution size.
        subsampling_factor: Subsampling factor.
        output_size: Block output dimension.
    """

    def __init__(
        self,
        input_size: int,
        conv_size: Union[int, Tuple],
        subsampling_factor: int = 4,
        conv_kernel_size: int = 3,
        output_size: Optional[int] = None,
    ) -> None:
        """Construct a ConvInput object."""
        super().__init__()
        if subsampling_factor == 1:
            conv_size1, conv_size2, conv_size3 = conv_size

            self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(1, conv_size1, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size1, conv_size1, conv_kernel_size, stride=[1, 2], padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size1, conv_size2, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size2, conv_size2, conv_kernel_size, stride=[1, 2], padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size2, conv_size3, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size3, conv_size3, conv_kernel_size, stride=[1, 2], padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
            )

            output_proj = conv_size3 * ((input_size // 2) // 2)

            self.subsampling_factor = 1

            self.stride_1 = 1

            self.create_new_mask = self.create_new_vgg_mask

        else:
            conv_size1, conv_size2, conv_size3 = conv_size

            kernel_1 = int(subsampling_factor / 2)

            self.conv = torch.nn.Sequential(
                    torch.nn.Conv2d(1, conv_size1, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size1, conv_size1, conv_kernel_size, stride=[kernel_1, 2], padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size1, conv_size2, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size2, conv_size2, conv_kernel_size, stride=[2, 2], padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size2, conv_size3, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
                    torch.nn.Conv2d(conv_size3, conv_size3, conv_kernel_size, stride=1, padding=(conv_kernel_size-1)//2),
                    torch.nn.ReLU(),
            )

            output_proj = conv_size3 * ((input_size // 2) // 2)

            self.subsampling_factor = subsampling_factor

            self.create_new_mask = self.create_new_vgg_mask

            self.stride_1 = kernel_1

        self.min_frame_length = 7

        if output_size is not None:
            self.output = torch.nn.Linear(output_proj, output_size)
            self.output_size = output_size
        else:
            self.output = None
            self.output_size = output_proj

    def forward(
        self, x: torch.Tensor, mask: Optional[torch.Tensor], chunk_size: Optional[torch.Tensor]
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Encode input sequences.
        Args:
            x: ConvInput input sequences. (B, T, D_feats)
            mask: Mask of input sequences. (B, 1, T)
        Returns:
            x: ConvInput output sequences. (B, sub(T), D_out)
            mask: Mask of output sequences. (B, 1, sub(T))
        """
        if mask is not None:
            mask = self.create_new_mask(mask)
            olens = max(mask.eq(0).sum(1))

        b, t, f = x.size()
        x = x.unsqueeze(1) # (b. 1. t. f)

        if chunk_size is not None:
            max_input_length = int(
                chunk_size * self.subsampling_factor * (math.ceil(float(t) / (chunk_size * self.subsampling_factor) ))
            )
            x = map(lambda inputs: pad_to_len(inputs, max_input_length, 1), x)
            x = list(x)
            x = torch.stack(x, dim=0)
            N_chunks = max_input_length // ( chunk_size * self.subsampling_factor)
            x = x.view(b * N_chunks, 1, chunk_size * self.subsampling_factor, f)

        x = self.conv(x)

        _, c, _, f = x.size()
        if chunk_size is not None:
            x = x.transpose(1, 2).contiguous().view(b, -1, c * f)[:,:olens,:]
        else:
            x = x.transpose(1, 2).contiguous().view(b, -1, c * f)

        if self.output is not None:
            x = self.output(x)

        return x, mask[:,:olens][:,:x.size(1)]

    def create_new_vgg_mask(self, mask: torch.Tensor) -> torch.Tensor:
        """Create a new mask for VGG output sequences.
        Args:
            mask: Mask of input sequences. (B, T)
        Returns:
            mask: Mask of output sequences. (B, sub(T))
        """
        if self.subsampling_factor > 1:
            return mask[:, ::2][:, ::self.stride_1]
        else:
            return mask 

    def get_size_before_subsampling(self, size: int) -> int:
        """Return the original size before subsampling for a given size.
        Args:
            size: Number of frames after subsampling.
        Returns:
            : Number of frames before subsampling.
        """
        return size * self.subsampling_factor