xinference 1.9.1__py3-none-any.whl → 1.10.1__py3-none-any.whl

xinference/thirdparty/indextts/utils/maskgct/models/codec/speechtokenizer/modules/quantization/distrib.py

@@ -0,0 +1,135 @@
+# Copyright (c) 2023 Amphion.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+# This source file is copied from https://github.com/facebookresearch/encodec
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Torch distributed utilities."""
+
+import typing as tp
+
+import torch
+
+
+def rank():
+    if torch.distributed.is_initialized():
+        return torch.distributed.get_rank()
+    else:
+        return 0
+
+
+def world_size():
+    if torch.distributed.is_initialized():
+        return torch.distributed.get_world_size()
+    else:
+        return 1
+
+
+def is_distributed():
+    return world_size() > 1
+
+
+def all_reduce(tensor: torch.Tensor, op=torch.distributed.ReduceOp.SUM):
+    if is_distributed():
+        return torch.distributed.all_reduce(tensor, op)
+
+
+def _is_complex_or_float(tensor):
+    return torch.is_floating_point(tensor) or torch.is_complex(tensor)
+
+
+def _check_number_of_params(params: tp.List[torch.Tensor]):
+    # utility function to check that the number of params in all workers is the same,
+    # and thus avoid a deadlock with distributed all reduce.
+    if not is_distributed() or not params:
+        return
+    # print('params[0].device ', params[0].device)
+    tensor = torch.tensor([len(params)], device=params[0].device, dtype=torch.long)
+    all_reduce(tensor)
+    if tensor.item() != len(params) * world_size():
+        # If not all the workers have the same number, for at least one of them,
+        # this inequality will be verified.
+        raise RuntimeError(
+            f"Mismatch in number of params: ours is {len(params)}, "
+            "at least one worker has a different one."
+        )
+
+
+def broadcast_tensors(tensors: tp.Iterable[torch.Tensor], src: int = 0):
+    """Broadcast the tensors from the given parameters to all workers.
+    This can be used to ensure that all workers have the same model to start with.
+    """
+    if not is_distributed():
+        return
+    tensors = [tensor for tensor in tensors if _is_complex_or_float(tensor)]
+    _check_number_of_params(tensors)
+    handles = []
+    for tensor in tensors:
+        # src = int(rank()) # added code
+        handle = torch.distributed.broadcast(tensor.data, src=src, async_op=True)
+        handles.append(handle)
+    for handle in handles:
+        handle.wait()
+
+
+def sync_buffer(buffers, average=True):
+    """
+    Sync grad for buffers. If average is False, broadcast instead of averaging.
+    """
+    if not is_distributed():
+        return
+    handles = []
+    for buffer in buffers:
+        if torch.is_floating_point(buffer.data):
+            if average:
+                handle = torch.distributed.all_reduce(
+                    buffer.data, op=torch.distributed.ReduceOp.SUM, async_op=True
+                )
+            else:
+                handle = torch.distributed.broadcast(buffer.data, src=0, async_op=True)
+            handles.append((buffer, handle))
+    for buffer, handle in handles:
+        handle.wait()
+        if average:
+            buffer.data /= world_size
+
+
+def sync_grad(params):
+    """
+    Simpler alternative to DistributedDataParallel, that doesn't rely
+    on any black magic. For simple models it can also be as fast.
+    Just call this on your model parameters after the call to backward!
+    """
+    if not is_distributed():
+        return
+    handles = []
+    for p in params:
+        if p.grad is not None:
+            handle = torch.distributed.all_reduce(
+                p.grad.data, op=torch.distributed.ReduceOp.SUM, async_op=True
+            )
+            handles.append((p, handle))
+    for p, handle in handles:
+        handle.wait()
+        p.grad.data /= world_size()
+
+
+def average_metrics(metrics: tp.Dict[str, float], count=1.0):
+    """Average a dictionary of metrics across all workers, using the optional
+    `count` as unormalized weight.
+    """
+    if not is_distributed():
+        return metrics
+    keys, values = zip(*metrics.items())
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    tensor = torch.tensor(list(values) + [1], device=device, dtype=torch.float32)
+    tensor *= count
+    all_reduce(tensor)
+    averaged = (tensor[:-1] / tensor[-1]).cpu().tolist()
+    return dict(zip(keys, averaged))

xinference/thirdparty/indextts/utils/maskgct/models/codec/speechtokenizer/modules/quantization/vq.py

@@ -0,0 +1,125 @@
+# Copyright (c) 2023 Amphion.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+# This source file is copied from https://github.com/facebookresearch/encodec
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Residual vector quantizer implementation."""
+
+from dataclasses import dataclass, field
+import math
+import typing as tp
+
+import torch
+from torch import nn
+
+from .core_vq import ResidualVectorQuantization
+
+
+@dataclass
+class QuantizedResult:
+    quantized: torch.Tensor
+    codes: torch.Tensor
+    bandwidth: torch.Tensor  # bandwidth in kb/s used, per batch item.
+    penalty: tp.Optional[torch.Tensor] = None
+    metrics: dict = field(default_factory=dict)
+
+
+class ResidualVectorQuantizer(nn.Module):
+    """Residual Vector Quantizer.
+    Args:
+        dimension (int): Dimension of the codebooks.
+        n_q (int): Number of residual vector quantizers used.
+        bins (int): Codebook size.
+        decay (float): Decay for exponential moving average over the codebooks.
+        kmeans_init (bool): Whether to use kmeans to initialize the codebooks.
+        kmeans_iters (int): Number of iterations used for kmeans initialization.
+        threshold_ema_dead_code (int): Threshold for dead code expiration. Replace any codes
+            that have an exponential moving average cluster size less than the specified threshold with
+            randomly selected vector from the current batch.
+    """
+
+    def __init__(
+        self,
+        dimension: int = 256,
+        n_q: int = 8,
+        bins: int = 1024,
+        decay: float = 0.99,
+        kmeans_init: bool = True,
+        kmeans_iters: int = 50,
+        threshold_ema_dead_code: int = 2,
+    ):
+        super().__init__()
+        self.n_q = n_q
+        self.dimension = dimension
+        self.bins = bins
+        self.decay = decay
+        self.kmeans_init = kmeans_init
+        self.kmeans_iters = kmeans_iters
+        self.threshold_ema_dead_code = threshold_ema_dead_code
+        self.vq = ResidualVectorQuantization(
+            dim=self.dimension,
+            codebook_size=self.bins,
+            num_quantizers=self.n_q,
+            decay=self.decay,
+            kmeans_init=self.kmeans_init,
+            kmeans_iters=self.kmeans_iters,
+            threshold_ema_dead_code=self.threshold_ema_dead_code,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        n_q: tp.Optional[int] = None,
+        layers: tp.Optional[list] = None,
+    ) -> QuantizedResult:
+        """Residual vector quantization on the given input tensor.
+        Args:
+            x (torch.Tensor): Input tensor.
+            n_q (int): Number of quantizer used to quantize. Default: All quantizers.
+            layers (list): Layer that need to return quantized. Defalt: None.
+        Returns:
+            QuantizedResult:
+                The quantized (or approximately quantized) representation with
+                the associated numbert quantizers and layer quantized required to return.
+        """
+        n_q = n_q if n_q else self.n_q
+        if layers and max(layers) >= n_q:
+            raise ValueError(
+                f"Last layer index in layers: A {max(layers)}. Number of quantizers in RVQ: B {self.n_q}. A must less than B."
+            )
+        quantized, codes, commit_loss, quantized_list = self.vq(
+            x, n_q=n_q, layers=layers
+        )
+        return quantized, codes, torch.mean(commit_loss), quantized_list
+
+    def encode(
+        self, x: torch.Tensor, n_q: tp.Optional[int] = None, st: tp.Optional[int] = None
+    ) -> torch.Tensor:
+        """Encode a given input tensor with the specified sample rate at the given bandwidth.
+        The RVQ encode method sets the appropriate number of quantizer to use
+        and returns indices for each quantizer.
+        Args:
+            x (torch.Tensor): Input tensor.
+            n_q (int): Number of quantizer used to quantize. Default: All quantizers.
+            st (int): Start to encode input from which layers. Default: 0.
+        """
+        n_q = n_q if n_q else self.n_q
+        st = st or 0
+        codes = self.vq.encode(x, n_q=n_q, st=st)
+        return codes
+
+    def decode(self, codes: torch.Tensor, st: int = 0) -> torch.Tensor:
+        """Decode the given codes to the quantized representation.
+        Args:
+            codes (torch.Tensor): Input indices for each quantizer.
+            st (int): Start to decode input codes from which layers. Default: 0.
+        """
+        quantized = self.vq.decode(codes, st=st)
+        return quantized

xinference/thirdparty/indextts/utils/maskgct/models/codec/speechtokenizer/modules/seanet.py

@@ -0,0 +1,414 @@
+# Copyright (c) 2023 Amphion.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+# This source file is copied from https://github.com/facebookresearch/encodec
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Encodec SEANet-based encoder and decoder implementation."""
+
+import typing as tp
+
+import numpy as np
+import torch.nn as nn
+import torch
+
+from . import SConv1d, SConvTranspose1d, SLSTM
+
+
+@torch.jit.script
+def snake(x, alpha):
+    shape = x.shape
+    x = x.reshape(shape[0], shape[1], -1)
+    x = x + (alpha + 1e-9).reciprocal() * torch.sin(alpha * x).pow(2)
+    x = x.reshape(shape)
+    return x
+
+
+class Snake1d(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.alpha = nn.Parameter(torch.ones(1, channels, 1))
+
+    def forward(self, x):
+        return snake(x, self.alpha)
+
+
+class SEANetResnetBlock(nn.Module):
+    """Residual block from SEANet model.
+    Args:
+        dim (int): Dimension of the input/output
+        kernel_sizes (list): List of kernel sizes for the convolutions.
+        dilations (list): List of dilations for the convolutions.
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function
+        norm (str): Normalization method.
+        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
+        causal (bool): Whether to use fully causal convolution.
+        pad_mode (str): Padding mode for the convolutions.
+        compress (int): Reduced dimensionality in residual branches (from Demucs v3)
+        true_skip (bool): Whether to use true skip connection or a simple convolution as the skip connection.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        kernel_sizes: tp.List[int] = [3, 1],
+        dilations: tp.List[int] = [1, 1],
+        activation: str = "ELU",
+        activation_params: dict = {"alpha": 1.0},
+        norm: str = "weight_norm",
+        norm_params: tp.Dict[str, tp.Any] = {},
+        causal: bool = False,
+        pad_mode: str = "reflect",
+        compress: int = 2,
+        true_skip: bool = True,
+    ):
+        super().__init__()
+        assert len(kernel_sizes) == len(
+            dilations
+        ), "Number of kernel sizes should match number of dilations"
+        act = getattr(nn, activation) if activation != "Snake" else Snake1d
+        hidden = dim // compress
+        block = []
+        for i, (kernel_size, dilation) in enumerate(zip(kernel_sizes, dilations)):
+            in_chs = dim if i == 0 else hidden
+            out_chs = dim if i == len(kernel_sizes) - 1 else hidden
+            block += [
+                act(**activation_params) if activation != "Snake" else act(in_chs),
+                SConv1d(
+                    in_chs,
+                    out_chs,
+                    kernel_size=kernel_size,
+                    dilation=dilation,
+                    norm=norm,
+                    norm_kwargs=norm_params,
+                    causal=causal,
+                    pad_mode=pad_mode,
+                ),
+            ]
+        self.block = nn.Sequential(*block)
+        self.shortcut: nn.Module
+        if true_skip:
+            self.shortcut = nn.Identity()
+        else:
+            self.shortcut = SConv1d(
+                dim,
+                dim,
+                kernel_size=1,
+                norm=norm,
+                norm_kwargs=norm_params,
+                causal=causal,
+                pad_mode=pad_mode,
+            )
+
+    def forward(self, x):
+        return self.shortcut(x) + self.block(x)
+
+
+class SEANetEncoder(nn.Module):
+    """SEANet encoder.
+    Args:
+        channels (int): Audio channels.
+        dimension (int): Intermediate representation dimension.
+        n_filters (int): Base width for the model.
+        n_residual_layers (int): nb of residual layers.
+        ratios (Sequence[int]): kernel size and stride ratios. The encoder uses downsampling ratios instead of
+            upsampling ratios, hence it will use the ratios in the reverse order to the ones specified here
+            that must match the decoder order
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function
+        norm (str): Normalization method.
+        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
+        kernel_size (int): Kernel size for the initial convolution.
+        last_kernel_size (int): Kernel size for the initial convolution.
+        residual_kernel_size (int): Kernel size for the residual layers.
+        dilation_base (int): How much to increase the dilation with each layer.
+        causal (bool): Whether to use fully causal convolution.
+        pad_mode (str): Padding mode for the convolutions.
+        true_skip (bool): Whether to use true skip connection or a simple
+            (streamable) convolution as the skip connection in the residual network blocks.
+        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
+        lstm (int): Number of LSTM layers at the end of the encoder.
+    """
+
+    def __init__(
+        self,
+        channels: int = 1,
+        dimension: int = 128,
+        n_filters: int = 32,
+        n_residual_layers: int = 1,
+        ratios: tp.List[int] = [8, 5, 4, 2],
+        activation: str = "ELU",
+        activation_params: dict = {"alpha": 1.0},
+        norm: str = "weight_norm",
+        norm_params: tp.Dict[str, tp.Any] = {},
+        kernel_size: int = 7,
+        last_kernel_size: int = 7,
+        residual_kernel_size: int = 3,
+        dilation_base: int = 2,
+        causal: bool = False,
+        pad_mode: str = "reflect",
+        true_skip: bool = False,
+        compress: int = 2,
+        lstm: int = 2,
+        bidirectional: bool = False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.dimension = dimension
+        self.n_filters = n_filters
+        self.ratios = list(reversed(ratios))
+        del ratios
+        self.n_residual_layers = n_residual_layers
+        self.hop_length = np.prod(self.ratios)  # 计算乘积
+
+        act = getattr(nn, activation) if activation != "Snake" else Snake1d
+        mult = 1
+        model: tp.List[nn.Module] = [
+            SConv1d(
+                channels,
+                mult * n_filters,
+                kernel_size,
+                norm=norm,
+                norm_kwargs=norm_params,
+                causal=causal,
+                pad_mode=pad_mode,
+            )
+        ]
+        # Downsample to raw audio scale
+        for i, ratio in enumerate(self.ratios):
+            # Add residual layers
+            for j in range(n_residual_layers):
+                model += [
+                    SEANetResnetBlock(
+                        mult * n_filters,
+                        kernel_sizes=[residual_kernel_size, 1],
+                        dilations=[dilation_base**j, 1],
+                        norm=norm,
+                        norm_params=norm_params,
+                        activation=activation,
+                        activation_params=activation_params,
+                        causal=causal,
+                        pad_mode=pad_mode,
+                        compress=compress,
+                        true_skip=true_skip,
+                    )
+                ]
+
+            # Add downsampling layers
+            model += [
+                (
+                    act(**activation_params)
+                    if activation != "Snake"
+                    else act(mult * n_filters)
+                ),
+                SConv1d(
+                    mult * n_filters,
+                    mult * n_filters * 2,
+                    kernel_size=ratio * 2,
+                    stride=ratio,
+                    norm=norm,
+                    norm_kwargs=norm_params,
+                    causal=causal,
+                    pad_mode=pad_mode,
+                ),
+            ]
+            mult *= 2
+
+        if lstm:
+            model += [
+                SLSTM(mult * n_filters, num_layers=lstm, bidirectional=bidirectional)
+            ]
+
+        mult = mult * 2 if bidirectional else mult
+        model += [
+            (
+                act(**activation_params)
+                if activation != "Snake"
+                else act(mult * n_filters)
+            ),
+            SConv1d(
+                mult * n_filters,
+                dimension,
+                last_kernel_size,
+                norm=norm,
+                norm_kwargs=norm_params,
+                causal=causal,
+                pad_mode=pad_mode,
+            ),
+        ]
+
+        self.model = nn.Sequential(*model)
+
+    def forward(self, x):
+        return self.model(x)
+
+
+class SEANetDecoder(nn.Module):
+    """SEANet decoder.
+    Args:
+        channels (int): Audio channels.
+        dimension (int): Intermediate representation dimension.
+        n_filters (int): Base width for the model.
+        n_residual_layers (int): nb of residual layers.
+        ratios (Sequence[int]): kernel size and stride ratios
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function
+        final_activation (str): Final activation function after all convolutions.
+        final_activation_params (dict): Parameters to provide to the activation function
+        norm (str): Normalization method.
+        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
+        kernel_size (int): Kernel size for the initial convolution.
+        last_kernel_size (int): Kernel size for the initial convolution.
+        residual_kernel_size (int): Kernel size for the residual layers.
+        dilation_base (int): How much to increase the dilation with each layer.
+        causal (bool): Whether to use fully causal convolution.
+        pad_mode (str): Padding mode for the convolutions.
+        true_skip (bool): Whether to use true skip connection or a simple
+            (streamable) convolution as the skip connection in the residual network blocks.
+        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
+        lstm (int): Number of LSTM layers at the end of the encoder.
+        trim_right_ratio (float): Ratio for trimming at the right of the transposed convolution under the causal setup.
+            If equal to 1.0, it means that all the trimming is done at the right.
+    """
+
+    def __init__(
+        self,
+        channels: int = 1,
+        dimension: int = 128,
+        n_filters: int = 32,
+        n_residual_layers: int = 1,
+        ratios: tp.List[int] = [8, 5, 4, 2],
+        activation: str = "ELU",
+        activation_params: dict = {"alpha": 1.0},
+        final_activation: tp.Optional[str] = None,
+        final_activation_params: tp.Optional[dict] = None,
+        norm: str = "weight_norm",
+        norm_params: tp.Dict[str, tp.Any] = {},
+        kernel_size: int = 7,
+        last_kernel_size: int = 7,
+        residual_kernel_size: int = 3,
+        dilation_base: int = 2,
+        causal: bool = False,
+        pad_mode: str = "reflect",
+        true_skip: bool = False,
+        compress: int = 2,
+        lstm: int = 2,
+        trim_right_ratio: float = 1.0,
+        bidirectional: bool = False,
+    ):
+        super().__init__()
+        self.dimension = dimension
+        self.channels = channels
+        self.n_filters = n_filters
+        self.ratios = ratios
+        del ratios
+        self.n_residual_layers = n_residual_layers
+        self.hop_length = np.prod(self.ratios)
+
+        act = getattr(nn, activation) if activation != "Snake" else Snake1d
+        mult = int(2 ** len(self.ratios))
+        model: tp.List[nn.Module] = [
+            SConv1d(
+                dimension,
+                mult * n_filters,
+                kernel_size,
+                norm=norm,
+                norm_kwargs=norm_params,
+                causal=causal,
+                pad_mode=pad_mode,
+            )
+        ]
+
+        if lstm:
+            model += [
+                SLSTM(mult * n_filters, num_layers=lstm, bidirectional=bidirectional)
+            ]
+
+        # Upsample to raw audio scale
+        for i, ratio in enumerate(self.ratios):
+            # Add upsampling layers
+            model += [
+                (
+                    act(**activation_params)
+                    if activation != "Snake"
+                    else act(mult * n_filters)
+                ),
+                SConvTranspose1d(
+                    mult * n_filters,
+                    mult * n_filters // 2,
+                    kernel_size=ratio * 2,
+                    stride=ratio,
+                    norm=norm,
+                    norm_kwargs=norm_params,
+                    causal=causal,
+                    trim_right_ratio=trim_right_ratio,
+                ),
+            ]
+            # Add residual layers
+            for j in range(n_residual_layers):
+                model += [
+                    SEANetResnetBlock(
+                        mult * n_filters // 2,
+                        kernel_sizes=[residual_kernel_size, 1],
+                        dilations=[dilation_base**j, 1],
+                        activation=activation,
+                        activation_params=activation_params,
+                        norm=norm,
+                        norm_params=norm_params,
+                        causal=causal,
+                        pad_mode=pad_mode,
+                        compress=compress,
+                        true_skip=true_skip,
+                    )
+                ]
+
+            mult //= 2
+
+        # Add final layers
+        model += [
+            act(**activation_params) if activation != "Snake" else act(n_filters),
+            SConv1d(
+                n_filters,
+                channels,
+                last_kernel_size,
+                norm=norm,
+                norm_kwargs=norm_params,
+                causal=causal,
+                pad_mode=pad_mode,
+            ),
+        ]
+        # Add optional final activation to decoder (eg. tanh)
+        if final_activation is not None:
+            final_act = getattr(nn, final_activation)
+            final_activation_params = final_activation_params or {}
+            model += [final_act(**final_activation_params)]
+        self.model = nn.Sequential(*model)
+
+    def forward(self, z):
+        y = self.model(z)
+        return y
+
+
+def test():
+    import torch
+
+    encoder = SEANetEncoder()
+    decoder = SEANetDecoder()
+    x = torch.randn(1, 1, 24000)
+    z = encoder(x)
+    print("z ", z.shape)
+    assert 1 == 2
+    assert list(z.shape) == [1, 128, 75], z.shape
+    y = decoder(z)
+    assert y.shape == x.shape, (x.shape, y.shape)
+
+
+if __name__ == "__main__":
+    test()