aphrodite-engine/aphrodite/modeling/megatron/tensor_parallel/layers.py

# Copyright 2023 The PygmalionAI team.
# Copyright 2023 The vLLM team.
# Adapted from https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/tensor_parallel/layers.py
# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.

# Parts of the code here are adapted from PyTorch
# repo: https://github.com/pytorch/pytorch


import torch
import torch.nn.functional as F
import torch.nn.init as init
from torch.nn.parameter import Parameter

from aphrodite.modeling.megatron.parallel_state import (
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
)
from .mappings import (
    gather_from_tensor_model_parallel_region,
    reduce_from_tensor_model_parallel_region,
    scatter_to_tensor_model_parallel_region,
)

from .utils import (
    divide,
    VocabUtility,
)

_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {'tensor_model_parallel': False,
                                      'partition_dim': -1,
                                      'partition_stride': 1}

def param_is_not_tensor_parallel_duplicate(param):
    return (hasattr(param, 'tensor_model_parallel') and
            param.tensor_model_parallel) or (
                get_tensor_model_parallel_rank() == 0)


def set_tensor_model_parallel_attributes(tensor, is_parallel, dim, stride):
    # Make sure the attributes are not set.
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        assert not hasattr(tensor, attribute)
    # Set the attributes.
    setattr(tensor, 'tensor_model_parallel', is_parallel)
    setattr(tensor, 'partition_dim', dim)
    setattr(tensor, 'partition_stride', stride)


def set_defaults_if_not_set_tensor_model_parallel_attributes(tensor):
    def maybe_set(attribute, value):
        if not hasattr(tensor, attribute):
            setattr(tensor, attribute, value)
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        maybe_set(attribute, _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS[attribute])


def copy_tensor_model_parallel_attributes(destination_tensor, source_tensor):
    def maybe_copy(attribute):
        if hasattr(source_tensor, attribute):
            setattr(destination_tensor, attribute,
                    getattr(source_tensor, attribute))
    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
        maybe_copy(attribute)


class VocabParallelEmbedding(torch.nn.Module):
    """Embedding parallelized in the vocabulary dimension.

    This is mainly adapted from torch.nn.Embedding and all the default
    values are kept.
    Arguments:
        num_embeddings: vocabulary size.
        embedding_dim: size of hidden state.

    Keyword Arguments:
        init_method: method to initialize weights.
        params_dtype
        use_cpu_initialization
        perform_initialization
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, *,
                 init_method=init.xavier_normal_,
                 params_dtype: torch.dtype=None,
                 use_cpu_initialization: bool=False,
                 perform_initialization: bool=True):
        super(VocabParallelEmbedding, self).__init__()
        assert not perform_initialization
        assert not use_cpu_initialization

        # Keep the input dimensions.
        self.num_embeddings = num_embeddings
        self.embedding_dim = embedding_dim
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()

        # Set the defaults for compatibility.
        self.padding_idx = None
        self.max_norm = None
        self.norm_type = 2.
        self.scale_grad_by_freq = False
        self.sparse = False
        self._weight = None
        self.tensor_model_parallel_size = get_tensor_model_parallel_world_size()
        # Divide the weight matrix along the vocaburaly dimension.
        self.vocab_start_index, self.vocab_end_index = \
            VocabUtility.vocab_range_from_global_vocab_size(
                self.num_embeddings, get_tensor_model_parallel_rank(),
                self.tensor_model_parallel_size)
        self.num_embeddings_per_partition = self.vocab_end_index - \
            self.vocab_start_index

        self.weight = Parameter(torch.empty(
            self.num_embeddings_per_partition, self.embedding_dim,
            device=torch.cuda.current_device(), dtype=params_dtype))
 
    def forward(self, input_):
        if self.tensor_model_parallel_size > 1:
            # Build the mask.
            input_mask = (input_ < self.vocab_start_index) | \
                         (input_ >= self.vocab_end_index)
            # Mask the input.
            masked_input = input_.clone() - self.vocab_start_index
            masked_input[input_mask] = 0
        else:
            masked_input = input_
            # Get the embeddings.
        output_parallel = F.embedding(masked_input, self.weight,
                                      self.padding_idx, self.max_norm,
                                      self.norm_type, self.scale_grad_by_freq,
                                      self.sparse)
        # Mask the output embedding.
        if self.tensor_model_parallel_size > 1:
            output_parallel[input_mask, :] = 0.0
        # Reduce across all the model parallel GPUs.
        output = reduce_from_tensor_model_parallel_region(output_parallel)
        return output


class ColumnParallelLinear(torch.nn.Module):
    """Linear layer with column parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its second dimension as A = [A_1, ..., A_p].

    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.

    Keyword Arguments
        bias: If true, add bias
        gather_output: If true, call all-gather on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y_i = XA_i
        init_method: method to initialize weights. Note that bias is always set
                     to zero.
        stride: For the strided linear layers.
        keep_master_weight_for_test: This was added for testing and should be
                                     set to False. It returns the master weights
                                     used for initialization.
        skip_bias_add: This was added to enable performance optimations where bias
                       can be fused with other elementwise operations. we skip
                       adding bias but instead return it.
        params_dtype:
        use_cpu_initialization:
    """

    def __init__(self, input_size, output_size, *,
                 bias=True, gather_output=True,
                 init_method=init.xavier_normal_, stride=1,
                 keep_master_weight_for_test=False,
                 skip_bias_add=False,
                 params_dtype=None,
                 use_cpu_initialization=False,
                 perform_initialization=True,
                 quant_config=None,
                 ):
        super(ColumnParallelLinear, self).__init__()
        assert not perform_initialization
        assert not use_cpu_initialization

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.gather_output = gather_output
        # Divide the weight matrix along the last dimension.
        self.world_size = get_tensor_model_parallel_world_size()
        self.output_size_per_partition = divide(output_size, self.world_size)
        self.skip_bias_add = skip_bias_add
        self.quant_config = quant_config

        if params_dtype is None:
            params_dtype = torch.get_default_dtype()

        # Parameters.
        # Note: torch.nn.functional.linear performs XA^T + b and as a result
        # we allocate the transpose.
        self.create_weights(params_dtype)

        if bias:
            self.bias = Parameter(torch.empty(
                self.output_size_per_partition,
                device=torch.cuda.current_device(),
                dtype=params_dtype))
            set_tensor_model_parallel_attributes(self.bias, True, 0, stride)
            # Always initialize bias to zero.
            with torch.no_grad():
                self.bias.zero_()
        else:
            self.register_parameter('bias', None)

    def create_weights(self, dtype: torch.dtype) -> None:
        self.weight = Parameter(torch.empty(
            self.output_size_per_partition, self.input_size,
            device=torch.cuda.current_device(), dtype=dtype))

    def apply_weights(
        self,
        x: torch.Tensor,
        bias: Optional[torch.Tensor],
    ) -> torch.Tensor:
        return F.linear(x, self.weight, bias)

    def forward(self, input_):
        """Forward of ColumnParallelLinear

        Args:
            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]

        Returns:
            - output
            - bias
        """
        bias = self.bias if not self.skip_bias_add else None

        input_parallel = input_
        # Matrix multiply.
        output_parallel = self.apply_weights(input_parallel, bias)
        if self.gather_output:
            # All-gather across the partitions.
            output = gather_from_tensor_model_parallel_region(output_parallel)
        else:
            output = output_parallel
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias


class RowParallelLinear(torch.nn.Module):
    """Linear layer with row parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its first dimension and X along its second dimension as:
               -   -
              | A_1 |
              | .   |
          A = | .   |        X = [X_1, ..., X_p]
              | .   |
              | A_p |
               -   -
    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.

    Keyword Arguments:
        bias: If true, add bias. Note that bias is not parallelized.
        input_is_parallel: If true, we assume that the input is already
                           split across the GPUs and we do not split
                           again.
        init_method: method to initialize weights. Note that bias is always set
                     to zero.
        stride: For the strided linear layers.
        keep_master_weight_for_test: This was added for testing and should be
                                     set to False. It returns the master weights
                                     used for initialization.
        skip_bias_add: This was added to enable performance optimization where bias
                       can be fused with other elementwise operations. We skip
                       adding bias but instead return it.
        params_dtype:
        use_cpu_initialization:
        perform_initialization:
        reduce_results:
    """

    def __init__(self, input_size, output_size, *,
                 bias=True, input_is_parallel=False,
                 init_method=init.xavier_normal_, stride=1,
                 keep_master_weight_for_test=False,
                 skip_bias_add=False,
                 params_dtype=None,
                 use_cpu_initialization=False,
                 perform_initialization=True,
                 reduce_results=True,
                 quant_config=None,
                 ):
        super(RowParallelLinear, self).__init__()
        assert not perform_initialization
        assert not use_cpu_initialization

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()

        # Divide the weight matrix along the last dimension.
        self.world_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.world_size)
        self.skip_bias_add = skip_bias_add
        self.quant_config = quant_config

        self.create_weights(params_dtype)

        if not reduce_results and (bias and not skip_bias_add):
            raise ValueError("When not reduce the results, adding bias to the "
                             "results can lead to incorrect results")

        if bias:
            self.bias = Parameter(torch.empty(
                self.output_size, device=torch.cuda.current_device(),
                dtype=params_dtype))

            # Always initialize bias to zero.
            with torch.no_grad():
                self.bias.zero_()
        else:
            self.register_parameter('bias', None)

    def create_weights(self, dtype: torch.dtype) -> None:
        self.weight = Parameter(torch.empty(
                self.output_size, self.input_size_per_partition,
                device=torch.cuda.current_device(), dtype=dtype))

    def apply_weights(self, x: torch.Tensor) -> torch.Tensor:
        return F.linear(x, self.weight)

    def forward(self, input_):
        """Forward of RowParallelLinear

        Args:
            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]

        Returns:
            - output
            - bias
        """
        # Set up backprop all-reduce.
        if self.input_is_parallel:
            input_parallel = input_
        else:
            input_parallel = scatter_to_tensor_model_parallel_region(input_)
        # Matrix multiply.
        output_parallel = self.apply_weights(input_parallel)
        if self.reduce_results and self.world_size > 1:
            output_ = reduce_from_tensor_model_parallel_region(output_parallel)
        else:
            output_ = output_parallel

        if not self.skip_bias_add:
            output = output_ + self.bias if self.bias is not None else output_
            output_bias = None
        else:
            output = output_
            output_bias = self.bias
        return output, output_bias