modules.py

# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from typing import Optional, TypeVar, Union, overload

import torch
import torch.nn.functional as F
from torch import Tensor, device, dtype, nn

import bitsandbytes as bnb
from bitsandbytes.optim import GlobalOptimManager

T = TypeVar("T", bound="torch.nn.Module")


class StableEmbedding(torch.nn.Embedding):
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: Optional[int] = None,
        max_norm: Optional[float] = None,
        norm_type: float = 2.0,
        scale_grad_by_freq: bool = False,
        sparse: bool = False,
        _weight: Optional[Tensor] = None,
        device=None,
        dtype=None,
    ) -> None:
        super().__init__(
            num_embeddings,
            embedding_dim,
            padding_idx,
            max_norm,
            norm_type,
            scale_grad_by_freq,
            sparse,
            _weight,
            device,
            dtype,
        )
        self.norm = torch.nn.LayerNorm(embedding_dim, device=device)
        GlobalOptimManager.get_instance().register_module_override(
            self, "weight", {"optim_bits": 32}
        )

    def reset_parameters(self) -> None:
        torch.nn.init.xavier_uniform_(self.weight)
        self._fill_padding_idx_with_zero()

    """ !!! This is a redefinition of _fill_padding_idx_with_zero in torch.nn.Embedding
        to make the Layer compatible with Pytorch < 1.9.
        This means that if this changes in future PyTorch releases this need to change too
        which is cumbersome. However, with this we can ensure compatibility with previous
        PyTorch releases.
    """

    def _fill_padding_idx_with_zero(self) -> None:
        if self.padding_idx is not None:
            with torch.no_grad():
                self.weight[self.padding_idx].fill_(0)

    def forward(self, input: Tensor) -> Tensor:
        emb = F.embedding(
            input,
            self.weight,
            self.padding_idx,
            self.max_norm,
            self.norm_type,
            self.scale_grad_by_freq,
            self.sparse,
        )

        # always apply layer norm in full precision
        emb = emb.to(torch.get_default_dtype())

        return self.norm(emb).to(self.weight.dtype)


class Embedding(torch.nn.Embedding):
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: Optional[int] = None,
        max_norm: Optional[float] = None,
        norm_type: float = 2.0,
        scale_grad_by_freq: bool = False,
        sparse: bool = False,
        _weight: Optional[Tensor] = None,
    ) -> None:
        super().__init__(
            num_embeddings,
            embedding_dim,
            padding_idx,
            max_norm,
            norm_type,
            scale_grad_by_freq,
            sparse,
            _weight,
        )
        GlobalOptimManager.get_instance().register_module_override(
            self, "weight", {"optim_bits": 32}
        )

    def reset_parameters(self) -> None:
        torch.nn.init.xavier_uniform_(self.weight)
        self._fill_padding_idx_with_zero()

    """ !!! This is a redefinition of _fill_padding_idx_with_zero in torch.nn.Embedding
        to make the Layer compatible with Pytorch < 1.9.
        This means that if this changes in future PyTorch releases this need to change too
        which is cumbersome. However, with this we can ensure compatibility with previous
        PyTorch releases.
    """

    def _fill_padding_idx_with_zero(self) -> None:
        if self.padding_idx is not None:
            with torch.no_grad():
                self.weight[self.padding_idx].fill_(0)

    def forward(self, input: Tensor) -> Tensor:
        emb = F.embedding(
            input,
            self.weight,
            self.padding_idx,
            self.max_norm,
            self.norm_type,
            self.scale_grad_by_freq,
            self.sparse,
        )

        return emb


class Int8Params(torch.nn.Parameter):
    def __new__(
        cls,
        data=None,
        requires_grad=True,
        has_fp16_weights=False,
        CB=None,
        SCB=None,
    ):
        cls.has_fp16_weights = has_fp16_weights
        cls.CB = None
        cls.SCB = None
        if data is None:
            data = torch.empty(0)
        return torch.Tensor._make_subclass(cls, data, requires_grad)

    def cuda(self, device):
        if self.has_fp16_weights:
            return super().cuda(device)
        else:
            # we store the 8-bit rows-major weight
            # we convert this weight to the turning/ampere weight during the first inference pass
            B = self.data.contiguous().half().cuda(device)
            CB, CBt, SCB, SCBt, coo_tensorB = bnb.functional.double_quant(B)
            del CBt
            del SCBt
            self.data = CB
            setattr(self, "CB", CB)
            setattr(self, "SCB", SCB)

        return self

    @overload
    def to(
        self: T,
        device: Optional[Union[int, device]] = ...,
        dtype: Optional[Union[dtype, str]] = ...,
        non_blocking: bool = ...,
    ) -> T:
        ...

    @overload
    def to(self: T, dtype: Union[dtype, str], non_blocking: bool = ...) -> T:
        ...

    @overload
    def to(self: T, tensor: Tensor, non_blocking: bool = ...) -> T:
        ...

    def to(self, *args, **kwargs):
        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(
            *args, **kwargs
        )

        if (
            device is not None
            and device.type == "cuda"
            and self.data.device.type == "cpu"
        ):
            return self.cuda(device)
        else:
            new_param = Int8Params(
                super().to(
                    device=device, dtype=dtype, non_blocking=non_blocking
                ),
                requires_grad=self.requires_grad,
                has_fp16_weights=self.has_fp16_weights,
            )
            new_param.CB = self.CB
            new_param.SCB = self.SCB

            return new_param


class Linear8bitLt(nn.Linear):
    def __init__(self, input_features, output_features, bias=True, has_fp16_weights=True,
                       memory_efficient_backward=False, threshold=0.0, index=None):
        super().__init__(input_features, output_features, bias)
        assert not memory_efficient_backward, "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
        self.state = bnb.MatmulLtState()
        self.index = index

        self.state.threshold = threshold
        self.state.has_fp16_weights = has_fp16_weights
        self.state.memory_efficient_backward = memory_efficient_backward
        if threshold > 0.0 and not has_fp16_weights:
            self.state.use_pool = True

        self.weight = Int8Params(self.weight.data, has_fp16_weights=has_fp16_weights, requires_grad=has_fp16_weights)

    def _save_to_state_dict(self, destination, prefix, keep_vars):
        super()._save_to_state_dict(destination, prefix, keep_vars)

        # we only need to save SCB as extra data, because CB for quantized weights is already stored in weight.data
        weight_name = "SCB"

        # case 1: .cuda was called, SCB is in self.weight
        param_from_weight = getattr(self.weight, weight_name)
        # case 2: self.init_8bit_state was called, SCB is in self.state
        param_from_state = getattr(self.state, weight_name)

        key_name = prefix + f"{weight_name}"
        if param_from_weight is not None:
            destination[key_name] = param_from_weight if keep_vars else param_from_weight.detach()
        elif not self.state.has_fp16_weights and param_from_state is not None:
            destination[key_name] = param_from_state if keep_vars else param_from_state.detach()

    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
                              missing_keys, unexpected_keys, error_msgs):
        super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
                                      error_msgs)
        for key in unexpected_keys:
            input_name = key[len(prefix):]
            if input_name == "SCB":
                input_param = state_dict[key]
                self.weight.SCB.copy_(input_param)
                unexpected_keys.remove(key)

    def init_8bit_state(self):
        self.state.CB = self.weight.CB
        self.state.SCB = self.weight.SCB
        self.weight.CB = None
        self.weight.SCB = None

    def forward(self, x: torch.Tensor):
        self.state.is_training = self.training
        if self.weight.CB is not None:
            self.init_8bit_state()

        # weights are cast automatically as Int8Params, but the bias has to be cast manually
        if self.bias is not None and self.bias.dtype != x.dtype:
            self.bias.data = self.bias.data.to(x.dtype)

        out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
        if not self.state.has_fp16_weights:
            if self.state.CB is not None and self.state.CxB is not None:
                # we converted 8-bit row major to turing/ampere format in the first inference pass
                # we no longer need the row-major weight
                del self.state.CB
                self.weight.data = self.state.CxB
        return out