hao-ai-lab-FastVideo/fastvideo/platforms/interface.py

# SPDX-License-Identifier: Apache-2.0

import enum
import platform
import random
from platform import uname
from typing import TYPE_CHECKING, NamedTuple, Optional, Tuple, Union

import numpy as np
import torch

from fastvideo.logger import init_logger

logger = init_logger(__name__)


class _Backend(enum.Enum):
    FLASH_ATTN = enum.auto()
    TORCH_SDPA = enum.auto()
    NO_ATTENTION = enum.auto()


class PlatformEnum(enum.Enum):
    CUDA = enum.auto()
    OOT = enum.auto()
    UNSPECIFIED = enum.auto()


class DeviceCapability(NamedTuple):
    major: int
    minor: int

    def as_version_str(self) -> str:
        return f"{self.major}.{self.minor}"

    def to_int(self) -> int:
        """
        Express device capability as an integer ``<major><minor>``.

        It is assumed that the minor version is always a single digit.
        """
        assert 0 <= self.minor < 10
        return self.major * 10 + self.minor


class Platform:
    _enum: PlatformEnum
    device_name: str
    device_type: str

    # available dispatch keys:
    # check https://github.com/pytorch/pytorch/blob/313dac6c1ca0fa0cde32477509cce32089f8532a/torchgen/model.py#L134 # noqa
    # use "CPU" as a fallback for platforms not registered in PyTorch
    dispatch_key: str = "CPU"

    supported_quantization: list[str] = []

    def is_cuda(self) -> bool:
        return self._enum == PlatformEnum.CUDA

    def is_out_of_tree(self) -> bool:
        return self._enum == PlatformEnum.OOT

    def is_cuda_alike(self) -> bool:
        """Stateless version of :func:`torch.cuda.is_available`."""
        return self._enum in (PlatformEnum.CUDA, PlatformEnum.ROCM)

    @classmethod
    def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
                             dtype: torch.dtype, kv_cache_dtype: Optional[str],
                             block_size: int, use_v1: bool,
                             use_mla: bool) -> str:
        """Get the attention backend class of a device."""
        return ""

    @classmethod
    def get_device_capability(
        cls,
        device_id: int = 0,
    ) -> Optional[DeviceCapability]:
        """Stateless version of :func:`torch.cuda.get_device_capability`."""
        return None

    @classmethod
    def has_device_capability(
        cls,
        capability: Union[Tuple[int, int], int],
        device_id: int = 0,
    ) -> bool:
        """
        Test whether this platform is compatible with a device capability.

        The ``capability`` argument can either be:

        - A tuple ``(major, minor)``.
        - An integer ``<major><minor>``. (See :meth:`DeviceCapability.to_int`)
        """
        current_capability = cls.get_device_capability(device_id=device_id)
        if current_capability is None:
            return False

        if isinstance(capability, tuple):
            return current_capability >= capability

        return current_capability.to_int() >= capability

    @classmethod
    def get_device_name(cls, device_id: int = 0) -> str:
        """Get the name of a device."""
        raise NotImplementedError

    @classmethod
    def get_device_uuid(cls, device_id: int = 0) -> str:
        """Get the uuid of a device, e.g. the PCI bus ID."""
        raise NotImplementedError

    @classmethod
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        """Get the total memory of a device in bytes."""
        raise NotImplementedError

    @classmethod
    def is_async_output_supported(cls, enforce_eager: Optional[bool]) -> bool:
        """
        Check if the current platform supports async output.
        """
        raise NotImplementedError

    @classmethod
    def inference_mode(cls):
        """A device-specific wrapper of `torch.inference_mode`.

        This wrapper is recommended because some hardware backends such as TPU
        do not support `torch.inference_mode`. In such a case, they will fall
        back to `torch.no_grad` by overriding this method.
        """
        return torch.inference_mode(mode=True)

    @classmethod
    def seed_everything(cls, seed: Optional[int] = None) -> None:
        """
        Set the seed of each random module.
        `torch.manual_seed` will set seed on all devices.

        Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
        """
        if seed is not None:
            random.seed(seed)
            np.random.seed(seed)
            torch.manual_seed(seed)

    @classmethod
    def verify_model_arch(cls, model_arch: str) -> None:
        """
        Verify whether the current platform supports the specified model
        architecture.

        - This will raise an Error or Warning based on the model support on
        the current platform.
        - By default all models are considered supported.
        """
        pass

    @classmethod
    def verify_quantization(cls, quant: str) -> None:
        """
        Verify whether the quantization is supported by the current platform.
        """
        if cls.supported_quantization and \
            quant not in cls.supported_quantization:
            raise ValueError(
                f"{quant} quantization is currently not supported in "
                f"{cls.device_name}.")

    @classmethod
    def get_current_memory_usage(cls,
                                 device: Optional[torch.types.Device] = None
                                 ) -> float:
        """
        Return the memory usage in bytes.
        """
        raise NotImplementedError

    @classmethod
    def get_device_communicator_cls(cls) -> str:
        """
        Get device specific communicator class for distributed communication.
        """
        return "fastvideo.distributed.device_communicators.base_device_communicator.DeviceCommunicatorBase"  # noqa


class UnspecifiedPlatform(Platform):
    _enum = PlatformEnum.UNSPECIFIED
    device_type = ""