cuda.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Code inside this file can safely assume cuda platform, e.g. importing
pynvml. However, it should not initialize cuda context.
"""

from __future__ import annotations

import os
from collections.abc import Callable
from datetime import timedelta
from functools import cache, lru_cache, wraps
from typing import TYPE_CHECKING, TypeVar

import torch
from torch.distributed import PrefixStore, ProcessGroup
from torch.distributed.distributed_c10d import is_nccl_available
from typing_extensions import ParamSpec

# import custom ops, trigger op registration
import vllm._C  # noqa
import vllm._C_stable_libtorch  # noqa
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.utils.import_utils import import_pynvml
from vllm.v1.attention.backends.registry import AttentionBackendEnum

from .interface import DeviceCapability, Platform, PlatformEnum

if TYPE_CHECKING:
    from vllm.config import VllmConfig
    from vllm.config.cache import CacheDType
    from vllm.v1.attention.selector import AttentionSelectorConfig
else:
    VllmConfig = None
    CacheDType = None

logger = init_logger(__name__)

_P = ParamSpec("_P")
_R = TypeVar("_R")

pynvml = import_pynvml()

# pytorch 2.5 uses cudnn sdpa by default, which will cause crash on some models
# see https://github.com/huggingface/diffusers/issues/9704 for details
torch.backends.cuda.enable_cudnn_sdp(False)


@lru_cache(maxsize=8)
def _cuda_device_count_stateless(cuda_visible_devices: str | None = None) -> int:
    """Get number of CUDA devices, caching based on the value of CUDA_VISIBLE_DEVICES
    at the time of call.

    This should be used instead of torch.accelerator.device_count() unless
    CUDA_VISIBLE_DEVICES has already been set to the desired value.

    # This can be removed and simply replaced with torch.cuda.get_device_count
    # after https://github.com/pytorch/pytorch/pull/122815 is released."""
    # Note: cuda_visible_devices is not used, but we keep it as an argument for
    # LRU Cache purposes.

    # Code below is based on
    # https://github.com/pytorch/pytorch/blob/
    # c1cd946818442aca8c7f812b16d187ce1586c3bc/
    # torch/cuda/__init__.py#L831C1-L831C17
    import torch.cuda

    if not torch.cuda._is_compiled():
        return 0
    raw_count = torch.cuda._device_count_nvml()
    r = torch._C._cuda_getDeviceCount() if raw_count < 0 else raw_count
    return r


@cache
def _get_backend_priorities(
    use_mla: bool,
    device_capability: DeviceCapability,
    num_heads: int | None = None,
    kv_cache_dtype: CacheDType | None = None,
) -> list[AttentionBackendEnum]:
    """Get backend priorities with lazy import to avoid circular dependency."""
    if use_mla:
        if device_capability.major == 10:
            # Sparse MLA backend priorities
            # See https://github.com/vllm-project/vllm/issues/35807 for
            # benchmark results
            if kv_cache_dtype is not None and kv_cache_dtype.startswith("fp8"):
                # Prefer FlashInfer for fp8 kv cache
                sparse_backends = [
                    AttentionBackendEnum.FLASHINFER_MLA_SPARSE,
                    AttentionBackendEnum.FLASHMLA_SPARSE,
                ]
            else:
                # BF16 KV Cache
                # Prefer FlashInfer at low head counts (FlashMLA uses padding)
                if num_heads is not None and num_heads <= 16:
                    sparse_backends = [
                        AttentionBackendEnum.FLASHINFER_MLA_SPARSE,
                        AttentionBackendEnum.FLASHMLA_SPARSE,
                    ]
                else:
                    sparse_backends = [
                        AttentionBackendEnum.FLASHMLA_SPARSE,
                        AttentionBackendEnum.FLASHINFER_MLA_SPARSE,
                    ]

            return [
                AttentionBackendEnum.FLASHINFER_MLA,
                AttentionBackendEnum.CUTLASS_MLA,
                AttentionBackendEnum.FLASH_ATTN_MLA,
                AttentionBackendEnum.FLASHMLA,
                AttentionBackendEnum.TRITON_MLA,
                *sparse_backends,
            ]
        else:
            return [
                AttentionBackendEnum.FLASH_ATTN_MLA,
                AttentionBackendEnum.FLASHMLA,
                AttentionBackendEnum.FLASHINFER_MLA,
                AttentionBackendEnum.TRITON_MLA,
                AttentionBackendEnum.FLASHMLA_SPARSE,
            ]
    else:
        if device_capability.major == 10:
            return [
                AttentionBackendEnum.FLASHINFER,
                AttentionBackendEnum.FLASH_ATTN,
                AttentionBackendEnum.TRITON_ATTN,
                AttentionBackendEnum.FLEX_ATTENTION,
            ]
        else:
            return [
                AttentionBackendEnum.FLASH_ATTN,
                AttentionBackendEnum.FLASHINFER,
                AttentionBackendEnum.TRITON_ATTN,
                AttentionBackendEnum.FLEX_ATTENTION,
            ]


def with_nvml_context(fn: Callable[_P, _R]) -> Callable[_P, _R]:
    @wraps(fn)
    def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> _R:
        pynvml.nvmlInit()
        try:
            return fn(*args, **kwargs)
        finally:
            pynvml.nvmlShutdown()

    return wrapper


class CudaPlatformBase(Platform):
    _enum = PlatformEnum.CUDA
    device_name: str = "cuda"
    device_type: str = "cuda"
    dispatch_key: str = "CUDA"
    ray_device_key: str = "GPU"
    dist_backend: str = "nccl"
    device_control_env_var: str = "CUDA_VISIBLE_DEVICES"
    ray_noset_device_env_vars: list[str] = [
        "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES",
    ]

    @property
    def supported_dtypes(self) -> list[torch.dtype]:
        if self.has_device_capability(80):
            # Ampere and Hopper or later NVIDIA GPUs.
            return [torch.bfloat16, torch.float16, torch.float32]
        if self.has_device_capability(60):
            # Pascal, Volta and Turing NVIDIA GPUs, BF16 is not supported
            return [torch.float16, torch.float32]
        # Kepler and Maxwell NVIDIA GPUs, only FP32 is supported,
        # though vLLM doesn't support these GPUs.
        return [torch.float32]

    @classmethod
    def set_device(cls, device: torch.device) -> None:
        """
        Set the device for the current platform.
        """
        torch.cuda.set_device(device)
        # With this trick we can force the device to be set eagerly
        # see https://github.com/pytorch/pytorch/issues/155668
        # for why and when it is needed
        _ = torch.zeros(1, device=device)

    @classmethod
    def get_device_capability(cls, device_id: int = 0) -> DeviceCapability | None:
        raise NotImplementedError

    @classmethod
    def get_device_name(cls, device_id: int = 0) -> str:
        raise NotImplementedError

    @classmethod
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        raise NotImplementedError

    @classmethod
    def is_fully_connected(cls, device_ids: list[int]) -> bool:
        raise NotImplementedError

    @classmethod
    def log_warnings(cls):
        pass

    @classmethod
    def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
        parallel_config = vllm_config.parallel_config
        model_config = vllm_config.model_config

        if parallel_config.worker_cls == "auto":
            parallel_config.worker_cls = "vllm.v1.worker.gpu_worker.Worker"

        scheduler_config = vllm_config.scheduler_config
        # Note: model_config may be None during testing
        if (
            model_config is not None
            and model_config.is_mm_prefix_lm
            and scheduler_config.is_multimodal_model
            and not scheduler_config.disable_chunked_mm_input
        ):
            logger.warning(
                "Forcing --disable_chunked_mm_input for models "
                "with multimodal-bidirectional attention."
            )
            scheduler_config.disable_chunked_mm_input = True

    @classmethod
    def get_current_memory_usage(
        cls, device: torch.types.Device | None = None
    ) -> float:
        torch.cuda.empty_cache()
        torch.cuda.reset_peak_memory_stats(device)
        return torch.cuda.max_memory_allocated(device)

    @classmethod
    def get_valid_backends(
        cls,
        device_capability: DeviceCapability,
        attn_selector_config: AttentionSelectorConfig,
        num_heads: int | None = None,
    ) -> tuple[
        list[tuple[AttentionBackendEnum, int]],
        dict[AttentionBackendEnum, tuple[int, list[str]]],
    ]:
        valid_backends_priorities = []
        invalid_reasons: dict[AttentionBackendEnum, tuple[int, list[str]]] = {}

        backend_priorities = _get_backend_priorities(
            attn_selector_config.use_mla,
            device_capability,
            num_heads,
            attn_selector_config.kv_cache_dtype,
        )
        for priority, backend in enumerate(backend_priorities):
            try:
                backend_class = backend.get_class()
                invalid_reasons_i = backend_class.validate_configuration(
                    device_capability=device_capability,
                    **attn_selector_config._asdict(),
                )
            except ImportError:
                invalid_reasons_i = ["ImportError"]
            if invalid_reasons_i:
                invalid_reasons[backend] = (priority, invalid_reasons_i)
            else:
                valid_backends_priorities.append((backend, priority))

        return valid_backends_priorities, invalid_reasons

    @classmethod
    def get_attn_backend_cls(
        cls,
        selected_backend: AttentionBackendEnum | None,
        attn_selector_config: AttentionSelectorConfig,
        num_heads: int | None = None,
    ) -> str:
        device_capability = cls.get_device_capability()
        assert device_capability is not None

        # First try checking just the selected backend, if there is one.
        if selected_backend is not None:
            try:
                backend_class = selected_backend.get_class()
                invalid_reasons = backend_class.validate_configuration(
                    device_capability=device_capability,
                    **attn_selector_config._asdict(),
                )
            except ImportError:
                invalid_reasons = ["ImportError"]
            if invalid_reasons:
                raise ValueError(
                    f"Selected backend {selected_backend} is not valid for "
                    f"this configuration. Reason: {invalid_reasons}"
                )
            else:
                logger.info("Using %s backend.", selected_backend)
                return selected_backend.get_path()

        # No selected backend or the selected backend is invalid,
        # so we try finding a valid backend.
        valid_backends_priorities, all_invalid_reasons = cls.get_valid_backends(
            device_capability=device_capability,
            attn_selector_config=attn_selector_config,
            num_heads=num_heads,
        )
        reasons_str = (
            "{"
            + ", ".join(
                f"{backend.name}: [{', '.join(reasons)}]"
                for backend, (_, reasons) in all_invalid_reasons.items()
            )
            + "}"
        )
        config_str = attn_selector_config.__repr__()
        logger.debug_once(
            f"Some attention backends are not valid for {cls.device_name} with "
            f"{config_str}. Reasons: {reasons_str}."
        )
        if len(valid_backends_priorities) == 0:
            raise ValueError(
                f"No valid attention backend found for {cls.device_name} "
                f"with {config_str}. Reasons: {reasons_str}."
            )

        # We have found some valid backends. Select the one with the
        # highest priority.
        sorted_indices = sorted(
            range(len(valid_backends_priorities)),
            key=lambda i: valid_backends_priorities[i][1],
        )
        selected_index = sorted_indices[0]
        selected_backend = valid_backends_priorities[selected_index][0]
        selected_priority = valid_backends_priorities[selected_index][1]

        # If the user specified --block-size (but not --attention-backend),
        # check whether that constraint precluded any higher-priority backends.
        if attn_selector_config.block_size is not None:
            excluded = [
                backend
                for backend, (priority, reasons) in all_invalid_reasons.items()
                if priority < selected_priority
                and reasons == ["block_size not supported"]
            ]
            if excluded:
                names = ", ".join(b.name for b in excluded)
                logger.warning(
                    "--block-size %d precluded higher-priority backend(s) "
                    "%s. Using %s instead, which may result in reduced "
                    "performance. Consider removing --block-size to "
                    "auto-select the optimal block size.",
                    attn_selector_config.block_size,
                    names,
                    selected_backend.name,
                )

        logger.info_once(
            "Using %s attention backend out of potential backends: %s.",
            selected_backend.name,
            "[" + ", ".join(f"'{b[0].name}'" for b in valid_backends_priorities) + "]",
            scope="local",
        )

        return selected_backend.get_path()

    @classmethod
    def get_supported_vit_attn_backends(cls) -> list[AttentionBackendEnum]:
        if cls.has_device_capability(80):
            return [
                AttentionBackendEnum.FLASH_ATTN,
                AttentionBackendEnum.TRITON_ATTN,
                AttentionBackendEnum.TORCH_SDPA,
                AttentionBackendEnum.FLASHINFER,
            ]
        else:
            return [
                AttentionBackendEnum.FLASH_ATTN,
                AttentionBackendEnum.TORCH_SDPA,
                AttentionBackendEnum.TRITON_ATTN,
                AttentionBackendEnum.FLASHINFER,
            ]

    @classmethod
    def get_vit_attn_backend(
        cls,
        head_size: int,
        dtype: torch.dtype,
        backend: AttentionBackendEnum | None = None,
    ) -> AttentionBackendEnum:
        if backend is not None:
            assert backend in cls.get_supported_vit_attn_backends(), (
                f"Backend {backend} is not supported for vit attention. "
                f"Supported backends are: {cls.get_supported_vit_attn_backends()}"
            )
            logger.info_once(f"Using backend {backend} for vit attention")
            return backend

        cc = cls.get_device_capability()
        for vit_attn_backend in cls.get_supported_vit_attn_backends():
            if vit_attn_backend == AttentionBackendEnum.TORCH_SDPA:
                return vit_attn_backend
            try:
                backend_class = vit_attn_backend.get_class()
                is_backend_supported = backend_class.supports_head_size(
                    head_size
                ) and backend_class.supports_dtype(dtype)
                if cc is not None:
                    is_backend_supported = (
                        is_backend_supported
                        and backend_class.supports_compute_capability(cc)
                    )
                if is_backend_supported:
                    logger.info_once(
                        f"Using backend {vit_attn_backend} for vit attention",
                        scope="local",
                    )
                    return vit_attn_backend
            except ImportError:
                pass

        return AttentionBackendEnum.TORCH_SDPA

    @classmethod
    def get_punica_wrapper(cls) -> str:
        return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"

    @classmethod
    def get_device_communicator_cls(cls) -> str:
        return (
            "vllm.distributed.device_communicators.cuda_communicator.CudaCommunicator"  # noqa
        )

    @classmethod
    def supports_fp8(cls) -> bool:
        return cls.has_device_capability(89)

    @classmethod
    def use_custom_allreduce(cls) -> bool:
        return True

    @classmethod
    def opaque_attention_op(cls) -> bool:
        return True

    @classmethod
    def get_static_graph_wrapper_cls(cls) -> str:
        return "vllm.compilation.cuda_graph.CUDAGraphWrapper"

    @classmethod
    def stateless_init_device_torch_dist_pg(
        cls,
        backend: str,
        prefix_store: PrefixStore,
        group_rank: int,
        group_size: int,
        timeout: timedelta,
    ) -> ProcessGroup:
        assert is_nccl_available()
        pg: ProcessGroup = ProcessGroup(
            prefix_store,
            group_rank,
            group_size,
        )
        from torch.distributed.distributed_c10d import ProcessGroupNCCL

        backend_options = ProcessGroupNCCL.Options()
        backend_options._timeout = timeout

        backend_class = ProcessGroupNCCL(
            prefix_store, group_rank, group_size, backend_options
        )
        backend_type = ProcessGroup.BackendType.NCCL
        device = torch.device("cuda")
        pg._set_default_backend(backend_type)
        backend_class._set_sequence_number_for_group()

        pg._register_backend(device, backend_type, backend_class)
        return pg

    @classmethod
    def device_count(cls) -> int:
        return _cuda_device_count_stateless(envs.CUDA_VISIBLE_DEVICES)

    @classmethod
    def check_if_supports_dtype(cls, dtype: torch.dtype):
        if dtype == torch.bfloat16:  # noqa: SIM102
            if not cls.has_device_capability(80):
                capability = cls.get_device_capability()
                gpu_name = cls.get_device_name()

                if capability is None:
                    compute_str = "does not have a compute capability"
                else:
                    version_str = capability.as_version_str()
                    compute_str = f"has compute capability {version_str}"

                raise ValueError(
                    "Bfloat16 is only supported on GPUs "
                    "with compute capability of at least 8.0. "
                    f"Your {gpu_name} GPU {compute_str}. "
                    "You can use float16 instead by explicitly setting the "
                    "`dtype` flag in CLI, for example: --dtype=half."
                )

    @classmethod
    def insert_blocks_to_device(
        cls,
        src_cache: torch.Tensor,
        dst_cache: torch.Tensor,
        src_block_indices: torch.Tensor,
        dst_block_indices: torch.Tensor,
    ) -> None:
        """Copy blocks from src_cache to dst_cache on GPU."""
        _src_cache = src_cache[:, src_block_indices]
        dst_cache[:, dst_block_indices] = _src_cache.to(dst_cache.device)

    @classmethod
    def swap_out_blocks_to_host(
        cls,
        src_cache: torch.Tensor,
        dst_cache: torch.Tensor,
        src_block_indices: torch.Tensor,
        dst_block_indices: torch.Tensor,
    ) -> None:
        """Copy blocks from GPU to host (CPU)."""
        _src_cache = src_cache[:, src_block_indices]
        dst_cache[:, dst_block_indices] = _src_cache.cpu()

    @classmethod
    def support_hybrid_kv_cache(cls) -> bool:
        return True

    @classmethod
    def support_static_graph_mode(cls) -> bool:
        return True

    @classmethod
    def support_deep_gemm(cls) -> bool:
        """Currently, only Hopper and Blackwell GPUs are supported."""
        return cls.is_device_capability(90) or cls.is_device_capability_family(100)

    @classmethod
    def num_compute_units(cls, device_id: int = 0) -> int:
        return torch.cuda.get_device_properties(device_id).multi_processor_count

    @classmethod
    def use_custom_op_collectives(cls) -> bool:
        return True


# NVML utils
# Note that NVML is not affected by `CUDA_VISIBLE_DEVICES`,
# all the related functions work on real physical device ids.
# the major benefit of using NVML is that it will not initialize CUDA
class NvmlCudaPlatform(CudaPlatformBase):
    @classmethod
    @cache
    @with_nvml_context
    def get_device_capability(cls, device_id: int = 0) -> DeviceCapability | None:
        try:
            physical_device_id = cls.device_id_to_physical_device_id(device_id)
            handle = pynvml.nvmlDeviceGetHandleByIndex(physical_device_id)
            major, minor = pynvml.nvmlDeviceGetCudaComputeCapability(handle)
            return DeviceCapability(major=major, minor=minor)
        except RuntimeError:
            return None

    @classmethod
    @with_nvml_context
    def has_device_capability(
        cls,
        capability: tuple[int, int] | int,
        device_id: int = 0,
    ) -> bool:
        try:
            return super().has_device_capability(capability, device_id)
        except RuntimeError:
            return False

    @classmethod
    @with_nvml_context
    def get_device_name(cls, device_id: int = 0) -> str:
        physical_device_id = cls.device_id_to_physical_device_id(device_id)
        return cls._get_physical_device_name(physical_device_id)

    @classmethod
    @with_nvml_context
    def get_device_uuid(cls, device_id: int = 0) -> str:
        physical_device_id = cls.device_id_to_physical_device_id(device_id)
        handle = pynvml.nvmlDeviceGetHandleByIndex(physical_device_id)
        return pynvml.nvmlDeviceGetUUID(handle)

    @classmethod
    @with_nvml_context
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        physical_device_id = cls.device_id_to_physical_device_id(device_id)
        handle = pynvml.nvmlDeviceGetHandleByIndex(physical_device_id)
        return int(pynvml.nvmlDeviceGetMemoryInfo(handle).total)

    @classmethod
    @with_nvml_context
    def is_fully_connected(cls, physical_device_ids: list[int]) -> bool:
        """
        query if the set of gpus are fully connected by nvlink (1 hop)
        """
        handles = [pynvml.nvmlDeviceGetHandleByIndex(i) for i in physical_device_ids]
        for i, handle in enumerate(handles):
            for j, peer_handle in enumerate(handles):
                if i < j:
                    try:
                        p2p_status = pynvml.nvmlDeviceGetP2PStatus(
                            handle,
                            peer_handle,
                            pynvml.NVML_P2P_CAPS_INDEX_NVLINK,
                        )
                        if p2p_status != pynvml.NVML_P2P_STATUS_OK:
                            return False
                    except pynvml.NVMLError:
                        logger.exception(
                            "NVLink detection failed. This is normal if"
                            " your machine has no NVLink equipped."
                        )
                        return False
        return True

    @classmethod
    def _get_physical_device_name(cls, device_id: int = 0) -> str:
        handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
        return pynvml.nvmlDeviceGetName(handle)

    @classmethod
    @with_nvml_context
    def log_warnings(cls):
        device_ids: int = pynvml.nvmlDeviceGetCount()
        if device_ids > 1:
            device_names = [cls._get_physical_device_name(i) for i in range(device_ids)]
            if (
                len(set(device_names)) > 1
                and os.environ.get("CUDA_DEVICE_ORDER") != "PCI_BUS_ID"
            ):
                logger.warning(
                    "Detected different devices in the system: %s. Please"
                    " make sure to set `CUDA_DEVICE_ORDER=PCI_BUS_ID` to "
                    "avoid unexpected behavior.",
                    ", ".join(device_names),
                )


class NonNvmlCudaPlatform(CudaPlatformBase):
    @classmethod
    @cache
    def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:
        major, minor = torch.cuda.get_device_capability(device_id)
        return DeviceCapability(major=major, minor=minor)

    @classmethod
    def get_device_name(cls, device_id: int = 0) -> str:
        return torch.cuda.get_device_name(device_id)

    @classmethod
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        device_props = torch.cuda.get_device_properties(device_id)
        return device_props.total_memory

    @classmethod
    def is_fully_connected(cls, physical_device_ids: list[int]) -> bool:
        logger.exception(
            "NVLink detection not possible, as context support was"
            " not found. Assuming no NVLink available."
        )
        return False


# Autodetect either NVML-enabled or non-NVML platform
# based on whether NVML is available.
nvml_available = False
try:
    try:
        pynvml.nvmlInit()
        nvml_available = True
    except Exception:
        # On Jetson, NVML is not supported.
        nvml_available = False
finally:
    if nvml_available:
        pynvml.nvmlShutdown()

CudaPlatform = NvmlCudaPlatform if nvml_available else NonNvmlCudaPlatform

CudaPlatform.log_warnings()