rocm.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import os
from datetime import timedelta
from functools import cache, lru_cache, wraps
from typing import TYPE_CHECKING, Optional

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

import vllm.envs as envs
from vllm.logger import init_logger
from vllm.utils import cuda_device_count_stateless

from .interface import DeviceCapability, Platform, PlatformEnum, _Backend


if TYPE_CHECKING:
    from vllm.config import ModelConfig, VllmConfig

logger = init_logger(__name__)

try:
    from amdsmi import (AmdSmiException, amdsmi_get_gpu_asic_info,
                        amdsmi_get_processor_handles, amdsmi_init,
                        amdsmi_shut_down, amdsmi_topo_get_link_type)
except ImportError as e:
    logger.warning("Failed to import from amdsmi with %r", e)

try:
    import vllm._C  # noqa: F401
except ImportError as e:
    logger.warning("Failed to import from vllm._C with %r", e)

# import custom ops, trigger op registration
# try:
#     import vllm._rocm_C  # noqa: F401
# except ImportError as e:
#     logger.warning("Failed to import from vllm._rocm_C with %r", e)


# Models not supported by ROCm.
_ROCM_UNSUPPORTED_MODELS: list[str] = []

# Models partially supported by ROCm.
# Architecture -> Reason.
# _ROCM_SWA_REASON = ("Sliding window attention (SWA) is not yet supported in "
#                     "Triton flash attention. For half-precision SWA support, "
#                     "please use CK flash attention by setting "
#                     "`VLLM_USE_TRITON_FLASH_ATTN=0`")
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {
    # "Qwen2ForCausalLM":
    # _ROCM_SWA_REASON,
    # "MistralForCausalLM":
    # _ROCM_SWA_REASON,
    # "MixtralForCausalLM":
    # _ROCM_SWA_REASON,
    # "PaliGemmaForConditionalGeneration":
    # ("ROCm flash attention does not yet "
    #  "fully support 32-bit precision on PaliGemma"),
    # "Phi3VForCausalLM":
    # ("ROCm Triton flash attention may run into compilation errors due to "
    #  "excessive use of shared memory. If this happens, disable Triton FA "
    #  "by setting `VLLM_USE_TRITON_FLASH_ATTN=0`")
}
_ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
    "0x74a0": "AMD_Instinct_MI300A",
    "0x74a1": "AMD_Instinct_MI300X",
    "0x74b5": "AMD_Instinct_MI300X",  # MI300X VF
    "0x74a5": "AMD_Instinct_MI325X",
    "0x74b9": "AMD_Instinct_MI325X",  # MI325X VF
    "0x74a9": "AMD_Instinct_MI300X_HF",
    "0x74bd": "AMD_Instinct_MI300X_HF",
}

# Prevent use of clashing `{CUDA/HIP}_VISIBLE_DEVICES``
# if "HIP_VISIBLE_DEVICES" in os.environ:
#     val = os.environ["HIP_VISIBLE_DEVICES"]
#     if cuda_val := os.environ.get("CUDA_VISIBLE_DEVICES", None):
#         assert val == cuda_val
#     else:
#         os.environ["CUDA_VISIBLE_DEVICES"] = val

# AMDSMI utils
# Note that NVML is not affected by `{CUDA/HIP}_VISIBLE_DEVICES`,
# all the related functions work on real physical device ids.
# the major benefit of using AMDSMI is that it will not initialize CUDA


def with_amdsmi_context(fn):

    @wraps(fn)
    def wrapper(*args, **kwargs):
        amdsmi_init()
        try:
            return fn(*args, **kwargs)
        finally:
            amdsmi_shut_down()

    return wrapper


def device_id_to_physical_device_id(device_id: int) -> int:
    if "CUDA_VISIBLE_DEVICES" in os.environ:
        device_ids = os.environ["CUDA_VISIBLE_DEVICES"].split(",")
        physical_device_id = device_ids[device_id]
        return int(physical_device_id)
    else:
        return device_id


@cache
def on_gfx1x() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])


@cache
def on_mi3xx() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx942", "gfx950"])


@cache
def on_gfx9() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936", "gfx938"])


@cache
def on_gfx950() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx950"])


@cache
def use_rocm_custom_paged_attention(
        qtype: torch.dtype,
        head_size: int,
        block_size: int,
        gqa_ratio: int,
        max_seq_len: int,
        sliding_window: int,
        kv_cache_dtype: str,
        alibi_slopes: Optional[torch.Tensor] = None,
        sinks: Optional[torch.Tensor] = None) -> bool:

    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936"])
    ON_GFX11_GFX12 = any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])

    # custom paged attn always supported on V0. On V1, requires sliding window
    # disabled due to observed numerical discrepancy.
    # if ON_GFX9:
    #     return ((not envs.VLLM_USE_V1 or sliding_window == 0
    #              or sliding_window == (-1, -1))
    #             and (qtype == torch.half or qtype == torch.bfloat16)
    #             and (head_size == 64 or head_size == 128)
    #             and (block_size == 16 or block_size == 32)
    #             and (gqa_ratio >= 1 and gqa_ratio <= 16)
    #             and max_seq_len <= 128 * 1024
    #             and (envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)
    #             and not (envs.VLLM_ROCM_USE_AITER_PAGED_ATTN
    #                      and envs.VLLM_ROCM_USE_AITER) and sinks is None)

    # else:
    #     return (ON_GFX11_GFX12 and (not envs.VLLM_USE_V1 or sliding_window == 0
    #                                 or sliding_window == (-1, -1))
    #             and (qtype == torch.half or qtype == torch.bfloat16)
    #             and head_size == 128 and block_size == 16
    #             and (gqa_ratio >= 3 and gqa_ratio <= 16)
    #             and max_seq_len <= 128 * 1024 and alibi_slopes is None
    #             and kv_cache_dtype == "auto"
    #             and envs.VLLM_ROCM_CUSTOM_PAGED_ATTN and sinks is None)
    return False


class RocmPlatform(Platform):
    _enum = PlatformEnum.ROCM
    device_name: str = "rocm"
    device_type: str = "cuda"
    dispatch_key: str = "CUDA"
    ray_device_key: str = "GPU"
    dist_backend: str = "nccl"
    # rocm shares the same device control env var as CUDA
    device_control_env_var: str = "CUDA_VISIBLE_DEVICES"

    supported_quantization: list[str] = [
        "awq", "gptq", "fp8", "compressed-tensors", "fbgemm_fp8", "gguf",
        "quark", "ptpc_fp8", "mxfp4", "petit_nvfp4", "torchao",
        "moe_wna16", "slimquant_w4a8", "w8a8_int8", "awq_marlin", "slimquant_w4a8_marlin", "slimquant_compressed_tensors_marlin"
    ]

    @classmethod
    def get_vit_attn_backend(cls, head_size: int,
                             dtype: torch.dtype) -> _Backend:
        # if (envs.VLLM_ROCM_USE_AITER and envs.VLLM_ROCM_USE_AITER_MHA
        #         and on_gfx9()):
        #     # Note: AITER FA is only supported for Qwen-VL models.
        #     # TODO: Add support for other VL models in their model class.
        #     return _Backend.ROCM_AITER_FA
        if on_gfx9():
            return _Backend.FLASH_ATTN
        return _Backend.TORCH_SDPA

    @classmethod
    def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
                             kv_cache_dtype, block_size, use_v1, use_mla,
                             has_sink, use_sparse) -> str:
        # if use_sparse:
        #     raise NotImplementedError(
        #         "Sparse Attention is not supported on ROCm.")
        if use_mla:
            if not use_v1:
                raise RuntimeError(
                    "MLA attention backends require the V1 engine. "
                    "Set VLLM_USE_V1=1 to enable them.")
                
            
            from vllm.attention.ops.flashmla import is_flashmla_supported
            from vllm.attention.utils.fa_utils import flash_attn_supports_mla

            if use_sparse:
                logger.info_once("Using Sparse MLA backend on V1 engine.")
                return ("vllm.v1.attention.backends.mla.flashmla_sparse."
                        "FlashMLASparseBackend")
                
            use_flashmla = selected_backend == _Backend.FLASHMLA or envs.VLLM_USE_FLASH_MLA or (
                selected_backend is None and is_flashmla_supported()[0])
            use_triton = selected_backend == _Backend.TRITON_MLA or (
                selected_backend is None)
            
            if use_flashmla: 
                if block_size != 64:
                    logger.warning(
                        "FlashMLA backend is not supported for block size %d"
                        " (currently only supports block size 64).",
                        block_size)
                else:
                    logger.info_once("Using FlashMLA backend on V1 engine.")
                    return ("vllm.v1.attention.backends.mla."
                            "flashmla.FlashMLABackend")
                    
            if use_triton:
                logger.info_once("Using Triton MLA backend on V1 engine.")
                return ("vllm.v1.attention.backends.mla."
                        "triton_mla.TritonMLABackend")
        
        if envs.VLLM_USE_V1:
            TRITON_ATTN = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend"  # noqa: E501
            FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"  # noqa: E501
            
            if envs.VLLM_USE_FLASH_ATTN_PA and block_size == 64:
                logger.info_once("Using Flash Attention backend on V1 engine. (only supports block size 64)")
                return FLASH_ATTN_V1
            else:
                os.environ['VLLM_USE_FLASH_ATTN_PA'] = '0'
                logger.info_once("Using Triton backend on V1 engine.")
                return TRITON_ATTN
            
        raise RuntimeError(
            "V0 attention backends have been removed. Set VLLM_USE_V1=1 "
            "to select a supported backend.")

    @classmethod
    def set_device(cls, device: torch.device) -> None:
        """
        Set the device for the current platform.
        """
        torch.cuda.set_device(device)

    @classmethod
    @lru_cache(maxsize=8)
    def get_device_capability(cls,
                              device_id: int = 0
                              ) -> Optional[DeviceCapability]:
        major, minor = torch.cuda.get_device_capability(device_id)
        return DeviceCapability(major=major, minor=minor)

    @classmethod
    @with_amdsmi_context
    def is_fully_connected(cls, physical_device_ids: list[int]) -> bool:
        """
        Query if the set of gpus are fully connected by xgmi (1 hop)
        """
        handles = [
            amdsmi_get_processor_handles()[i] for i in physical_device_ids
        ]
        for i, handle in enumerate(handles):
            for j, peer_handle in enumerate(handles):
                if i < j:
                    try:
                        link_type = amdsmi_topo_get_link_type(
                            handle, peer_handle)
                        # type is 2 for XGMI
                        if link_type["hops"] != 1 or link_type["type"] != 2:
                            return False
                    except AmdSmiException as error:
                        logger.error("AMD 1 hop XGMI detection failed.",
                                     exc_info=error)
                        return False
        return True

    @classmethod
    @with_amdsmi_context
    @lru_cache(maxsize=8)
    def get_device_name(cls, device_id: int = 0) -> str:
        physical_device_id = device_id_to_physical_device_id(device_id)
        handle = amdsmi_get_processor_handles()[physical_device_id]
        # return amdsmi_get_gpu_asic_info(handle)["market_name"]
        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 check_and_update_config(cls, vllm_config: "VllmConfig") -> None:
        from vllm.config.compilation import CUDAGraphMode

        cache_config = vllm_config.cache_config
        compilation_config = vllm_config.compilation_config
        parallel_config = vllm_config.parallel_config
        is_eager_execution = compilation_config == CUDAGraphMode.NONE

        use_v1 = envs.VLLM_USE_V1
        use_aiter_rms_norm = envs.VLLM_ROCM_USE_AITER and \
             envs.VLLM_ROCM_USE_AITER_RMSNORM

        if cache_config and cache_config.block_size is None:
            cache_config.block_size = 16

        if parallel_config.worker_cls == "auto":
            if vllm_config.speculative_config:
                if not use_v1:
                    raise NotImplementedError(
                        "Speculative decoding is not supported on vLLM V0.")
                parallel_config.worker_cls = "vllm.v1.worker.gpu_worker.Worker"
            else:
                if use_v1:
                    parallel_config.worker_cls = \
                        "vllm.v1.worker.gpu_worker.Worker"
                else:
                    parallel_config.worker_cls = "vllm.worker.worker.Worker"
        #  Aiter rms norm perform best when CUDA Graph capture is enabled.
        if (use_v1 and use_aiter_rms_norm and not is_eager_execution
                and "-rms_norm" not in compilation_config.custom_ops):
            compilation_config.custom_ops.append("+rms_norm")

    @classmethod
    def verify_model_arch(cls, model_arch: str) -> None:
        if model_arch in _ROCM_UNSUPPORTED_MODELS:
            raise ValueError(f"Model architecture '{model_arch}' is not "
                             "supported by ROCm for now.")

        if model_arch in _ROCM_PARTIALLY_SUPPORTED_MODELS:
            msg = _ROCM_PARTIALLY_SUPPORTED_MODELS[model_arch]
            logger.warning(
                "Model architecture '%s' is partially "
                "supported by ROCm: %s", model_arch, msg)

    @classmethod
    def verify_quantization(cls, quant: str) -> None:
        super().verify_quantization(quant)
        if quant == "awq" and not envs.VLLM_USE_TRITON_AWQ:
            logger.warning(
                "Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ"
                " is not set, disabling VLLM_USE_TRITON_AWQ.")
            envs.VLLM_USE_TRITON_AWQ = False

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

    @classmethod
    def get_current_memory_usage(cls,
                                 device: Optional[torch.types.Device] = None
                                 ) -> float:
        torch.cuda.reset_peak_memory_stats(device)
        # return torch.cuda.mem_get_info(device)[1] - torch.cuda.mem_get_info(
        #     device)[0]
        return torch.cuda.max_memory_allocated(device)

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

    @classmethod
    def supports_mx(cls) -> bool:
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        return any(gfx in gcn_arch for gfx in ["gfx95"])

    @classmethod
    def supports_fp8(cls) -> bool:
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        return any(gfx in gcn_arch for gfx in ['gfx94', 'gfx95', 'gfx12'])

    @classmethod
    def is_fp8_fnuz(cls) -> bool:
        # only device 0 is checked, this assumes MI300 platforms are homogeneous
        return 'gfx94' in torch.cuda.get_device_properties(0).gcnArchName

    @classmethod
    def fp8_dtype(cls) -> torch.dtype:
        if cls.is_fp8_fnuz():
            return torch.float8_e4m3fnuz
        else:
            return torch.float8_e4m3fn

    @classmethod
    def use_custom_allreduce(cls) -> bool:
        # We only enable custom allreduce for MI300 series
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        supported_archs = ['gfx94', 'gfx95']
        return any(gfx in gcn_arch for gfx in supported_archs)

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

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

    @classmethod
    def is_navi(cls) -> bool:
        return 'gfx1' in torch.cuda.get_device_properties(0).gcnArchName

    @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()

    @classmethod
    def is_kv_cache_dtype_supported(cls, kv_cache_dtype: str,
                                    model_config: "ModelConfig") -> bool:
        return True

    @classmethod
    def check_if_supports_dtype(cls, torch_dtype: torch.dtype):
        if torch_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 support_hybrid_kv_cache(cls) -> bool:
        return True

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