vllm-omni/vllm_omni/platforms/xpu/platform.py

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

import os

import torch
from vllm.config import VllmConfig
from vllm.config.kernel import IrOpPriorityConfig
from vllm.logger import init_logger
from vllm.platforms.xpu import XPUPlatform

from vllm_omni.diffusion.attention.backends.registry import DiffusionAttentionBackendEnum
from vllm_omni.platforms.interface import OmniPlatform, OmniPlatformEnum

logger = init_logger(__name__)


# Use the native query; vLLM's XPU custom op reports free=0 in spawned workers.
torch.accelerator.get_memory_info = lambda device=None: torch.xpu.mem_get_info(device)

# The XPU sampler kernel is broken for omni on the current vLLM; use the native path.
os.environ.setdefault("VLLM_XPU_USE_SAMPLER_KERNEL", "0")


class XPUOmniPlatform(OmniPlatform, XPUPlatform):
    """XPU/Intel GPU implementation of OmniPlatform.

    Inherits all XPU-specific implementations from vLLM's XPUPlatform,
    and adds Omni-specific interfaces from OmniPlatform.
    """

    _omni_enum = OmniPlatformEnum.XPU

    @classmethod
    def get_omni_ar_worker_cls(cls) -> str:
        return "vllm_omni.platforms.xpu.worker.xpu_ar_worker.XPUARWorker"

    @classmethod
    def get_omni_generation_worker_cls(cls) -> str:
        return "vllm_omni.platforms.xpu.worker.xpu_generation_worker.XPUGenerationWorker"

    @classmethod
    def get_diffusion_attn_backend_cls(
        cls,
        selected_backend: str | None,
        head_size: int,
        allow_trtllm_default: bool = False,
    ) -> str:
        # XPU has no TRTLLM backend; arg accepted for signature parity but unused.
        compute_capability = torch.xpu.get_device_capability()
        # Intel Max 1100 and 1550 will not support flash_attn currently
        flash_attn_supported = compute_capability["architecture"] not in [13136561920]

        if selected_backend is not None:
            backend_upper = selected_backend.upper()
            cls.validate_diffusion_attn_backend(backend_upper)
            if backend_upper in ("FLASH_ATTN_HUB", "FLASH_ATTN_3_HUB"):
                logger.warning(
                    "HuggingFace kernels-backed FlashAttention is "
                    "not supported on XPU. Falling back to local "
                    "FLASH_ATTN."
                )
                backend_upper = "FLASH_ATTN"

            backend = DiffusionAttentionBackendEnum[backend_upper]
            logger.debug("Using diffusion attention backend '%s'", backend_upper)
            return backend.get_path()

        if flash_attn_supported:
            logger.debug("Defaulting to diffusion attention backend FLASH_ATTN")
            return DiffusionAttentionBackendEnum.FLASH_ATTN.get_path()

        logger.debug("Defaulting to diffusion attention backend SDPA")
        return DiffusionAttentionBackendEnum.TORCH_SDPA.get_path()

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

    @classmethod
    def get_default_stage_config_path(cls) -> str:
        return "vllm_omni/platforms/xpu/stage_configs"

    @classmethod
    def get_torch_device(cls, local_rank: int | None = None) -> torch.device:
        if local_rank is None:
            return torch.device("xpu")
        return torch.device("xpu", local_rank)

    @classmethod
    def get_device_count(cls) -> int:
        return torch.xpu.device_count()

    @classmethod
    def get_device_version(cls) -> str | None:
        # XPU does not have a version string like CUDA
        return None

    @classmethod
    def synchronize(cls) -> None:
        torch.xpu.synchronize()

    @classmethod
    def record_device_event(cls) -> torch.Event | None:
        """Record an XPU event on the current stream to mark tensor readiness.

        Deliberately a device-agnostic ``torch.Event`` rather than a
        ``torch.xpu.Event``. The consumer (the async diffusion output thread)
        waits with ``torch.Stream.wait_event`` on a generic ``torch.Stream``,
        and that C-level binding silently no-ops for a ``torch.xpu.Event``
        instead of enqueuing the dependency — the side stream then starts its
        D2H copy while the compute stream is still writing the tensor, so the
        host reads a partially-written image (garbage rows at the bottom of the
        output). ``torch.Event`` dispatches through the accelerator hooks and
        the wait is honored, which is the actual fix.
        """
        try:
            event = torch.Event()
            event.record()
            return event
        except Exception:
            logger.warning("Failed to record XPU device event for cross-stream sync")
            return None

    @classmethod
    def get_free_memory(cls, device: torch.device | None = None) -> int:
        free, _ = torch.xpu.mem_get_info(device)
        return free

    @classmethod
    def get_device_memory(cls, device: torch.device | None = None) -> tuple[int, int]:
        free, total = torch.xpu.mem_get_info(device)
        return free, total

    @classmethod
    def get_profiler_cls(cls) -> str:
        """Return XPU-specific profiler that handles XPU events."""
        return "vllm_omni.platforms.xpu.profiler.XPUTorchProfilerWrapper"

    @classmethod
    def get_default_ir_op_priority(cls, vllm_config: VllmConfig) -> IrOpPriorityConfig:
        """Copied from upstream XPUPlatform with inductor-aware logic.

        When inductor is active (compiling) use native as the default;
        otherwise prefer vllm_c where available.
        """
        from vllm.config.compilation import CompilationMode

        cc = vllm_config.compilation_config
        using_inductor = cc.backend == "inductor" and cc.mode != CompilationMode.NONE
        default = ["native"] if using_inductor else ["vllm_c", "native"]

        return IrOpPriorityConfig.with_default(default)