Skip to content

vllm.v1.attention.ops.flashmla

Functions:

is_flashmla_dense_supported()

Return: is_supported_flag, unsupported_reason (optional).

Source code in vllm/v1/attention/ops/flashmla.py 
51
52
53
54
55
56
57
58
59
60
def is_flashmla_dense_supported() -> tuple[bool, str | None]:
    """
    Return: is_supported_flag, unsupported_reason (optional).
    """
    is_available, maybe_reason = _is_flashmla_available()
    if not is_available:
        return False, maybe_reason
    if not current_platform.is_device_capability_family(90):
        return False, "FlashMLA Dense is only supported on Hopper devices."
    return True, None

is_flashmla_sparse_supported()

Return: is_supported_flag, unsupported_reason (optional).

Source code in vllm/v1/attention/ops/flashmla.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
def is_flashmla_sparse_supported() -> tuple[bool, str | None]:
    """
    Return: is_supported_flag, unsupported_reason (optional).
    """
    is_available, maybe_reason = _is_flashmla_available()
    if not is_available:
        return False, maybe_reason
    if not (
        current_platform.is_device_capability_family(90)
        or current_platform.is_device_capability_family(100)
    ):
        return (
            False,
            "FlashMLA Sparse is only supported on Hopper and Blackwell devices.",
        )
    return True, None