Skip to content

vllm.model_executor.layers.quantization.kernels.scaled_mm

Modules:

Name Description
ScaledMMLinearKernel
aiter
cutlass
triton
xla

_POSSIBLE_KERNELS module-attribute 

_POSSIBLE_KERNELS: dict[
    PlatformEnum, list[type[ScaledMMLinearKernel]]
] = {
    CPU: [CutlassScaledMMLinearKernel],
    CUDA: [CutlassScaledMMLinearKernel],
    ROCM: [
        AiterScaledMMLinearKernel,
        TritonScaledMMLinearKernel,
    ],
    TPU: [XLAScaledMMLinearKernel],
}

choose_scaled_mm_linear_kernel 

choose_scaled_mm_linear_kernel(
    config: ScaledMMLinearLayerConfig,
    compute_capability: Optional[int] = None,
) -> type[ScaledMMLinearKernel]

Choose an ScaledMMLinearKernel that can implement the given config for the given compute capability. Attempts to choose the best kernel in terms of performance.

Parameters:

Name Type Description Default
config ScaledMMLinearLayerConfig

Description of the linear layer to be implemented.

 required
compute_capability Optional[int]

The compute capability of the target device, if None uses current_platform to get the compute capability. Defaults to None.

 None

Raises:

Type Description
ValueError

If no kernel can implement the given config.

Returns:

Type Description
type[ScaledMMLinearKernel]

type[ScaledMMLinearKernel]: Chosen kernel.
Source code in vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
def choose_scaled_mm_linear_kernel(
        config: ScaledMMLinearLayerConfig,
        compute_capability: Optional[int] = None
) -> type[ScaledMMLinearKernel]:
    """
    Choose an ScaledMMLinearKernel that can implement the given config for the 
    given compute capability. Attempts to choose the best kernel in terms of 
    performance.

    Args:
        config (ScaledMMLinearLayerConfig): Description of the linear layer 
            to be implemented.
        compute_capability (Optional[int], optional): The compute capability of
            the target device, if None uses `current_platform` to get the 
            compute capability. Defaults to None.

    Raises:
        ValueError: If no kernel can implement the given config.

    Returns:
        type[ScaledMMLinearKernel]: Chosen kernel.
    """

    if compute_capability is None:
        _cc = current_platform.get_device_capability()
        if _cc is not None:
            compute_capability = _cc[0] * 10 + _cc[1]

    failure_reasons = []
    for kernel in _POSSIBLE_KERNELS[current_platform._enum]:
        if kernel.__name__ in os.environ.get("VLLM_DISABLED_KERNELS", "")\
            .split(","):
            failure_reasons.append(
                f' {kernel.__name__} disabled by environment variable')
            continue

        # If the current platform uses compute_capability,
        # make sure the kernel supports the compute cability.
        if compute_capability is not None:
            kernel_min_capability = kernel.get_min_capability()
            if (kernel_min_capability is not None
                    and kernel_min_capability > compute_capability):
                failure_reasons.append(
                    f"{kernel.__name__} requires capability "
                    f"{kernel_min_capability}, current compute capability "
                    f"is {compute_capability}")
                continue

        can_implement, failure_reason = kernel.can_implement(config)
        if can_implement:
            return kernel
        else:
            failure_reasons.append(
                f' {kernel.__name__} cannot implement due to: {failure_reason}'
            )

    raise ValueError(
        "Failed to find a kernel that can implement the "\
        "ScaledMM linear layer. Reasons: \n"
        + '\n'.join(failure_reasons))