Skip to content

vllm.kernels.oink_ops

This file registers Oink implementations for vLLM IR ops.

vLLM does not depend on the external Oink repository/package. When an external plugin registers torch.library.custom_op entrypoints under the oink:: namespace (e.g. via vLLM's general_plugins mechanism), these ops will be marked as supported. To dispatch to those ops, set kernel_config.ir_op_priority. to oink. Alternatively, VLLM_USE_OINK_OPS=1 will add this to priority by default.

Functions:

  • has_oink_op –

    Check if a specific oink op is registered.

Attributes:

oink_add_rms_supported = lambda x, x_residual, weight, epsilon, variance_size=None: variance_size is None and weight is not None and x.dim() >= 2 and x.dtype == weight.dtype and weight.is_contiguous() and _can_view_as_2d(x) and _is_oink_stride_compatible_2d(x.view(-1, x.shape[-1])) and x.dtype == x_residual.dtype and x.shape == x_residual.shape and _can_view_as_2d(x_residual) and _is_oink_stride_compatible_2d(x_residual.view(-1, x_residual.shape[-1])) module-attribute

Oink fused_add_rms_norm has the same constraints as rms_norm, and residual must be 2d-like with compatible strides.

oink_rms_supported = lambda x, weight, epsilon, variance_size=None: variance_size is None and weight is not None and x.dim() >= 2 and x.dtype == weight.dtype and weight.is_contiguous() and _can_view_as_2d(x) and _is_oink_stride_compatible_2d(x.view(-1, x.shape[-1])) module-attribute

Oink rms only supports 2d-like inputs with contiguous weight and no variance_size override.

_can_view_as_2d(x)

Return True if x.view(-1, x.shape[-1]) is viewable (no copy).

Source code in vllm/kernels/oink_ops.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
def _can_view_as_2d(x: Tensor) -> bool:
    """Return True if x.view(-1, x.shape[-1]) is viewable (no copy)."""
    if x.dim() < 2:
        return False
    if x.dim() == 2:
        return True
    # For a view(-1, N) to be valid, all leading dims must be contiguous with
    # respect to each other (size-1 dims are ignored).
    for dim in range(x.dim() - 1):
        # Strides for size-1 dims are irrelevant and can be arbitrary.
        if x.size(dim + 1) != 1 and x.stride(dim) != x.stride(dim + 1) * x.size(
            dim + 1
        ):
            return False
    return True

_is_oink_stride_compatible_2d(x_2d)

Return True if x_2d meets Oink's pointer-path stride constraints.

Source code in vllm/kernels/oink_ops.py 
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def _is_oink_stride_compatible_2d(x_2d: Tensor) -> bool:
    """Return True if x_2d meets Oink's pointer-path stride constraints."""
    if x_2d.dim() != 2:
        return False
    if x_2d.stride(1) != 1:
        return False
    # Match Oink's vectorization constraint: stride(0) divisible by 256b.
    if x_2d.dtype in (torch.float16, torch.bfloat16):
        divby = 16
    elif x_2d.dtype == torch.float32:
        divby = 8
    else:
        return False
    return (x_2d.stride(0) % divby) == 0

has_oink_op(name)

Check if a specific oink op is registered.

Source code in vllm/kernels/oink_ops.py 
23
24
25
def has_oink_op(name: str) -> bool:
    """Check if a specific oink op is registered."""
    return OINK_AVAILABLE and hasattr(torch.ops.oink, name)