evox.triton_kernels.op_register

模块内容

函数

register_triton_op

Register an operator with a PyTorch fallback and an optional Triton CUDA kernel.

API

evox.triton_kernels.op_register.register_triton_op(*, fake_fn, triton_fn, vmap_fn=None, fake_vmap_fn=None, vmap_wrap_inputs=None, vmap_out_dims=0, max_vmap_level=None, name=None, mutates_args=(), device_types=None, triton_device_types: str | Sequence[str] | None = None, schema=None)[源代码]

Register an operator with a PyTorch fallback and an optional Triton CUDA kernel.

This decorator wraps :func:evox.utils.register_vmap_op, registering the decorated function (fallback_fn) as the default implementation with fake / vmap support, and additionally registering triton_fn as a CUDA-specific kernel via

Func:

torch.library.register_kernel.

PyTorch's dispatcher automatically selects the correct backend at call time:

  • CUDA tensors → Triton kernel (triton_fn)

  • CPU / MPS / other tensors → PyTorch fallback (the decorated function)

If Triton is not installed, only the PyTorch fallback is registered and the operation works on all devices without Triton.

参数:

  • fake_fn -- The fake (abstract evaluation) function for the op. Required.

  • triton_fn -- The Triton CUDA kernel launcher. Called with the same arguments as the decorated function when dispatched on CUDA. Must return outputs matching the fake function's shapes/dtypes.

  • vmap_fn -- Optional vmap implementation. See :func:register_vmap_op.

  • fake_vmap_fn -- Optional fake vmap function. See :func:register_vmap_op.

  • vmap_wrap_inputs -- Optional input wrapper for vmap. See :func:register_vmap_op.

  • vmap_out_dims -- Output vmap dimensions. See :func:register_vmap_op.

  • max_vmap_level -- Maximum vmap nesting level. See :func:register_vmap_op.

  • name -- Custom op name. Default "evox::_custom_op_" + fn.__name__.

  • mutates_args -- Args mutated by the op. See :func:register_vmap_op.

  • device_types -- Supported device types for the op registration.

  • triton_device_types -- Device type(s) for which the Triton kernel is registered via :func:torch.library.register_kernel. When None (default), uses the globally registered device types (:func:~evox.triton_kernels.backend.triton_device_types, default {"cuda"}). Accepts a single string (e.g. "npu") or a sequence of strings (e.g. ["cuda", "npu"]) to extend or restrict coverage.

  • schema -- Op schema string.

样例

@triton.jit
def _add_kernel(x_ptr, y_ptr, out_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
    pid = tl.program_id(0)
    offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
    mask = offsets < n_elements
    x = tl.load(x_ptr + offsets, mask=mask)
    y = tl.load(y_ptr + offsets, mask=mask)
    tl.store(out_ptr + offsets, x + y, mask=mask)

def _triton_add(x, y):
    out = torch.empty_like(x)
    n = x.numel()
    grid = lambda meta: (triton.cdiv(n, meta["BLOCK_SIZE"]),)
    _add_kernel[grid](x, y, out, n, BLOCK_SIZE=1024)
    return out

def _add_fake(x, y):
    return torch.empty_like(x)

@register_triton_op(fake_fn=_add_fake, triton_fn=_triton_add)
def fused_add(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    return x + y