Cubin producer API
The following names are available from tensoracle.producer. Install tensoracle[producer] to compile cubins with Triton 3.8.0, ptxas 12.9.86, and cuobjdump 13.3. compile_bundle finds the NVIDIA wheel executables when their paths are omitted; explicit ptxas and cuobjdump paths override the defaults. Normal imports, engine building, and engine reuse do not import the producer or Triton.
KernelSpec
dataclass
One caller-owned Triton entry and its public launch arguments.
name
instance-attribute
function
instance-attribute
signature
instance-attribute
constexprs
instance-attribute
divisible_by = field(default_factory=dict)
class-attribute
instance-attribute
__init__(name, function, signature, constexprs, divisible_by=dict())
compile_bundle(specs, output, *, ptxas=None, cuobjdump=None)
Produce and inspect sm86, sm89, and sm120 cubins without importing torch.
Omitted tool paths use the NVIDIA wheels installed by tensoracle[producer]. Explicit ptxas and cuobjdump paths override those defaults independently.