tvm_tirx.testing.bench

Benchmark and profile GPU launch callables, including distributed execution.

Classes

DistributedBenchContext(rank, world_size, ...)	Rank-local synchronization needed by the distributed Kineto timer.
EventType(*values)
CudaProfiler(*args, **kwargs)	A lightweight wrapper around T.timer_* CUDA intrinsics.

Functions

is_running_under_pytest()	Check if the code is being executed within a pytest session.
setup()	Parse benchmark compile settings for explicit forwarding by the caller.
bench(funcs, *[, warmup, repeat, ...])	Benchmark pure-launch implementations using Triton-standard timing.
decode_tag(tag, num_groups)
export_to_perfetto_trace(profiler_buffer, ...)