tilelang.profiler.torch_bench
=============================

.. py:module:: tilelang.profiler.torch_bench

.. autoapi-nested-parse::

   PyTorch GPU timing implementations.



Classes
-------

.. autoapisummary::

   tilelang.profiler.torch_bench.suppress_stdout_stderr


Functions
---------

.. autoapisummary::

   tilelang.profiler.torch_bench.bench_with_cuda_events
   tilelang.profiler.torch_bench.bench_with_cupti
   tilelang.profiler.torch_bench.bench_with_cudagraph


Module Contents
---------------

.. py:class:: suppress_stdout_stderr

   Context manager to suppress stdout and stderr output.

   Source: https://github.com/deepseek-ai/DeepGEMM/blob/main/deep_gemm/testing/bench.py


   .. py:method:: __enter__()


   .. py:method:: __exit__(*_)


.. py:function:: bench_with_cuda_events(fn, cache, n_repeat, quantiles, return_mode, device_idx)

   Benchmark using CUDA/HIP or MPS events for timing.


.. py:function:: bench_with_cupti(fn, cache, n_repeat)

   Benchmark using CUPTI profiler for detailed kernel timing.


.. py:function:: bench_with_cudagraph(fn, cache, n_repeat, quantiles, return_mode, device_idx)

   Benchmark using CUDA graph for minimal launch overhead.

   This implementation follows triton.testing.do_bench_cudagraph.
   It captures the kernel execution in a CUDA graph and replays it multiple
   times to minimize host overhead and provide accurate timing measurements.

   Note: Cache flushing is done before graph replay, not within the graph,
   since CUDA graphs require fixed execution patterns.