Kai Londenberg

Addition of HiLU (Hinge Linear Unit ) parametric nonlinearity

1

A flexible and simple parametric invertitble nonlinearity, which separately parameterizes the slope for negative and positive elements. Efficient calculation of forward/inverse and logabsdet due to parametrization in log-space.

Full-rank orthogonal transforms with settable parameter matrix

1

The difference to your existing implementation ( HouseholderSequence ) is, that this implementation allows to reverse engineer a sequence of householder reflections given an orthogonal matrix. This allows to use...

Ensure safe build dir name

8

Summary: In D44642328 we saw a test failure related to some test generating a test name which contains characters which might be unsafe for file or directory names ( e.g....

Hotfix: Prevent race conditions in unit tests by locking the build directory.

5

Summary: During recent CI runs, strange test failures unrelated to the diff being under test were encountered. It turned out that these were also happening on the main branch, possibly...

Fix build dir naming in test_activation.py

8

Summary: This fixes the naming of build directories in test_activation.py, so that each invocation of compile_model or safe_compile_model gets passed a unique build dir. Stacked on top of D44421367 but...

Introduce safe_compile_model with locking

1

Summary: This diff extends D44374161 in the following ways: one criticism with respect to compile_model locking it's build dir is, that this is kind of unexpected behavior and not part...

[Inductor Cutlass backend] 1 of 2 - Enabling flexible EVT-based pointwise fusions with additional tensor input

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #119701 * #119600 * #119009 * #119008 * #119601 * __->__ #119007 * #119598 * #119597 * #119006 * #119004 This diff enables...

[Inductor Cutlass backend] Development feature flag

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #119701 * #119600 * #119009 * #119008 * #119601 * #119007 * #119598 * __->__ #119597 * #119006 * #119004 Introducing a feature...

[inductor max autotune] Detailed autotuning result logs ( machine-readable )

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #119701 * #119600 * #119009 * #119008 * #119601 * #119007 * #119598 * #119597 * #119006 * __->__ #119004 This diff introduces...

[BUG] Misaligned address when running GEMM with SM90 EVT-based epilogue

10

When running a standalone Cutlass GEMM with a generated SM90 EVT-based epilogue which loads two auxiliary inputs ( one broadcasted, one with full dimensionality), I get a CUDA error about...