Pearu Peterson

The exception is raised in https://github.com/openxla/xla/blob/958489d78808fc152c192ab003b405a46c61aba8/xla/pjrt/c_api_client/pjrt_c_api_client.cc#L798 When considering the definition of HostBufferSemantics: https://github.com/openxla/xla/blob/958489d78808fc152c192ab003b405a46c61aba8/xla/pjrt/pjrt_client.h#L878 the value of host_buffer_semantics must be kMutableZeroCopy. For a non-CPU platform kMutableZeroCopy is identical to kImmutableUntilTransferCompletes. So,...

> Furthermore, an exception occurred in the torch.mm within the fbank method, ultimately leading to an unexpected segmentation fault. What was the exception and which torch version are you using?...

JAX arrays on CPU use FTZ mode, that is, subnormal numbers (like 1e-45) are flushed to zeros: ```python >>> with jax.default_device(jax.devices('cpu')[0]): ... jax.numpy.array(numpy.array([-0.0, 1.401298464324817e-45, -0.0], numpy.float32)) == 0 ... Array([...

In processors that support the FTZ flag, enabling FTZ is optional. PyTorch obviously does not enable FTZ mode while JAX (read: some of its underlying component) does on CPU. >...

FWIW, handling subnormals in a device-dependent way complicates testing on samples with subnormals. Here's another example of the issue reported here: https://github.com/pearu/functional_algorithms/issues/38#issuecomment-2366843504 where the results of evaluating math functions near...

Quoting [Wikipedia article on ULP](https://en.wikipedia.org/wiki/Unit_in_the_last_place): "Reputable libraries compute the basic transcendental functions to between 0.5 and about 1 ulp." So, I think seldom 1 ULP difference in PyTorch and JAX...

Closing as currently irrelevant (docs/source/build.ffmpeg.rst does not exist in main)