Stella Laurenzo

With CUDA disabled, the `benchmark_suites/TF/vmfb/minilm-l12-h384-uncased-seqlen128` benchmark is attempting to be compiled anyway. This causes an error like: ``` iree-compile: Unknown command line argument '--iree-hal-cuda-llvm-target-arch=sm_80'. Try: '/usr/local/google/home/laurenzo/src/iree-build/tools/iree-compile --help' iree-compile: Did you...

When --iree-llvmcpu-sanitize=address is enabled, we should auto set --iree-llvmcpu-link-embedded=false

12

The test suite futzes with this at the CMake level but that still leaves the footgun in the compiler. Specifying the first flag causes an opaque link error about missing...

Pending LLVM Dependent Patches

3

We keep this issue open to track local patches that should be applied once integration has caught up. This issue is pinned to the #llvm-integrate channel on Discord. In order...

Compilation errors in CUDA HAL Driver with tracing features enabled

4

``` cmake -DIREE_BUILD_COMPILER=OFF -DIREE_ENABLE_RUNTIME_TRACING=ON -DIREE_TRACING_PROVIDER=console -DIREE_HAL_DRIVER_CUDA=ON . ninja ``` Compilation errors: ``` FAILED: runtime/src/iree/hal/drivers/cuda/CMakeFiles/iree_hal_drivers_cuda_cuda.objects.dir/tracing.c.o ccache /usr/bin/clang -DIREE_TRACING_MODE=2 -DIREE_TRACING_PROVIDER_H=\"iree/base/tracing/console.h\" -I/home/stella/src/iree -I/home/stella/src/iree-build -I/home/stella/src/iree/runtime/src -I/home/stella/src/iree-build/runtime/src -I/home/stella/src/iree-build/build_tools/third_party/cuda/12.2.1/linux-x86_64/include -I/home/stella/src/iree/third_party/nccl -I/home/stella/src/iree-build/runtime/src/iree/base/internal/flatcc -I/home/stella/src/iree-build/runtime/src/iree/schemas -isystem /home/stella/src/iree/third_party/flatcc/include -O2 -g...

XRT build drops a `kernel-doc` directory in the PWD

2

When I invoke the build from an arbitrary location, it drops a `kernel-doc` directory in the current PWD. Recommend finding this and scoping it to the build dir (or something).

DISABLE_ABI_CHECK could probably be dropped and still have cache friendliness

1

I recently spent a little time scratching my head on why ccache seems ineffective with XRT. I noted that the hashes were changing run to run. I haven't tracked through...

[fx] Use module Operations instead of Module.

2

This was only used in certain advanced uses of the API that want to build into their own module. The MLIR `Module` class is an awkward/restrictive way to require this...

Remove build automation that is no longer used.

3

Most of this can't actually work given the token/security issues, so even if we re-introduce it, we are better off fetching from history vs carrying things that will lead folks...

torch version: 2.2.0.dev20231117 torch commit hash: 0603dcb449ccf52baea0cc6981f8fc587566f1b3 torchvision version: 0.17.0.dev20231117

Fall cleaning: Re-organize sources so that each target type is contained and optional.

3

While working on https://github.com/openxla/iree/pull/14917, I noticed that it is somewhat hard to take a dependency on torch-mlir such that one only builds deps for the target(s) of interest (in this...