Question: Will there be plans for non-binary / source distribution of TileRT CUDA runtime (libtilert.so)

[A-suozhang]

Thanks for the great work on this project! I’m interested in adapting and extending the runtime to support additional GPU architectures and models. While exploring the repository, I noticed that the current runtime is distributed in binary form, and the corresponding CUDA / C++ source files (e.g., .cu, .cpp) do not appear to be included.

I was wondering whether the runtime is intended to remain binary-only, or if there might be a plan or roadmap to open-source this part in the future. Having access to the runtime source would be very helpful for evaluating and potentially integrating TileRT into our stack.

Current State

• The repository provides MIT-licensed Python source code (under the python/ directory).
• The CUDA/runtime implementation appears to be distributed as a precompiled binary libtilert.so inside the PyPI wheel / Docker image.
• I was not able to find the corresponding CUDA / C++ source files (e.g., .cu, .cpp) for libtilert.so in this repository.

Question

Are there any plans to provide a non-binary distribution option for the TileRT runtime, e.g.:

• open-sourcing the CUDA / C++ source code behind libtilert.so, or
• offering a “from-source” build path for the runtime (even if it is marked as experimental)?

[tunglinwood]

Same question. But I think Tilelang would prefer not to because they might want to monetize.