[Feature Request] Provide pre-built toolchains and plugins compatible with CentOS 7 (GLIBC 2.17)

[Zhang-kg]

🚀 The feature, motivation and pitch

Environment

OS: CentOS 7.9

GLIBC version: 2.17

Compiler: devtoolset-7 (GCC 7.3.1)

Backend: Iluvatar BI-V100 version 3.0.1

Issue Description

Currently, the recommended pre-built toolchains (iluvatar-llvm18-x86_64_v0.4.0.tar.gz) and plugins (iluvatarTritonPlugin-...-glibc2.35-...) are compiled against GLIBC 2.32/2.35. When building flagtree from source on CentOS 7, the process fails because mlir-tblgen cannot find the required GLIBC versions:

/root/.flagtree/iluvatar/iluvatar-llvm18-x86_64/bin/mlir-tblgen: /lib64/libc.so.6: version `GLIBC_2.33' not found
/root/.flagtree/iluvatar/iluvatar-llvm18-x86_64/bin/mlir-tblgen: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.29' not found

Request

For the Iluvatar BI-V100 GPU, the official Docker images provided by the vendor (Tianshu) are primarily based on CentOS 7 (GLIBC 2.17) and Ubuntu 20.04 (GLIBC 2.31).

Currently, the provided iluvatarTritonPlugin requires GLIBC 2.35, which is unavailable even in the official Ubuntu 20.04 container. This creates a significant environment mismatch for BI-V100 users.

Could you please provide a version of the Iluvatar toolchain and Triton plugin built with an older GLIBC (e.g., compatible with GLIBC 2.17 for CentOS 7 or GLIBC 2.31 for Ubuntu 20.04)?

Alternatives

No response

Additional context

No response

[Salamanca001]

Thank you for the detailed report and for clearly describing the environment and the GLIBC compatibility issue.

Regarding BI-V100, we would like to clarify an important point first. BI-V100 is a relatively older GPU architecture, and at the moment Triton is not fully supported on this platform. As a result, Triton is not included in the BI-V100 SDK releases. The overall software stack on BI-V100 is also relatively old (for example, PyTorch support is typically limited to around version 1.13).

Because of these limitations, FlagTree is not officially recommended to be installed or used on BI-V100. For a smoother experience and full feature support (including Triton), we strongly recommend using MR-V100 or BI-V150 GPUs instead.

That said, your feedback about the GLIBC version mismatch across different operating systems is valid and very helpful. We understand that CentOS 7 (GLIBC 2.17) and Ubuntu 20.04 (GLIBC 2.31) are still widely used, especially in official BI-V100 Docker images. We will further investigate, verify, and test this issue internally. After confirmation, we plan to provide plugin builds compatible with multiple operating systems and GLIBC versions.

Thank you again for your valuable feedback and for bringing this to our attention.

[Zhang-kg]

Thank you for your reply. I would like to correct some earlier information regarding BI-V100 support based on my direct communication with Iluvatar (Tianshu) engineers.

1. Updated Support Info: According to the official engineers, BI-V100 can support PyTorch 2.x by upgrading the firmware and SDK to version 3.2.1. They have provided an SFTP server with the following installers (located in /BI_100/3.2.1_BI100/x86):

corex-installer-linux64-3.2.1_x86_64_10.2.run

corex-docker-installer-3.2.1-10.2-centos7.8.2003-py3.10-x86_64.run

corex-docker-installer-3.2.1-10.2-ubuntu20.04-py3.10-x86_64.run

If the FlagTree team is interested in verifying this or needs to test against this newer software stack, I am happy to share the SFTP login credentials and detailed download instructions.

2. The Triton Question: Interestingly, when I inquired about Triton, the Tianshu team stated that there is currently no "official" Triton support for BI-V100. This leads to an interesting technical question: if PyTorch 2.x is supported, how is the integration handled without a Triton backend (especially for torch.compile)? It is possible they are using a custom C++ backend or focus primarily on Eager mode.

3. Next Steps: I haven't performed the full upgrade to SDK 3.2.1 and PyTorch 2.x yet, but I plan to do so shortly. I will monitor how the system behaves and whether the GLIBC issue persists in this newer environment.

I believe that aligning FlagTree's Triton-based approach with Tianshu's updated SDK could unlock modern compiler capabilities for the BI-V100 install base.

[Salamanca001]

Thank you for the additional details and for sharing your understanding of the current BI-V100 software stack.

To clarify our position more clearly: at this time, BI-V100 itself does not have official Triton support. Since FlagTree’s compiler path relies on Triton, supporting FlagTree on BI-V100 would first require Triton to be properly enabled, validated, and supported on that platform.

While support for newer PyTorch versions (e.g. PyTorch 2.x) is certainly a necessary prerequisite, it is not a sufficient condition for Triton support. Enabling Triton on a given GPU architecture also depends on the compatibility and maturity of the surrounding SDK components (runtime, drivers, codegen paths, stability and correctness validation, etc.), and this is typically not something that can be addressed in a short time frame.

For these reasons, Triton (and consequently FlagTree) is currently supported and validated on MR-V100 and newer architectures, but not on BI-V100.

Whether BI-V100 should support Triton in the future is a broader platform-level decision and is best discussed directly with the Iluvatar/Tianshu engineering teams responsible for the BI-V100 SDK and roadmap. Once Triton support on BI-V100 is clearly defined and established at that level, we would be happy to re-evaluate what makes sense to support from the FlagTree side.

Thanks again for the thoughtful discussion and for your interest in FlagTree.