Closed llv22 closed 2 years ago
@TomHeaven some updates about cracking tensorflow 2.9 on macOS 10.13.6, CUDA 10.1, cudnn 7.6.3 via Xcode 10.1. Recently I spent about 2 weeks to identify the root cause of why tensorflow 2.9 can't work on the aforementioned environment.
Besides the normal compilation issues, the major show-blocker lies at the bazel building system, because on macOS 10.13.6 the ARG_MAX is about 262144, which is far less than the setting on 10.14. But we can crack bazel system for its default linking behavior in https://github.com/llv22/bazel-macOS-10.13.6/blob/r5.2.1/src/main/java/com/google/devtools/build/lib/rules/cpp/CppLinkAction.java. Basically it merges object files to intermediate objects, then further link them together, which can workaround to avoid the issue over the pre-setting of input arguments.
Great work!
@TomHeaven actually there is still one minor feature to enable auto-switch between GPU and CPU mode, esp. for GPU tensorflow installer. But not a big issue.
as title, curious about this.