[cyyever]

Draft until reviewed.

Agent details (AI-generated)

Problem. ROCm wheels built with PYTORCH_ROCM_EXTERNAL_SDK=1 declare no ROCm dependency and ship no torch/_rocm_init.py. import torch therefore finds ROCm only through LD_LIBRARY_PATH or the host's ld.so.conf. On a Breadboard guest that means the accelerator pack's ROCm 10.0 rather than our patched 10.1 SDK. Without LD_LIBRARY_PATH the import fails on librocm-openblas.so.0.

Change.

• The build reads the SDK wheels' version from PYTORCH_ROCM_SDK_VERSION.
• tools/metadata/dependencies.py pins rocm[libraries] to that version.
• torch/CMakeLists.txt writes the _rocm_init.py that the existing install rule and torch/__init__.py already expect, so import torch preloads the SDK libraries through rocm_sdk.initialize_process and checks the version. The preload list is TheRock's build_prod_wheels.py list without hipdnn, which this SDK does not build.
• PYTORCH_ROCM_EXTERNAL_SDK is removed.

Needs a breadkernel change. accelerator-wheels-rocm.toml must set PYTORCH_ROCM_SDK_VERSION to the SDK wheel version and drop PYTORCH_ROCM_EXTERNAL_SDK. Without it, the fallback pins whatever rocm is installed in the build venv, which is 10.0.0 today.

Tested. The CMake block was run with and without the variable; the dependency provider was exercised. On a gfx1201 sandbox, with the generated file added to the published gdb71170c wheel, import torch works without LD_LIBRARY_PATH and all ROCm libraries load from the rocm_sdk wheels. A full wheel build with the change has not been run.

🤖 Generated with Claude Code