馃摎 The doc issue
Recently, AMD released the Radeon PRO W7900 (and other AI-focused variants) with 32 GB VRAM, while Intel expanded its Arc Pro lineup with models like the B65/B70, also featuring 32 GB VRAM. These cards represent affordable, high-VRAM alternatives for LLM inference and computer vision workloads, especially given the ongoing constraints in the supply chain for high-end NVIDIA GPUs.
However, the OpenMMLab ecosystem (MMDetection, MMSegmentation, MMDetection3D, etc.) relies heavily on MMCV as its foundation. Currently, there is no official documentation covering the installation and building of MMCV for AMD ROCm or Intel oneAPI.
Problem Statement
-
The official MMCV installation guide only covers Linux, Windows, and macOS with CUDA support.
-
There is no mention of ROCm or Intel GPU support in the official docs.
-
The only relevant community resource I found is an outdated third-party ROCm fork, which is not maintained and does not reflect the current MMCV codebase.
This creates a barrier for researchers and engineers who wish to evaluate or deploy OpenMMLab tools on non-NVIDIA hardware.
Questions / Request
-
Is MMCV currently compatible with ROCm and/or Intel oneAPI? If yes, what are the exact build flags and dependencies?
-
If not officially supported, are there plans to add support? If so, what is the expected timeline?
-
Can the official documentation be updated to include:
-
A dedicated section for ROCm-based installation (with HIP and rocBLAS dependencies)
-
A dedicated section for Intel Arc GPU installation (using oneAPI, Level Zero, or OpenCL)
-
Known limitations or performance considerations for each platform
Suggest a potential alternative/fix
No response
馃摎 The doc issue
Recently, AMD released the Radeon PRO W7900 (and other AI-focused variants) with 32 GB VRAM, while Intel expanded its Arc Pro lineup with models like the B65/B70, also featuring 32 GB VRAM. These cards represent affordable, high-VRAM alternatives for LLM inference and computer vision workloads, especially given the ongoing constraints in the supply chain for high-end NVIDIA GPUs.
However, the OpenMMLab ecosystem (MMDetection, MMSegmentation, MMDetection3D, etc.) relies heavily on MMCV as its foundation. Currently, there is no official documentation covering the installation and building of MMCV for AMD ROCm or Intel oneAPI.
Problem Statement
The official MMCV installation guide only covers Linux, Windows, and macOS with CUDA support.
There is no mention of ROCm or Intel GPU support in the official docs.
The only relevant community resource I found is an outdated third-party ROCm fork, which is not maintained and does not reflect the current MMCV codebase.
This creates a barrier for researchers and engineers who wish to evaluate or deploy OpenMMLab tools on non-NVIDIA hardware.
Questions / Request
Is MMCV currently compatible with ROCm and/or Intel oneAPI? If yes, what are the exact build flags and dependencies?
If not officially supported, are there plans to add support? If so, what is the expected timeline?
Can the official documentation be updated to include:
A dedicated section for ROCm-based installation (with HIP and rocBLAS dependencies)
A dedicated section for Intel Arc GPU installation (using oneAPI, Level Zero, or OpenCL)
Known limitations or performance considerations for each platform
Suggest a potential alternative/fix
No response