[0xDELUXA]

On ROCm for Windows, torch.cuda.mem_get_info only counts the VRAM used by the current process, so the desktop and other programs look like free VRAM to vram_limit. With the desktop idle, torch reports 0.15 GiB used while the GPU has 2.54 GiB in use.

check_free_vram now goes through diffsynth.core.device.mem_get_info, which on Windows ROCm takes the VRAM used by other processes from ADLX (amd-adlx, new [amd] extra), so vram_limit behaves as it does on NVIDIA GPUs. The other-process share is refreshed at most once a second because an ADLX query takes about half a millisecond and check_free_vram runs on every layer forward; the call averages about 18 us. Without amd-adlx it prints a one-time hint and keeps the torch numbers. Other platforms are unchanged.

On ROCm, TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 is now set by default, otherwise PyTorch SDPA has no flash or memory-efficient kernel on GPUs where AOTriton support is still experimental and fails or falls back to the math path. The setup docs (en/zh) also gain Windows install steps for AMD's ROCm 10 torch wheels.

[Artiprocher]

@0xDELUXA Thank you for your contribution! We don't currently have a Windows AMD device available, so we'll test this feature once we get one.

[0xDELUXA]

@Artiprocher Thanks! ADLX (AMD Device Library eXtra) is AMD's official SDK for reading GPU state on Windows. It ships with the Adrenalin driver (amdadlx64.dll), so there is nothing extra to install on the driver side: https://gpuopen.com/adlx/

The Python binding amd-adlx (https://pypi.org/project/amd-adlx/) is published by AMD itself (author adlx@amd.com). Its homepage is the GPUOpen page above, and its source is in AMD's ADLX repo under https://github.com/GPUOpen-LibrariesAndSDKs/ADLX/tree/main/ADLXPybind. AMD's own ADLX MCP servers in that repo depend on amd-adlx>=1.0.3.

Projects that already read AMD GPU memory and metrics through ADLX:

• kijai/ComfyUI-MemoryVisualization, via amd-adlx (Add AMD telemetry on Windows ROCm kijai/ComfyUI-MemoryVisualization#16)
• darksidewalker/ComfyUI-DaSiWa-Nodes, via amd-adlx (System Monitor: AMD telemetry on Windows through ADLX darksidewalker/ComfyUI-DaSiWa-Nodes#56)
• AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio, via amd-adlx (Read AMD VRAM through ADLX on Windows and list AMD GPUs in the selector AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio#34)
• facefusion/facefusion, ADLX device handling merged in Support for AMD ADLX and refactor device handles facefusion/facefusion#1251
• ROCm/rocm-cli, ADLX for the Windows dashboard telemetry (open, feat(dash-tui): collect GPU/memory metrics in-process on Windows ROCm/rocm-cli#715)
• ostris/ai-toolkit (open, Add AMD ROCm support on Windows and Linux ostris/ai-toolkit#1081)

Nothing changes on other setups. mem_get_info only calls ADLX when sys.platform == "win32" and torch.version.hip is set. Everywhere else it returns torch.cuda.mem_get_info unchanged. On Windows ROCm without amd-adlx, it prints a one-time hint and uses the torch numbers as before.