Skip to content

AMD.md

Latest commit

 

History

History
81 lines (60 loc) · 3.96 KB

AMD.md

File metadata and controls

81 lines (60 loc) · 3.96 KB
Raw

AMD ROCm Windows Port — Summary

Intent

The upstream ComfyUI-XPUSYS-Monitor was designed primarily for Intel Arc (XPU) and NVIDIA (CUDA). Its AMD GPU support depended on rocm_smi_lib — a Linux-only Python package that wraps librocm_smi64.so. On Windows with AMD ROCm, rocm_smi_lib cannot be installed via pip because the native library (.so) does not exist on Windows.

This fork (ComfyUI-XPUSYS-Monitor-AMD) replaces the AMD provider's GPU monitoring with Windows-native APIs that work on a standard Windows ROCm stack — no rocm_smi_lib required.

Changes Made

providers/amd.py — Rewritten AMD provider

Before After
Used rocm_smi.getMemFreeVdev() / getMemSizeVdev() for VRAM Uses torch.cuda.mem_get_info(0) — returns (free_bytes, total_bytes) from the AMD driver directly
Used rocm_smi.getGpuBusyVdev() for GPU load Uses typeperf (Windows built-in) to read WDDM GPU engine counters averaged via max()
Used rocm_smi.getSingleClockSpeed() for core clock Returns 0 — driver does not expose via Windows API
Used rocm_smi.getTempVdev() for temperature Returns -1 — driver does not expose via Windows API
Used rocm_smi.getPowerVdev() for power draw Returns (-1, 0, False) — driver does not expose via Windows API
Imported shared utils from providers/intel.py Imports from providers/_utils.py (extracted shared module)

providers/_utils.py — New shared utilities module

Extracted from providers/intel.py so all providers can share CPU/RAM polling without depending on the Intel Level Zero provider:

  • _get_cpu_info() — CPU model name and thread count
  • _read_cpu_ram_stats() — CPU load, frequency, RAM total/used/free
  • _read_commit_charge() — Windows commit charge (virtual memory)
  • _PdhQuery — PDH API ctypes wrapper (not used on AMD — kept for reference)
  • _TypeperfGpuQuery — GPU engine utilisation via typeperf.exe (CSV parse, uses max across all engines)
  • _AmdSmiGpuQuery — Official AMD SMI library probe (gracefully skipped on Windows where libamd_smi.so is unavailable)

providers/__init__.py — Detection fix

_is_amd_rocme() now uses getattr(torch.version, 'roc', None) instead of torch.version.roc directly. Windows ROCm builds may lack the roc attribute entirely — getattr avoids the AttributeError and falls through to GPU name matching ("amd", "radeon", "advanced micro devices").

providers/nvidia.py — Updated imports

NvidiaProvider now imports shared utilities from _utils.py instead of intel.py (no functional change).

requirements.txt — Removed rocm_smi_lib

rocm_smi_lib is commented out with a note explaining it is Linux-only. No replacement dependency needed — all monitoring uses torch (already installed with ROCm) and typeperf (Windows built-in).

pyproject.toml

Project name updated to ComfyUI-XPUSYS-Monitor-AMD, publisher set to forkless.

web/ directory — Frontend

Added web/xpu_monitor.js from upstream (56 KB). This is the JavaScript toolbar extension that renders the seven-capsule status bar in ComfyUI's UI.

End Result

Metric Source Status
VRAM free / total torch.cuda.mem_get_info(0) ✅
VRAM allocated / reserved torch.cuda.memory_allocated() / memory_reserved() ✅
GPU utilisation typeperf WDDM engine counters (max across engines) ✅
Device name torch.cuda.get_device_name(0) ✅
GPU core temperature Driver does not expose via Windows API ❌
Power draw Driver does not expose via Windows API ❌
Core clock Driver does not expose via Windows API ❌
GPU load on idle max() across all WDDM engines → ~0% ✅
GPU load under workflow max() reflects busy 3D/compute engine ✅
rocm_smi_lib dependency Removed; no replacement needed ✅
Windows ROCm detection getattr(torch.version, 'roc', None) + GPU name fallback ✅