Script 'mail_helper' called by obssrc Hello community, here is the log from the commit of package comfyui for openSUSE:Factory checked in at 2026-08-24 15:43:50 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Comparing /work/SRC/openSUSE:Factory/comfyui (Old) and /work/SRC/openSUSE:Factory/.comfyui.new.1258 (New) ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "comfyui" Mon Aug 24 15:43:50 2026 rev:3 rq:1373274 version:0.33.3 Changes: -------- --- /work/SRC/openSUSE:Factory/comfyui/comfyui.changes 2026-08-22 21:37:45.447681427 +0200 +++ /work/SRC/openSUSE:Factory/.comfyui.new.1258/comfyui.changes 2026-08-24 15:44:03.464649450 +0200 @@ -1,0 +2,18 @@ +Sun Aug 23 17:27:06 UTC 2026 - Martin Pluskal <[email protected]> + +- Start on the CPU when the machine has no usable accelerator, instead + of aborting during import: add + comfyui-fall-back-to-cpu-without-accelerator.patch + * comfy/model_management.py assumed a GPU unless --cpu was given, so + on a CPU-only installation importing it raised "AssertionError: + Torch not compiled with CUDA enabled" and the server never started + * openSUSE builds PyTorch without CUDA and without ROCm, so this + affected every user who did not know to pass --cpu + * the fallback logs a warning naming the cause, because the same + condition also holds for a GPU-capable PyTorch whose device is + currently unusable, where a silent CPU run would be far slower + than the user expects +- Require python-comfy-aimdo 0.4.14 instead of 0.4.13, so the exact + pin keeps resolving once that update is published + +------------------------------------------------------------------- New: ---- comfyui-fall-back-to-cpu-without-accelerator.patch ----------(New B)---------- New: of aborting during import: add comfyui-fall-back-to-cpu-without-accelerator.patch * comfy/model_management.py assumed a GPU unless --cpu was given, so ----------(New E)---------- ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Other differences: ------------------ ++++++ comfyui.spec ++++++ --- /var/tmp/diff_new_pack.iIIo0D/_old 2026-08-24 15:44:04.153674030 +0200 +++ /var/tmp/diff_new_pack.iIIo0D/_new 2026-08-24 15:44:04.155674101 +0200 @@ -30,6 +30,8 @@ Source2: comfyui-packaged-paths.yaml # PATCH-FIX-UPSTREAM comfyui-create-custom-nodes-directory.patch gh#Comfy-Org/ComfyUI#8110 gh#Comfy-Org/ComfyUI#8434 [email protected] -- create custom_nodes under --base-directory, main.py lists it before the server starts Patch0: comfyui-create-custom-nodes-directory.patch +# PATCH-FIX-UPSTREAM comfyui-fall-back-to-cpu-without-accelerator.patch [email protected] -- start on the CPU when no accelerator is present, instead of raising AssertionError while importing +Patch1: comfyui-fall-back-to-cpu-without-accelerator.patch BuildRequires: %{python_module base} BuildRequires: fdupes BuildRequires: python-rpm-macros @@ -43,7 +45,7 @@ Requires: %{primary_python}-alembic Requires: %{primary_python}-av >= 16.0.0 Requires: %{primary_python}-blake3 -Requires: %{primary_python}-comfy-aimdo = 0.4.13 +Requires: %{primary_python}-comfy-aimdo = 0.4.14 Requires: %{primary_python}-comfy-kitchen = 0.2.31 Requires: %{primary_python}-comfyui-embedded-docs >= 0.5.9 Requires: %{primary_python}-comfyui-frontend-package = 1.50.6 ++++++ comfyui-fall-back-to-cpu-without-accelerator.patch ++++++ Fall back to the CPU device when no accelerator is present. comfy/model_management.py sets cpu_state = CPUState.GPU unconditionally at import time and the only assignment back to CPUState.CPU is behind the --cpu command line flag. There is no autodetection of "this machine has no usable GPU at all", so on a CPU-only installation the module-level statement total_vram = get_total_memory(get_torch_device()) / (1024 * 1024) reaches torch.device(torch.cuda.current_device()) and raises AssertionError: Torch not compiled with CUDA enabled while ComfyUI is still importing, i.e. the server never starts. openSUSE builds PyTorch without CUDA and without ROCm, so this is what every user of the packaged ComfyUI gets unless they know to pass --cpu. Fall back to the CPU device when none of the accelerator backends the module already probes is available. The condition mirrors the flags that are bound above it, and it deliberately leaves cpu_state alone when it is CPUState.MPS, so Apple silicon is unaffected. ixuca_available is included in the check so that an Iluvatar CoreX PyTorch keeps exactly the behaviour it has today: get_torch_device() has no ixuca branch, so such a build must continue to take the torch.cuda.* path rather than being silently downgraded to the CPU by this patch. That hardware is not available for testing, so it is preserved rather than changed. The message is a warning rather than an informational line on purpose. The same condition is true both for a PyTorch built without any GPU support and for a PyTorch that has GPU support whose device is unusable right now (driver not loaded, CUDA_VISIBLE_DEVICES empty, a container without device passthrough). In the second case the user previously got a loud failure and would now get a silent run that is one to two orders of magnitude slower, so the fallback has to announce itself and name the cause. --- ComfyUI-0.33.3.orig/comfy/model_management.py +++ ComfyUI-0.33.3/comfy/model_management.py @@ -156,6 +156,14 @@ if args.cpu: cpu_state = CPUState.CPU +elif cpu_state == CPUState.GPU and not (directml_enabled or xpu_available + or npu_available or mlu_available + or ixuca_available + or torch.cuda.is_available()): + logging.warning("No usable accelerator found (this PyTorch reports no CUDA " + "or ROCm device), falling back to the CPU. Generation will " + "be very slow.") + cpu_state = CPUState.CPU def is_intel_xpu(): global cpu_state
