Back to Home

ComfyUI Killed DirectML, Exactly as It Warned

Somewhere inside the ComfyUI 0.46.0 changelog, wedged between a router migration and a workflow template bump, sits a line that ends a small era: "Remove directml support." No blog post. No banner. Just a merged commit that deleted the --directml startup flag and 62 lines of cautious optimism from comfy/model_management.py.

The removal landed on November 12, 2026, and shipped to everyone with version 0.46.0 on November 18. The reasoning is boringly sensible. The last PyTorch release that carried a DirectML plugin was version 2.5, and ComfyUI no longer supports that version. So the backend was sitting on a foundation nobody was maintaining, in a building where the elevator had been broken for a year and the sign on the door said so.

A Program That Apologized For Itself

If you ever launched ComfyUI with that flag, you probably remember the console message. It was not subtle:

"WARNING: torch-directml barely works, is very slow, has not been updated in over 1 year and might be removed soon, please don't use it, there are better options."

Read that again. A tool called its own code path barely functional, slow, abandoned, and possibly about to disappear, then asked politely that you not use it. ComfyUI put its failure in a log line and hoped you would take the hint. People did not take the hint, because for a lot of Windows machines the flag was the only thing standing between them and a weekend of WSL driver archaeology.

The deleted code was a small museum of workarounds. DirectML had no reliable way to report free VRAM, so low-VRAM mode was simply switched off. When ComfyUI asked the device how much memory it had, the answer was a hardcoded one gigabyte with a TODO comment next to it. Your memory planner was told your card had the capacity of a 2012 netbook, then blamed when sampling stalled.

What Actually Breaks

The flag is the visible casualty. The quieter problem is the symbol. Model management used to expose a global named is_directml_enabled, and a chunk of the custom node ecosystem imported it to decide which device path to take. That name is gone, so nodes that grep for it now fall over on import. Kijai's KJNodes has an open issue for exactly this on the RifleXRoPE nodes. Furkan Gozukara pushed guard patches to RES4LYF and the WanVideoWrapper within days.

If you are wondering whether you are affected, here is the short checklist:

  • You launched with --directml. The argument no longer exists, so your launcher will complain or silently ignore it.
  • You run a Radeon or an Intel integrated GPU on Windows. That was the whole demographic of this flag.
  • You use custom nodes that touch device detection. Search custom_nodes for is_directml_enabled before your next session. The crash arrives as an import error.

What The Docs Tell You Now

ComfyUI's documentation moved first. One pull request stripped the flag from the developer docs. A second rewrote the AMD Windows row across five languages, and the new wording is direct: ComfyUI no longer supports DirectML, install the ROCm PyTorch build instead.

Here is the detail that deserves a raised eyebrow. In those same tables, the column describing what the install command does still lists torch-directml. The docs can now warn you, in the same table, that the thing they are about to install is unsupported. Somewhere a technical writer is very tired.

The practical path forward, roughly in order of effort:

  1. Strip --directml from every batch file, shell script, desktop shortcut, and that one launcher with fourteen flags you copied from a video in 2024.
  2. Update ComfyUI rather than pinning 0.45 to keep the flag. The flag bought you a slow backend, not stability.
  3. Audit custom nodes for DirectML checks, and update the ones that break. Guards are landing quickly.
  4. Pick a real backend: ROCm wheels, WSL2 with ROCm, or an NVIDIA card. The docs now point AMD Windows at ROCm, and that is a bigger jump than deleting an argument.

The uncomfortable part: DirectML was slow, badly maintained, and could not even count its own memory, but it worked on almost anything with a shader core. ROCm's support matrix is narrower, and cards that used to limp along at four seconds per step may now have nowhere to go inside core. ComfyUI traded universality for a cleaner codebase, a defensible trade that still stings on a two-year-old laptop.

The part that is genuinely funny is the mismatch between the drama and the diff. Sixty-two lines, one boolean, one hardcoded gigabyte. The ecosystem will absorb it in a week.

So pour one out for the worst-supported backend in generative AI. It told you exactly what it was. You used it anyway, because the alternative was reinstalling your operating system to run a node graph. "Remove directml support" is changelog language for a very human sentence: we tried, it did not work, and we are finally allowed to stop pretending.

Comments

No comments yet. Be the first to share your thoughts!