Reference only

TorchScale conflicts with timm: exact consumer pins are not fixed by upgrading everything

Two nodes can require different timm versions. Map the consumers and their supported ranges before deciding whether one environment can serve both.

Scope and symptoms

Two nodes can require different timm versions. Map the consumers and their supported ranges before deciding whether one environment can serve both.

Search fragments; wording and context vary:

torchscale
timm==0.6.13

Source-supported context

The upstream TorchScale main-branch setup.py still labels itself 0.2.0 and declares timm==0.6.13 and fairscale==0.4.0. The separately published PyPI 0.3.0 distribution is newer than that source version, yet its current package metadata also declares those exact dependencies. The source branch and PyPI release are not the same artifact; verify the installed Requires-Dist and origin. timm's installation guide recommends a virtual environment to avoid dependency conflicts, and its generic pip install timm command does not override a consumer's exact requirement.

Distinguish these cases

1. The installed TorchScale distribution declares an exact timm version outside another node's documented supported range. A resolver conflict is different from a later runtime API failure.

2. The installed wheel/source package or a third-party fork differs from the upstream branch the user inspected; package name alone does not identify its requirement set.

3. Metadata appears consistent but a consumer still calls an incompatible API at runtime.

Suggested diagnostic sequence

This sequence is editorial guidance, not a diagnosis of your machine.

1. With the Python executable that starts ComfyUI, record installed torchscale, timm and fairscale versions and each owning node revision. Read installed Requires-Dist plus node constraints, note wheel/VCS origins, and calculate range intersections rather than comparing which version is newer.

2. Run a dependency consistency check. If a mutually supported combination exists, pin it in a recoverable copy and test both consumers. If no intersection exists, use a maintainer-supported implementation or separate environments instead of globally forcing one node's pin.

3. Removing == from requirements or installing the latest timm with --no-deps only bypasses resolution; it is not compatibility evidence. Consumer-code changes require a maintained fix or explicit API/model tests, kept outside the stable installation.

4. In the copy, import each consumer's exact API, run its smallest supported model, and regress the other timm user. Check model registration names, weight source/load and input/output behavior; package import alone does not prove both paths.

Completion checks

Dependency checks and both consumers' minimal operations pass. A successful final pip install does not prove the shared environment is consistent.

Limits and cautions

PyPI 0.3.0 metadata currently retains the pin, but this does not make timm==0.6.13 a site-wide recommendation. A fork or future distribution may differ, and model/hardware requirements remain separate.

Source review: 2026-09-26. No installation, dependency resolution, model call, GPU inference or user-environment repair was executed.

Original sources

Didn't solve it?

Check the next possible cause

The same symptom can come from a different cause. Work through these related entries in order.

  1. ModuleNotFoundError in ComfyUI even though the package is installedModuleNotFoundErrorA successful installation in a terminal does not show that ComfyUI uses the same Python. Separate the import name, distribution name and running interpreter before changing dependencies.
  2. timm is missing, a model name is unknown, or weights will not downloadNo module named 'timm'A node imports the library, chooses an implementation and may then obtain pretrained weights. Determine which stage failed instead of treating every message as a missing package.
Paste your full log into the error finder

Was this page helpful?

Anonymous. Only a yes/no counter is stored; no account, IP address or device details.

Sources & references

2026-09-26 English review of TorchScale main setup.py (0.2.0), PyPI 0.3.0 Requires-Dist and timm installation guidance; no install, resolution, model or GPU run.

01TorchScale setup.pySource checked: 2026-09-2602TorchScale PyPI distributionSource checked: 2026-09-2603timm installation documentationSource checked: 2026-09-26

Report an issue · ad8bb20e-95a6-54c8-a0d2-644d986b6c68