SciPy, einops and torchsde: distinguish missing imports from failed operations

Keep imports separate from operations

SciPy, einops and torchsde have different roles: scientific computing, tensor rearrangement and differentiable SDE solving respectively.123 SciPy recommends managed project or environment installs; einops patterns must match actual tensor dimensions; torchsde examples specify drift/diffusion functions, state and time-grid shapes. A name in a traceback does not itself establish a missing package.

python -m pip show scipy einops torchsde numpy torch

Use the Python executable that starts ComfyUI and save the first complete traceback and node revision. First confirm which import is actually missing; if import reaches a compiled extension and fails there, record the interpreter, NumPy version, wheel/platform and first native error. Test a genuinely missing dependency against the node's declared range in a recoverable copy, rather than globally replacing a shared numerical stack.

After import succeeds, test the original operation separately. For an einops pattern error, keep the full dimension expression, axis sizes, tensor shape and dtype. For a torchsde runtime error, preserve model drift/diffusion signatures, initial-state and time-grid shapes, and solver options rather than silently substituting a different operation.

Build a small reproduction and compare one change at a time. Report a working import separately from a working node. A new package with a similar name is not a justified repair, and an untested version combination remains unverified.

Related checks

• Keep a reproducible ComfyUI repair record without exposing secrets
• Where to run ComfyUI commands: PowerShell, CMD and virtual environments

Source review and separate English editorial pass: 2026-09-26. No local installation, numerical operation, model or GPU run occurred; no universal version combination is asserted.