DLL load failed: identify the native extension instead of downloading loose DLLs

Symptoms and scope

The named Python module can depend on another missing or incompatible runtime library. Native import failure is different from a top-level Python package simply being absent.

Error fragments for search; IDs, values and filenames may vary:

DLL load failed
The specified module could not be found
WinError 126

Source-supported context

ComfyUI reports include _fused and custom_rasterizer_kernel examples, while NumPy and xFormers document binary compatibility. Windows error 126 says a module could not be found; that message alone does not identify whether the named extension or one of its dependent DLLs is absent. Windows searches dependencies by module name, so another copy on the search path can also matter. These sources support extension-specific investigation, not one universal missing-DLL answer. Windows error code DLL search order Source 1 Source 2 Source 3 Source 4

Cases to distinguish

1. Python architecture, Torch or the extension ABI does not match.

2. A required runtime/backend is missing or another same-named library wins on the search path.

3. Installation is incomplete or the node/package combination is unsupported.

Diagnostic sequence

The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.

Step 1. Preserve the first failed native module and full traceback. Record the executable path and architecture of the Python process actually running ComfyUI, plus Python/Torch and extension versions and installation source; a system python may be a different environment.

Step 2. Check whether the named extension file is present in that environment. For a trusted extension, test its import with the same interpreter in an environment copy; imports can execute package code. If Visual Studio developer tools are already available, dumpbin /DEPENDENTS <extension.pyd> can list statically imported DLL names, but it does not prove which dependency Windows actually loaded. Microsoft DUMPBIN

Step 3. Match any identified missing dependency or incompatible binary to the extension publisher's supported wheel, Torch/CUDA combination or vendor runtime. Test a complete supported distribution in an environment copy; do not mix wheels or loose DLLs from unrelated tutorials.

Step 4. For a source build, confirm compiler, SDK and version prerequisites. Otherwise restore a known environment or choose a path without the extension.

Completion check

The native import and a small operation succeed before returning to the full workflow.

Limitations and cautions

Do not copy arbitrary files from DLL-download sites into System32, Python or the project. That can introduce malicious code while hiding the actual version conflict.

Original sources

• SageAttention fused DLL import failure — checked 2026-09-26.
• custom_rasterizer_kernel DLL import failure — checked 2026-09-26.
• NumPy troubleshooting import errors — checked 2026-09-26.
• xFormers README — checked 2026-09-26.
• Microsoft Windows error codes — checked 2026-09-26.
• Microsoft DLL search order — checked 2026-09-26.
• Microsoft DUMPBIN /DEPENDENTS — checked 2026-09-26.

Source and English-language review: 2026-09-26. No native import, DLL inspection, GPU run or runtime guarantee was performed by this site.

Related troubleshooting and guides

• NumPy ABI errors: _ARRAY_API and dtype size changed
• xFormers is installed but unavailable: inspect its native extension and operators
• SageAttention or Triton import failure: match the platform and build combination