Hi @PRADEEP DABAS ,
Thank you for the detailed description of your configuration. Please find my answers below, in the order of your questions.
1. Can LocalDumps capture a user-mode dump with the current AeDebug configuration?
Yes, I expect it can. Microsoft documents that local dumps are configured and controlled independently of the rest of WER, and that the dump is collected after the crash and before the process terminates. The documented exception is that a crash dump is not collected when automatic debugging is set for application crashes. Automatic debugging is enabled when AeDebug\Auto is "1", which suppresses the dialog and launches the debugger directly. Because Auto is absent in your configuration, automatic debugging is not enabled, so it should not prevent LocalDumps.
A few points about what LocalDumps can and cannot capture:
- A dump is written only for a native crash that reaches WER as an unhandled exception. A Python-level exception that prints a traceback and exits will not produce a dump.
- If your historical traceback was preceded by a line such as
Windows fatal exception: access violation, a native fault did occur, and Python's fault handler printed the Python stack at that moment. If there was no such line, the traceback was most likely a regular Python exception. - A native exception that is handled inside the process (for example,
ctypesconverting an access violation into a PythonOSError) does not reach WER either. - Applications that do their own custom crash reporting are not supported by this feature. Please check whether your ComfyUI launcher or any installed extension adds its own native crash handler.
2. Is a configuration change required, and how does it affect Visual Studio JIT debugging?
No AeDebug change is required. I suggest leaving the existing Debugger entries (in both the 64-bit and WOW6432Node views) unchanged and not adding Auto = "1". Without Auto, the system shows a dialog before invoking the registered debugger, so Visual Studio JIT debugging remains as it is today.
If Auto is ever set to "1" in the future, you can exclude python.exe from automatic debugging by adding a REG_DWORD value named python.exe with data 1 under HKLM\SOFTWARE\Microsoft\Windows NT\CurrentVersion\AeDebug\AutoExclusionList.
3. Registry value types and settings
Create a per-application key, which overrides the global settings:
HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\Windows Error Reporting\LocalDumps\python.exe [Collecting...soft Learn]
Then add these values under that key:
- DumpFolder is a
REG_EXPAND_SZvalue. It sets where dumps are written. For example, useC:\CrashDumps\ComfyUI. If it's absent, the default is%LOCALAPPDATA%\CrashDumps. [Collecting...soft Learn] - DumpType is a
REG_DWORDvalue. Use1(mini dump) to start, or2(full dump) if more memory is needed for analysis. The default is1. [Collecting...soft Learn] - DumpCount is a
REG_DWORDvalue. It sets the maximum number of dump files kept, for example5. When the limit is exceeded, the oldest dump is replaced. The default is10. [Collecting...soft Learn] - CustomDumpFlags is a
REG_DWORDvalue. It is only used whenDumpTypeis0, so you can leave it out. The default is0x00000121.
A mini dump is usually enough to identify the faulting module and the native call stack. A full dump of a PyTorch process can be very large, so I suggest switching to 2 only if the mini dump doesn't provide enough information.
Please also confirm the crashing process's image name in Task Manager (Details tab). The key must match the executable name exactly, for example python.exe and not pythonw.exe.
4. Permissions, privacy and rollback
- Permissions: Administrator rights are required to create the key under HKLM. If you use a custom DumpFolder, its ACL must allow the account running ComfyUI to write to it. Please also make sure the drive has enough free space.
- Privacy: The setting applies to every crashing
python.exeon the machine, not just ComfyUI. Dumps, and full dumps in particular, can contain file paths, prompts, model data, tokens or credentials. Store dumps in a folder restricted to the relevant user and administrators, share them only through a secure channel, and do not attach them to a public forum thread. - Rollback: If the
LocalDumpskey already exists, you can back it up first (for example,reg export "HKLM\SOFTWARE\Microsoft\Windows\Windows Error Reporting\LocalDumps" LocalDumps_backup.reg). If it doesn't exist, there is nothing to back up. To roll back, delete theLocalDumps\python.exesubkey (and theLocalDumpskey itself if you created it) and remove any dump files collected. No AeDebug changes are involved, so Visual Studio JIT settings are unaffected.
After the next native crash, check the dump folder for a new .dmp file. If none is written, the crash most likely did not reach WER as an unhandled native exception, which is useful diagnostic information in itself.
References:
I hope this helps. Let me know if you have any further questions.
If this instruction is applicable to your situation, I would greatly appreciate it if you could follow the instruction here so others experiencing similar behavior can benefit from it as well.