Extremely Slow Import in Conda Environments (torch / torchvision / cupy / PyMuPDF, etc.) with Almost Zero CPU/Disk Usage

[Li122]

Hello, I am encountering an unusual and severe performance problem when using Conda environments on my system. I would greatly appreciate help from the community to determine whether this issue is related to my CPU (Ultra 7 265K) or something else in my environment.

---

My System Configuration

• CPU: Intel Ultra 7 265K

• GPU: NVIDIA RTX 5060 Ti

• RAM: 32 GB

• Storage: NVMe SSD

• OS: Windows 11 23H2

• Conda: Fresh installation of both Anaconda and Miniconda (issue occurs on both)

---

Problem Description

In brand-new Conda environments, I encounter the following:

1. First-time imports of certain libraries are extremely slow, including:

   • torch

   • torchvision

   • cupy

   • PyMuPDF

   • (possibly other large packages)

   The initial import takes several minutes to over ten minutes.

2. If I terminate Python and wait more than ~10 minutes, the imports become slow again

   • If I re-import immediately (within ~1 minute) after the first import, the import is fast (less than 1 second).

   • But if I wait longer (10+ minutes), the import slows down again dramatically. Importing libraries like OS, sys, and numpy is running smoothly.

3. During these slow imports, CPU usage is near 0%, and disk activity is also near 0%
   This does not look like normal compilation, I/O, or cache-building behavior.

4. On another machine with the same OS and similar configuration, everything works normally:

   Reference machine:

   • CPU: Intel 14650

   • GPU: RTX 4060

   • RAM: 32 GB

   • OS: Windows 11 23H2

   → Imports are always fast on this machine.
   The exact same code, same Conda environment, same package versions.

---

What I Have Already Tried (none resolved the issue)

• Reinstalled Anaconda / Miniconda

• Created multiple fresh environments

• Tried different versions of PyTorch and CUDA

• Updated GPU drivers

• Cleared __pycache__ and Conda caches

• Disabled antivirus (including Windows Defender)

• Checked Windows Indexing

• Verified SSD health

• Disabled Windows Real-Time Protection and Core Isolation

• Used python -v to trace imports → shows stalls but no clear reason

The issue persists across all attempts.

---

My Questions

Could this be related to the Intel Ultra 7 265K architecture or compatibility with Python/Conda?
Under what circumstances would Python show:

• extremely slow first-time imports

• very low CPU usage

• almost no disk I/O

Is there any known compatibility issue with this CPU series (Meteor Lake / Intel Ultra)?

I am unsure whether this is:

• a CPU architecture issue

• some Windows security mechanism

• a Conda/Python dynamic library loading issue

• filesystem or signature verification issue

• or something else entirely

---

Any help is greatly appreciated!

I can provide additional data if needed:

• python -v import logs

• Process Monitor traces

• Conda environment details

• System event logs

Thank you in advance — I’m really hoping to identify the root cause of this unusual behavior.

[Li122]

Thanks everyone for the insights.
I want to post the final explanation and solution, in case other users encounter the same issue.

---

Root Cause Identified: Network-Based Reputation / Security Verification Blocked by Campus Firewall

After repeated tests, I discovered that the extremely slow first-time import of packages such as torch, torchvision, cupy, PyMuPDF, etc., was not related to Conda, Python, or my hardware.

The root cause turned out to be:

Windows’ DLL reputation / security verification being blocked by my campus network firewall.

Here’s what happened:

• When Python imports large packages for the first time, it loads many DLLs.

• Windows performs online reputation checks / security validation for these DLLs (SmartScreen + Cloud-based Protection).

• My campus network has strict firewall rules, which caused these online checks to hang indefinitely.

• During this time:

  • CPU = 0%

  • Disk = 0%

  • import takes several minutes to more than ten minutes

  • no visible error messages

This exactly matched the behavior of SmartScreen or Windows Defender waiting for remote validation that never completes.

---

How I Confirmed the Issue

When I disconnected the internet:

• Imports became fast immediately.

When I connected to a different network (mobile hotspot):

• First-time imports became normal.

After this successful check, switching back to the original network:

• The imports remained fast, because the reputation/cache was already stored locally.

This confirmed that initial DLL reputation checks were blocked by my campus network, and once they passed once, Windows cached the results.

---

Working Solution

Two things helped solve the issue permanently:

1. Connect to a different network once

(e.g., mobile hotspot)
→ Allow Windows to complete initial DLL reputation checks
→ Results get cached locally
→ No more slow imports afterward

and/or

2. Disable SmartScreen reputation checks

Windows Security → App & Browser Control → Reputation-based protection settings
Disable:

• Check apps and files

• SmartScreen for Microsoft Edge

• SmartScreen for Microsoft Store apps

• (and any other SmartScreen-related options)

After this, Python imports became consistently fast even on the campus network.

---

Final Result

• The issue had nothing to do with Conda

• Not caused by Intel Ultra 7 265K, GPU, Python, or package versions

• The slowdown was caused by Windows online reputation scanning being blocked by firewall

• Once verification succeeded ONCE, everything became normal

---

Hope this helps others

If your Python imports are:

• extremely slow only on the first import

• CPU/disk activity near zero

• and the machine is connected to a restricted network

then the issue may be caused by:

Windows SmartScreen / Defender cloud-based scanning blocking DLL validation

The workaround is either:

1. Try a different network once, or

2. Disable Reputation-based protection.