SystemResourcesObserver Crashes on Startup — GUI Unreliable, All Models Fail to Load

[Gongman514]

Any Help would be much appreciated.

main.log

LM Studio Version

0.4.12 (Build 1)

Operating System

Windows 11 Pro, Version 25H2, Build 26200.8246

---

Hardware

Component	Details
CPU	Intel Core Ultra 9 285K
RAM	64.0 GB
GPU	NVIDIA RTX A5000 (24 GB VRAM)
GPU Driver	NVIDIA Enterprise Driver 595.97
CUDA Version	13.2 (confirmed via nvidia-smi)
Storage (C:)	Local Disk, sufficient free space
OS Architecture	x86_64 (AVX, AVX2)

---

Bug Description

There are two related bugs present, both confirmed via main.log. They occur on every launch without exception.

Bug 1 — SystemResourcesObserver crashes immediately on startup (PRIMARY BUG)

On every launch, the SystemResourcesObserver child process exits immediately with either exit code 1 or exit code 1073807364 before it can initialize. LM Studio attempts one restart, which also fails, and then gives up. Because this observer never starts, pullReport is always undefined, which cascades into a complete inability to load any model regardless of runtime (CUDA, Vulkan, or CPU).

Exit code 1073807364 (hex 0xC0150004) is a Windows side-by-side (SxS) assembly error, suggesting a possible missing or mismatched Visual C++ runtime dependency — though this has not been confirmed as the root cause.

Bug 2 — GUI fails to appear on launch (~70% of the time)

When LM Studio is launched, only 5 of the expected 10 processes spawn (visible in Task Manager → Details tab). The GUI never appears. All 5 processes show 0K working set — they have launched but stalled and allocated no memory. The only workaround is to repeatedly kill individual LM Studio processes from the Details tab until the GUI appears, which succeeds approximately 30% of the time per attempt. This behavior is consistent across every launch and has been reproduced dozens of times.

Bug 3 — EBUSY / EPERM lock on Python runtime directory (SECONDARY BUG)

On first run after install, LM Studio fails to update its bundled Python (amphibian) runtime due to a file lock on cpython3.11-win-x86@6\Lib\site.py. This causes the runtime directory rename to fail and falls back to a broken merge, leaving the Python layer in a corrupt state.

---

Symptoms

• GUI does not appear on the majority of launch attempts
• Task Manager → Details shows exactly 5 LM Studio.exe processes, all with 0K working set
• lms.exe runtime ls hangs indefinitely on "Waking up LM Studio service..." — the backend service never starts
• All model load attempts fail with: Failed to load the model — Attempt to pull a snapshot of system resources failed. Error: 'Cannot read properties of undefined (reading 'pullReport')'
• This failure occurs with all runtimes: CUDA 12, CUDA, Vulkan, and CPU llama.cpp
• Hardware page correctly detects GPU via Vulkan (1 GPU detected, NVIDIA RTX A5000, 26.64 GB VRAM) — GPU detection itself is not the issue
• server-logs only ever contain [INFO] Server stopped. — the server never starts
• Memory estimate shows "unavailable" for all models

---

Steps to Reproduce

1. Install LM Studio 0.4.12 (Build 1) on Windows 11 Pro Build 26200.8246, selecting "Install for me only"
2. Launch LM Studio.exe from C:\Users\<user>\AppData\Local\Programs\LM Studio\
3. Observe Task Manager → Details tab: exactly 5 LM Studio.exe processes appear, all with 0K working set, GUI does not appear
4. Kill processes from Details tab one at a time until GUI appears (requires multiple attempts)
5. Once GUI appears, navigate to Settings → Hardware — GPU is correctly detected via Vulkan
6. Navigate to Settings → Runtime — switch GGUF runtime to any option (Vulkan, CUDA, CPU)
7. Attempt to load any model
8. Observe failure: "Failed to load the model — Cannot read properties of undefined (reading 'pullReport')"

---

Tests Performed / Troubleshooting Already Done

The following were all attempted and did not resolve the issue:

• Clean uninstall and reinstall (multiple times), including deletion of AppData\Local\LM Studio, AppData\Local\lm-studio-updater, C:\Users\<user>\.lmstudio, and C:\Users\<user>\.cache\lm-studio
• Reinstalled using "Install for me only" (not "all users")
• Ran installer as administrator
• Ran LM Studio.exe as administrator
• Disabled Windows Defender real-time protection entirely
• Switched GGUF runtime to all available options: Vulkan llama.cpp v2.13.0, CUDA 12 llama.cpp v2.13.0, CUDA llama.cpp v2.13.0, CPU llama.cpp v2.13.0 — all fail identically
• Set GPU Offload to 0 before loading model
• Deleted C:\Users\<user>\.lmstudio\extensions\backends\vendor\_amphibian folder to force re-download of Python runtime
• Verified GPU health via nvidia-smi — RTX A5000 reports correctly with 23028 MiB VRAM, driver 595.97, CUDA 13.2
• Confirmed no third-party antivirus installed; Windows Defender is the only security software
• Confirmed no GPU-related startup applications (no GeForce Experience overlay, no MSI Afterburner, etc.)
• Toggled "Auto-update selected Runtime Extension Packs" off
• Attempted to downgrade Vulkan runtime via CLI (lms.exe runtime ls) — CLI hangs indefinitely because backend service never starts

---

Log Evidence

main.log location: C:\Users\Bob\AppData\Roaming\LM Studio\main.log

Key entries from main.log (full log attached):

[2026-04-22 19:10:05.403] [error] [SystemResourcesObserverProxyObject] System resources observer process exited unexpectedly (pid 10400, exitCode 1073807364).
[2026-04-22 19:10:05.406] [error] [SystemResourcesObserverProxyObject] Failed to initialize system resources observer: Error: System resources observer process exited during startup.. Production build: yes

[2026-04-23 15:58:13.477] [error] [SystemResourcesObserverProxyObject] System resources observer process exited unexpectedly (pid 56656, exitCode 1).
[2026-04-23 15:58:13.479] [error] [SystemResourcesObserverProxyObject] Failed to initialize system resources observer: Error: System resources observer process exited during startup.. Production build: yes
[2026-04-23 15:58:13.697] [error] [SystemResourcesObserverProxyObject] Restart attempt 1 failed. Error: System resources observer process exited during startup.

[2026-04-23 15:58:22.685] [error] [RuntimeIndexerTarGzExtractionPostDownloadActionExecutor] Failed to delete target directory: Error: EBUSY: resource busy or locked, unlink 'C:\Users\Bob\.lmstudio\extensions\backends\vendor\_amphibian\cpython3.11-win-x86@6\Lib\site.py'
[2026-04-23 15:58:22.697] [error] [RuntimeIndexerTarGzExtractionPostDownloadActionExecutor] Failed to rename target directory: Error: EPERM: operation not permitted, rename 'C:\Users\Bob\.lmstudio\extensions\backends\vendor\_amphibian\cpython3.11-win-x86@6' -> 'C:\Users\Bob\.lmstudio\extensions\backends\vendor\_amphibian\.to-delete-70707358323300'. Falling back to merge.

[2026-04-23 16:02:48.095] [error] [ModelLoadingProvider] Error checking model guardrails, assuming can load Error: Attempt to pull a snapshot of system resources failed. Error: 'Cannot read properties of undefined (reading 'pullReport')'

[2026-04-24 07:52:27.184] [error] [SystemResourcesObserverProxyObject] System resources observer process exited unexpectedly (pid 49692, exitCode 1).
[2026-04-24 07:52:27.186] [error] [SystemResourcesObserverProxyObject] Failed to initialize system resources observer: Error: System resources observer process exited during startup.. Production build: yes

server-logs contents (all entries across all sessions):

[2026-04-24 07:28:53][INFO] Server stopped.
[2026-04-24 07:56:14][INFO] Server stopped.
[2026-04-24 08:17:20][INFO] Server stopped.

The server never starts — only stop events are recorded.

---

nvidia-smi Output

NVIDIA-SMI 595.97    Driver Version: 595.97    CUDA Version: 13.2

GPU  Name                  Driver-Model  VRAM
  0  NVIDIA RTX A5000      WDDM          2981MiB / 23028MiB    GPU-Util: 12%

---

Additional Notes

• This bug appears related to open issue Some processes do not start and the GUI does not appear at all #1364 (5 processes / no GUI on Windows 11 26200.x). However, the SystemResourcesObserver crash is a separate and more critical issue not addressed in that thread.
• The pullReport error is a downstream consequence of the observer crash, not an independent bug.
• The GPU is correctly detected by LM Studio via Vulkan (Hardware page shows "1 GPU detected with Vulkan, NVIDIA RTX A5000, 26.64 GB VRAM"). The issue is not GPU detection — it is the observer subprocess failing before it can serve resource data to the model loader.
• The NVIDIA Enterprise driver (595.97) is in use rather than the consumer GeForce/Studio driver. This may or may not be relevant to the observer crash.
• Exit code 1073807364 (hex 0xC0150004 = STATUS_SXS_CANT_GEN_ACTCTX) on the first crash session suggests a Windows side-by-side assembly activation context failure, which typically indicates a missing Visual C++ redistributable or a DLL manifest mismatch. Subsequent sessions show exit code 1 (generic failure). It is unclear whether these represent the same root cause.

---

Full main.log attached.

[christophergoltz]

Spent the morning on this, same setup as OP (Win 11 26200, LM Studio 0.4.12 Build 1) but on AMD instead of NVIDIA - RX 9070 XT / RDNA 4. So this isn't GPU-vendor specific.

Two of the three things I had to fix were Windows-side, one is the actual LM Studio bug:

1. The SystemResourcesObserver exit code 0xC0150004 (SXS_ASSEMBLY_NOT_FOUND) was Windows component-store corruption. Fixed by DISM /Online /Cleanup-Image /RestoreHealth and then sfc /scannow. SFC actually found and
   repaired corrupted files on my box. The signature to look for in the Event Log is WindowsWcpOtherFailure3 with 0x800f0983 (CBS_E_IDENTITY_MISMATCH).

2. The EPERM rename failures during _amphibian extraction are Windows Defender holding a handle on the freshly-extracted python.exe for its real-time / IOAV scan, in the exact moment LM Studio tries to rename .tmp-XXX -> cpython3.11-win-x86@6. Nothing ever gets quarantined, no threat is flagged. It's a pure file-handle race. Set Defender exclusions on ~/.lmstudio, ~/AppData/Roaming/LM Studio, ~/AppData/Local/Programs/LM Studio before
   the first launch. Setting them afterwards doesn't help, because the half-state from the first run sticks around.

3. After both of those are sorted, _amphibian extracts cleanly. Both app-harmony-win-x86@7 and cpython3.11-win-x86@6 are present, python.exe and share/venv/metadata/venvstacks_layer.json in place, no .tmp-* or
   .to-delete-* residue. But this still shows up in main.log:

[error] [amphibianUtils] Error getting Python path for amphibian runtime layer:
Error: Failed to parse venvstacks for ...\_amphibian\.tmp-1231189619800:
ENOENT: no such file or directory, open '...\_amphibian\.tmp-1231189619800\share\venv\metadata\venvstacks_layer.json' 
at ... indexVendorLibs ... reindexVendorLibs ... performInitialIndex ...

Note the path: It's the pre-rename temp directory .tmp-1231189619800, not the final cpython3.11-win-x86@6 that's actually on disk. The indexer cached the path from before the atomic rename and never re-resolves. That lookup fails, the post-install script for the harmony layer then errors with ENOENT: ...cpython3.11-win-x86@6\python.exe, the layer is marked failing-health, and LM Studio then deletes its own cpython3.11-win-x86@6 directory. I literally watched the folder disappear between two consecutive Get-ChildItem calls.

That's why reinstalls don't stick on 0.4.12 for anyone. Every fresh install hits the same caching race on first launch, the same self-deletion follows, and the persistent end state is always "cpython3.11 missing, Harmony shows Missing libraries, all runtimes fall back to Download ready". I went through full uninstall → wipe ~/.lmstudio and Roaming/LM Studio → fresh install twice, both ended in the same broken state.

---

The "Fix" button on Harmony re-pulls the harmony layer but not cpython3.11, so it doesn't break the loop.

Fix area looks like indexVendorLibs / performInitialIndex in amphibianUtils. Needs to re-resolve the layer path after the rename completes, not hold the source-side temp path through what should be an atomic step.

[egal-bburr]

TLDR for a possible fix: In my case: AVX2 was disabled in my BIOS.
Fix: reboot → BIOS → set AVX2 to Enabled/Auto → save & exit.

Similar symptoms as this thread — Hardware tab shows CPU Incompatible / Unknown, 0 GPUs detected, 0 GB RAM/VRAM, no model loads. main.log shows:

• Failed to load liblmstudio ...\bin\liblmstudio\cpu\liblmstudio_bindings.node. A dynamic link library (DLL) initialization routine failed.
• bundled vulkanAND cpu hardware survey both fail: child process. exited with code 1
• downstream: [BackendDownloadProvider] Failed to refresh backends master list TypeError: Cannot read properties of undefined (reading 'result')

Might be worth a try.