LLThreadSafeQueue error during startup on Strix Halo (Windows)
Deviant Scientist
On a new Strix Halo PC (AMD RYZEN AI MAX+ w/ Radeon 8060s) running Windows 11 Pro 25H2, the Second Life Viewer (along with Firestorm, Alchemy, and Megapahit viewers) hangs during startup during the 'Initializing Texture Cache dialog'.
Cool VL Viewer (1.x based) and ApertureOS (skips startup checks) do not have this crash.
The last entry (only apparent error) in the log is the line
> WARNING #ThreadPool# llcommon/llthreadsafequeue.h(314) LLThreadSafeQueue<struct tagMSG,class std::queue<struct tagMSG,class std::deque<struct tagMSG,class std::allocator<struct tagMSG> > > >::push_ : ONCE: Threadsafe queue push_(lockacquired) queue full 1024 >= 1024
Multiple reinstalls of the viewer, exceptions added to windows defender whitelist, as well as a reinstall of Windows 11 have all been performed to no avail.
For reference, this PC features 128GB 8000MT/s unified memory with 32GB allocated to the system and 96GB allocated to VRAM. SL and windows are installed and running from a 2TB Crucial Gen4 NVMe SSD.
While this is probably currently an edge-case as these machines are not currently common in the market, AMD are looking to push more APU based computers in the near future and if this is due to the APU-based platform (relatively high core count at 16 Core / 32 thread, VRAM amount, or nature of unified memory) this could warrant investigation.
Attached log from latest Firestorm initialization attempt.
Log In
Deviant Scientist
Thinking maybe the issue was SL not liking the large amount of VRAM I changed the allocation to 32GB in the bios and the issue remained, so (hopefully) ruling that out as the reason.
Deviant Scientist
And in further testing, the issue still persists even with the beta driver.
What seems to permit the SL viewer to initialize is performing an AI workload in comfy UI - once it loads the model into memory and starts generating a video (sample wan 2.2 workflow) the initializing texture cache step completes near instantly.
Unfortunately, the moment this workload is terminated, the viewer then crashes.
Deviant Scientist
For reference, this appeared to be an AMD driver issue.
Installing the Pytorch beta driver - https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html - which appears to be based on an older version of the GPU driver fixed the issue.