Shared GPU memory Detection error on with AMD Ryzen 395

[nosdregamon]

I got to test a AMD Ryzen 395 with 128GB of Ram and even if I dedicate 96GB of that to the GPU, llmfit detects "only" 32 GB, which changes the results a bit. Auto-Detection:

vs forced memory:

PS C:\Users\AI HX> llmfit system --json
{
"system": {
"available_ram_gb": 20.46,
"backend": "Vulkan",
"cpu_cores": 32,
"cpu_name": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S ",
"gpu_count": 1,
"gpu_name": "AMD Radeon(TM) 8060S Graphics",
"gpu_vram_gb": 31.65,
"gpus": [
{
"backend": "Vulkan",
"count": 1,
"name": "AMD Radeon(TM) 8060S Graphics",
"unified_memory": true,
"vram_gb": 31.65
}
],
"has_gpu": true,
"total_ram_gb": 31.65,
"unified_memory": true
}
}

PS C:\Users\AI HX> llmfit --version
llmfit 0.6.5

[three-foxes-in-a-trenchcoat]

Hey @nosdregamon 👋

Great catch! The Ryzen AI MAX+ 395 uses unified memory (shared between CPU and iGPU), and llmfit is only detecting the default GPU allocation (~32GB) rather than what you've configured in BIOS/UEFI.

The issue is that Vulkan (the backend being used here) only reports the GPU's currently allocated VRAM, not the maximum configurable amount. On unified memory APUs like the Ryzen AI MAX+ series, the actual usable VRAM depends on BIOS settings — and that info isn't exposed through standard GPU APIs.

Workaround for now: Use --vram to override:

llmfit fit --vram 96

Proper fix ideas:

1. On AMD APUs with unified memory, detect total system RAM and note that VRAM is configurable up to a portion of it
2. Read the BIOS-configured GPU memory allocation (platform-specific — may need AMD's ADLX SDK or reading from sysfs on Linux)
3. At minimum, show a hint in the TUI: "Unified memory detected — use --vram to set your configured allocation"

The total_ram_gb: 31.65 in your system output confirms it's reporting the GPU allocation as total RAM too, which compounds the issue. On a 128GB system, that's way off.

Thanks for testing on this hardware — these new AMD APUs are increasingly popular for local LLM inference and getting the detection right matters! 🦊

[bytesizedsolutions-jfox]

@nosdregamon can you share how you "forced" the memory for your second screenshot. I'm working with similar hardware

[nosdregamon]

@nosdregamon can you share how you "forced" the memory for your second screenshot. I'm working with similar hardware

I opened the terminal and typed in

llmfit --memory 96G
followed by "Enter"

[iYoungblood]

I still see this bug with an AI HX 370. 32GB of total memory, 24 to GPU and 8 to CPU pool. Showing only 8GB and models reflect that.

[AlexsJones]

Not forgotten about this, I think I see the toxic combination causing this, on holiday but will work on a fix soon