← VRAM Lab

INSIDE THE LAB

Start here: local AI on WSL2 and Apple Silicon

Find tested WSL2 disk and cache fixes, CUDA compatibility checks, and Apple Silicon prompt-latency and memory measurements.

Start with the symptom you have now. These reports cover Windows disk space, the CUDA environment, inference on an 8 GB GPU, and long prompts on a 16GB M4. Every linked experiment names the environment it tested and includes a session log.

What is stopping you?Start with
C: is still full after deleting files inside WSLWSL2 storage and cache fixes
Dataset files remain outside the cache directory you configuredHub downloads versus generated Arrow caches
WSL starts, but --resize returns 0xc03a001aThe measured resize failure and recovery
PyTorch says CUDA is unavailableFour reproduced causes on WSL2
PyTorch sees the GPU, but a CUDA operation failsCUDA and GPU compatibility
llama.cpp runs unusually slowly with mixed KV cacheThe measured CPU/CUDA dispatch difference
A long prompt on a 16GB M4 takes time before any outputFirst response, RSS, KV memory and swap
You want the performance numbers and their limitsBenchmarks and measurement conditions

Before applying a fix

Match the operation and the tested versions to your own setup. A VHDX that starts normally can still fail during resize. A GPU listed by PyTorch can still lack the kernel needed by the next operation. The reports preserve these distinctions so that an error from one step does not become a diagnosis of the entire system.

Storage commands deserve particular care: find the exact distribution and disk path, read the backup instructions, and verify the result before deleting an original copy. Each storage report explains which changes it actually tested.

How to read a result

The environment box states where and when a result was measured. Tables describe the operations and samples in that experiment. The linked log contains commands and output; the limitations section identifies what remains untested.

For example, a raw file read and the first model inference measure different parts of model loading. Likewise, the llama.cpp throughput report uses synthetic tokens and does not measure answer quality. Follow the metric definition as well as the number.

Browse all reports by date, read how the lab works, or search for an exact error.