These reports measure a specific operation under recorded conditions. Choose a report by the metric you need, then read its test setup before carrying a number into another system.
Long prompts on Apple Silicon
Qwen3-8B on a 16GB M4 measures first-content latency, completion time, process RSS, runtime KV allocation, compression and swap. Nine primary runs cover 2K/4K/8K capacities with progressively longer input and 128 output tokens. A separate 16K scout reached the 120-second first-content cutoff; it is not an OOM or maximum-context result.
The public records include every repeat and distinguish process, runtime and system memory. This is one AC-powered Mac with ordinary apps retained, not a cross-device performance comparison.
Inference throughput and KV memory
llama.cpp mixed KV cache on RTX 5060 Ti compares prompt processing and token generation for several K/V cache types and context depths. The experiment changes one CUDA build option and uses separate scheduler traces to verify where mixed-cache Flash Attention runs.
The report includes llama.cpp’s KV buffer sizes, build cost, and timing samples. Its synthetic-token measurements exclude tokenization and sampling. It does not compare model quality, and process-wide GPU telemetry is not presented as the exact memory use of an individual timed step.
Model files on ext4 versus /mnt/c
Hugging Face cache location on WSL2 measures sequential file reads, tokenizer loading, model loading, and loading followed by a first forward pass. The paths were on one physical SSD so the comparison did not also change storage hardware.
The first forward pass matters: memory-mapped model loading can defer reading weight data until inference. The experiment clears Linux page cache between runs but does not claim control of Windows caching or Defender scanning.
Windows disk allocation
WSL2 sparse VHD and diskpart measurements track logical VHDX length and allocated size through a fill, delete, and reclaim sequence. The resize error experiment separately measures compressed and sparse VHDX behavior during expansion and recovery.
These are storage measurements. A smaller allocation or a successful launch does not establish faster inference or long-term disk integrity.
Compatibility results have a different purpose
The PyTorch 2.13 wheel matrix records which operations completed correctly on an RTX 5060 Ti. It does not rank wheel performance. Use Compatibility for those pass/fail results and About for the lab’s reporting and correction policy.