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.
| Question | Metric | Tested environment | Report |
|---|---|---|---|
| How long does a long prompt take before content appears? | First-content latency, completion time, process RSS, runtime KV allocation, compression and swap | 16GB M4, macOS 26.6.2, Qwen3-8B Q4_K_M, llama.cpp Metal | M4 long prompts |
| What remains usable after a Qwen3-8B context batch hits a host RAM guard? | Six n=1 scouts, sampled memory, host safety trace, and one separate fresh-build replay; zero eligible primary results | RTX 5060 Ti 8 GB, Windows 11 / Ubuntu 24.04.4 WSL2, Qwen3-8B Q4_K_M, F16 K/V | Qwen3-8B host RAM stop |
| Where does mixed KV Flash Attention run, and what does that change? | Synthetic prompt-processing and token-generation throughput, KV buffer sizes, scheduler placement | RTX 5060 Ti 8 GB, WSL2, llama.cpp f280b2698, Qwen3 1.7B Q4_K_M | Mixed KV cache |
| Does a second streaming process reuse the same files and responses? | Persistent file lengths, first-row time, RSS high-water at exit, HTTP response-body bytes | WSL2 ext4, datasets 5.0.1, CSV and Parquet served by a loopback server | Streaming and repeated reads |
| Does moving a model to /mnt/c change loading costs? | Sequential file reads, tokenizer and model loading, loading plus first forward pass | WSL2 Ubuntu 26.04, ext4 and /mnt/c on one physical Samsung NVMe SSD | Model cache location |
| How much Windows disk allocation returns after WSL files are deleted? | VHDX logical length and allocated size during fill, delete and reclaim | Windows 11, WSL 2.7.3, fresh Ubuntu 26.04 | Sparse VHD and diskpart |
| Which file attributes prevent VHDX expansion? | Resize failure and recovery checks, logical length and allocated size | Windows 11, WSL 2.7.3, compressed and sparse VHDX conditions | Resize error 0xc03a001a |
The rows measure different operations and environments. They are not a shared hardware ranking. For pass/fail CUDA operation results, use Compatibility.
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
Qwen3-8B on WSL2: a host RAM guard stop preserves six individually validated scouts at 2K/4K/8K capacities with all-offloaded or --gpu-layers 18. Each condition has one observation. The later eighteen-trial primary plan stopped during startup with zero final eligible results. The report includes the host safety trace and one separate fresh-build 2K replay; it does not establish a maximum context, CUDA OOM, or a hardware ranking.
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.
Dataset streaming and repeated reads
Hugging Face streaming on WSL2 measures persistent cache files, first-row time, RSS high-water at atexit and repeated HTTP responses. Sixty fresh processes compare normal and streaming reads, CSV and Parquet, a 1,000-row prefix and all 200,000 rows. The response counters are from a loopback server; streaming’s time and memory tradeoffs differ by format.
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.