Qwen3-8B on WSL2: My Context Test Stopped at the Host RAM Guard
Six single-sample Qwen3-8B scouts on an 8GB RTX 5060 Ti, followed by an incomplete primary batch stopped below 2 GiB of available host RAM.
LOCAL AI · INDEPENDENT EXPERIMENTS
Small hardware.
Deep experiments.
What can the hardware on my desk actually do?
Experiments in file paths, memory, and compatibility.
Tested conditions. Commands you can repeat.
10 REPORTSWINDOWS / WSL2CONSUMER HARDWARE
STUDY 01 / LAYERSCONCEPTUAL 3D
COMPUTEMEMORYSTORAGE
The layers behind a local AI experiment.
Not a hardware model or live measurement.
LATEST / PUBLISHED REPORTS
Recently recorded.
Hugging Face Streaming on WSL2: Cache and Repeated Reads
streaming=True on WSL2: actual cache files, first-row latency, RAM and repeated HTTP reads. CSV and Parquet, 60 measured processes.
01 / THE NOTEBOOK
Follow the question.
One dataset.
Two cache roots.
Changing HF_DATASETS_CACHE changed the destination of newly prepared Arrow files—not the downloaded CSVs.
17 conditions × 3 runsRead the experiment
OBSERVED FILE LOCATIONS WSL2
CSV
DOWNLOADED SOURCE
HF_HOME/hubARROW
PREPARED DATA
HF_DATASETS_CACHEChanging a variable does not move existing files.
Explore the cache map ARCHIFY ↗A CUDA buffer.
CPU attention.
The same mixed KV cache took different execution paths with one build flag changed.
011,412
RTX 5060 Ti · Qwen3 1.7B Q4_K_M · q8_0/q4_0
llama.cpp f280b2698 · median of 3 selected samples
One GPU. Two builds.Read report
Detected isn't
the same as working.
Three PyTorch wheels detected the GPU. Actual operations told a different story.
| Wheel | Detected | fill_ |
|---|---|---|
| cu126 | 3/3 | 0/3 |
| cu130 | 3/3 | 3/3 |
| cu132 | 3/3 | 3/3 |
Completed probes / 3 fresh processes. Not a general compatibility score.
Check the operation.Read report
02 / THE METHOD
The result matters.
So does the route.
Each report keeps the environment, commands, evidence, and limits together.
Find your starting point- 01
Pin the conditions.
Hardware, versions, revisions, and the question being tested.
- 02
Run and record.
Real operations and output. Keep failure paths visible.
- 03
Check the evidence.
Trace file locations and execution—not just status flags.
- 04
State the limits.
Explain what this setup does and does not establish.
03 / ON THE DESK
The hardware at hand.
A small workbench.
Not a GPU leaderboard.
BENCH / 01PUBLISHED EXPERIMENTS
RTX 5060 Ti 8 GB
Windows + WSL2
Cache · CUDA · llama.cppView reports ↗
BENCH / 02PUBLISHED EXPERIMENT
MacBook M4 16 GB
macOS · system memory
Long prompts · Metal · memoryView report ↗
04 / FIELD NOTES
From the notebook.
Qwen3-8B on WSL2: My Context Test Stopped at the Host RAM Guard
Six single-sample Qwen3-8B scouts on an 8GB RTX 5060 Ti, followed by an incomplete primary batch stopped below 2 GiB of available host RAM.
Hugging Face Streaming on WSL2: Cache and Repeated Reads
streaming=True on WSL2: actual cache files, first-row latency, RAM and repeated HTTP reads. CSV and Parquet, 60 measured processes.
Long Prompts on a 16GB M4: Memory, Swap, and First Response
Qwen3-8B on a 16GB M4: nine Metal runs measure first response, RSS, KV memory, compression and swap. A separate 16K scout hit a 120-second cutoff.
HF_DATASETS_CACHE on WSL2: Where the Other Files Go
Measured Hub downloads, Arrow files, import order, saved datasets, and cleanup on WSL2. Why HF_DATASETS_CACHE and cache_dir leave other files behind.
llama.cpp Mixed KV Cache on RTX 5060 Ti: q8_0/q4_0 Test
On an RTX 5060 Ti, llama.cpp's default CUDA build sent mixed q8_0/q4_0 Flash Attention to CPU; one build flag restored CUDA and an 86x PP gap.
PyTorch 2.13 on RTX 5060 Ti: cu126 vs cu130 vs cu132
All three PyTorch 2.13 wheels saw my RTX 5060 Ti, but cu126 mixed working dot/GEMM paths with missing native CUDA kernels.
torch.cuda.is_available() Returns False on WSL2: 4 Tested Causes
I tested four reasons torch.cuda.is_available() returns False on WSL2: a CUDA 13 driver mismatch, CPU-only wheel, hidden GPU, and wrong Python.
Fix WSL2 Resize Error Wsl/Service/0xc03a001a (2.7.3)
On WSL 2.7.3, compressed and sparse VHDX files booted but --resize failed with 0xc03a001a; clearing the attribute restored 9/9 retries.
Hugging Face Cache on WSL2: HF_HOME, /mnt/c, and Cleanup
I tested HF_HOME, HF_HUB_CACHE, symlinks, and /mnt/c on WSL2, then measured cleanup behavior with the current hf cache commands.
Shrink WSL2 ext4.vhdx in 2026: Sparse VHD vs diskpart, Measured
Deleted 20 GB inside WSL but ext4.vhdx stayed huge? Measured on WSL 2.7.3: fstrim no longer needed, sparse VHD gated as unsafe, diskpart rejects sparse files.