[Phantomn]

Problem

On WSL2 with CUDA, qmd embed sometimes aborts with exit 134:

ggml-cuda.cu:106: CUDA error
CUDA error: out of memory
  current device: 0, in function alloc at ggml-cuda.cu:593
  cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0)

This happens with plenty of VRAM free (RTX 4060 Ti 8 GB, about 6.8 GB free, embeddinggemma-300M).

Root cause

resolveSafeParallelism already serializes CUDA to one context on Windows because of #519. It only checks process.platform === "win32". WSL reports linux, but its CUDA runs on the Windows GPU driver: nvidia-smi inside WSL shows the Windows driver version. So WSL gets the full computed pool (8 contexts here), and the same class of multi-context CUDA failure appears.

Change

• Move isWSL() from store.ts to llm.ts and export it. store.ts already imports from llm.ts, so this avoids an import cycle and keeps one WSL detector.
• resolveSafeParallelism: treat WSL + CUDA like Windows + CUDA and default to 1 context. Vulkan and CPU are unchanged. QMD_EMBED_PARALLELISM still overrides it.
• Add an optional wsl option so tests are deterministic on any host.
• Add a CHANGELOG entry under Unreleased.

Evidence (WSL2 Ubuntu 24.04, driver 616.92, qmd 2.8.3)

A single ~1700-chunk document, re-embedded repeatedly with other CUDA processes holding the GPU:

QMD_EMBED_PARALLELISM	runs	aborts
8 (the auto value on this card)	7	1, identical cuMemAddressReserve abort
1	7	0

The same abort also hit 2 real qmd embed runs before the investigation. Cost of serializing: 31s → 42s for that document.

Ruled out along the way: real VRAM exhaustion (about 6.8 GB free), other processes holding CUDA (embed still succeeded with 5 holders), and address-space leaks from context create/dispose cycles (VmSize flat over 40 cycles).

Tests

• test/llm.test.ts: the new WSL case fails on main and passes here.
• Full vitest run test/: 1247 passed, 79 skipped. Run with WSL_DISTRO_NAME/WSL_INTEROP unset, as CI does.
• tsc -p tsconfig.build.json --noEmit and oxlint are clean.

Note: 4 Git Bash tests in test/store-paths.test.ts fail when run inside WSL on both main and this branch, because isWSL() is true there. That behavior is pre-existing and unchanged.

Related: #519, #640

🤖 Generated with Claude Code

[Phantomn]

Friendly ping: main still gates the CUDA one-context default on win32 only (resolveSafeParallelism in src/llm.ts), so WSL2 + CUDA users can still hit the intermittent cuMemAddressReserve abort. This PR is rebased-clean against current main. Happy to rebase, split, or gather more logs if that helps review.

[Phantomn]

Rebased onto current main (86d2365, after #1072) — no conflicts. Re-checked against the stricter type-check that #1072 added:

• tsc -p tsconfig.json --noEmit (whole project, tests included) and tsc -p tsconfig.build.json --noEmit: clean
• oxlint: clean
• test/llm.test.ts: 53 passed under vitest, 53 passed under bun. The new WSL test fails on main and passes on this branch.

The CI and Nix workflows on this PR show action_required, so they haven't run yet. Could you approve them when you get a chance?