[chmjkb]

Description

Closes the kernel uniform leaks #60 found. Stacked on #61.

Eight kernels made their dims uniform with root.createBuffer(...) inside their handle, so it lived until the device did, past executor.dispose(): attention (batch, rows, rows split), attnDecode (decode, decode split), layerNorm, swigluChunk, conv1d, argmaxDot, gatherRowsFrom and writeRows. They now take it from cachedUniform and their bind group from cachedBindGroup, like attnFlash, rope and the other kernels already do. The uniforms are per device and keyed by value, so a new executor reuses them instead of making more.

attnDecodeHandle now lists its bind group entries one by one: the shared attention buffers carry a segs the decode layout has no slot for, which createBindGroup ignored and cachedBindGroup would key on.

The leak checks in eval.gpu.test.ts and zoo/leaks.gpu.test.ts no longer expect any uniform left over.

Kernel A/B against #61

pnpm --filter runntime bench:ab kernels -b @chmjkb/gpu-leak-fixes -k attention,reduce,shape,elementwise,conv,matmul: 150 of 150 cases same, max abs(A−B) 0 on every case.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Performance (kernel tuning, graph building, anything related to performance)
• Other (chores, tests, code style improvements etc.)

Testing instructions

1. pnpm --filter runntime exec playwright install chromium (once per machine)
2. pnpm --filter runntime test

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

[xxpopielxx]

LGTM once the stack is rebased.