[chmjkb]

Description

GPU suites for attn, attnDecode, attnFlash and rope, with a shared f64 reference in checkSdpa.ts (windows, GQA, sinks, segments, packed qkv).

kernel	covered
attn	rowsSubgroupSplit, rowsSplit, rows, batch, each with every option; headDim 1/200/512, chunk boundaries (96/97 keys, past 16 chunks), maxSegment on both sides of the chunk, past 65535 row-heads, split scratch past the binding limit, rows that see no key
attnDecode	decode and decodeSplit, with and without subgroups, f32/f16; chunk thresholds, headDim 1/37/128, windows, sinks, MQA, 64 heads, two split calls in one submit
attnFlash	headDim 16 to 512 (both lane counts and tile sizes), every option, partial workgroups, fallbacks on the no-subgroups device
rope	slab slices, a hand-worked rotation, past GRID_Y_STRIDE

Known bugs found

• decodeSplit with more than 64 query heads: heads 64+ come out 0 (scratch holds 64 heads, the merge runs one 64-thread workgroup). Repro: 72 q heads, 8 kv heads, kvLen 600.
• attnFlash past 65535·16 rows: grid x passes the per-dimension limit, output is all zeros, no error. Repro: qLen 1048561, headDim 16.
• batch attention past 2^24 row-heads and rope f16 past 2^24 pairs: the flat index is split with d.u32(a / b), which TypeGPU compiles to an f32 division.

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:gpu: 787 passed, 15 expected fail at this point in the stack

Screenshots

Related issues

Stack: 1. reduce → 2. attention → 3. shape + elementwise → 4. matmul → 5. conv + ssm. Related: #49 (empty tensor fix), #44 (same f32 index bug in tiled matmul).

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

Additional notes

• Tests only, no src/ changes. Each case says in a comment which route it takes and why; nothing asserts the route yet.
• Every suite checks against a plain CPU loop written from the formula (or torch values where noted), plus wrong-dtype and bad-shape refusals.
• Each suite has an empty-input case as it.fails. fix(core): reject an empty tensor with INVALID_ARGUMENT #49 fixes that bug; once it merges these flip to it in a follow-up commit at the bottom of the stack.
• Known bugs found are it.fails with a comment naming the bug. None are fixed here.