Make KVarN usable on ROCm RDNA3 (toward #122: D256 race, portable prefill, f32 WMMA prototype) - #149

Merged

Anbeeld merged 17 commits into

Anbeeld:v0.4.7from

raufaser:kvarn-rocm-prompt-precision

Sep 12, 2026

Commits on Sep 7, 2026

kvarn/rocm: fix D256 k00-combine race and enable RDNA D256 route
raufaser
committed
42f2250
kvarn/rocm: route HIP prompt-prefill through portable-native attention
raufaser
committed
27c130c
kvarn/portable: warp-shuffle reduction and fp32 rescale
raufaser
committed
a6f9fc5
kvarn/tests: production-shape accuracy ladder harness
raufaser
committed
e502e19
kvarn/portable: batched queries, shuffle reduction, hoisted resolve
raufaser
committed
78b63a4

Commits on Sep 8, 2026

kvarn/rocm: fp32 WMMA accumulator for DV=128/256 prefill (opt-in)
raufaser
committed
27639c9

Commits on Sep 9, 2026

Commits on Sep 10, 2026

Merge upstream v0.4.7 (D128 low-GQA decode opt, QSA cell-alignment) into kvarn-rocm-v0.4.7
raufaser
committed
799bc58

Commits on Sep 11, 2026