Files
foxhunt/crates
jgrusewski 74f63f80b2 perf: attention + IQL block 32→256 (occupancy 3.1%→25%)
Same optimization pattern as IQN: increase block size from 32 (1 warp)
to 256 (8 warps) for all attention and IQL kernels.

Attention (forward + backward):
- Loop strides 32→256, shared_concat dynamic shmem
- Added bf16_block_max/bf16_block_sum for 8-warp reduction
- Replaces warp-only shuffle with shared memory cross-warp reduce

IQL (forward+loss, backward, forward-only):
- Loop strides 32→256, added iql_bf16_block_sum
- grad_norm + adam kernels already at 256, unchanged

Expected: ~15-20ms/step savings from attention+IQL combined.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-06 23:37:36 +02:00
..