muh-engine
5fbcfff7f3
[ENGINE] fused_moe.py: CCCL kernel_transform_tile assume_divisible
Applied CCCL kernel_transform_tile.cuh patterns to MoE config:
1. assume_divisible<16> principle: BLOCK_SIZE_M always a multiple of 16
so moe_align_block_size produces token counts compatible with
vectorized LDG.E.128 loads (128-bit aligned memory access).
2. partition_view pattern: moe_align_block_size already implements
CCCL's auto-partitioning (pad tokens to BLOCK_SIZE_M boundary),
added comments linking this to kernel_transform_tile.cuh.
3. GridEvenShare + spread_out_items sizing: added numel 256-1024 tier
(was collapsing 64→1024 into single BLOCK_SIZE_M=64). For large
prefill (numel>1024), use 256 to amortize launch overhead.
CCCL file: cub/device/dispatch/kernels/kernel_transform_tile.cuh
2026-08-05 09:30:53 +00:00
..
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-08-05 09:30:53 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00
2026-07-30 16:06:20 +00:00