Amir Ghamarian 5b591af58e Add block_size=32 support and fix int32 overflow in unified attention ...

- Add BlockSize_ template parameter to all kernel trait structs,
  enabling block_size=32 instances for d64/GQA-8 (medium/small/tiny tiers).
- 12 new instance files for block_size=32 (bf16/fp16 x mask/nmask x 3 tiers).
- Dispatch branches on page_blk_size to select bs32 vs bs64 instances.
- Fix potential int32 overflow in KV cache tensor view by using long_index_t
  for sizes and strides (overflow threshold was ~66K blocks for d64/GQA-8).

Made-with: Cursor

2026-03-30 14:16:17 +00:00

..

ck

Implement batched gemm bias permute for RDNA4 (#3534)

2026-01-17 08:30:27 +01:00

ck_tile

Add block_size=32 support and fix int32 overflow in unified attention

2026-03-30 14:16:17 +00:00

rapidjson

Update pre-commit to fixed versions, run remod for ck_tile (#2895)

2025-10-16 15:29:17 -07:00