project_6 5ca49d0e7c [docs] GridEvenShare work distribution — BI-V100 attention reduce needs only 9 CTAs ...

From reading cccl_upstream/cub/cub/grid/grid_even_share.cuh:

Key finding: For Qwen3.6 attention score reduction (100K seq_len),
with tile_items=12288 (512 threads × 24 items), only 9 CTAs are
needed. All fit in one wave on 16 SMs.

This means reduce tuning (items/threads) matters less than the V1/V2
dispatch choice in paged_attn.py. V1 uses a single CTA iterating
sequentially over all KV blocks, completely bypassing GridEvenShare's
parallel distribution. V2 would enable partition-based parallelism.

Also documents: RAKE (scan) vs STRIP_MINE (reduce) strategies,
'big shares' load balancing, and the SingleTile fast path for
short sequences.

2026-08-05 03:36:45 +00:00

..

server_recon

docs: document Qwen3.6-35B-A3B bootstrap failure and architecture analysis

2026-08-01 13:16:01 +00:00

CCCL_REDUCE_ARCHITECTURE_NOTES.md

[docs] GridEvenShare work distribution — BI-V100 attention reduce needs only 9 CTAs

2026-08-05 03:36:45 +00:00

CCCL_TO_TRITON_METHODOLOGY.md

docs: CCCL → Triton methodology transfer — parameter search translation table

2026-08-03 04:37:23 +00:00

paged_attention_kernel_architecture.md

[ARCH] CCCL-derived paged attention kernel architecture + Triton rewrite

2026-07-31 04:13:07 +00:00

SPECIALIZATION_ANALYSIS.md

[muh] kernel_map: full vllm→CCCL mapping with SMEM overflow detection

2026-07-31 11:13:33 +00:00