ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 05:01:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] moe sorting ex kernel to support expert > 128 (#1840)

Browse Source 
* moe sorting ex

* fix bug for race condition

* fix bug and optimze large expert

* fix

* optimize with sub_token_oneshot

* support skip empty tokens for expert sorting

* update moe_sorting

* tidy code
This commit is contained in:
carlushuang
2025-02-11 17:49:17 +08:00
committed by GitHub
parent 2312f4aaf9
commit c0adab4850
12 changed files with 936 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/15_fused_moe/README.md
View File

@@ -42,7 +42,7 @@ summary of the key design of this fused-moe operator:
// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5
// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
//
// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated)
// * this could be larger than actual, since actual tokens are on GPU
//
// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]