chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[opt kimi k2 4 / n] Delete useless pad kernel in sgl_moe_align_block_size (#13587)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-11-21 13:16:42 +08:00
committed by GitHub
parent fb04d43428
commit bfcf15a129
1 changed files with 1 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py
+1 -6
View File
@@ -69,11 +69,6 @@ def moe_align_block_size(
(num_experts + 2,), dtype=torch.int32, device=topk_ids.device (num_experts + 2,), dtype=torch.int32, device=topk_ids.device
) )
# Threshold based on benchmark results
fuse_sorted_ids_padding = sorted_ids.shape[0] <= 4096
if not fuse_sorted_ids_padding:
sorted_ids.fill_(topk_ids.numel())
sgl_moe_align_block_size( sgl_moe_align_block_size(
topk_ids, topk_ids,
num_experts + 1, num_experts + 1,
@@ -82,6 +77,6 @@ def moe_align_block_size(
expert_ids, expert_ids,
num_tokens_post_pad, num_tokens_post_pad,
cumsum_buffer, cumsum_buffer,
fuse_sorted_ids_padding, True,
) )
return sorted_ids, expert_ids, num_tokens_post_pad return sorted_ids, expert_ids, num_tokens_post_pad