EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix: use torch.sum for compatible (#2161)

Browse Source
This commit is contained in:
Yineng Zhang
2024-11-24 22:37:04 +08:00
committed by GitHub
parent 5652c56535
commit fa27161380
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/triton_fused_moe/fused_moe.py
View File

@@ -766,9 +766,10 @@ def fused_experts_impl(
use_int8_w8a16=use_int8_w8a16, use_int8_w8a16=use_int8_w8a16,
) )
ops.moe_sum( torch.sum(
intermediate_cache3.view(*intermediate_cache3.shape), intermediate_cache3.view(*intermediate_cache3.shape),
out_hidden_states[begin_chunk_idx:end_chunk_idx], dim=1,
out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
) )
return out_hidden_states return out_hidden_states