minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

fix(moe): MoRI EP init_mori_op missing BF16 dispatch branch (#28469)

Browse Source 
Signed-off-by: Chengze Fan <fancz2002@gmail.com>
This commit is contained in:
Chengze Fan
2026-06-16 21:02:06 -07:00
committed by GitHub
parent b8a73bfba0
commit 66ac385f52
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/token_dispatcher/moriep.py
+4 -1
View File
@@ -257,7 +257,10 @@ def init_mori_op(
data_type = fp8_dtype data_type = fp8_dtype
scale_type_size = torch.float32.itemsize scale_type_size = torch.float32.itemsize
if dispatch_dtype == DispatchDtype.fp8: if dispatch_dtype == DispatchDtype.bf16:
data_type = params_dtype
scale_dim = 0
elif dispatch_dtype == DispatchDtype.fp8:
scale_dim = hidden_size // FP8_BLOCK_SIZE scale_dim = hidden_size // FP8_BLOCK_SIZE
elif dispatch_dtype == DispatchDtype.fp4: elif dispatch_dtype == DispatchDtype.fp4:
# FP4 kernel still takes the original hidden size and do quantization # FP4 kernel still takes the original hidden size and do quantization