[2/2] Deepseek deterministic: support deepseek v3 deterministic inference on 8 x H200 (#12095)

This commit is contained in:
Minglei Zhu
2025-10-29 11:49:04 -07:00
committed by GitHub
parent bacb3825fe
commit e39628fd07
3 changed files with 19 additions and 0 deletions

View File

@@ -515,6 +515,9 @@ class MoEGate(nn.Module):
True, # is_vnni
)
if get_global_server_args().enable_deterministic_inference:
return F.linear(hidden_states, self.weight, None)
# NOTE: For some unknown reason, router_gemm seems degrade accept length.
if (
_is_cuda