Support FA3 backend for gpt-oss (#9028)

This commit is contained in:
Ke Bao
2025-08-14 01:41:50 +08:00
committed by GitHub
parent 4a16a71c36
commit 0ff6d1fce1
4 changed files with 24 additions and 6 deletions

View File

@@ -294,7 +294,7 @@ class GptOssAttention(nn.Module):
)
self.sinks = nn.Parameter(
torch.empty(self.num_heads, dtype=torch.float32), requires_grad=False
torch.empty(self.num_heads, dtype=torch.bfloat16), requires_grad=False
)
self.o_proj = RowParallelLinear(