[feature] add deepspeed, rallio dialogue dataset and codegen parameters

2026-09-11 11:50:17 +08:00 · 2023-01-06 16:47:58 +00:00
parent f2db4220bb
commit b67181776a
8 changed files with 169 additions and 51 deletions
@@ -7,7 +7,7 @@ class CrossEntropyLoss(nn.CrossEntropyLoss):

    def forward(self, input, target, mask=None):
        if mask is not None:
-            mask = mask.view(-1)
+            mask = mask.view(-1).bool()
            input = input.view(-1, input.size(-1))
            target = target.view(-1)
            input = input[mask]