Always convert soft prompt to float32 if using TPU backend

TPUs do not support float16. Attempting to use a float16 soft prompt throws an error.
2021-11-21 18:22:10 -05:00 · 2021-11-21 18:22:10 -05:00 · 9b8bcb5516
parent e068aa9f26
commit 9b8bcb5516
1 changed files with 1 additions and 1 deletions
--- a/aiserver.py
+++ b/aiserver.py
@ -2572,7 +2572,7 @@ def spRequest(filename):
            -1,
            tpu_mtj_backend.params["d_model"],
        )
-        vars.sp = tensor
+        vars.sp = np.float32(tensor)
    else:
        vars.sp = torch.from_numpy(tensor)