fix: Properly determine seq_len for TextTransformer processing

2023-11-18 23:04:56 -08:00
parent 2fa1b95045
commit 0de2ab3f9a
1 changed files with 1 additions and 2 deletions
@@ -193,8 +193,7 @@ class PrompLangCLIPTextTransformer(CLIPTextTransformer):
        hidden_states = self.embeddings(input_dicts=input_ids)

        bsz = len(input_ids)
-        # TODO: Properly gather this
-        seq_len = 77
+        seq_len = hidden_states.shape[1]

        causal_attention_mask, attention_mask = self.process_attention_mask(hidden_states, attention_mask, bsz, seq_len)