fix unnecessary pos_emb for RoPE (#2963)

shucongzhang · Shucong Zhang/Embedded AI /SRUK/Engineer/Samsung Electronics · web-flow · commit 7d214a38861b · 2025-08-13T10:47:19.000+02:00
Co-authored-by: Shucong Zhang/Embedded AI /SRUK/Engineer/Samsung Electronics &lt;s1.zhang@sruk-ccn4.eu.corp.samsungelectronics.net&gt;
diff --git a/speechbrain/lobes/models/transformer/Conformer.py b/speechbrain/lobes/models/transformer/Conformer.py
@@ -802,10 +802,7 @@ def forward_streaming(
             The attention values.
         """
 
-        if (
-            self.attention_type == "RelPosMHAXL"
-            or self.attention_type == "RoPEMHA"
-        ):
+        if self.attention_type == "RelPosMHAXL":
             if pos_embs is None:
                 raise ValueError(
                     f"The chosen attention type for the Conformer is {self.attention_type}. For this attention type, the positional embeddings are mandatory"
