fix encoder/decoder errors

python/ctranslate2/converters/eole_ct2.py

@@ -29,10 +29,16 @@ def _get_model_spec_seq2seq(
         getattr(config.embeddings, "position_encoding_type", None)
         == PositionEncodingType.Rotary
     )
+    if with_rotary:
+        raise ValueError(
+            "Rotary embeddings are not supported yet for encoder/decoder models"
+        )
     with_alibi = (
         getattr(config.embeddings, "position_encoding_type", None)
         == PositionEncodingType.Alibi
     )
+    if with_alibi:
+        raise ValueError("Alibi is not supported yet for encoder/decoder models")
     activation_fn = getattr(config, "mlp_activation_fn", "relu")
 
     # Return the first head of the last layer unless the model was trained with alignments.
@@ -44,26 +50,30 @@ def _get_model_spec_seq2seq(
         alignment_heads = config.decoder.alignment_heads
 
     num_heads = getattr(config.decoder, "heads", 8)
-    num_kv = getattr(config.decoder, "heads_kv", 0)
-    if num_kv == num_heads or num_kv == 0:
-        num_kv = None
-    rotary_dim = 0 if with_rotary else None
-    rotary_interleave = getattr(config.rope_config, "rotary_interleave", True)
+    # num_kv = getattr(config.decoder, "heads_kv", 0)
+    # if num_kv == num_heads or num_kv == 0:
+    #    num_kv = None
+    # rotary_dim = 0 if with_rotary else None
+    # rotary_interleave = getattr(config.rope_config, "rotary_interleave", True)
     ffn_glu = activation_fn == "gated-silu"
     sliding_window = getattr(config, "sliding_window", 0)
+    if sliding_window != 0:
+        raise ValueError(
+            "Sliding window is not suported yet for encoder/decoder models"
+        )
 
     model_spec = transformer_spec.TransformerSpec.from_config(
         (config.encoder.layers, config.decoder.layers),
         num_heads,
         with_relative_position=with_relative_position,
-        alibi=with_alibi,
+        # alibi=with_alibi,
         activation=_SUPPORTED_ACTIVATIONS[activation_fn],
         ffn_glu=ffn_glu,
         rms_norm=config.layer_norm == "rms",
-        rotary_dim=rotary_dim,
-        rotary_interleave=rotary_interleave,
-        num_heads_kv=num_kv,
-        sliding_window=sliding_window,
+        # rotary_dim=rotary_dim,
+        # rotary_interleave=rotary_interleave,
+        # num_heads_kv=num_kv,
+        # sliding_window=sliding_window,
         alignment_layer=alignment_layer,
         alignment_heads=alignment_heads,
         num_source_embeddings=num_source_embeddings,
@@ -202,7 +212,7 @@ def set_transformer_spec(spec, variables):
 
 
 def set_transformer_encoder(spec, variables):
-    set_input_layers(spec, variables, "encoder")
+    set_input_layers(spec, variables, "src_emb")
     set_layer_norm(spec.layer_norm, variables, "encoder.layer_norm")
     for i, layer in enumerate(spec.layer):
         set_transformer_encoder_layer(layer, variables, "encoder.transformer.%d" % i)
@@ -227,7 +237,7 @@ def set_input_layers(spec, variables, scope):
         set_position_encodings(
             spec.position_encodings,
             variables,
-            "%s.embeddings.pe" % scope,
+            "%s.pe" % scope,
         )
     else:
         spec.scale_embeddings = False
@@ -236,14 +246,19 @@ def set_input_layers(spec, variables, scope):
 
 
 def set_transformer_encoder_layer(spec, variables, scope):
-    set_ffn(spec.ffn, variables, "%s.feed_forward" % scope)
     set_multi_head_attention(
         spec.self_attention,
         variables,
         "%s.self_attn" % scope,
         self_attention=True,
     )
-    set_layer_norm(spec.self_attention.layer_norm, variables, "%s.layer_norm" % scope)
+    set_layer_norm(
+        spec.self_attention.layer_norm, variables, "%s.input_layernorm" % scope
+    )
+    set_layer_norm(
+        spec.ffn.layer_norm, variables, "%s.post_attention_layernorm" % scope
+    )
+    set_ffn(spec.ffn, variables, "%s.mlp" % scope)
 
 
 def set_transformer_decoder_layer(spec, variables, scope, with_encoder_attention=True):
@@ -258,7 +273,9 @@ def set_transformer_decoder_layer(spec, variables, scope, with_encoder_attention
     )
     if with_encoder_attention:
         set_multi_head_attention(spec.attention, variables, "%s.context_attn" % scope)
-        set_layer_norm(spec.attention.layer_norm, variables, "%s.layer_norm_2" % scope)
+        set_layer_norm(
+            spec.attention.layer_norm, variables, "%s.precontext_layernorm" % scope
+        )
     set_layer_norm(
         spec.ffn.layer_norm, variables, "%s.post_attention_layernorm" % scope
     )