Only cut decoder_input_ids if past model output

src/models.js

@@ -409,10 +409,17 @@ function seq2seqStartBeams(self, inputTokenIds, generation_config, numOutputToke
 async function seq2seqRunBeam(self, beam) {
     const input_name = self.main_input_name;
 
+    let decoder_input_ids = beam.output_token_ids;
+    if (beam.prev_model_outputs) {
+        // After the first step, `prev_model_outputs` won't be null.
+        // So, we cut decoder_input_ids if past is used
+        decoder_input_ids = decoder_input_ids.slice(-1);
+    }
+
     // 1. Prepare
     let model_inputs = {
         [input_name]: beam.inputs,
-        decoder_input_ids: toI64Tensor(beam.output_token_ids.slice(-1)),
+        decoder_input_ids: toI64Tensor(decoder_input_ids),
         encoder_outputs: beam.encoder_outputs,
         past_key_values: beam.prev_model_outputs?.past_key_values,
     }
@@ -3294,15 +3301,6 @@ export class MarianMTModel extends MarianPreTrainedModel {
         this.num_encoder_heads = this.config.encoder_attention_heads;
         this.encoder_dim_kv = this.config.d_model / this.num_encoder_heads;
     }
-
-
-    /**
-     * @param {any} model_inputs
-     * @returns {Promise<Seq2SeqLMOutput>}
-     */
-    async forward(model_inputs) {
-        return await seq2seqForward(this, model_inputs);
-    }
 }
 //////////////////////////////////////////////////
 
@@ -3335,13 +3333,6 @@ export class M2M100ForConditionalGeneration extends M2M100PreTrainedModel {
         this.encoder_dim_kv = this.config.d_model / this.num_encoder_heads;
     }
 
-    /**
-     * @param {any} model_inputs
-     * @returns {Promise<Seq2SeqLMOutput>}
-     */
-    async forward(model_inputs) {
-        return await seq2seqForward(this, model_inputs);
-    }
 }
 //////////////////////////////////////////////////