added eos token id handling for starcoder models, as they use a different EOS ID

gpttype_adapter.cpp

@@ -1173,6 +1173,16 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
                             int topid = std::min_element(logits.begin(),logits.end())-logits.begin();
                             logits[eosID] = (logits[topid] < 0 ? logits[topid] : 0);
                         }
+                        else
+                        {
+                            //special case, starcoder models use ID 0 for EOS
+                            if (file_format == FileFormat::GPT2_3 || file_format == FileFormat::GPT2_4)
+                            {
+                                eosID = 0;
+                                int topid = std::min_element(logits.begin(), logits.end()) - logits.begin();
+                                logits[eosID] = (logits[topid] < 0 ? logits[topid] : 0);
+                            }
+                        }
                     }
 
                      // set the logit of the eos token (0) to minimum to avoid sampling it