Fix softprompts when deepspeed is active (#112)

modules/text_generation.py

@@ -37,7 +37,7 @@ def generate_softprompt_input_tensors(input_ids):
     inputs_embeds = shared.model.transformer.wte(input_ids)
     inputs_embeds = torch.cat((shared.soft_prompt_tensor, inputs_embeds), dim=1)
     filler_input_ids = torch.zeros((1, inputs_embeds.shape[1]), dtype=input_ids.dtype).to(shared.model.device)
-    filler_input_ids += shared.model.config.bos_token_id # setting dummy input_ids to bos tokens
+    #filler_input_ids += shared.model.config.bos_token_id # setting dummy input_ids to bos tokens
     return inputs_embeds, filler_input_ids
 
 # Removes empty replies from gpt4chan outputs