fix llm_runner for tinyllama (#407)

models/turbine_models/custom_models/llm_runner.py

@@ -50,9 +50,7 @@
 parser.add_argument(
     "--prompt",
     type=str,
-    default="""<s>[INST] <<SYS>>
-Be concise. You are a helpful, respectful and honest assistant. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. <</SYS>> hi what are you? [/INST]
-""",
+    default="hi what are you?",
     help="prompt for llm model",
 )
 parser.add_argument(
@@ -183,14 +181,14 @@ def run_llm(
         streaming_llm=streaming_llm,
     )
     if not chat_mode:
+        prompt = append_user_prompt(chat_sys_prompt, prompt)
         initial_input = tokenizer(prompt, return_tensors="pt")
         example_input_id = initial_input.input_ids
         turbine_results = llm.generate(example_input_id)
         return tokenizer.decode(turbine_results)
-    prompt = chat_sys_prompt
     while True:
         user_prompt = input("User prompt: ")
-        prompt = append_user_prompt(prompt, user_prompt)
+        prompt = append_user_prompt(chat_sys_prompt, user_prompt)
         initial_input = tokenizer(prompt, return_tensors="pt")
         example_input_id = initial_input.input_ids
         result = llm.generate(example_input_id)
@@ -199,7 +197,13 @@ def run_llm(
         prompt = append_bot_prompt(prompt, bot_response)
 
 
-def run_torch_llm(hf_model_name, hf_auth_token, prompt, streaming_llm=False):
+def run_torch_llm(
+    hf_model_name,
+    hf_auth_token,
+    prompt,
+    streaming_llm=False,
+    chat_sys_prompt=DEFAULT_CHAT_SYS_PROMPT,
+):
     from turbine_models.model_builder import HFTransformerBuilder
     from transformers import AutoModelForCausalLM
 
@@ -210,13 +214,13 @@ def run_torch_llm(hf_model_name, hf_auth_token, prompt, streaming_llm=False):
         hf_auth_token=hf_auth_token,
         auto_tokenizer=AutoTokenizer,
     )
-    model_builder.build_model()
     if streaming_llm is True:
         enable_llama_pos_shift_attention(model_builder.model)
 
     def get_token_from_logits(logits):
         return torch.argmax(logits[:, -1, :], dim=1)
 
+    prompt = append_user_prompt(chat_sys_prompt, prompt)
     initial_input = model_builder.tokenizer(prompt, return_tensors="pt")
     example_input_id = initial_input.input_ids
 
@@ -256,6 +260,10 @@ def get_token_from_logits(logits):
     if args.compare_vs_torch:
         print("generating torch output: ")
         torch_output = run_torch_llm(
-            args.hf_model_name, args.hf_auth_token, args.prompt
+            args.hf_model_name,
+            args.hf_auth_token,
+            args.prompt,
+            args.streaming_llm,
+            args.chat_sys_prompt,
         )
         print(torch_output)