HF-Reference LLM mode + Update test result to match latest Turbine. (#2080)

* HF-Reference LLM mode.

* Fixup test to match current output from Turbine.

* lint

* Fix test error message + Only initialize HF torch model when used.

* Remove redundant format_out change.

apps/shark_studio/api/llm.py

@@ -9,7 +9,7 @@ from itertools import chain
 import gc
 import os
 import torch
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
 llm_model_map = {
     "llama2_7b": {
@@ -109,6 +109,7 @@ class LanguageModel:
         self.global_iter = 0
         self.prev_token_len = 0
         self.first_input = True
+        self.hf_auth_token = hf_auth_token
         if self.external_weight_file is not None:
             if not os.path.exists(self.external_weight_file):
                 print(
@@ -164,6 +165,8 @@ class LanguageModel:
                 use_auth_token=hf_auth_token,
             )
             self.compile()
+        # Reserved for running HF torch model as reference.
+        self.hf_mod = None
 
     def compile(self) -> None:
         # this comes with keys: "vmfb", "config", and "temp_file_to_unlink".
@@ -267,6 +270,50 @@ class LanguageModel:
         self.global_iter += 1
         return result_output, total_time
 
+    # Reference HF model function for sanity checks.
+    def chat_hf(self, prompt):
+        if self.hf_mod is None:
+            self.hf_mod = AutoModelForCausalLM.from_pretrained(
+                self.hf_model_name,
+                torch_dtype=torch.float,
+                token=self.hf_auth_token,
+            )
+        prompt = self.sanitize_prompt(prompt)
+
+        input_tensor = self.tokenizer(prompt, return_tensors="pt").input_ids
+        history = []
+        for iter in range(self.max_tokens):
+            token_len = input_tensor.shape[-1]
+            if self.first_input:
+                st_time = time.time()
+                result = self.hf_mod(input_tensor)
+                token = torch.argmax(result.logits[:, -1, :], dim=1)
+                total_time = time.time() - st_time
+                token_len += 1
+                pkv = result.past_key_values
+                self.first_input = False
+
+            history.append(int(token))
+            while token != llm_model_map["llama2_7b"]["stop_token"]:
+                dec_time = time.time()
+                result = self.hf_mod(token.reshape([1, 1]), past_key_values=pkv)
+                history.append(int(token))
+                total_time = time.time() - dec_time
+                token = torch.argmax(result.logits[:, -1, :], dim=1)
+                pkv = result.past_key_values
+                yield self.tokenizer.decode(history), total_time
+
+            self.prev_token_len = token_len + len(history)
+
+            if token == llm_model_map["llama2_7b"]["stop_token"]:
+                break
+        for i in range(len(history)):
+            if type(history[i]) != int:
+                history[i] = int(history[i])
+        result_output = self.tokenizer.decode(history)
+        self.global_iter += 1
+        return result_output, total_time
+
 
 if __name__ == "__main__":
     lm = LanguageModel(

apps/shark_studio/tests/api_test.py

@@ -20,14 +20,15 @@ class LLMAPITest(unittest.TestCase):
             quantization="None",
         )
         count = 0
+        label = "Turkishoure Turkish"
         for msg, _ in lm.chat("hi, what are you?"):
             # skip first token output
             if count == 0:
                 count += 1
                 continue
             assert (
-                msg.strip(" ") == "Turkish Turkish Turkish"
-            ), f"LLM API failed to return correct response, expected 'Turkish Turkish Turkish', received {msg}"
+                msg.strip(" ") == label
+            ), f"LLM API failed to return correct response, expected '{label}', received {msg}"
             break
         del lm
         gc.collect()