fix: chatglm not working in doc qa, meta qa and plugin (#318)

Close #316

pilot/configs/config.py

@@ -28,7 +28,7 @@ def __init__(self) -> None:
         self.skip_reprompt = False
         self.temperature = float(os.getenv("TEMPERATURE", 0.7))
 
-        self.NUM_GPUS = int(os.getenv("NUM_GPUS",1))
+        self.NUM_GPUS = int(os.getenv("NUM_GPUS", 1))
 
         self.execute_local_commands = (
             os.getenv("EXECUTE_LOCAL_COMMANDS", "False") == "True"

pilot/model/adapter.py

@@ -73,6 +73,7 @@ def loader(self, model_path: str, from_pretrained_kwagrs: dict):
         )
         return model, tokenizer
 
+
 def auto_configure_device_map(num_gpus):
     """handling multi gpu calls"""
     # transformer.word_embeddings occupying 1 floors
@@ -81,18 +82,18 @@ def auto_configure_device_map(num_gpus):
     # Allocate a total of 30 layers to number On gpus cards
     num_trans_layers = 28
     per_gpu_layers = 30 / num_gpus
-    #Bugfix: call torch.embedding in Linux and the incoming weight and input are not on the same device, resulting in a RuntimeError
-    #Under Windows, model. device will be set to transformer. word_ Embeddings. device
-    #Under Linux, model. device will be set to lm_ Head.device
-    #When calling chat or stream_ During chat, input_ IDS will be placed on model. device
-    #If transformer. word_ If embeddings. device and model. device are different, it will cause a RuntimeError
-    #Therefore, here we will transform. word_ Embeddings, transformer. final_ Layernorm, lm_ Put all the heads on the first card
+    # Bugfix: call torch.embedding in Linux and the incoming weight and input are not on the same device, resulting in a RuntimeError
+    # Under Windows, model. device will be set to transformer. word_ Embeddings. device
+    # Under Linux, model. device will be set to lm_ Head.device
+    # When calling chat or stream_ During chat, input_ IDS will be placed on model. device
+    # If transformer. word_ If embeddings. device and model. device are different, it will cause a RuntimeError
+    # Therefore, here we will transform. word_ Embeddings, transformer. final_ Layernorm, lm_ Put all the heads on the first card
     device_map = {
-        'transformer.embedding.word_embeddings': 0,
-        'transformer.encoder.final_layernorm': 0,
-        'transformer.output_layer': 0,
-        'transformer.rotary_pos_emb': 0,
-        'lm_head': 0
+        "transformer.embedding.word_embeddings": 0,
+        "transformer.encoder.final_layernorm": 0,
+        "transformer.output_layer": 0,
+        "transformer.rotary_pos_emb": 0,
+        "lm_head": 0,
     }
 
     used = 2
@@ -102,7 +103,7 @@ def auto_configure_device_map(num_gpus):
             gpu_target += 1
             used = 0
         assert gpu_target < num_gpus
-        device_map[f'transformer.encoder.layers.{i}'] = gpu_target
+        device_map[f"transformer.encoder.layers.{i}"] = gpu_target
         used += 1
 
     return device_map
@@ -114,7 +115,13 @@ class ChatGLMAdapater(BaseLLMAdaper):
     def match(self, model_path: str):
         return "chatglm" in model_path
 
-    def loader(self, model_path: str, from_pretrained_kwargs: dict, device_map=None, num_gpus=CFG.NUM_GPUS):
+    def loader(
+        self,
+        model_path: str,
+        from_pretrained_kwargs: dict,
+        device_map=None,
+        num_gpus=CFG.NUM_GPUS,
+    ):
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 
         if DEVICE != "cuda":
@@ -125,10 +132,8 @@ def loader(self, model_path: str, from_pretrained_kwargs: dict, device_map=None,
         else:
             model = (
                 AutoModel.from_pretrained(
-                    model_path, trust_remote_code=True,
-                    **from_pretrained_kwargs
-                )
-                .half()
+                    model_path, trust_remote_code=True, **from_pretrained_kwargs
+                ).half()
                 # .cuda()
             )
             from accelerate import dispatch_model

pilot/model/llm_out/chatglm_llm.py

@@ -1,5 +1,8 @@
 #!/usr/bin/env python3
 # -*- coding:utf-8 -*-
+
+from typing import List
+import re
 import copy
 
 import torch
@@ -33,34 +36,36 @@ def chatglm_generate_stream(
     messages = prompt.split(stop)
     #
     # # Add history conversation
-    hist = []
-    once_conversation = []
+    hist = [HistoryEntry()]
+    system_messages = []
     for message in messages[:-2]:
         if len(message) <= 0:
             continue
-
         if "human:" in message:
-            once_conversation.append(message.split("human:")[1])
-        # elif "system:" in message:
-        #     once_conversation.append(f"""###system:{message.split("system:")[1]} """)
+            hist[-1].add_question(message.split("human:")[1])
+        elif "system:" in message:
+            msg = message.split("system:")[1]
+            hist[-1].add_question(msg)
+            system_messages.append(msg)
         elif "ai:" in message:
-            once_conversation.append(message.split("ai:")[1])
-            last_conversation = copy.deepcopy(once_conversation)
-            hist.append(last_conversation)
-            once_conversation = []
-        # else:
-        #     once_conversation.append(f"""###system:{message} """)
+            hist[-1].add_answer(message.split("ai:")[1])
+            hist.append(HistoryEntry())
+        else:
+            # TODO
+            # hist[-1].add_question(message.split("system:")[1])
+            # once_conversation.append(f"""###system:{message} """)
+            pass
 
     try:
         query = messages[-2].split("human:")[1]
     except IndexError:
-        # fix doc qa: https://github.com/csunny/DB-GPT/issues/274
-        doc_qa_message = messages[-2]
-        if "system:" in doc_qa_message:
-            query = doc_qa_message.split("system:")[1]
-        else:
-            query = messages[-3].split("human:")[1]
+        query = messages[-3].split("human:")[1]
+    hist = build_history(hist)
+    if not hist:
+        # No history conversation, but has system messages, merge to user`s query
+        query = prompt_adaptation(system_messages, query)
     print("Query Message: ", query)
+    print("hist: ", hist)
     # output = ""
     # i = 0
 
@@ -75,3 +80,43 @@ def chatglm_generate_stream(
         yield output
 
     yield output
+
+
+class HistoryEntry:
+    def __init__(self, question: str = "", answer: str = ""):
+        self.question = question
+        self.answer = answer
+
+    def add_question(self, question: str):
+        self.question += question
+
+    def add_answer(self, answer: str):
+        self.answer += answer
+
+    def to_list(self):
+        if self.question == "" or self.answer == "":
+            return None
+        return [self.question, self.answer]
+
+
+def build_history(hist: List[HistoryEntry]) -> List[List[str]]:
+    return list(filter(lambda hl: hl is not None, map(lambda h: h.to_list(), hist)))
+
+
+def prompt_adaptation(system_messages: List[str], human_message: str) -> str:
+    if not system_messages:
+        return human_message
+    system_messages_str = " ".join(system_messages)
+    adaptation_rules = [
+        r"Question:\s*{}\s*",  # chat_db scene
+        r"Goals:\s*{}\s*",  # chat_execution
+        r"问题:\s*{}\s*",  # chat_knowledge zh
+        r"question:\s*{}\s*",  # chat_knowledge en
+    ]
+    # system message has include human question
+    for rule in adaptation_rules:
+        pattern = re.compile(rule.format(re.escape(human_message)))
+        if re.search(pattern, system_messages_str):
+            return system_messages_str
+    # https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
+    return f"{system_messages_str}\n\n问：{human_message}\n\n答："