Xr llama dev (#68)

* add text generation requirement * add huggingface api key input * add llamanager file as manager * add llama analysis skills * add llama prompt builder * add llama write prd skills * add test llama analysis skill * fixed the skill class names * remove the configuration * change the test file Productanalysis name
AI-Citizen · Oct 19, 2023 · f2bf371 · f2bf371
1 parent 88937ad
commit f2bf371
Show file tree

Hide file tree

Showing 7 changed files with 276 additions and 1 deletion.
diff --git a/requirements.txt b/requirements.txt
@@ -33,3 +33,4 @@ redis==4.6.0
 rq==1.15.1
 eventlet==0.33.3
 tiktoken==0.5.1
+text-generation==0.6.1
diff --git a/solidgpt/src/configuration/Configuration.yaml b/solidgpt/src/configuration/Configuration.yaml
@@ -6,4 +6,8 @@ openai_model: gpt-3.5-turbo-16k
 notion_api_key : 
 notion_page_id :
 
-azure_blob_connection_string:
+azure_blob_connection_string:
+
+#Huggingface API
+HF_API_LLAMA2_BASE : 
+HF_API_KEY : 
diff --git a/solidgpt/src/manager/llamanager.py b/solidgpt/src/manager/llamanager.py
@@ -0,0 +1,67 @@
+import os
+from solidgpt.src.configuration.configreader import ConfigReader
+from text_generation import Client
+from solidgpt.src.manager.promptresource import llama_v2_prompt
+
+class LLAManager:
+    _instance = None
+
+    def __new__(cls):
+        if cls._instance is None:
+            cls._instance = super(LLAManager, cls).__new__(cls)
+        return cls._instance
+
+    def __init__(self, if_show_reply=False):
+        self.llama2_base_url = ConfigReader().get_property("HF_API_LLAMA2_BASE")
+        self.llama2_api_key = ConfigReader().get_property("HF_API_KEY")
+        self.llama_models_container = {}
+        self.if_show_reply = if_show_reply
+
+    def create_model(self, prompt, llama_api, llama_model_label, temperature=1, model=None):
+        if model is None:
+            model = self.llama2_base_url  # Use LLAMA2 base URL as the model
+        llama_model = LLamaModel(prompt, self.llama2_api_key,  self.llama2_base_url, self.if_show_reply, temperature)
+        self.llama_models_container[llama_model_label] = llama_model
+        return llama_model
+
+    def create_and_chat_with_model(self, prompt, llama_model_label, input_message, temperature=0.1, model=None):
+        llama_model = self.create_model(prompt, llama_model_label, temperature, model)
+        return llama_model.chat_with_model(input_message)
+
+    def get_llama_model(self, llama_model_label):
+        return self.llama_models_container.get(llama_model_label)
+
+    def remove_llama_model(self, llama_model_label):
+        self.llama_models_container.pop(llama_model_label, None)
+
+class LLamaModel:
+    def __init__(self, prompt, api, model, if_show_reply=True, temperature=0.1):
+        self.prompt = prompt
+        self.api = api
+        self.model = model
+        self.messages = [{"role": "system", "content": self.prompt}]
+        self.last_reply = None
+        self.if_show_reply = if_show_reply
+        self.temperature = temperature
+
+    def chat_with_model(self, input_message):
+        self.messages.append({"role": "user", "content": input_message})
+        self._run_model()
+        return self.last_reply
+
+    def _run_model(self):
+        client = Client(self.model, headers={"Authorization": f"Bearer {self.api}"}, timeout=120)
+        chat = client.generate(
+            llama_v2_prompt(self.messages),  # Convert messages to LLAMA2 prompt
+            temperature=self.temperature,
+            max_new_tokens=1000
+        )
+        reply = chat.generated_text
+        if self.if_show_reply:
+            print(f"LLAMA2: {reply}")
+        self.messages.append({"role": "assistant", "content": reply})
+        self.last_reply = reply
+
+    def add_background(self, background_message):
+        self.messages.append({"role": "assistant", "content": background_message})
+
diff --git a/solidgpt/src/manager/promptresource.py b/solidgpt/src/manager/promptresource.py
@@ -1,3 +1,4 @@
+
 PRODUCT_MANAGER_PRD_OUTPUT_TEMPLATE = f'''Base on the information help me generate Markdown PRD follow this format. 
 Here is the output template and explain each sections mean. Always output with this template in Markdown format.
 
@@ -1183,3 +1184,35 @@ def build_custom_skill_gpt_prompt(role_assumption: str, instruction: str, princi
     Here are principles you need to always follow when give the response: {principles}\n\n 
     Here are the prompt and completion examples: {few_shots}
     If no suitable content then response base on your professional knowledge. '''
+
+def llama_v2_prompt(messages):
+    """
+    Convert the messages in list of dictionary format to Llama2 compliant format.
+    """
+    B_INST, E_INST = "[INST]", "[/INST]"
+    B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
+    BOS, EOS = "<s>", "</s>"
+    DEFAULT_SYSTEM_PROMPT = f"""You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Please ensure that your responses are socially unbiased and positive in nature. 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."""
+
+    if messages[0]["role"] != "system":
+        messages = [
+            {
+                "role": "system",
+                "content": DEFAULT_SYSTEM_PROMPT,
+            }
+        ] + messages
+    messages = [
+        {
+            "role": messages[1]["role"],
+            "content": B_SYS + messages[0]["content"] + E_SYS + messages[1]["content"],
+        }
+    ] + messages[2:]
+
+    messages_list = [
+        f"{BOS}{B_INST} {(prompt['content']).strip()} {E_INST} {(answer['content']).strip()} {EOS}"
+        for prompt, answer in zip(messages[::2], messages[1::2])
+    ]
+    messages_list.append(
+        f"{BOS}{B_INST} {(messages[-1]['content']).strip()} {E_INST}")
+
+    return "".join(messages_list)
diff --git a/solidgpt/src/workskill/skills/llama_analysis.py b/solidgpt/src/workskill/skills/llama_analysis.py
@@ -0,0 +1,66 @@
+from solidgpt.src.manager.llamanager import LLAManager
+from solidgpt.src.manager.promptresource import PRODUCT_MANAGER_5H2W_OUTPUT_TEMPLATE, PRODUCT_MANAGER_ANALYSIS_ROLE_ASSUMPTION, PRODUCT_MANAGER_BRAINSTORM_OUTPUT_TEMPLATE, PRODUCT_MANAGER_BRAINSTORM_ROLE_ASSUMPTION, PRODUCT_MANAGER_PRD_OUTPUT_TEMPLATE, PRODUCT_MANAGER_PRD_ROLE_ASSUMPTION, build_gpt_prompt
+from solidgpt.src.util.util import *
+from solidgpt.src.workskill.workskill import *
+
+class ProductAnalysisLlama(WorkSkill):
+    def __init__(self):
+        super().__init__()
+        self.llm_manager = LLAManager._instance
+        self.name = SKILL_NAME_ANALYSIS_PRODUCT
+        self.repo_summary = SkillInput(
+            "Product Analysis Repo Summary",
+            SkillIOParamCategory.PlainText,
+        )
+        self.additional_info = SkillInput(
+            "Product Analysis Additional Info",
+            SkillIOParamCategory.PlainText,
+        )
+        self.requirements = SkillInput(
+            "Product Analysis Requirements",
+            SkillIOParamCategory.PlainText,
+        )
+        self.add_input(self.repo_summary)
+        self.add_input(self.additional_info)
+        self.add_input(self.requirements)
+        self.output_md = SkillOutput(
+            "Requirments Analysis Markdown",
+            SkillIOParamCategory.PlainText,
+        )
+        self.add_output(self.output_md)
+        self.additional_info_content = None
+        self.repo_summary_content = None
+        self.requirements_content = None
+
+    def _read_input(self):
+        if self.additional_info is None:
+            self.additional_info_content = self.additional_info.content
+        if self.repo_summary_content is None:
+            self.repo_summary_content = self.__get_input_content(self.repo_summary)
+        if self.requirements_content is None:
+            self.requirements_content = self.requirements.content
+
+    def __get_input_content(self, skill_input : SkillInput):
+        return load_from_text(self.get_input_path(skill_input), extension=".txt")
+
+    def execution_impl(self):
+        print("Generate product analysis here...")
+        product_analysis = self._run_product_analysis_model()
+        save_to_md2(self.output_md.param_path, product_analysis)
+        self._save_to_result_cache(self.output_md, product_analysis)
+        return
+
+    def _run_product_analysis_model(self):
+        logging.info("Running product analysis model...")
+        prompt = build_gpt_prompt(PRODUCT_MANAGER_ANALYSIS_ROLE_ASSUMPTION, PRODUCT_MANAGER_5H2W_OUTPUT_TEMPLATE)
+        model = self.llm_manager.create_model(
+            prompt=prompt,
+            llama_model_label="product_brainstorm",
+            temperature=0.01,
+        )
+        analysis = model.chat_with_model(self.__get_model_input())
+        logging.info("Product analysis report: %s", analysis)
+        return analysis
+
+    def __get_model_input(self):
+        return f'''Requirements: {self.requirements_content} \n Product Instruction: {self.repo_summary_content} \n Product additional background information: {self.additional_info_content}'''
diff --git a/solidgpt/src/workskill/skills/llama_write_prd.py b/solidgpt/src/workskill/skills/llama_write_prd.py
@@ -0,0 +1,59 @@
+from solidgpt.src.manager.llamanager import LLAManager
+from solidgpt.src.manager.promptresource import PRODUCT_MANAGER_BRAINSTORM_OUTPUT_TEMPLATE, PRODUCT_MANAGER_BRAINSTORM_ROLE_ASSUMPTION, PRODUCT_MANAGER_PRD_OUTPUT_TEMPLATE, PRODUCT_MANAGER_PRD_ROLE_ASSUMPTION, build_gpt_prompt
+from solidgpt.src.util.util import *
+from solidgpt.src.workskill.workskill import *
+
+class WritePRDLlama(WorkSkill):
+    def __init__(self):
+        super().__init__()
+        self.llm_manager = LLAManager._instance
+        self.name = SKILL_NAME_WRITE_PRODUCT_REQUIREMENTS_DOCUMENTATION
+        self.input_product_key_info = SkillInput(
+            "Design Doc",
+            SkillIOParamCategory.PlainText,
+        )
+        self.add_input(self.input_product_key_info)
+        self.output_md = SkillOutput(
+            "Write prd Model PRD Result",
+            SkillIOParamCategory.ProductRequirementsDocument,
+        )
+        self.add_output(self.output_md)
+        self.input_content = None
+
+    def _read_input(self):
+        input_path = self.get_input_path(self.input_product_key_info)
+
+        # if input is not a path, infer it as a string content
+        try:
+            self.input_content = load_from_text(input_path, extension=".md")
+        except Exception as e:
+            self.input_content = self.input_product_key_info.content
+
+    def execution_impl(self):
+        print("Printing PRD result here...")
+        brain_storm_product_info = self._run_product_brainstorm_model()
+        prd = self.__run_write_prd_model(brain_storm_product_info)
+        self._save_to_result_cache(self.output_md, prd)
+        save_to_md2(self.output_md.param_path, prd)
+        return
+
+    def __run_write_prd_model(self, brain_storm_product_info):
+        logging.info("Running write prd model...")
+        prompt = build_gpt_prompt(PRODUCT_MANAGER_PRD_ROLE_ASSUMPTION, PRODUCT_MANAGER_PRD_OUTPUT_TEMPLATE)
+        return self.llm_manager.create_and_chat_with_model(
+            prompt=prompt,
+            llama_model_label="write_prd",
+            input_message=brain_storm_product_info
+        )
+
+    def _run_product_brainstorm_model(self):
+        logging.info("Running product brainstorm model...")
+        prompt = build_gpt_prompt(PRODUCT_MANAGER_BRAINSTORM_ROLE_ASSUMPTION, PRODUCT_MANAGER_BRAINSTORM_OUTPUT_TEMPLATE)
+        model = self.llm_manager.create_model(
+            prompt=prompt,
+            llama_model_label="product_brainstorm",
+            temperature=0.01,
+        )
+        brainstorm = model.chat_with_model(self.input_content)
+        logging.info("Brainstorm result: %s", brainstorm)
+        return brainstorm
diff --git a/solidgpt/test/workskill/skills/test_llama_analysis.py b/solidgpt/test/workskill/skills/test_llama_analysis.py
@@ -0,0 +1,45 @@
+from solidgpt.src.manager.initializer import Initializer
+from solidgpt.src.workgraph.workgraph import *
+from solidgpt.src.workskill.skills.llama_analysis import ProductAnalysisLlama
+
+
+TEST_SKILL_WORKSPACE = os.path.join(TEST_DIR, "workskill", "skills", "workspace")
+
+def run_test():
+    Initializer()
+    app = WorkGraph()
+    skill: WorkSkill = ProductAnalysisLlama()
+    requirement_input_path = os.path.join(TEST_SKILL_WORKSPACE, "in", "ProductRequirements.txt")
+    schema_input_path = os.path.join(TEST_SKILL_WORKSPACE, "in", "RepoSchema.txt")
+    introduction_input_path = os.path.join(TEST_SKILL_WORKSPACE, "in", "ProductIntroduction.txt")
+    skill.init_config(
+        [
+            {
+                "param_path": introduction_input_path,
+                "loading_method": "SkillInputLoadingMethod.LOAD_FROM_STRING",
+                "load_from_output_id": -1
+            },
+            {
+                "param_path": schema_input_path,
+                "loading_method": "SkillInputLoadingMethod.LOAD_FROM_STRING",
+                "load_from_output_id": -1
+            },
+            {
+                "param_path": requirement_input_path,
+                "loading_method": "SkillInputLoadingMethod.LOAD_FROM_STRING",
+                "load_from_output_id": -1
+            },
+        ],
+        [
+            {
+                "id": 1
+            }
+        ])
+    node = WorkNode("1", skill)
+    app.add_node(node)
+    app.init_node_dependencies()
+    # app.save_data(os.path.join(TEST_SKILL_WORKSPACE, "config", "config_data.json"))
+    app.execute()
+
+if __name__ == "__main__":
+    run_test()