Spaces:
Sleeping
Sleeping
refactor: 重构项目结构并优化模型加载方式
Browse files将应用逻辑拆分为三个独立的模块:
- `app.py`: Gradio 界面及应用入口。
- `graph.py`: LangGraph 状态及工作流定义。
- `comp.py`: 模型加载及推理逻辑。
此次变更还在 `comp.py` 中更新了模型加载方式,使用 `device_map="auto"` 和 `torch_dtype="auto"` 以实现硬件自动优化,提高可移植性。
app.py
CHANGED
|
@@ -1,67 +1,10 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import spaces
|
| 3 |
-
import
|
| 4 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 5 |
|
| 6 |
-
|
| 7 |
-
from
|
| 8 |
-
from typing_extensions import TypedDict
|
| 9 |
|
| 10 |
-
from langchain_core.messages import AIMessage, AnyMessage, SystemMessage, HumanMessage, ToolMessage
|
| 11 |
-
from langgraph.graph import StateGraph, END
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
# 定义图的状态
|
| 15 |
-
class GraphState(TypedDict):
|
| 16 |
-
messages: Annotated[list[AnyMessage], operator.add]
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
# 只加载一次模型和分词器
|
| 20 |
-
MODEL_NAME = "inclusionAI/Ring-mini-2.0"
|
| 21 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 22 |
-
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
| 23 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 24 |
-
MODEL_NAME,
|
| 25 |
-
torch_dtype=torch.float16 if device == "cuda" else torch.float32,
|
| 26 |
-
trust_remote_code=True
|
| 27 |
-
).to(device)
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
# 定义图的节点
|
| 31 |
-
def call_model(state: GraphState):
|
| 32 |
-
"""模型调用节点"""
|
| 33 |
-
messages = state["messages"]
|
| 34 |
-
|
| 35 |
-
# 拼接 prompt
|
| 36 |
-
prompt = ""
|
| 37 |
-
for msg in messages:
|
| 38 |
-
if msg.type == "system":
|
| 39 |
-
prompt += f"{msg.content}\n"
|
| 40 |
-
elif msg.type == "human":
|
| 41 |
-
prompt += f"User: {msg.content}\n"
|
| 42 |
-
elif msg.type == "ai":
|
| 43 |
-
prompt += f"Assistant: {msg.content}\n"
|
| 44 |
-
prompt += "Assistant:"
|
| 45 |
-
|
| 46 |
-
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
| 47 |
-
output_ids = model.generate(
|
| 48 |
-
input_ids,
|
| 49 |
-
max_new_tokens=512, # 暂时硬编码
|
| 50 |
-
do_sample=True,
|
| 51 |
-
pad_token_id=tokenizer.eos_token_id,
|
| 52 |
-
)
|
| 53 |
-
output = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
|
| 54 |
-
|
| 55 |
-
return {"messages": [AIMessage(content=output)]}
|
| 56 |
-
|
| 57 |
-
# 构建图
|
| 58 |
-
workflow = StateGraph(GraphState)
|
| 59 |
-
workflow.add_node("llm", call_model)
|
| 60 |
-
workflow.set_entry_point("llm")
|
| 61 |
-
workflow.add_edge("llm", END)
|
| 62 |
-
|
| 63 |
-
# 编译图
|
| 64 |
-
app = workflow.compile()
|
| 65 |
@spaces.GPU
|
| 66 |
def respond(message, history, system_message, hf_token: gr.OAuthToken = None):
|
| 67 |
"""Gradio 接口的响应函数,调用 LangGraph 应用"""
|
|
@@ -106,4 +49,4 @@ with gr.Blocks() as demo:
|
|
| 106 |
|
| 107 |
|
| 108 |
if __name__ == "__main__":
|
| 109 |
-
demo.launch()
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import spaces
|
| 3 |
+
from langchain_core.messages import AIMessage, SystemMessage, HumanMessage
|
|
|
|
| 4 |
|
| 5 |
+
# 导入已编译的 LangGraph 应用
|
| 6 |
+
from graph import app
|
|
|
|
| 7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
@spaces.GPU
|
| 9 |
def respond(message, history, system_message, hf_token: gr.OAuthToken = None):
|
| 10 |
"""Gradio 接口的响应函数,调用 LangGraph 应用"""
|
|
|
|
| 49 |
|
| 50 |
|
| 51 |
if __name__ == "__main__":
|
| 52 |
+
demo.launch()
|
comp.py
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 3 |
+
from langchain_core.messages import AIMessage
|
| 4 |
+
from typing import TypedDict, Annotated, List
|
| 5 |
+
import operator
|
| 6 |
+
|
| 7 |
+
# 定义此组件操作的图状态的子集
|
| 8 |
+
class GraphState(TypedDict):
|
| 9 |
+
messages: Annotated[List[AIMessage], operator.add]
|
| 10 |
+
|
| 11 |
+
# --- 模型加载 ---
|
| 12 |
+
# 使用 "auto" 模式加载模型和分词器,Hugging Face Accelerate 会自动处理设备和精度
|
| 13 |
+
MODEL_NAME = "inclusionAI/Ring-mini-2.0"
|
| 14 |
+
|
| 15 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
| 16 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 17 |
+
MODEL_NAME,
|
| 18 |
+
torch_dtype="auto",
|
| 19 |
+
device_map="auto",
|
| 20 |
+
trust_remote_code=True
|
| 21 |
+
)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def completion_node(state: GraphState) -> dict:
|
| 25 |
+
"""
|
| 26 |
+
一个调用语言模型以获取响应的节点。
|
| 27 |
+
|
| 28 |
+
Args:
|
| 29 |
+
state (GraphState): 图的当前状态,包含消息历史。
|
| 30 |
+
|
| 31 |
+
Returns:
|
| 32 |
+
dict: 一个包含新 AI 消息的字典,该消息将被添加到状态中。
|
| 33 |
+
"""
|
| 34 |
+
messages = state["messages"]
|
| 35 |
+
|
| 36 |
+
# --- 提示工程 ---
|
| 37 |
+
# 从消息历史中组装提示。
|
| 38 |
+
prompt = ""
|
| 39 |
+
for msg in messages:
|
| 40 |
+
if msg.type == "system":
|
| 41 |
+
prompt += f"{msg.content}\n"
|
| 42 |
+
elif msg.type == "human":
|
| 43 |
+
prompt += f"User: {msg.content}\n"
|
| 44 |
+
elif msg.type == "ai":
|
| 45 |
+
prompt += f"Assistant: {msg.content}\n"
|
| 46 |
+
prompt += "Assistant:"
|
| 47 |
+
|
| 48 |
+
# --- 模型调用 ---
|
| 49 |
+
# 使用 device_map="auto" 时,我们无需手动将张量移动到特定设备
|
| 50 |
+
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
|
| 51 |
+
output_ids = model.generate(
|
| 52 |
+
input_ids,
|
| 53 |
+
max_new_tokens=512, # 暂时硬编码
|
| 54 |
+
do_sample=True,
|
| 55 |
+
pad_token_id=tokenizer.eos_token_id,
|
| 56 |
+
)
|
| 57 |
+
output = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
|
| 58 |
+
|
| 59 |
+
# 以 AIMessage 的形式返回响应,以添加到图的状态中。
|
| 60 |
+
return {"messages": [AIMessage(content=output)]}
|
graph.py
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import operator
|
| 2 |
+
from typing import Annotated, List
|
| 3 |
+
from typing_extensions import TypedDict
|
| 4 |
+
|
| 5 |
+
from langchain_core.messages import AnyMessage
|
| 6 |
+
from langgraph.graph import StateGraph, END
|
| 7 |
+
|
| 8 |
+
# 从我们的组件文件中导入模型补全节点
|
| 9 |
+
from comp import completion_node
|
| 10 |
+
|
| 11 |
+
# --- 图状态定义 ---
|
| 12 |
+
# 状态是我们图的内存或上下文。它是一个字典,
|
| 13 |
+
# 保存了对话过程中交换的所有消息。
|
| 14 |
+
class GraphState(TypedDict):
|
| 15 |
+
"""
|
| 16 |
+
表示我们图的状态。
|
| 17 |
+
|
| 18 |
+
Attributes:
|
| 19 |
+
messages: 一个随时间自动累积的消息列表。
|
| 20 |
+
`operator.add` 注解告诉 LangGraph 将新消息附加到此列表,
|
| 21 |
+
而不是覆盖它。这就是图如何维护对话历史(上下文)的方式。
|
| 22 |
+
"""
|
| 23 |
+
messages: Annotated[List[AnyMessage], operator.add]
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
# --- 图工作流构建 ---
|
| 27 |
+
# 使用我们定义的状态创建一个新的状态图
|
| 28 |
+
workflow = StateGraph(GraphState)
|
| 29 |
+
|
| 30 |
+
# 将补全节点添加到图中。我们将其命名为 “llm”。
|
| 31 |
+
# 这个节点负责调用语言模型。
|
| 32 |
+
workflow.add_node("llm", completion_node)
|
| 33 |
+
|
| 34 |
+
# 设置图的入口点。第一个被调用的节点是 “llm”。
|
| 35 |
+
workflow.set_entry_point("llm")
|
| 36 |
+
|
| 37 |
+
# 从 “llm” 节点到 END 添加一条简单的边。
|
| 38 |
+
# 这意味着在调用 LLM 后,图的执行就完成了。
|
| 39 |
+
workflow.add_edge("llm", END)
|
| 40 |
+
|
| 41 |
+
# 将工作流编译成一个可运行的应用。
|
| 42 |
+
app = workflow.compile()
|