优化LLM和Embedding模型运行设备配置，可设为auto自动检测；修复：重建知识库时FAISS未保存 (#1330)

* 避免configs对torch的依赖；
* webui自动从configs获取api地址(close #1319)
* bug fix: 重建知识库时FAISS未保存
* 优化LLM和Embedding模型运行设备配置，可设为auto自动检测

.gitignore

@@ -3,7 +3,7 @@
 logs
 .idea/
 __pycache__/
-knowledge_base/
+/knowledge_base/
 configs/*.py
 .vscode/
 .pytest_cache/

configs/model_config.py.example

@@ -7,19 +7,6 @@ logger.setLevel(logging.INFO)
 logging.basicConfig(format=LOG_FORMAT)
 
 
-# 分布式部署时，不运行LLM的机器上可以不装torch
-def default_device():
-    try:
-        import torch
-        if torch.cuda.is_available():
-            return "cuda"
-        if torch.backends.mps.is_available():
-            return "mps"
-    except:
-        pass
-    return "cpu"
-
-
 # 在以下字典中修改属性值，以指定本地embedding模型存储位置
 # 如将 "text2vec": "GanymedeNil/text2vec-large-chinese" 修改为 "text2vec": "User/Downloads/text2vec-large-chinese"
 # 此处请写绝对路径
@@ -44,8 +31,8 @@ embedding_model_dict = {
 # 选用的 Embedding 名称
 EMBEDDING_MODEL = "m3e-base"
 
-# Embedding 模型运行设备
+# Embedding 模型运行设备。设为"auto"会自动检测，也可手动设定为"cuda","mps","cpu"其中之一。
-EMBEDDING_DEVICE = default_device()
+EMBEDDING_DEVICE = "auto"
 
 
 llm_model_dict = {
@@ -94,8 +81,8 @@ LLM_MODEL = "chatglm2-6b"
 # 历史对话轮数
 HISTORY_LEN = 3
 
-# LLM 运行设备
+# LLM 运行设备。可选项同Embedding 运行设备。
-LLM_DEVICE = default_device()
+LLM_DEVICE = "auto"
 
 # 日志存储路径
 LOG_PATH = os.path.join(os.path.dirname(os.path.dirname(__file__)), "logs")

server/knowledge_base/kb_service/base.py

@@ -18,11 +18,12 @@ from server.db.repository.knowledge_file_repository import (
 )
 
 from configs.model_config import (kbs_config, VECTOR_SEARCH_TOP_K, SCORE_THRESHOLD,
-                                  EMBEDDING_DEVICE, EMBEDDING_MODEL)
+                                  EMBEDDING_MODEL)
 from server.knowledge_base.utils import (
     get_kb_path, get_doc_path, load_embeddings, KnowledgeFile,
     list_kbs_from_folder, list_files_from_folder,
 )
+from server.utils import embedding_device
 from typing import List, Union, Dict
 
 
@@ -45,7 +46,7 @@ class KBService(ABC):
         self.doc_path = get_doc_path(self.kb_name)
         self.do_init()
 
-    def _load_embeddings(self, embed_device: str = EMBEDDING_DEVICE) -> Embeddings:
+    def _load_embeddings(self, embed_device: str = embedding_device()) -> Embeddings:
         return load_embeddings(self.embed_model, embed_device)
 
     def create_kb(self):

server/knowledge_base/kb_service/faiss_kb_service.py

@@ -5,7 +5,6 @@ from configs.model_config import (
     KB_ROOT_PATH,
     CACHED_VS_NUM,
     EMBEDDING_MODEL,
-    EMBEDDING_DEVICE,
     SCORE_THRESHOLD
 )
 from server.knowledge_base.kb_service.base import KBService, SupportedVSType
@@ -15,7 +14,7 @@ from langchain.vectorstores import FAISS
 from langchain.embeddings.base import Embeddings
 from typing import List
 from langchain.docstore.document import Document
-from server.utils import torch_gc
+from server.utils import torch_gc, embedding_device
 
 
 _VECTOR_STORE_TICKS = {}
@@ -25,10 +24,10 @@ _VECTOR_STORE_TICKS = {}
 def load_faiss_vector_store(
         knowledge_base_name: str,
         embed_model: str = EMBEDDING_MODEL,
-        embed_device: str = EMBEDDING_DEVICE,
+        embed_device: str = embedding_device(),
         embeddings: Embeddings = None,
         tick: int = 0,  # tick will be changed by upload_doc etc. and make cache refreshed.
-):
+) -> FAISS:
     print(f"loading vector store in '{knowledge_base_name}'.")
     vs_path = get_vs_path(knowledge_base_name)
     if embeddings is None:
@@ -74,13 +73,18 @@ class FaissKBService(KBService):
     def get_kb_path(self):
         return os.path.join(KB_ROOT_PATH, self.kb_name)
 
-    def load_vector_store(self):
+    def load_vector_store(self) -> FAISS:
         return load_faiss_vector_store(
             knowledge_base_name=self.kb_name,
             embed_model=self.embed_model,
             tick=_VECTOR_STORE_TICKS.get(self.kb_name, 0),
         )
 
+    def save_vector_store(self, vector_store: FAISS = None):
+        vector_store = vector_store or self.load_vector_store()
+        vector_store.save_local(self.vs_path)
+        return vector_store
+
     def refresh_vs_cache(self):
         refresh_vs_cache(self.kb_name)
 
@@ -117,11 +121,11 @@ class FaissKBService(KBService):
         if not kwargs.get("not_refresh_vs_cache"):
             vector_store.save_local(self.vs_path)
             self.refresh_vs_cache()
+        return vector_store
 
     def do_delete_doc(self,
                       kb_file: KnowledgeFile,
                       **kwargs):
-        embeddings = self._load_embeddings()
         vector_store = self.load_vector_store()
 
         ids = [k for k, v in vector_store.docstore._dict.items() if v.metadata["source"] == kb_file.filepath]
@@ -133,7 +137,7 @@ class FaissKBService(KBService):
             vector_store.save_local(self.vs_path)
             self.refresh_vs_cache()
 
-        return True
+        return vector_store
 
     def do_clear_vs(self):
         shutil.rmtree(self.vs_path)

server/knowledge_base/kb_service/pg_kb_service.py

@@ -6,16 +6,16 @@ from langchain.vectorstores import PGVector
 from langchain.vectorstores.pgvector import DistanceStrategy
 from sqlalchemy import text
 
-from configs.model_config import EMBEDDING_DEVICE, kbs_config
+from configs.model_config import kbs_config
 from server.knowledge_base.kb_service.base import SupportedVSType, KBService, EmbeddingsFunAdapter, \
     score_threshold_process
 from server.knowledge_base.utils import load_embeddings, KnowledgeFile
-
+from server.utils import embedding_device as get_embedding_device
 
 class PGKBService(KBService):
     pg_vector: PGVector
 
-    def _load_pg_vector(self, embedding_device: str = EMBEDDING_DEVICE, embeddings: Embeddings = None):
+    def _load_pg_vector(self, embedding_device: str = get_embedding_device(), embeddings: Embeddings = None):
         _embeddings = embeddings
         if _embeddings is None:
             _embeddings = load_embeddings(self.embed_model, embedding_device)

server/knowledge_base/migrate.py

@@ -69,6 +69,7 @@ def folder2db(
                 print(result)
 
         if kb.vs_type() == SupportedVSType.FAISS:
+            kb.save_vector_store()
             kb.refresh_vs_cache()
     elif mode == "fill_info_only":
         files = list_files_from_folder(kb_name)
@@ -85,6 +86,7 @@ def folder2db(
             kb.update_doc(kb_file, not_refresh_vs_cache=True)
 
         if kb.vs_type() == SupportedVSType.FAISS:
+            kb.save_vector_store()
             kb.refresh_vs_cache()
     elif mode == "increament":
         db_files = kb.list_files()
@@ -102,6 +104,7 @@ def folder2db(
                 print(result)
 
         if kb.vs_type() == SupportedVSType.FAISS:
+            kb.save_vector_store()
             kb.refresh_vs_cache()
     else:
         print(f"unspported migrate mode: {mode}")
@@ -131,7 +134,10 @@ def prune_db_files(kb_name: str):
         files = list(set(files_in_db) - set(files_in_folder))
         kb_files = file_to_kbfile(kb_name, files)
         for kb_file in kb_files:
-            kb.delete_doc(kb_file)
+            kb.delete_doc(kb_file, not_refresh_vs_cache=True)
+        if kb.vs_type() == SupportedVSType.FAISS:
+            kb.save_vector_store()
+            kb.refresh_vs_cache()
         return kb_files
 
 def prune_folder_files(kb_name: str):

server/llm_api.py

@@ -4,8 +4,8 @@ import sys
 import os
 
 sys.path.append(os.path.dirname(os.path.dirname(__file__)))
-from configs.model_config import llm_model_dict, LLM_MODEL, LLM_DEVICE, LOG_PATH, logger
+from configs.model_config import llm_model_dict, LLM_MODEL, LOG_PATH, logger
-from server.utils import MakeFastAPIOffline, set_httpx_timeout
+from server.utils import MakeFastAPIOffline, set_httpx_timeout, llm_device
 
 
 host_ip = "0.0.0.0"
@@ -34,7 +34,7 @@ def create_model_worker_app(
         worker_address=base_url.format(model_worker_port),
         controller_address=base_url.format(controller_port),
         model_path=llm_model_dict[LLM_MODEL].get("local_model_path"),
-        device=LLM_DEVICE,
+        device=llm_device(),
         gpus=None,
         max_gpu_memory="20GiB",
         load_8bit=False,

server/utils.py

@@ -5,8 +5,8 @@ import torch
 from fastapi import FastAPI
 from pathlib import Path
 import asyncio
-from configs.model_config import LLM_MODEL
+from configs.model_config import LLM_MODEL, LLM_DEVICE, EMBEDDING_DEVICE
-from typing import Any, Optional
+from typing import Literal, Optional
 
 
 class BaseResponse(BaseModel):
@@ -201,6 +201,7 @@ def get_model_worker_config(model_name: str = LLM_MODEL) -> dict:
     config = FSCHAT_MODEL_WORKERS.get("default", {}).copy()
     config.update(llm_model_dict.get(model_name, {}))
     config.update(FSCHAT_MODEL_WORKERS.get(model_name, {}))
+    config["device"] = llm_device(config.get("device"))
     return config
 
 
@@ -256,3 +257,28 @@ def set_httpx_timeout(timeout: float = None):
     httpx._config.DEFAULT_TIMEOUT_CONFIG.connect = timeout
     httpx._config.DEFAULT_TIMEOUT_CONFIG.read = timeout
     httpx._config.DEFAULT_TIMEOUT_CONFIG.write = timeout
+
+
+# 自动检查torch可用的设备。分布式部署时，不运行LLM的机器上可以不装torch
+def detect_device() -> Literal["cuda", "mps", "cpu"]:
+    try:
+        import torch
+        if torch.cuda.is_available():
+            return "cuda"
+        if torch.backends.mps.is_available():
+            return "mps"
+    except:
+        pass
+    return "cpu"
+
+
+def llm_device(device: str = LLM_DEVICE) -> Literal["cuda", "mps", "cpu"]:
+    if device not in ["cuda", "mps", "cpu"]:
+        device = detect_device()
+    return device
+
+
+def embedding_device(device: str = EMBEDDING_DEVICE) -> Literal["cuda", "mps", "cpu"]:
+    if device not in ["cuda", "mps", "cpu"]:
+        device = detect_device()
+    return device

startup.py

@@ -14,12 +14,13 @@ except:
     pass
 
 sys.path.append(os.path.dirname(os.path.dirname(__file__)))
-from configs.model_config import EMBEDDING_DEVICE, EMBEDDING_MODEL, llm_model_dict, LLM_MODEL, LLM_DEVICE, LOG_PATH, \
+from configs.model_config import EMBEDDING_MODEL, llm_model_dict, LLM_MODEL, LOG_PATH, \
     logger
 from configs.server_config import (WEBUI_SERVER, API_SERVER, OPEN_CROSS_DOMAIN, FSCHAT_CONTROLLER, FSCHAT_MODEL_WORKERS,
                                    FSCHAT_OPENAI_API, )
 from server.utils import (fschat_controller_address, fschat_model_worker_address,
-                        fschat_openai_api_address, set_httpx_timeout)
+                        fschat_openai_api_address, set_httpx_timeout,
+                        llm_device, embedding_device, get_model_worker_config)
 from server.utils import MakeFastAPIOffline, FastAPI
 import argparse
 from typing import Tuple, List
@@ -195,7 +196,7 @@ def run_model_worker(
 ):
     import uvicorn
 
-    kwargs = FSCHAT_MODEL_WORKERS[model_name].copy()
+    kwargs = get_model_worker_config(model_name)
     host = kwargs.pop("host")
     port = kwargs.pop("port")
     model_path = llm_model_dict[model_name].get("local_model_path", "")
@@ -331,9 +332,9 @@ def dump_server_info(after_start=False):
     print(f"项目版本：{VERSION}")
     print(f"langchain版本：{langchain.__version__}. fastchat版本：{fastchat.__version__}")
     print("\n")
-    print(f"当前LLM模型：{LLM_MODEL} @ {LLM_DEVICE}")
+    print(f"当前LLM模型：{LLM_MODEL} @ {llm_device()}")
     pprint(llm_model_dict[LLM_MODEL])
-    print(f"当前Embbedings模型： {EMBEDDING_MODEL} @ {EMBEDDING_DEVICE}")
+    print(f"当前Embbedings模型： {EMBEDDING_MODEL} @ {embedding_device()}")
     if after_start:
         print("\n")
         print(f"服务端运行信息：")