refactor: remove torch.

Author: liqiang-fit2cloud

apps/models_provider/impl/local_model_provider/model/reranker.py

@@ -9,11 +9,10 @@
 from typing import Sequence, Optional, Dict, Any, ClassVar
 
 import requests
-import torch
 from langchain_core.callbacks import Callbacks
 from langchain_core.documents import BaseDocumentCompressor, Document
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
-
+import numpy as np
 from models_provider.base_model_provider import MaxKBBaseModel
 from maxkb.const import CONFIG
 
@@ -90,13 +89,16 @@ def compress_documents(self, documents: Sequence[Document], query: str, callback
             Sequence[Document]:
         if documents is None or len(documents) == 0:
             return []
-        with torch.no_grad():
-            inputs = self.tokenizer([[query, document.page_content] for document in documents], padding=True,
-                                    truncation=True, return_tensors='pt', max_length=512)
-            scores = [torch.sigmoid(s).float().item() for s in
-                      self.client(**inputs, return_dict=True).logits.view(-1, ).float()]
-            result = [Document(page_content=documents[index].page_content, metadata={'relevance_score': scores[index]})
-                      for index
-                      in range(len(documents))]
-            result.sort(key=lambda row: row.metadata.get('relevance_score'), reverse=True)
-            return result
+        inputs = self.tokenizer([[query, document.page_content] for document in documents], padding=True,
+                                truncation=True, return_tensors='pt', max_length=512)
+        scores = [self.sigmoid(s).float().item() for s in
+                  self.client(**inputs, return_dict=True).logits.view(-1, ).float()]
+        result = [Document(page_content=documents[index].page_content, metadata={'relevance_score': scores[index]})
+                  for index
+                  in range(len(documents))]
+        result.sort(key=lambda row: row.metadata.get('relevance_score'), reverse=True)
+        return result
+
+    def sigmoid(x):
+        x = np.asarray(x, dtype=np.float64)
+        return 1 / (1 + np.exp(-x))

installer/Dockerfile

@@ -24,7 +24,6 @@ RUN rm -rf /opt/maxkb-app/ui && \
     pip install poetry==2.0.0 --break-system-packages && \
     poetry config virtualenvs.create false && \
     . /opt/py3/bin/activate && \
-    if [ "$(uname -m)" = "x86_64" ]; then sed -i 's/^torch.*/torch = {version = "2.7.1+cpu", source = "pytorch"}/g' pyproject.toml; fi && \
     poetry install && \
     find /opt/maxkb-app  -depth \( -name ".git*" -o -name ".docker*" -o -name ".idea*" -o -name ".editorconfig*" -o -name ".prettierrc*" -o -name "README.md" -o -name "poetry.lock" -o -name "pyproject.toml"  \) -exec rm -rf {} + && \
     export MAXKB_CONFIG_TYPE=ENV && python3 /opt/maxkb-app/apps/manage.py compilemessages && \

pyproject.toml

@@ -37,7 +37,6 @@ langchain-mcp-adapters = "0.1.9"
 langchain-huggingface = "0.3.0"
 langchain-ollama = "0.3.4"
 langgraph = "0.5.3"
-torch = "2.7.1"
 sentence-transformers = "5.0.0"
 
 # 云服务SDK
@@ -80,9 +79,4 @@ pylint = "3.3.7"
 
 [build-system]
 requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"
-
-[[tool.poetry.source]]
-name = "pytorch"
-url = "https://download.pytorch.org/whl/cpu"
-priority = "explicit"
+build-backend = "poetry.core.masonry.api"