Spaces:
Sleeping
Sleeping
fix: add @spaces.GPU decorator to load_chain_components for ZeroGPU
Browse filesThe HF Spaces runtime emits 'No @spaces.GPU function detected during
startup' when no function is decorated with @spaces.GPU, and may tear
down the SSR/Node proxy as a result.
Fix:
- Decorate load_chain_components (the function that instantiates the
torch-based HuggingFaceEmbeddings model) with @spaces.GPU
- Move import spaces to the very top of rag_chain.py, before the
langchain_huggingface β sentence-transformers β torch import chain
(preserves the CUDA init order fix from the previous commit)
- Keep import spaces in app.py as belt-and-suspenders safety
- app.py +4 -8
- rag_chain.py +19 -0
app.py
CHANGED
|
@@ -1,12 +1,8 @@
|
|
| 1 |
-
# Import
|
| 2 |
-
#
|
| 3 |
-
#
|
| 4 |
-
# has already initialized CUDA, that lazy import raises:
|
| 5 |
-
# RuntimeError: CUDA has been initialized before importing the `spaces` package.
|
| 6 |
-
# Importing it here (top of app.py, before rag_chain pulls in torch) pre-empts
|
| 7 |
-
# the error. The try/except keeps local dev working (spaces isn't installed outside HF Spaces).
|
| 8 |
try:
|
| 9 |
-
import spaces # noqa: F401
|
| 10 |
except ImportError:
|
| 11 |
pass
|
| 12 |
|
|
|
|
| 1 |
+
# Import `spaces` first β must happen before any torch/CUDA initialization
|
| 2 |
+
# (torch is pulled in transitively via rag_chain β langchain_huggingface).
|
| 3 |
+
# The try/except keeps local dev working (spaces is HF-Spaces-only).
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
try:
|
| 5 |
+
import spaces # noqa: F401 β imported for GPU-context registration, not called directly
|
| 6 |
except ImportError:
|
| 7 |
pass
|
| 8 |
|
rag_chain.py
CHANGED
|
@@ -1,4 +1,22 @@
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
from dotenv import load_dotenv
|
| 3 |
from langchain_community.vectorstores import FAISS
|
| 4 |
from langchain_huggingface import HuggingFaceEmbeddings
|
|
@@ -30,6 +48,7 @@ Question: {question}
|
|
| 30 |
Answer (with inline [n] citations):"""
|
| 31 |
|
| 32 |
|
|
|
|
| 33 |
def load_chain_components():
|
| 34 |
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
|
| 35 |
vectorstore = FAISS.load_local(INDEX_DIR, embeddings, allow_dangerous_deserialization=True)
|
|
|
|
| 1 |
import os
|
| 2 |
+
|
| 3 |
+
# HF Spaces ZeroGPU: the `spaces` package scans for @spaces.GPU-decorated
|
| 4 |
+
# functions so the runtime knows which code path needs GPU access.
|
| 5 |
+
# When no GPU function is registered, Spaces emits:
|
| 6 |
+
# "No @spaces.GPU function detected during startup"
|
| 7 |
+
# and may tear down the SSR proxy.
|
| 8 |
+
#
|
| 9 |
+
# CRITICAL: this import MUST come before any torch/CUDA-initializing import
|
| 10 |
+
# (e.g. langchain_huggingface β sentence-transformers β torch). If torch
|
| 11 |
+
# is imported first and initializes CUDA, `import spaces` raises:
|
| 12 |
+
# RuntimeError: CUDA has been initialized before importing the `spaces` package.
|
| 13 |
+
try:
|
| 14 |
+
import spaces
|
| 15 |
+
GPU_DECORATOR = spaces.GPU
|
| 16 |
+
except ImportError:
|
| 17 |
+
def GPU_DECORATOR(func):
|
| 18 |
+
return func
|
| 19 |
+
|
| 20 |
from dotenv import load_dotenv
|
| 21 |
from langchain_community.vectorstores import FAISS
|
| 22 |
from langchain_huggingface import HuggingFaceEmbeddings
|
|
|
|
| 48 |
Answer (with inline [n] citations):"""
|
| 49 |
|
| 50 |
|
| 51 |
+
@GPU_DECORATOR
|
| 52 |
def load_chain_components():
|
| 53 |
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
|
| 54 |
vectorstore = FAISS.load_local(INDEX_DIR, embeddings, allow_dangerous_deserialization=True)
|