Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
Commit
•
5becf31
1
Parent(s):
f82711d
Update app.py
Browse files
app.py
CHANGED
@@ -23,20 +23,44 @@ HF_TOKEN = os.getenv("HF_TOKEN")
|
|
23 |
RANKER_URL = os.getenv("RANKER_URL")
|
24 |
RETRIEVER_URL = os.getenv("RETRIEVER_URL")
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
-
|
34 |
-
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
-
# if RANKER_IE.status != "running":
|
38 |
-
# RANKER_IE.resume()
|
39 |
-
# RANKER_IE.wait()
|
40 |
|
41 |
|
42 |
def post(url, payload):
|
@@ -189,6 +213,14 @@ pipe.add_node(component=ranker, name="Ranker", inputs=["Retriever"])
|
|
189 |
|
190 |
|
191 |
def run(query: str) -> dict:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
pipe_output = pipe.run(query=query)
|
193 |
|
194 |
output = f"""<h2>Top {TOP_K} Documents</h2>"""
|
|
|
23 |
RANKER_URL = os.getenv("RANKER_URL")
|
24 |
RETRIEVER_URL = os.getenv("RETRIEVER_URL")
|
25 |
|
26 |
+
RETRIEVER_IE = get_inference_endpoint(
|
27 |
+
"fastrag-retriever", namespace="optimum-intel", token=HF_TOKEN
|
28 |
+
)
|
29 |
+
RANKER_IE = get_inference_endpoint(
|
30 |
+
"fastrag-ranker", namespace="optimum-intel", token=HF_TOKEN
|
31 |
+
)
|
32 |
+
|
33 |
+
def check_inference_endpoints():
|
34 |
+
RETRIEVER_IE.update()
|
35 |
+
RANKER_IE.update()
|
36 |
+
|
37 |
+
messages = []
|
38 |
+
|
39 |
+
if RETRIEVER_IE.status in ["initializing", "pending"]:
|
40 |
+
messages += [
|
41 |
+
f"Retriever Inference Endpoint is {RETRIEVER_IE.status}. Please wait a few seconds and try again."
|
42 |
+
]
|
43 |
+
elif RETRIEVER_IE.status in ["paused", "scaledToZero"]:
|
44 |
+
messages += [
|
45 |
+
f"Retriever Inference Endpoint is {RETRIEVER_IE.status}. Resuming it. Please wait a few seconds and try again."
|
46 |
+
]
|
47 |
+
RETRIEVER_IE.resume()
|
48 |
|
49 |
+
if RANKER_IE.status in ["initializing", "pending"]:
|
50 |
+
messages += [
|
51 |
+
f"Ranker Inference Endpoint is {RANKER_IE.status}. Please wait a few seconds and try again."
|
52 |
+
]
|
53 |
+
elif RANKER_IE.status in ["paused", "scaledToZero"]:
|
54 |
+
messages += [
|
55 |
+
f"Ranker Inference Endpoint is {RANKER_IE.status}. Resuming it. Please wait a few seconds and try again."
|
56 |
+
]
|
57 |
+
RANKER_IE.resume()
|
58 |
+
|
59 |
+
if len(messages) > 0:
|
60 |
+
return "<br>".join(messages)
|
61 |
+
else:
|
62 |
+
return None
|
63 |
|
|
|
|
|
|
|
64 |
|
65 |
|
66 |
def post(url, payload):
|
|
|
213 |
|
214 |
|
215 |
def run(query: str) -> dict:
|
216 |
+
message = check_inference_endpoints()
|
217 |
+
|
218 |
+
if message is not None:
|
219 |
+
return f"""
|
220 |
+
<h2>Service Unavailable</h2>
|
221 |
+
<p>{message}</p>
|
222 |
+
"""
|
223 |
+
|
224 |
pipe_output = pipe.run(query=query)
|
225 |
|
226 |
output = f"""<h2>Top {TOP_K} Documents</h2>"""
|