Spaces:
Running
Running
dev/vicuna (#11)
Browse files- support vicuna (32b5d08526a27b165a41b4ff3ab03ff27ed04b23)
app.py
CHANGED
@@ -3,6 +3,8 @@ from datetime import datetime, date, timedelta
|
|
3 |
from typing import Iterable
|
4 |
import streamlit as st
|
5 |
import torch
|
|
|
|
|
6 |
from langchain.embeddings import HuggingFaceEmbeddings
|
7 |
from langchain.vectorstores import Qdrant
|
8 |
from qdrant_client import QdrantClient
|
@@ -33,8 +35,44 @@ def llm_model(model="gpt-3.5-turbo", temperature=0.2):
|
|
33 |
return llm
|
34 |
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
EMBEDDINGS = load_embeddings()
|
37 |
LLM = llm_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
|
39 |
|
40 |
def make_filter_obj(options: list[dict[str]]):
|
@@ -78,7 +116,7 @@ def get_similay(query: str, filter: Filter):
|
|
78 |
return docs
|
79 |
|
80 |
|
81 |
-
def get_retrieval_qa(filter: Filter):
|
82 |
db_url, db_api_key, db_collection_name = DB_CONFIG
|
83 |
client = QdrantClient(url=db_url, api_key=db_api_key)
|
84 |
db = Qdrant(
|
@@ -90,7 +128,7 @@ def get_retrieval_qa(filter: Filter):
|
|
90 |
}
|
91 |
)
|
92 |
result = RetrievalQA.from_chain_type(
|
93 |
-
llm=
|
94 |
chain_type="stuff",
|
95 |
retriever=retriever,
|
96 |
return_source_documents=True,
|
@@ -143,6 +181,7 @@ def _get_query_str_filter(
|
|
143 |
|
144 |
|
145 |
def run_qa(
|
|
|
146 |
query: str,
|
147 |
repo_name: str,
|
148 |
query_options: str,
|
@@ -154,7 +193,7 @@ def run_qa(
|
|
154 |
query_str, filter = _get_query_str_filter(
|
155 |
query, repo_name, query_options, start_date, end_date, include_comments
|
156 |
)
|
157 |
-
qa = get_retrieval_qa(filter)
|
158 |
try:
|
159 |
result = qa(query_str)
|
160 |
except InvalidRequestError as e:
|
@@ -271,10 +310,37 @@ with st.form("my_form"):
|
|
271 |
st.divider()
|
272 |
with st.spinner("QA Searching..."):
|
273 |
results = run_qa(
|
274 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
275 |
)
|
276 |
answer, html = results
|
277 |
with st.container():
|
278 |
st.write(answer)
|
279 |
st.markdown(html, unsafe_allow_html=True)
|
280 |
st.divider()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
from typing import Iterable
|
4 |
import streamlit as st
|
5 |
import torch
|
6 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
7 |
+
from langchain.llms import HuggingFacePipeline
|
8 |
from langchain.embeddings import HuggingFaceEmbeddings
|
9 |
from langchain.vectorstores import Qdrant
|
10 |
from qdrant_client import QdrantClient
|
|
|
35 |
return llm
|
36 |
|
37 |
|
38 |
+
@st.cache_resource
|
39 |
+
def load_vicuna_model():
|
40 |
+
if torch.cuda.is_available():
|
41 |
+
model_name = "lmsys/vicuna-13b-v1.5"
|
42 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
|
43 |
+
model = AutoModelForCausalLM.from_pretrained(
|
44 |
+
model_name,
|
45 |
+
load_in_8bit=True,
|
46 |
+
torch_dtype=torch.float16,
|
47 |
+
device_map="auto",
|
48 |
+
)
|
49 |
+
return tokenizer, model
|
50 |
+
else:
|
51 |
+
return None, None
|
52 |
+
|
53 |
+
|
54 |
EMBEDDINGS = load_embeddings()
|
55 |
LLM = llm_model()
|
56 |
+
VICUNA_TOKENIZER, VICUNA_MODEL = load_vicuna_model()
|
57 |
+
|
58 |
+
|
59 |
+
@st.cache_resource
|
60 |
+
def _get_vicuna_llm(temperature=0.2) -> HuggingFacePipeline | None:
|
61 |
+
if VICUNA_MODEL is not None:
|
62 |
+
pipe = pipeline(
|
63 |
+
"text-generation",
|
64 |
+
model=VICUNA_MODEL,
|
65 |
+
tokenizer=VICUNA_TOKENIZER,
|
66 |
+
max_new_tokens=1024,
|
67 |
+
temperature=temperature,
|
68 |
+
)
|
69 |
+
llm = HuggingFacePipeline(pipeline=pipe)
|
70 |
+
else:
|
71 |
+
llm = None
|
72 |
+
return llm
|
73 |
+
|
74 |
+
|
75 |
+
VICUNA_LLM = _get_vicuna_llm()
|
76 |
|
77 |
|
78 |
def make_filter_obj(options: list[dict[str]]):
|
|
|
116 |
return docs
|
117 |
|
118 |
|
119 |
+
def get_retrieval_qa(filter: Filter, llm):
|
120 |
db_url, db_api_key, db_collection_name = DB_CONFIG
|
121 |
client = QdrantClient(url=db_url, api_key=db_api_key)
|
122 |
db = Qdrant(
|
|
|
128 |
}
|
129 |
)
|
130 |
result = RetrievalQA.from_chain_type(
|
131 |
+
llm=llm,
|
132 |
chain_type="stuff",
|
133 |
retriever=retriever,
|
134 |
return_source_documents=True,
|
|
|
181 |
|
182 |
|
183 |
def run_qa(
|
184 |
+
llm,
|
185 |
query: str,
|
186 |
repo_name: str,
|
187 |
query_options: str,
|
|
|
193 |
query_str, filter = _get_query_str_filter(
|
194 |
query, repo_name, query_options, start_date, end_date, include_comments
|
195 |
)
|
196 |
+
qa = get_retrieval_qa(filter, llm)
|
197 |
try:
|
198 |
result = qa(query_str)
|
199 |
except InvalidRequestError as e:
|
|
|
310 |
st.divider()
|
311 |
with st.spinner("QA Searching..."):
|
312 |
results = run_qa(
|
313 |
+
LLM,
|
314 |
+
query,
|
315 |
+
repo_name,
|
316 |
+
query_options,
|
317 |
+
start_date,
|
318 |
+
end_date,
|
319 |
+
include_comments,
|
320 |
)
|
321 |
answer, html = results
|
322 |
with st.container():
|
323 |
st.write(answer)
|
324 |
st.markdown(html, unsafe_allow_html=True)
|
325 |
st.divider()
|
326 |
+
if torch.cuda.is_available():
|
327 |
+
qa_searched_vicuna = submit_col2.form_submit_button("QA Search by Vicuna")
|
328 |
+
if qa_searched_vicuna:
|
329 |
+
st.divider()
|
330 |
+
st.header("QA Search Results by Vicuna-13b-v1.5")
|
331 |
+
st.divider()
|
332 |
+
with st.spinner("QA Searching..."):
|
333 |
+
results = run_qa(
|
334 |
+
VICUNA_LLM,
|
335 |
+
query,
|
336 |
+
repo_name,
|
337 |
+
query_options,
|
338 |
+
start_date,
|
339 |
+
end_date,
|
340 |
+
include_comments,
|
341 |
+
)
|
342 |
+
answer, html = results
|
343 |
+
with st.container():
|
344 |
+
st.write(answer)
|
345 |
+
st.markdown(html, unsafe_allow_html=True)
|
346 |
+
st.divider()
|