Add device support

Signed-off-by: GitHub <[email protected]>

.github/workflows/ci-backend.yaml

@@ -7,9 +7,7 @@ on:
       - 'v*'
   pull_request:
     branches:
-      - 'feature/**'
-      - 'feat/**'
-      - 'fix/**'
+      - 'main'
 
 jobs:
   build:

Makefile

@@ -8,4 +8,9 @@ build: env
 
 .PHONY: up
 up: env build
-	docker-compose up -d
+	docker-compose up -d
+
+
+.PHONY: stop
+stop:
+	docker-compose stop

backend/src/repository/rag/chat.py

@@ -4,22 +4,22 @@
 from sentence_transformers import SentenceTransformer
 from sentence_transformers.util import cos_sim
 from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline
+from utilities.devices.devices import get_device
 
 from src.config.settings.const import CHAT_COMTEXT, DEFAULT_MODEL, MAX_SQL_LENGTH, UPLOAD_FILE_PATH
 from src.repository.rag.base import BaseRAGRepository
 
 
 class RAGChatModelRepository(BaseRAGRepository):
-    # model = SentenceTransformer(DEFAULT_MODEL, "cuda")
-    # embeddings = model.encode([], convert_to_tensor=True).to("cuda")
     model_name = "deepset/roberta-base-squad2"
 
     nlp = pipeline("question-answering", model=model_name, tokenizer=model_name)
 
     async def load_model(self, session_id: int, model_name: str) -> bool:
         # Init model with input model_name
         try:
-            model = SentenceTransformer(model_name, "cuda")
+            # https://github.com/UKPLab/sentence-transformers/blob/85810ead37d02ef706da39e4a1757702d1b9f7c5/sentence_transformers/SentenceTransformer.py#L47
+            model = SentenceTransformer(model_name, device=get_device())
             model.max_seq_length = MAX_SQL_LENGTH
         except Exception as e:
             print(e)
@@ -29,16 +29,12 @@ async def load_model(self, session_id: int, model_name: str) -> bool:
     async def get_response(self, session_id: int, input_msg: str) -> str:
         # TODO use RAG framework to generate the response message @Aisuko
         # query_embedding = self.model.encode(input_msg, convert_to_tensor=True).to("cuda")
-        # print(self.embeddings)
-        # print(query_embedding)
         # we use cosine-similarity and torch.topk to find the highest 5 scores
         # cos_scores = cos_sim(query_embedding, self.embeddings)[0]
         # top_results = torch.topk(cos_scores, k=1)
         # response_msg = self.data[top_results[1].item()]
         QA_input = {"question": input_msg, "context": CHAT_COMTEXT}
         res = self.nlp(QA_input)
-        print(res)
-        # response_msg = "Oh, really? It's amazing !"
         return res["answer"]
 
     async def load_csv_file(self, file_name: str, model_name: str) -> bool:

backend/src/utilities/devices/devices.py

@@ -0,0 +1,9 @@
+import platform
+import torch
+
+def get_device():
+    if platform.system() == 'Darwin':
+        return 'mps'
+    elif torch.cuda.is_available():
+        return 'cuda'
+    return 'cpu'

backend/tests/unit_tests/test_src.py

@@ -14,3 +14,9 @@ def test_application_is_fastapi_instance() -> None:
     assert backend_app.docs_url == "/docs"
     assert backend_app.openapi_url == "/openapi.json"
     assert backend_app.redoc_url == "/redoc"
+
+
+def test_get_device() -> None:
+    from src.utilities.devices.devices import get_device
+
+    assert get_device() in ["mps", "cuda", "cpu"]

docker-compose.yaml

@@ -135,7 +135,7 @@ services:
       - "minio"
 
   frontend:
-    image: ghcr.io/skywardai/rebel:v0.0.5
+    image: ghcr.io/skywardai/rebel:v0.1.1
     container_name: frontend
     restart: always
     expose: