Merge branch 'kyegomez:main' into main

pyproject.toml

@@ -31,10 +31,11 @@ einops = "*"
 pydantic = ">2,<3"
 stripe = "*"
 transformers = "*"
-sse-starlette = "*" 
+sse-starlette = "2.0.0" 
 uvicorn = "*"
 shortuuid = "*"
 xformers = "*"
+exxa = "*"
 
 [tool.poetry.group.lint.dependencies]
 ruff = ">=0.1.6,<0.4.0"

requirements.txt

@@ -1,11 +1,11 @@
 skypilot
-fastapi
+fastapi==0.110.0
 supabase
 pytest 
 torch 
 einops
 tiktoken
-sse-starlette
+sse-starlette==2.0.0
 uvicorn
 loguru
 pydantic
@@ -18,8 +18,8 @@ peft
 accelerate
 transformers==4.37.2
 huggingface_hub[hf_transfer]
-huggingface-hub
 optimum
 auto-gptq
 whisperx
-shortuuid
+shortuuid
+exxa

servers/cogvlm/cogvlm.py

@@ -36,9 +36,8 @@
     TextContent,
     UsageInfo,
 )
-from swarms_cloud.utils.count_cores_for_workers import calculate_workers
-from starlette.middleware import Middleware
-from starlette.middleware.gzip import GZipMiddleware
+
+from exa import calculate_workers
 import torch.distributed as dist
 
 
@@ -52,12 +51,6 @@
 QUANT_ENABLED = os.environ.get("QUANT_ENABLED", True)
 
 
-# Middleware
-middleware = [
-    Middleware(GZipMiddleware, minimum_size=1000),
-]
-
-
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """
@@ -71,7 +64,7 @@ async def lifespan(app: FastAPI):
 
 
 # Create a FastAPI app
-app = FastAPI(lifespan=lifespan, debug=True, middleware=middleware)
+app = FastAPI(lifespan=lifespan, debug=True)
 
 
 # Load the middleware to handle CORS