allow create Ray cluster with reduced dependency (#125)

* add verl as dependency

* lazy import

Author: dongwang218

matrix/app_server/deploy_utils.py

@@ -416,9 +416,6 @@ def get_yaml_for_deployment(
 ):
     """deploy helper function.
     Return modified applications and yaml for deployment"""
-    from vllm.engine.arg_utils import AsyncEngineArgs
-
-    from matrix.app_server.llm.ray_serve_vllm import BaseDeployment
 
     temp_dir = cluster_info.temp_dir
     cache_dir = setup_native_cache_dirs()
@@ -463,6 +460,10 @@ def get_yaml_for_deployment(
                 app["max_replica"] = app["min_replica"]
 
             if app_type in ["llm", "sglang_llm", "fastgen"]:
+                from vllm.engine.arg_utils import AsyncEngineArgs
+
+                from matrix.app_server.llm.ray_serve_vllm import BaseDeployment
+
                 unknown = {
                     k: v
                     for k, v in app.items()

pyproject.toml

@@ -162,6 +162,17 @@ classifiers=[
       "iopath",
       "jsonlines",
   ]
+  ray = [
+      "submitit>=1.5.2",
+      "ray[default,serve]>=2.48.0",
+      "boto3",
+      "google-genai>=1.13.0",
+      "datasketch",
+      "s3fs",
+      "datasets",
+      "iopath",
+      "jsonlines",
+  ]
   # For video processing
   vision = [
       "submitit>=1.5.2",