A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 22:00 UTC
[BugFix] Avoid unnecessary Ray import warnings (#6079)
Nick Hill committed
Jul 2, 2024 at 23:09 UTC
d830656a9722bfc719426ce6bdd13b3d9d456304
3 files changed
+28
-9
vllm/config.py
+7
-2
@@ -682,11 +682,13 @@ class ParallelConfig:
682
683
from vllm.executor import ray_utils
684
backend = "mp"
685
- ray_found = ray_utils.ray is not None
685
+ ray_found = ray_utils.ray_is_available()
686
if cuda_device_count_stateless() < self.world_size:
687
if not ray_found:
688
raise ValueError("Unable to load Ray which is "
689
- "required for multi-node inference")
689
+ "required for multi-node inference, "
690
+ "please install Ray with `pip install "
691
+ "ray`.") from ray_utils.ray_import_err
692
backend = "ray"
693
elif ray_found:
694
if self.placement_group:
@@ -718,6 +720,9 @@ class ParallelConfig:
720
raise ValueError(
721
"Unrecognized distributed executor backend. Supported values "
722
"are 'ray' or 'mp'.")
723
+ if self.distributed_executor_backend == "ray":
724
+ from vllm.executor import ray_utils
725
+ ray_utils.assert_ray_available()
726
if not self.disable_custom_all_reduce and self.world_size > 1:
727
if is_hip():
728
self.disable_custom_all_reduce = True
vllm/engine/async_llm_engine.py
+5
@@ -380,6 +380,11 @@ class AsyncLLMEngine:
380
"""Creates an async LLM engine from the engine arguments."""
381
# Create the engine configs.
382
engine_config = engine_args.create_engine_config()
383
+
384
+ if engine_args.engine_use_ray:
385
+ from vllm.executor import ray_utils
386
+ ray_utils.assert_ray_available()
387
+
388
distributed_executor_backend = (
389
engine_config.parallel_config.distributed_executor_backend)
390
vllm/executor/ray_utils.py
+16
-7
@@ -42,14 +42,26 @@ try:
42
output = pickle.dumps(output)
43
return output
44
45
+ ray_import_err = None
46
+
47
except ImportError as e:
46
- logger.warning(
47
- "Failed to import Ray with %r. For multi-node inference, "
48
- "please install Ray with `pip install ray`.", e)
48
ray = None # type: ignore
49
+ ray_import_err = e
50
RayWorkerWrapper = None # type: ignore
51
52
53
+def ray_is_available() -> bool:
54
+ """Returns True if Ray is available."""
55
+ return ray is not None
56
+
57
+
58
+def assert_ray_available():
59
+ """Raise an exception if Ray is not available."""
60
+ if ray is None:
61
+ raise ValueError("Failed to import Ray, please install Ray with "
62
+ "`pip install ray`.") from ray_import_err
63
+
64
+
65
def initialize_ray_cluster(
66
parallel_config: ParallelConfig,
67
ray_address: Optional[str] = None,
@@ -65,10 +77,7 @@ def initialize_ray_cluster(
77
ray_address: The address of the Ray cluster. If None, uses
78
the default Ray cluster address.
79
"""
68
- if ray is None:
69
- raise ImportError(
70
- "Ray is not installed. Please install Ray to use multi-node "
71
- "serving.")
80
+ assert_ray_available()
81
82
# Connect to a ray cluster.
83
if is_hip() or is_xpu():