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 15, 21:00 UTC
[core] Bump ray to 2.43 (#13994)
Signed-off-by: Rui Qiao <ruisearch42@gmail.com>
Rui Qiao committed
Feb 28, 2025 at 13:47 UTC
084bbac8cc4c29b7dcd2098418168c61d3d42e9b
5 files changed
+8
-10
.github/dependabot.yml
+1
-1
@@ -23,7 +23,7 @@ updates:
23
- dependency-name: "lm-format-enforcer"
24
- dependency-name: "gguf"
25
- dependency-name: "compressed-tensors"
26
- - dependency-name: "ray[adag]"
26
+ - dependency-name: "ray[cgraph]" # Ray Compiled Graph
27
- dependency-name: "lm-eval"
28
groups:
29
minor-update:
requirements-cuda.txt
+1
-1
@@ -2,7 +2,7 @@
2
-r requirements-common.txt
3
4
# Dependencies for NVIDIA GPUs
5
-ray[adag] == 2.40.0 # Required for pipeline parallelism in V1.
5
+ray[cgraph] >= 2.43.0 # Ray Compiled Graph, required for pipeline parallelism in V1.
6
torch == 2.5.1
7
torchaudio==2.5.1
8
# These must be updated alongside torch
requirements-test.in
+1
-1
@@ -16,7 +16,7 @@ vector_quantize_pytorch # required for minicpmo_26 test
16
vocos # required for minicpmo_26 test
17
peft
18
pqdm
19
-ray[adag]==2.40.0
19
+ray[cgraph]>=2.43.0 # Ray Compiled Graph, required by pipeline parallelism tests
20
sentence-transformers # required for embedding tests
21
soundfile # required for audio tests
22
jiwer # required for audio tests
requirements-test.txt
+1
-1
@@ -472,7 +472,7 @@ pyyaml==6.0.2
472
# vocos
473
rapidfuzz==3.12.1
474
# via jiwer
475
-ray==2.40.0
475
+ray==2.43.0
476
# via -r requirements-test.in
477
redis==5.2.0
478
# via tensorizer
vllm/executor/ray_distributed_executor.py
+4
-6
@@ -500,7 +500,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
500
import pkg_resources
501
from packaging import version
502
503
- required_version = version.parse("2.40")
503
+ required_version = version.parse("2.43.0")
504
current_version = version.parse(
505
pkg_resources.get_distribution("ray").version)
506
if current_version < required_version:
@@ -512,20 +512,19 @@ class RayDistributedExecutor(DistributedExecutorBase):
512
"ray.experimental.compiled_dag_ref")
513
if cgraph_spec is None:
514
raise ValueError("Ray Compiled Graph is not installed. "
515
- "Run `pip install ray[adag]` to install it.")
515
+ "Run `pip install ray[cgraph]` to install it.")
516
517
cupy_spec = importlib.util.find_spec("cupy")
518
if cupy_spec is None and envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL:
519
raise ValueError(
520
"cupy is not installed but required since "
521
"VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL is set. "
522
- "Run `pip install ray[adag]` and check cupy installation.")
522
+ "Run `pip install ray[cgraph]` and check cupy installation.")
523
524
def _compiled_ray_dag(self, enable_asyncio: bool):
525
assert self.parallel_config.use_ray
526
self._check_ray_cgraph_installation()
527
from ray.dag import InputNode, MultiOutputNode
528
- from ray.experimental.channel.torch_tensor_type import TorchTensorType
528
529
logger.info("VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL = %s",
530
envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL)
@@ -574,8 +573,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
573
if envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL \
574
else "auto"
575
outputs = [
577
- output.with_type_hint(
578
- TorchTensorType(transport=transport))
576
+ output.with_tensor_transport(transport=transport)
577
for output in outputs
578
]
579