[Usage]: How to use existing Ray cluster

[baughmann]

Your current environment

PyTorch version: 2.1.2+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.29.1
Libc version: glibc-2.35

Python version: 3.11.9 (main, Apr  6 2024, 17:59:24) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-1019-oem-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA RTX A5500 Laptop GPU
Nvidia driver version: 535.161.07
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Address sizes:                      46 bits physical, 48 bits virtual
Byte Order:                         Little Endian
CPU(s):                             20
On-line CPU(s) list:                0-19
Vendor ID:                          GenuineIntel
Model name:                         12th Gen Intel(R) Core(TM) i9-12900H
CPU family:                         6
Model:                              154
Thread(s) per core:                 2
Core(s) per socket:                 14
Socket(s):                          1
Stepping:                           3
CPU max MHz:                        5000.0000
CPU min MHz:                        400.0000
BogoMIPS:                           5836.80
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq tme rdpid movdiri movdir64b fsrm md_clear serialize pconfig arch_lbr ibt flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          544 KiB (14 instances)
L1i cache:                          704 KiB (14 instances)
L2 cache:                           11.5 MiB (8 instances)
L3 cache:                           24 MiB (1 instance)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0-19
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:           Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; Enhanced / Automatic IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] onnx==1.16.0
[pip3] onnxruntime==1.15.1
[pip3] torch==2.1.2
[pip3] torchvision==0.16.2
[pip3] triton==2.1.0
[conda] No relevant packagesROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.4.0.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      0-19    0               N/A

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

How would you like to use vllm

I would like to pre-initialize my Ray cluster and have AsyncLLMEngine use it. I had thought that this would be as simple as passing engine_use_ray after creating my cluster, however something is going on after vLLM calls ray.init after I already have.

I end up getting messages from Ray along the lines of:

(autoscaler +5m21s) Error: No available node types can fulfill resource request {'CPU': 1.0, 'GPU': 0.5}. Add suitable node types to this cluster to resolve this issue.

Here's how I'm trying to use it:

# First, manually initialize the Ray cluster
import ray

if ray.is_initialized():
    ray.shutdown()

ray.init(
    num_gpus=1,
    include_dashboard=True,
    dashboard_host="0.0.0.0",
)

# Next, try to get vLLM to use the existing ray cluster
from vllm import AsyncEngineArgs, AsyncLLMEngine
args = AsyncEngineArgs(
    model="/your/model/path", # replace this
    engine_use_ray=True,
    gpu_memory_utilization=0.5
)

engine = AsyncLLMEngine.from_engine_args(args)

I'm running this in a local python notebook and would recommend you do the same when attempting to re-create it. You should see The error messages I saw after the Ray init codeblock. The first one Ray sent was +/- 6s after calling AsyncLLMEngine.from_engine_args().

I know that doing this must be possible, but I lost the notebook that I did it in. As a side-note, it may be nice to allow users to pass Ray options to engine initialization if it turns out that using an existing Ray cluster is more difficult than it seems.

[baughmann]

That error was resolved by adding worker_use_ray, e.g.:

ray.init(
    include_dashboard=True,
    dashboard_host="0.0.0.0",
)
# ...
args = AsyncEngineArgs(
    model="/your/model/path",
    engine_use_ray=True,
    worker_use_ray=True, # added this
    gpu_memory_utilization=0.5,
)

As such, while the errors given in my first post are resolved, I'm still not sure if this is the proper way to go about this and would appreciate some feedback and possibly some links to functions in the source that may be of interest.

For example, if nodes with GPUs are added to the cluster, do I really need to re-initialize the engine with an updated tensor_parallel_size?

Other than those questions, I think this can be closed.