[Bug]: Multinode serving with Ray not working due to cupy

[nikhil-tensorwave]

Your current environment

The output of python collect_env.py

INFO 07-02 22:38:50 [__init__.py:244] Automatically detected platform rocm.
Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version                : 19.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.4.1 25184 c87081df219c42dc27c5b6d86c0525bc7d01f727)
CMake version                : version 3.31.6
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.7.0+gitf717b2a
Is debug build               : False
CUDA used to build PyTorch   : N/A
ROCM used to build PyTorch   : 6.4.43483-a187df25c

==============================
      Python Environment
==============================
Python version               : 3.12.10 (main, Apr  9 2025, 08:55:05) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-6.8.0-60-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration : AMD Instinct MI325X (gfx942:sramecc+:xnack-)
Nvidia driver version        : Could not collect
cuDNN version                : Could not collect
HIP runtime version          : 6.4.43483
MIOpen runtime version       : 3.4.0
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               128
On-line CPU(s) list:                  0-127
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 9575F 64-Core Processor
CPU family:                           26
Model:                                2
Thread(s) per core:                   1
Core(s) per socket:                   64
Socket(s):                            2
Stepping:                             1
Frequency boost:                      enabled
CPU max MHz:                          5008.0068
CPU min MHz:                          1500.0000
BogoMIPS:                             6599.83
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc amd_ibpb_ret arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d debug_swap
L1d cache:                            6 MiB (128 instances)
L1i cache:                            4 MiB (128 instances)
L2 cache:                             128 MiB (128 instances)
L3 cache:                             512 MiB (16 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-63
NUMA node1 CPU(s):                    64-127
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 Reg file data sampling: 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; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] pyzmq==26.4.0
[pip3] torch==2.7.0+gitf717b2a
[pip3] torchvision==0.21.0+7af6987
[pip3] transformers==4.52.4
[pip3] triton==3.2.0+gite5be006a
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : 6.4.43483-a187df25c
Neuron SDK Version           : N/A
vLLM Version                 : 0.9.0.2.dev108+g71faa1880 (git sha: 71faa1880)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
  ============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            15           15           15           15           15           15           15           
GPU1   15           0            15           15           15           15           15           15           
GPU2   15           15           0            15           15           15           15           15           
GPU3   15           15           15           0            15           15           15           15           
GPU4   15           15           15           15           0            15           15           15           
GPU5   15           15           15           15           15           0            15           15           
GPU6   15           15           15           15           15           15           0            15           
GPU7   15           15           15           15           15           15           15           0            

================================= Hops between two GPUs ==================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            1            1            1            1            1            1            1            
GPU1   1            0            1            1            1            1            1            1            
GPU2   1            1            0            1            1            1            1            1            
GPU3   1            1            1            0            1            1            1            1            
GPU4   1            1            1            1            0            1            1            1            
GPU5   1            1            1            1            1            0            1            1            
GPU6   1            1            1            1            1            1            0            1            
GPU7   1            1            1            1            1            1            1            0            

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU1   XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU2   XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         
GPU3   XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         
GPU4   XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         
GPU5   XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         
GPU6   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         
GPU7   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            

======================================= Numa Nodes =======================================
GPU[0]          : (Topology) Numa Node: 0
GPU[0]          : (Topology) Numa Affinity: 0
GPU[1]          : (Topology) Numa Node: 0
GPU[1]          : (Topology) Numa Affinity: 0
GPU[2]          : (Topology) Numa Node: 0
GPU[2]          : (Topology) Numa Affinity: 0
GPU[3]          : (Topology) Numa Node: 0
GPU[3]          : (Topology) Numa Affinity: 0
GPU[4]          : (Topology) Numa Node: 1
GPU[4]          : (Topology) Numa Affinity: 1
GPU[5]          : (Topology) Numa Node: 1
GPU[5]          : (Topology) Numa Affinity: 1
GPU[6]          : (Topology) Numa Node: 1
GPU[6]          : (Topology) Numa Affinity: 1
GPU[7]          : (Topology) Numa Node: 1
GPU[7]          : (Topology) Numa Affinity: 1
================================== End of ROCm SMI Log ===================================

==============================
     Environment Variables
==============================
PYTORCH_TUNABLEOP_TUNING=0
NCCL_IB_PCI_RELAXED_ORDERING=1
NCCL_DEBUG=WARN
PYTORCH_TUNABLEOP_ENABLED=1
NCCL_SHM_DISABLE=1
NCCL_PXN_DISABLE=0
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
NCCL_MAX_NCHANNELS=48
PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx1100;gfx1101;gfx1200;gfx1201
NCCL_NET_GDR_READ=1
LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
VLLM_HOST_IP=10.31.69.48
NCCL_IGNORE_CPU_AFFINITY=1
PYTORCH_TUNABLEOP_FILENAME=/app/afo_tune_device_%d_full.csv
NCCL_CROSS_NIC=0
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

When running multinode inference with the Docker image rocm/vllm:rocm6.4.1_vllm_0.9.0.1_20250605 I am unable to actually have requests be processed correctly. The model loads properly (after applying the patch discussed in Issue #590, but when I try to send a request like the following:

curl -X POST http://localhost:8080/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "/root/models/Llama-3.1-70B-Instruct",
        "messages": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Who won the world series in 2020?"}
        ]
    }'

then I get the following error:

ray.exceptions.RayTaskError(ModuleNotFoundError): ray::RayWorkerWrapper.__ray_call__() (pid=1152, ip=10.31.69.48, actor_id=c43f3b996680f147308b8ebc02000000, repr=<vllm.executor.ray_utils.RayWorkerWrapper object at 0x7830b3a7c8c0>)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/ray/actor.py", line 1732, in __ray_call__
    return fn(self, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/ray/experimental/channel/torch_tensor_nccl_channel.py", line 683, in _do_get_unique_nccl_id
    from cupy.cuda import nccl
ModuleNotFoundError: No module named 'cupy'

cupy seems to be a CUDA-dependent package but there are ROCm build instructions and ROCm whls in their docs. Unfortunately, the build instructions do not work and the whls are for ROCm 4.3 and ROCm 5.0, and are not compatible with ROCm 6.3. Why does Ray have a dependency on a CUDA package and why is a ROCm version of this package not included in the vLLM image? Is there some additional setting I am missing?

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[ppanchad-amd]

Hi @nikhil-tensorwave. Internal ticket has been created to investigate this issue. Thanks!

[schung-amd]

Hi @nikhil-tensorwave, can you try this with rocm/vllm-dev:nightly? Also, can you try updating your Ray installation (either through pip or building from source)?

[nikhil-tensorwave]

I can try a nightly build, can you point to instructions for building from source? Are there any ROCm specific instructions or dependencies?

[schung-amd]

I recommend trying updating through pip first, the ray build from source instructions can be found at https://docs.ray.io/en/latest/ray-contribute/development.html#building-ray but I'm not sure if this is sufficient for ROCm systems. I'll update if I get a chance to try building from source myself.

[nikhil-tensorwave]

I tried with the newest vllm image rocm/vllm:rocm6.4.1_vllm_0.9.1_20250702 but I am unable to get it to work. I pip installed ray version 2.45 and 2.47, but both resulted in a new error in model loading:

ERROR 07-08 00:35:29 [core.py:519]   File "/usr/local/lib/python3.12/dist-packages/vllm/platforms/rocm.py", line 79, in <module>
ERROR 07-08 00:35:29 [core.py:519]     assert val == cuda_val
ERROR 07-08 00:35:29 [core.py:519]            ^^^^^^^^^^^^^^^
ERROR 07-08 00:35:29 [core.py:519] AssertionError

I am also unable to build ray from source but am running into issues, I see it installing cupy-cuda and then run into other issues.

[schung-amd]

Thanks for checking. We also have ROCm-compatible ray wheels at https://github.com/ROCm/ray/releases/tag/v3.0.0-dev0%2Brocm, but I don't see one for Python 3.12. Will try to get hardware to reproduce this.

[islwx]

same bug at RX7600 multi-node

[islwx]

Thanks for checking. We also have ROCm-compatible ray wheels at https://github.com/ROCm/ray/releases/tag/v3.0.0-dev0%2Brocm, but I don't see one for Python 3.12. Will try to get hardware to reproduce this.

I saw this whl and bug, i found this bug is raise by vllm v1 engine. when using vllm v0 will not raise.
Perhaps this anomaly was caused by updates to v1 and v0(vllm 0.9+, default using v1 engine)

[schung-amd]

One thing to try is installing cupy from source: https://docs.cupy.dev/en/stable/install.html#using-cupy-on-amd-gpu-experimental. While cupy binaries are only provided for ROCm 4 and 5, the install from source docs state experimental support for ROCm 6.x.

[schung-amd]

I've inquired internally about this and it seems like we might not actually have ray support in vllm at the moment; not that it won't work necessarily, but would explain why we don't include cupy in the Docker images. Again, I suggest trying the build from source instructions for cupy, as it looks like those should work on ROCm 6.x.