vllm: [Bug]: NCCL hangs and causes timeout

Your current environment

Collecting environment information...
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.24.4
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.10.112-005.ali5000.alios7.x86_64-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.3.107
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 515.105.01
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.4
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.4
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, 57 bits virtual
Byte Order:                      Little Endian
CPU(s):                          128
On-line CPU(s) list:             0-127
Vendor ID:                       GenuineIntel
BIOS Vendor ID:                  Intel(R) Corporation
Model name:                      Intel(R) Xeon(R) Platinum 8369B CPU @ 2.90GHz
BIOS Model name:                 Intel(R) Xeon(R) Platinum 8369B CPU @ 2.90GHz
CPU family:                      6
Model:                           106
Thread(s) per core:              2
Core(s) per socket:              32
Socket(s):                       2
Stepping:                        6
CPU max MHz:                     3500.0000
CPU min MHz:                     800.0000
BogoMIPS:                        5800.00
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 pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid fsrm md_clear pconfig flush_l1d arch_capabilities
Virtualization:                  VT-x
L1d cache:                       3 MiB (64 instances)
L1i cache:                       2 MiB (64 instances)
L2 cache:                        80 MiB (64 instances)
L3 cache:                        96 MiB (2 instances)
NUMA node(s):                    1
NUMA node0 CPU(s):               0-127
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Vulnerable: eIBRS with unprivileged eBPF
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected

Versions of relevant libraries:
[pip3] mypy==1.9.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.2
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] onnx==1.16.0
[pip3] onnx-graphsurgeon==0.3.27
[pip3] onnxruntime==1.16.3
[pip3] torch==2.3.0
[pip3] triton==2.3.0
[pip3] tritonclient==2.44.0
[conda] Could not collectROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.4.2
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	NIC0	NIC1	NIC2	NIC3	CPU Affinity	NUMA Affinity
GPU0	 X 	NV12	NV12	NV12	NV12	NV12	NV12	NV12	PXB	SYS	SYS	SYS	0-127		N/A
GPU1	NV12	 X 	NV12	NV12	NV12	NV12	NV12	NV12	PXB	SYS	SYS	SYS	0-127		N/A
GPU2	NV12	NV12	 X 	NV12	NV12	NV12	NV12	NV12	SYS	PXB	SYS	SYS	0-127		N/A
GPU3	NV12	NV12	NV12	 X 	NV12	NV12	NV12	NV12	SYS	PXB	SYS	SYS	0-127		N/A
GPU4	NV12	NV12	NV12	NV12	 X 	NV12	NV12	NV12	SYS	SYS	PXB	SYS	0-127		N/A
GPU5	NV12	NV12	NV12	NV12	NV12	 X 	NV12	NV12	SYS	SYS	PXB	SYS	0-127		N/A
GPU6	NV12	NV12	NV12	NV12	NV12	NV12	 X 	NV12	SYS	SYS	SYS	PXB	0-127		N/A
GPU7	NV12	NV12	NV12	NV12	NV12	NV12	NV12	 X 	SYS	SYS	SYS	PXB	0-127		N/A
NIC0	PXB	PXB	SYS	SYS	SYS	SYS	SYS	SYS	 X 	SYS	SYS	SYS
NIC1	SYS	SYS	PXB	PXB	SYS	SYS	SYS	SYS	SYS	 X 	SYS	SYS
NIC2	SYS	SYS	SYS	SYS	PXB	PXB	SYS	SYS	SYS	SYS	 X 	SYS
NIC3	SYS	SYS	SYS	SYS	SYS	SYS	PXB	PXB	SYS	SYS	SYS	 X

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

NIC Legend:

  NIC0: mlx5_bond_0
  NIC1: mlx5_bond_1
  NIC2: mlx5_bond_2
  NIC3: mlx5_bond_3

šŸ› Describe the bug

from vllm import LLM
import psutil
import random

llm = LLM($MODEL_PATH, trust_remote_code=True, tensor_parallel_size=4)

prompt_token_ids_list = [[random.randint(1, 40000) for _ in range(random.randint(1, 1000))] for _ in range(1000)]

for i in range(0, 10000):
    batch_size = random.randint(1, 64)
    print(f">>> Iteration: {i}, Batch Size: {batch_size}")
    output = llm.generate(prompt_token_ids=prompt_token_ids_list[:batch_size], use_tqdm=False)
    if i % 1000 == 0:
        cpu_percent = psutil.cpu_percent()
        memory_percent = psutil.virtual_memory().percent
        print(f"CPU utilization: {cpu_percent}%")
        print(f"Memory utilization: {memory_percent}%")
        print("=========================================")

print("Done!")

This is an occasional BUG, ​​which will cause the nccl timeout problem shown below image

cudagraph and custom all recude are enabled

related issue:https://github.com/vllm-project/vllm/issues/1726 https://github.com/vllm-project/vllm/issues/5360 https://github.com/vllm-project/vllm/issues/4653 https://github.com/vllm-project/vllm/issues/4430

About this issue

  • Original URL
  • State: closed
  • Created 2 years ago
  • Reactions: 13
  • Comments: 26 (10 by maintainers)

Most upvoted comments

so the final solution is NCCL_P2P_DISABLE=1 ?

I’m going to close this issue by https://github.com/vllm-project/vllm/pull/10236 .

Just keepin a bug log if it helps someone in the future.

tl;dr still doesn’t work reliably

I upgraded NCCL, restarted the machine, and ran with pipeline-parallel-size=2 and it worked on the first try.

BUT, then I terminated and tried again, and it hung at the usual spot, pinning 3 CPU threads at 100% indefinitely.

I noticed it was not picking up the new version of nccl (2.23.4), just the original old version (2.20.5).

I don’t know if vllm bakes in those dependencies at install time, so I tried:

pip uninstall vllm torch
pip install --no-cache-dir vllm torch

The dep tree is in a weird state, it tried installing decreasing versions of vllm and none were compatible, so, I blew out the whole env and installed everything from scratch, --no-cache-dir.

Again, it can’t find compatible versions. In a fresh env:

$ rm -r .env
$ python3 -m venv .env
$ source .env/bin/activate
$ pip install --no-cache-dir torch matplotlib numpy vllm transformers

<downloads a ton of versions of vllm, none work>

INFO: pip is looking at multiple versions of vllm to determine which version is compatible with other requirements. This could take a while.

Then I blow out the env again, and try just vllm on its own:

pip install --no-cache-dir vllm

That works. And then I can install successive packages successfully too.

But when I do vllm serve ... it still picks up the old nccl version. Torch picks up the wrong one too:

>>> torch.cuda.nccl.version()
(2, 20, 5)

If I do dpkg -L libnccl2 I see it’s the new version, but torch still gets an old version, so I guess it’s installing with its own version?

I notice pip install vllm forces install of the older torch==2.4.

The vllm main branch’s pyproject.toml says it wants torch==2.5, so then I tried git clone https://github.com/vllm-project/vllm and pip install -e vllm (takes ~10-15 min).

Now has an nccl version upgrade (finally!):

>>> import torch
>>> torch.cuda.nccl.version()
(2, 21, 5)

vllm serve ..., it picks up the new nccl version.

The first time running this, it successfully ran, using both cards.

The next 3 times trying this it hung at the same step.

I’m also facing the same issue,

Do we’ve a solution to this? Few other isses I read where they mentioned to install nvidia-fabricmanager.service, but even after installing that the fabricmanager service was not getting started https://forums.developer.nvidia.com/t/nvidia-fabricmanager-running-error-with-nv-warn-nothing-to-do/272899 Following above forum, I found that for NVSwitch we require to use fabricmanager for better intercommunication between nvidia GPUs.

I’m having 8 nvidia GPUs on my system.

vllm serve model.gguf --pipeline-parallel-size=2

the command is correct, but your gpus cannot talk to each other.

FYI i recently added a page of debugging tips for this problem: https://docs.vllm.ai/en/latest/getting_started/debugging.html

In addition to these two switches --disable-custom-all-reduce and --enforce-eager, is there any other solution to solve this problem? Thanks~