[BUG] Expert server fails to bootstrap up to 90% of the time

[Vectorrent]

Describe the bug
When using hivemind.moe.Server to host experts in a background thread, bootstrapping will fail over and over again, repeatedly - leading to a complete deadlock. I am forced to restart my application repeatedly, sometimes upwards of 10 times, before a single bootstrap will work.

As you can imagine, this is very frustrating - especially during development, when I need to iterate quickly.

If there is a better way to host experts, or if you could tell me how to resolve this problem - I would GREATLY appreciate it!

To Reproduce
Run this script a few times in a row. It will ALWAYS fail, eventually:

import time

import torch
import torch.nn as nn
import torch.nn.functional as F
from hivemind import DHT
from hivemind.moe import ModuleBackend, Server, register_expert_class
from hivemind.moe.server.layers import name_to_block
from hivemind.utils import BatchTensorDescriptor

sample_input = lambda batch_size, hidden_dim: torch.empty((batch_size, hidden_dim))


@register_expert_class("perceptron", sample_input)
class MultilayerPerceptron(nn.Module):
    def __init__(self, hidden_dim, num_classes=10):
        super().__init__()
        self.layer1 = nn.Linear(hidden_dim, 2 * hidden_dim)
        self.layer2 = nn.Linear(2 * hidden_dim, 2 * hidden_dim)
        self.layer3 = nn.Linear(2 * hidden_dim, num_classes)

    def forward(self, x):
        x = F.relu(self.layer1(x))
        x = F.relu(self.layer2(x))
        x = self.layer3(x)
        return x


PUBLIC_INITIAL_PEERS = [
    # IPv4 DNS addresses
    "/dns/bootstrap1.petals.dev/tcp/31337/p2p/QmedTaZXmULqwspJXz44SsPZyTNKxhnnFvYRajfH7MGhCY",
    "/dns/bootstrap2.petals.dev/tcp/31338/p2p/QmQGTqmM7NKjV6ggU1ZCap8zWiyKR89RViDXiqehSiCpY5",
    # IPv6 DNS addresses
    "/dns6/bootstrap1.petals.dev/tcp/31337/p2p/QmedTaZXmULqwspJXz44SsPZyTNKxhnnFvYRajfH7MGhCY",
    "/dns6/bootstrap2.petals.dev/tcp/31338/p2p/QmQGTqmM7NKjV6ggU1ZCap8zWiyKR89RViDXiqehSiCpY5",
    # Reserved IPs
    "/ip4/159.89.214.152/tcp/31337/p2p/QmedTaZXmULqwspJXz44SsPZyTNKxhnnFvYRajfH7MGhCY",
    "/ip4/159.203.156.48/tcp/31338/p2p/QmQGTqmM7NKjV6ggU1ZCap8zWiyKR89RViDXiqehSiCpY5",
]

dht = DHT(
    initial_peers=PUBLIC_INITIAL_PEERS,
    host_maddrs=["/ip4/0.0.0.0/tcp/0", "/ip4/0.0.0.0/udp/0/quic"],
    start=True,
)
hidden_schema = BatchTensorDescriptor(
    64,
)
backends = {}
num_layers = 3
for i in range(num_layers):
    expert_name = f"expert.{i}"
    expert = ModuleBackend(
        name=expert_name,
        module=name_to_block["perceptron"](64),
        args_schema=(hidden_schema,),
        outputs_schema=(hidden_schema),
        max_batch_size=64,
        start=True,
    )
    backends[expert_name] = expert
server = Server(
    dht,
    backends,
)
server.run_in_background(timeout=5)

while True:
    print("waiting...")
    time.sleep(3)

Environment

• python: v3.12.6
• pytorch: v2.4.1
• hivemind: https://github.com/learning-at-home/hivemind.git@213bff98a62accb91f254e2afdccbf1d69ebdea9
• collect_env.py:

Collecting environment information...
PyTorch version: 2.4.1+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Arch Linux (x86_64)
GCC version: (GCC) 14.2.1 20240910
Clang version: 18.1.8
CMake version: Could not collect
Libc version: glibc-2.40

Python version: 3.12.6 (main, Sep  8 2024, 13:18:56) [GCC 14.2.1 20240805] (64-bit runtime)
Python platform: Linux-6.11.1-arch1-1-x86_64-with-glibc2.40
Is CUDA available: True
CUDA runtime version: 12.6.68
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA GeForce GTX 960
GPU 1: NVIDIA GeForce GTX 1070

Nvidia driver version: 560.35.03
cuDNN version: Probably one of the following:
/usr/lib/libcudnn.so.9.2.1
/usr/lib/libcudnn_adv.so.9.2.1
/usr/lib/libcudnn_cnn.so.9.2.1
/usr/lib/libcudnn_engines_precompiled.so.9.2.1
/usr/lib/libcudnn_engines_runtime_compiled.so.9.2.1
/usr/lib/libcudnn_graph.so.9.2.1
/usr/lib/libcudnn_heuristic.so.9.2.1
/usr/lib/libcudnn_ops.so.9.2.1
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:                        39 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               8
On-line CPU(s) list:                  0-7
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
CPU family:                           6
Model:                                60
Thread(s) per core:                   2
Core(s) per socket:                   4
Socket(s):                            1
Stepping:                             3
CPU(s) scaling MHz:                   96%
CPU max MHz:                          4400.0000
CPU min MHz:                          800.0000
BogoMIPS:                             7999.08
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 arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm cpuid_fault epb pti ssbd ibrs ibpb stibp fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid xsaveopt dtherm ida arat pln pts md_clear flush_l1d
L1d cache:                            128 KiB (4 instances)
L1i cache:                            128 KiB (4 instances)
L2 cache:                             1 MiB (4 instances)
L3 cache:                             8 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-7
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          KVM: Mitigation: VMX unsupported
Vulnerability L1tf:                   Mitigation; PTE Inversion
Vulnerability Mds:                    Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown:               Mitigation; PTI
Vulnerability Mmio stale data:        Unknown: No mitigations
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; Retpolines; IBPB conditional; IBRS_FW; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Mitigation; Microcode
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] numpy==2.1.2
[pip3] nvidia-cublas-cu12==12.1.3.1
[pip3] nvidia-cuda-cupti-cu12==12.1.105
[pip3] nvidia-cuda-nvrtc-cu12==12.1.105
[pip3] nvidia-cuda-runtime-cu12==12.1.105
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.0.2.54
[pip3] nvidia-curand-cu12==10.3.2.106
[pip3] nvidia-cusolver-cu12==11.4.5.107
[pip3] nvidia-cusparse-cu12==12.1.0.106
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] nvidia-nvjitlink-cu12==12.6.77
[pip3] nvidia-nvtx-cu12==12.1.105
[pip3] pytorch-lightning==2.4.0
[pip3] pytorch_optimizer==3.1.2
[pip3] torch==2.4.1
[pip3] torchmetrics==1.4.3
[pip3] triton==3.0.0
[conda] Could not collect

[Vectorrent]

So, it appears that using Server.create() to launch experts is much more reliable. While I can probably work with this new method, I remember now why I moved-away from it in the first place:

Server.create() makes an assumption that experts may only accept hidden_dim as an argument during initialization:

expert = name_to_block[expert_cls](hidden_dim)

I find this annoying, but maybe I'll find a clean way to work round the limitation.

[Vectorrent]

Okay, yeah - inheriting from Server and overwriting the create() method seems like a decent, not terrible solution for my needs. The other code is still buggy, but perhaps we were never supposed to use it that way. Will close this issue for now.

[Vectorrent]

I spoke too soon. The bootstrapping problem returned, even when using my customized Server.create() method. However, I think I've found the source of the problem.

For whatever reason, this piece of code is REQUIRED, to prevent issues during bootstrapping:

visible_maddrs_str = [str(a) for a in dht.get_visible_maddrs()]

You don't actually have to use this data. The mere act of calling dht.get_visible_maddrs() from within the server creation thread is enough to prevent bootstrapping issues. Without this code, bootstrap can fail up to 90% of the time.

Clearly, this a bug - and one that should be somewhat easy to fix.