sglang ipv6 issue for https://github.com/sgl-project/sglang/issues/2892

issue.py

# modal run issue.py
import os
import subprocess
import time

import modal
import modal.experimental
import modal.gpu
import modal.volume

image = (
    modal.Image.debian_slim(python_version="3.12")
    .pip_install("torch", "numpy<2")
    .apt_install("curl", "git", "iproute2")
    .run_commands("git clone https://github.com/sgl-project/sglang.git")
    .run_commands(
        "cd sglang && pip install -e 'python[all]' --find-links https://flashinfer.ai/whl/cu124/torch2.4/flashinfer/"
    )
)

app = modal.App(
    os.path.basename(__file__),
    image=image,
    secrets=[modal.Secret.from_name("huggingface-secret", required_keys=["HF_TOKEN"])],
)

n_nodes = 2
n_proc_per_node = 2
total_gpus = n_nodes * n_proc_per_node


def shellbg(cmd: str):
    print(f"Running:\n\t$ {cmd}", flush=True)
    subprocess.Popen(cmd, shell=True)


def mysleep(seconds):
    "long sleeps block stdout network sync?"
    for _ in range(seconds * 1000):
        time.sleep(0.001)


@app.function(
    gpu=modal.gpu.H100(count=n_proc_per_node),
    timeout=3600 * 24,
    cpu=64,
    memory=344_064,  # in megabytes
)  # pyright: ignore
@modal.experimental.clustered(size=n_nodes)
def fire_it_up():
    cluster_info = modal.experimental.get_cluster_info()
    node_rank = cluster_info.rank
    main_addr = cluster_info.container_ips[0]
    my_addr = cluster_info.container_ips[node_rank]
    num_nodes = len(cluster_info.container_ips)
    task_id = os.environ["MODAL_TASK_ID"]
    print(f"{cluster_info=}")
    print(f"hello from {node_rank=}, {main_addr=}, {my_addr=}, {num_nodes=}, {task_id=}")

    shellbg("ip address show")

    tp = num_nodes * n_proc_per_node

    print("\n\n\n")
    mysleep(5)
    if node_rank == 0:
        shellbg("python3 -m sglang.check_env")
    mysleep(15)
    print("\n\n\n")

    shellbg(
        f"""
        export GLOO_SOCKET_IFNAME='eth1' MASTER_ADDR={main_addr} MASTER_PORT=1234;
        echo GLOO_SOCKET_IFNAME=$GLOO_SOCKET_IFNAME and MASTER_ADDR=$MASTER_ADDR and MASTER_PORT=$MASTER_PORT;
        python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B \
            --tp-size {tp}  --nnodes {num_nodes} --node-rank {node_rank} \
            --port 2242 --dist-init-addr [{main_addr}]:1234 \
            --disable-cuda-graph --trust-remote-code
    """
    )
    for i in range(10000000):
        shellbg("curl http://localhost:2242/v1/models")
        if i % 3 == 1:
            shellbg("nvidia-smi --query-gpu=power.draw,memory.used,memory.free --format=csv")
        mysleep(10)