[BugFix] Bind RayExecutorV2 TCPStore before publishing its port by aoshen02 · Pull Request #50969 · vllm-project/vllm

aoshen02 · 2026-08-04T05:03:49Z

Purpose

RayExecutorV2 initialized its torch.distributed world group in two separate steps:

driver (possibly another node)                 rank-0 Ray actor
-----------------------------                  ----------------
choose/probe a free integer port
close the probing socket
publish tcp://<rank-0-ip>:<port>  -----------> init_process_group(...)
                                                TCPStore binds <port>

This has two independent correctness problems:

On a multi-node placement, the driver can probe a port on the wrong host. Port occupancy is host-local state.
Even if the probe is moved to the rank-0 actor, get_open_port() closes its socket before returning. Another process can bind the port before init_process_group() creates TCPStore. That is still a TOCTOU race.

The second window includes a Ray RPC back to the driver and the dispatch of every worker initialization RPC. Concurrent engine cold starts make the collision much easier to hit.

Fix

Rank 0 now creates the actual server before publishing its address:

rank-0 Ray actor                               driver / other ranks
----------------                              --------------------
TCPStore(host, port=0, is_master=True,
         wait_for_workers=False,
         multi_tenant=True)
  kernel atomically selects + binds port
  actor retains Store ownership
return tcp://<actor-ip>:<store.port>  --------> initialize workers
                                                rank 0 reuses server
                                                remote ranks connect as clients

port=0 delegates selection and binding to the kernel in one operation. The actor retains the Store, so the listening socket is never released between selection and publication.

PyTorch documents multi_tenant=True as sharing the underlying TCPServer between TCPStore instances in the same process with the same host/port. Therefore the existing tcp:// rendezvous remains unchanged: rank 0's later master TCPStore reuses the already-bound server, while nonzero ranks create clients.

There is no fallback and no application-managed port window. The change is limited to RayExecutorV2; legacy Ray V1 is intentionally unchanged.

Why this is not a duplicate

[Bugfix] Seed RayExecutorV2 TCPStore port by DP rank to avoid collisions #45960 mitigated co-located DP collisions with deterministic port windows, but it still selected an unbound integer before TCPStore bind and did not make a driver-side probe valid on a remote rank-0 host. This PR replaces that window machinery for Ray V2.
[Bugfix] Fix get_open_port() livelock on DP-reserved ports and cover get_open_ports_list #50965 fixes reserved-range behavior inside get_open_port(); this PR removes get_open_port() from this call site.
[Bugfix] Fix ZMQ port TOCTOU race in shm_broadcast MessageQueue #50960 is already merged and removes the analogous ZMQ MessageQueue probe/bind window by binding XPUB to port 0 directly. The latest main containing that fix is merged into this branch.
[BugFix] Pick the DP world-group port at bind time via the coordination store #51018 covers the separate DP world-group coordination port lifecycle; it does not cover this Ray V2 executor TCPStore.
[BugFix] Use file:// rendezvous for single-node executors to eliminate startup port races #50999 handles single-node executor rendezvous with FileStore. Ray V2 may place rank 0 on another node because it calls initialize_ray_cluster(..., require_gpu_on_driver=False), so it requires a network-reachable TCPStore.
[Bugfix] Avoid TCPStore port collision for co-located non-DP Ray engines #53666 (merged after this PR was opened) patched the driver-side _select_tcpstore_port so co-located non-DP replicas pick different ports, with test_colocated_tcpstore_ports_differ. This PR removes _select_tcpstore_port entirely: the port is bound with port=0 on the rank-0 host and held in _dist_init_store, so distinct ports per co-located replica are guaranteed by the kernel. The merge of main therefore drops [Bugfix] Avoid TCPStore port collision for co-located non-DP Ray engines #53666's patch and its test deliberately — it is subsumed, not regressed.
RFC [RFC]: Race-free port management: pick ports at bind time, publish over an existing channel #51275 tracks the wider port-lifecycle work. The duplicate searches for 51275 in:body and RayExecutorV2 TCPStore port found no other open PR covering this Ray V2 call site.

Tests

Repository checks

.venv/bin/python -m pytest \
  tests/distributed/test_ray_v2_executor.py \
  -k dist_init_store_owns_port -q

pre-commit run --files \
  vllm/v1/executor/ray_executor_v2.py \
  tests/distributed/test_ray_v2_executor.py

pre-commit run mypy-3.12 \
  --files vllm/v1/executor/ray_executor_v2.py \
  --hook-stage manual

Results:

Targeted pytest in a CUDA vLLM image: 1 passed, 21 deselected.
All changed-file pre-commit hooks passed.
Python 3.10 and 3.12 mypy passed.
Same-process server reuse was also checked for all use_libuv pairs (True/True, True/False, False/True, and False/False); all reused the held server successfully on Linux.

The unit test asserts three observable invariants:

the published port is nonzero;
a competing socket gets Linux EADDRINUSE while the actor holds the Store;
a second master TCPStore with multi_tenant=True can reuse the same server and exchange a key.

Deterministic Linux A/B

This script forces the old race, validates the final handoff with a real two-process Gloo all-reduce, and stress-checks 128 concurrent ephemeral servers:

.venv/bin/python - <<'PY'
import multiprocessing as mp
import socket
from concurrent.futures import ThreadPoolExecutor
from datetime import timedelta

import torch
import torch.distributed as dist

# Old path: probe -> close -> later bind. A contender deterministically wins.
probe = socket.socket()
probe.bind(("127.0.0.1", 0))
old_port = probe.getsockname()[1]
probe.close()

contender = socket.socket()
contender.bind(("127.0.0.1", old_port))
contender.listen()
try:
    dist.TCPStore(
        "127.0.0.1", old_port, 1, True,
        wait_for_workers=False, multi_tenant=True,
    )
except Exception as exc:
    assert "EADDRINUSE" in str(exc) or "address already in use" in str(exc).lower()
    print("old: EADDRINUSE (expected)")
else:
    raise AssertionError("old path unexpectedly rebound a stolen port")
finally:
    contender.close()

# Final path: rank 0 binds port=0 and holds the Store before publication.
ctx = mp.get_context("fork")
port_q, result_q, start = ctx.Queue(), ctx.Queue(), ctx.Event()

def rank0():
    held = dist.TCPStore(
        "127.0.0.1", 0, 2, True, timedelta(seconds=15),
        wait_for_workers=False, multi_tenant=True,
    )
    port_q.put(held.port)
    start.wait()
    dist.init_process_group(
        "gloo", init_method=f"tcp://127.0.0.1:{held.port}",
        rank=0, world_size=2, timeout=timedelta(seconds=15),
    )
    value = torch.tensor([1])
    dist.all_reduce(value)
    result_q.put((0, value.item()))
    dist.destroy_process_group()

def rank1(port):
    start.wait()
    dist.init_process_group(
        "gloo", init_method=f"tcp://127.0.0.1:{port}",
        rank=1, world_size=2, timeout=timedelta(seconds=15),
    )
    value = torch.tensor([1])
    dist.all_reduce(value)
    result_q.put((1, value.item()))
    dist.destroy_process_group()

p0 = ctx.Process(target=rank0)
p0.start()
new_port = port_q.get(timeout=10)
with socket.socket() as sock:
    try:
        sock.bind(("127.0.0.1", new_port))
    except OSError as exc:
        assert exc.errno == 98
    else:
        raise AssertionError("published port was not held")

p1 = ctx.Process(target=rank1, args=(new_port,))
p1.start()
start.set()
p0.join(20)
p1.join(20)
assert p0.exitcode == p1.exitcode == 0
results = sorted([result_q.get(timeout=2), result_q.get(timeout=2)])
assert results == [(0, 2), (1, 2)]
print("new:", results)

with ThreadPoolExecutor(max_workers=32) as pool:
    stores = list(pool.map(
        lambda _: dist.TCPStore(
            "127.0.0.1", 0, 1, True,
            wait_for_workers=False, multi_tenant=True,
        ),
        range(128),
    ))
ports = [store.port for store in stores]
assert len(set(ports)) == 128
reused = [
    dist.TCPStore(
        "127.0.0.1", port, 1, True,
        wait_for_workers=False, multi_tenant=True,
    )
    for port in ports
]
assert [store.port for store in reused] == ports
print("stress: 128 unique ports, 128 reused servers")
PY

Observed:

old: EADDRINUSE (expected)
new: [(0, 2), (1, 2)]
stress: 128 unique ports, 128 reused servers

Real two-node Ray V2 integration

Tested with one H200 exposed on each of h200-0 and h200-1, forcing a TP=2 world group to span nodes. Ray reported two active nodes and two total GPUs. Both nodes used their routable interface:

export GLOO_SOCKET_IFNAME=ens7
export NCCL_SOCKET_IFNAME=ens7

# h200-0
ray start --head --node-ip-address=172.27.48.114 --port=26379 \
  --disable-usage-stats --num-gpus=1 --block

# h200-1
ray start --address=172.27.48.114:26379 \
  --node-ip-address=172.27.55.233 \
  --disable-usage-stats --num-gpus=1 --block

Driver command:

VLLM_USE_RAY_V2_EXECUTOR_BACKEND=1 \
RAY_ADDRESS=172.27.48.114:26379 \
VLLM_ALLOW_INSECURE_SERIALIZATION=1 \
.venv/bin/python - <<'PY'
from vllm import LLM, SamplingParams

llm = LLM(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    tensor_parallel_size=2,
    distributed_executor_backend="ray",
    enforce_eager=True,
    max_model_len=256,
    gpu_memory_utilization=0.05,
)
output = llm.generate(["Hello"], SamplingParams(max_tokens=1))
print(repr(output[0].outputs[0].text))
PY

Observed on the two workers:

rank 0: world_size=2 ... distributed_init_method=tcp://172.27.48.114:37991
rank 1 (172.27.55.233): world_size=2 ... distributed_init_method=tcp://172.27.48.114:37991
RESULT '，'

The model loaded on both nodes, NCCL/TP groups initialized, and generation completed. Model evaluation is not applicable because the change only affects startup rendezvous and does not alter model execution or outputs.

AI disclosure

AI assistance (Claude Code and OpenAI Codex) was used for investigation, implementation, testing, and drafting. The human submitter must review every changed line and can explain the change and tests end-to-end before merge.

…ds it The driver probed the TCPStore port on its own node, but the store is bound by rank 0 on rank 0's node; on multi-node placements the probe is invalid and concurrent cold starts fail intermittently with EADDRINUSE. Probe on rank 0's actor right before worker init instead, riding the existing deferred initialization; the per-rank port window machinery becomes dead code and is removed. Co-authored-by: Claude <[email protected]> Signed-off-by: aoshen02 <[email protected]>

Co-authored-by: Claude <[email protected]> Signed-off-by: aoshen02 <[email protected]>

claude

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

aoshen02 · 2026-08-07T14:51:12Z

/ci run

github-actions · 2026-08-07T14:51:30Z

✅ CI is already running for this commit: https://buildkite.com/vllm/ci/builds/82860

Co-authored-by: Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 · 2026-08-08T00:00:42Z

/ci run

github-actions · 2026-08-08T00:00:58Z

✅ Triggered Buildkite CI #82950 for commit 94028f4cc6f2.

Keep PR vllm-project#50969 scoped to RayExecutorV2, whose rank zero may be remote from the driver. Co-authored-by: Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]>

Create and retain the TCPStore server on rank zero before publishing its kernel-assigned port. Co-authored-by: OpenAI Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]>

…er-side-port

aoshen02 · 2026-08-08T13:59:52Z

/ci run

github-actions · 2026-08-08T14:00:09Z

✅ Triggered Buildkite CI #83002 for commit fc888a0442e3.

vllm/v1/executor/ray_executor_v2.py: keep binder-side TCPStore port selection on the rank-0 actor; drop main's driver-side _select_tcpstore_port call (vllm-project#53666), which the bind-and-hold approach subsumes. main's sharded_rdt actor options (vllm-project#43375) auto-merged. tests/distributed/test_ray_v2_executor.py: keep test_dist_init_store_owns_port; drop main's _select_tcpstore_port tests including test_colocated_tcpstore_ports_differ, since that helper no longer exists. Co-authored-by: Claude Fable 5 <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 · 2026-08-28T02:21:34Z

/ci run

github-actions · 2026-08-28T02:21:52Z

✅ Triggered Buildkite CI #85917 for commit e7da3bed0f5f.

AndreasKaratzas · 2026-08-28T17:01:15Z

/ci run

github-actions · 2026-08-28T17:02:00Z

✅ Triggered Buildkite CI #85994 for commit a092ffd95ce0.

AndreasKaratzas · 2026-08-28T19:27:05Z

@aoshen02 I m going to merge this PR to resolve a ci failure. However can you open a follow up one? The old port-selection-specific tests can reasonably disappear, but test_colocated_tcpstore_ports_differ captures an externally meaningful invariant rather than an implementation detail. Could we retain that coverage using the new create_dist_init_method() path (e.g. create two rank-0 workers/stores and assert their held ephemeral ports differ)? The new test verifies ownership/reuse for a single store, but no longer directly protects the co-located-engine collision scenario that motivated the earlier regression test.

aoshen02 · 2026-08-29T03:47:26Z

@aoshen02 I m going to merge this PR to resolve a ci failure. However can you open a follow up one? The old port-selection-specific tests can reasonably disappear, but test_colocated_tcpstore_ports_differ captures an externally meaningful invariant rather than an implementation detail. Could we retain that coverage using the new create_dist_init_method() path (e.g. create two rank-0 workers/stores and assert their held ephemeral ports differ)? The new test verifies ownership/reuse for a single store, but no longer directly protects the co-located-engine collision scenario that motivated the earlier regression test.

Thanks, will do.

…-project#50969) Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]> Co-authored-by: aoshen02 <[email protected]> Co-authored-by: Claude <[email protected]> Co-authored-by: OpenAI Codex <[email protected]> Co-authored-by: Andreas Karatzas <[email protected]>

…-project#50969) Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]> Co-authored-by: aoshen02 <[email protected]> Co-authored-by: Claude <[email protected]> Co-authored-by: OpenAI Codex <[email protected]> Co-authored-by: Andreas Karatzas <[email protected]> (cherry picked from commit 67e86d1)

…-project#50969) Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]> Co-authored-by: aoshen02 <[email protected]> Co-authored-by: Claude <[email protected]> Co-authored-by: OpenAI Codex <[email protected]> Co-authored-by: Andreas Karatzas <[email protected]>

…969 is in 0.33.0 (#2729) # What does this PR do? Deletes the `VLLM_DP_MASTER_PORT` workaround that gave every vLLM engine its own TCPStore probe window. vLLM 0.33.0, which SkyRL pins since #2604, includes vllm-project/vllm#50969, so the race it worked around no longer exists. TLDR: SkyRL stops setting `VLLM_DP_MASTER_PORT`. The helpers, constants and tests that existed only for it are removed, following the removal checklist in `inference_servers/common.py`. No behavior change on vLLM 0.33.0. ## Background Added with the vLLM 0.26.0 bump (#1854). Until vllm#50969, `RayExecutorV2._select_tcpstore_port` chose the workers' TCPStore port by probing `[VLLM_DP_MASTER_PORT + 100, +32)`, closing the probe socket, and binding the TCPStore afterwards. With DP disabled `VLLM_DP_MASTER_PORT` defaults to 0, so every engine on a node probed ports 100-131. In containers with `net.ipv4.ip_unprivileged_port_start=0`, two co-located engines both chose port 100 and the second died with `EADDRINUSE`. SkyRL avoided this by setting `VLLM_DP_MASTER_PORT` per server actor so that each engine probed a disjoint window inside its own `SERVER_PORT_STRIDE` block. In vLLM 0.33.0, `_select_tcpstore_port` is gone. `RayExecutorV2.create_dist_init_method` creates the TCPStore with `port=0` and keeps the bound store, so there is no probe/bind gap and `VLLM_DP_MASTER_PORT` no longer affects the port. What still reads `VLLM_DP_MASTER_PORT` in 0.33.0, and why none of it matters to SkyRL: - `ParallelConfig`'s env fallback copies it into `data_parallel_master_port` when DP is off; that field is only consumed on DP > 1 paths. - With DP > 1, SkyRL passes `data_parallel_size` / `data_parallel_rank` / `data_parallel_address` / `data_parallel_rpc_port` as CLI args, and vLLM takes the DP master port from `get_open_ports_list()` without reading the env var. - `get_open_port()` skips `[VLLM_DP_MASTER_PORT, +10)` while the variable is set; the kernel's random ports come from the ephemeral range, far from SkyRL's 8000-range blocks. ## Solution Exactly the removal checklist at `common.py:174-181`: - `common.py`: delete the `DP_TCPSTORE_*` constants and their assert, the checklist, `compute_dp_master_port` and `dp_tcpstore_probe_window`. - `vllm_server_actor.py`: delete the `VLLM_DP_MASTER_PORT` assignment in `VLLMServerActor.__init__`, and `_seed_dp_master_port` with its call in `_build_and_serve_vllm_server` (the standalone `python -m` path). - `test_common.py`: delete `TestDPMasterPort` and the imports only it used. - `setup.py`, `opd/teacher_launch.py`: drop "plus the DP TCPStore probe range" from the port-window docstrings. `SERVER_PORT_STRIDE` stays at 100: the HTTP, NIXL and Mooncake ports and the OPD teacher offset all stride by it, and changing it would move every port SkyRL assigns. Alternatives considered: - Keep `VLLM_DP_MASTER_PORT` set as a safety net. Rejected: nothing on 0.33.0 uses it to pick the TCPStore port, and keeping it means keeping docs and tests for a mechanism that no longer exists. - Replace `TestDPMasterPort` with a test that the variable is not set. Rejected: it would assert the absence of a line. The behavior that matters, co-located engines starting, is covered by `gpu_ci/inference_servers/test_inference_server_group.py`. Both training backends share this inference-server code, so Megatron and FSDP behave the same. ## Test plan - `tests/backends/skyrl_train/inference_servers/test_common.py` keeps covering `find_and_reserve_port` and the rest of `common.py`. - `gpu_ci/inference_servers/test_inference_server_group.py` starts two TP=2 servers on one node through `RayExecutorV2`, the co-located case the workaround existed for. ## Validation | Check | Hardware | Command | Result | |---|---|---|---| | CPU unit tests | macOS | `pytest --noconftest tests/backends/skyrl_train/inference_servers/test_common.py` | 18 passed (23 before, minus the 5 deleted) | | Co-located engines | 4×L4 (`l4_ci`, Anyscale) | `pytest tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_inference_server_group.py -v` | 4 passed in 101.6 s, no `EADDRINUSE` | | Lint | | `pre-commit run --all-files` | passed | The L4 node had `ip_unprivileged_port_start=1024`, so it could not reproduce the original container condition; the run confirms co-located engines start without `VLLM_DP_MASTER_PORT`, and the 0.33.0 source shows the probe is gone. 🤖 Generated with [Claude Code](https://claude.com/claude-code)  --- > [!NOTE] > **Low Risk** > Cleanup tied to a pinned vLLM release; port assignment behavior is unchanged aside from dropping obsolete env seeding. > > **Overview** > SkyRL **stops setting `VLLM_DP_MASTER_PORT`** now that vLLM 0.33.0 includes the upstream fix (vllm#50969) for TCPStore port probing on co-located Ray engines. > > The workaround is fully removed: `DP_TCPSTORE_*` constants, `compute_dp_master_port`, `dp_tcpstore_probe_window`, the env assignment in `VLLMServerActor.__init__`, and `_seed_dp_master_port` on the standalone `python -m` path. **`TestDPMasterPort`** and related imports are deleted from `test_common.py`. Docstrings in `setup.py` and `teacher_launch.py` no longer mention a DP TCPStore range inside each **`SERVER_PORT_STRIDE`** window—those windows still only reserve HTTP (and related) ports via **`find_and_reserve_port`**. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit da6ab9d. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup>  --------- Signed-off-by: Artur Niederfahrenhorst <[email protected]> Co-authored-by: Claude Opus 5.5 <[email protected]>

mergify Bot added the bug Something isn't working label Aug 4, 2026

Replace _select_tcpstore_port unit tests with binder-side probe test

ea21db3

Co-authored-by: Claude <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 marked this pull request as ready for review August 7, 2026 14:50

aoshen02 requested a review from njhill as a code owner August 7, 2026 14:50

Merge branch 'main' into fix/ray-tcpstore-binder-side-port

48860b0

claude Bot reviewed Aug 7, 2026

View reviewed changes

aoshen02 added the ready ONLY add when PR is ready to merge/full CI is needed label Aug 7, 2026

[BugFix] Select Ray v1 TCPStore endpoint on rank zero

94028f4

Co-authored-by: Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 changed the title ~~[BugFix] RayExecutorV2: choose the TCPStore port on the node that binds it~~ [BugFix] Choose Ray TCPStore endpoint on rank zero (v1/v2) Aug 7, 2026

Revert Ray v1 TCPStore endpoint change

6781e90

Keep PR vllm-project#50969 scoped to RayExecutorV2, whose rank zero may be remote from the driver. Co-authored-by: Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 changed the title ~~[BugFix] Choose Ray TCPStore endpoint on rank zero (v1/v2)~~ [BugFix] Choose RayExecutorV2 TCPStore port on rank zero Aug 8, 2026

Fix Ray V2 TCPStore port race

9a6732b

Create and retain the TCPStore server on rank zero before publishing its kernel-assigned port. Co-authored-by: OpenAI Codex <[email protected]> Signed-off-by: aoshen02 <[email protected]>

aoshen02 changed the title ~~[BugFix] Choose RayExecutorV2 TCPStore port on rank zero~~ [BugFix] Bind RayExecutorV2 TCPStore before publishing its port Aug 8, 2026

Merge remote-tracking branch 'origin/main' into fix/ray-tcpstore-bind…

fc888a0

…er-side-port

erictang000 mentioned this pull request Aug 12, 2026

[chore] bump vllm to 0.26.0 NovaSky-AI/SkyRL#1854

Merged

Winter-Dry mentioned this pull request Aug 14, 2026

[Bugfix][Ray] Strip brackets from IPv6 node IP for MessageQueue connect_ip #52307

Open

mergify Bot added the ray anything related with ray label Aug 19, 2026

github-project-automation Bot added this to Ray Aug 19, 2026

github-project-automation Bot moved this to Backlog in Ray Aug 19, 2026

Merge branch 'main' into fix/ray-tcpstore-binder-side-port

a092ffd

AndreasKaratzas approved these changes Aug 28, 2026

View reviewed changes

AndreasKaratzas merged commit 67e86d1 into vllm-project:main Aug 28, 2026
89 checks passed

github-project-automation Bot moved this from Backlog to Done in Ray Aug 28, 2026

aoshen02 deleted the fix/ray-tcpstore-binder-side-port branch August 29, 2026 03:47

aoshen02 mentioned this pull request Aug 29, 2026

[Test] Assert co-located RayExecutorV2 stores publish distinct ports #54310

Merged

seathatflowsinourveins mentioned this pull request Sep 25, 2026

Run Serena's own script from the MCP templates; attribute pin observations to their host; record sparse-VHD, gitleaks-order and WSL lessons native-agent-stack/native-agent-stack#230

Merged

ArturNiederfahrenhorst mentioned this pull request Nov 12, 2026

[vllm] Retire the DP TCPStore port-window workaround now that vllm#50969 is in 0.33.0 NovaSky-AI/SkyRL#2729

Merged

Conversation

aoshen02 commented Aug 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Fix

Why this is not a duplicate

Tests

Repository checks

Deterministic Linux A/B

Real two-node Ray V2 integration

AI disclosure

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Claude Code Review

Uh oh!

aoshen02 commented Aug 7, 2026

Uh oh!

github-actions Bot commented Aug 7, 2026

Uh oh!

aoshen02 commented Aug 8, 2026

Uh oh!

github-actions Bot commented Aug 8, 2026

Uh oh!

aoshen02 commented Aug 8, 2026

Uh oh!

github-actions Bot commented Aug 8, 2026

Uh oh!

aoshen02 commented Aug 28, 2026

Uh oh!

github-actions Bot commented Aug 28, 2026

Uh oh!

AndreasKaratzas commented Aug 28, 2026

Uh oh!

github-actions Bot commented Aug 28, 2026

Uh oh!

AndreasKaratzas commented Aug 28, 2026

Uh oh!

Uh oh!

aoshen02 commented Aug 29, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

aoshen02 commented Aug 4, 2026 •

edited

Loading