Repository navigation
Conversation
In pull mode, the decoder does not release a remote prefill's NIXL state on failed heartbeat (which is detectable from UCX disconnection error, but current code does not check it). When a prefill node crashes, its rkeys stay loaded on the decode nodes it recently communicated with, which results in MNNVL stale mappings. When the prefill node attempts to restart, it is unable to remap memory. The current implementation of heartbeat could detect this, but they are only sent by decoders when they need to extend KV leases for pending requests. This patch adds periodic sending of heartbeats, with detection and cleanup of disconnected peers. Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
This was referenced Oct 8, 2026
Contributor
|
This pull request has merge conflicts that must be resolved before it can be |
stu-cao
pushed a commit
to stu-cao/vllm
that referenced
this pull request
Oct 10, 2026
In pull mode, the decoder does not release a remote prefill's NIXL state on failed heartbeat (which is detectable from UCX disconnection error, but current code does not check it). When a prefill node crashes, its rkeys stay loaded on the decode nodes it recently communicated with, which results in MNNVL stale mappings. When the prefill node attempts to restart, it is unable to remap memory. The current implementation of heartbeat could detect this, but they are only sent by decoders when they need to extend KV leases for pending requests. This patch adds periodic sending of heartbeats, with detection and cleanup of disconnected peers. (cherry picked from vllm-project#60268, commit e629145) [prime: conflict in get_transfer_results resolved by keeping the fork's _cleanup_replaced_remote_engines() call and adding _release_unreachable_remote_engines() after it] Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Stu Cao <stu@primeintellect.ai>
stu-cao
added a commit
to stu-cao/vllm
that referenced
this pull request
Oct 10, 2026
…ct#60268 Three tests for the pull-mode heartbeat added by the previous commit: a peer whose heartbeat raises a remote-disconnect error is released, it is kept while a read from it is still in flight, and the per-engine heartbeat is rate limited. The connector fixtures that build workers with object.__new__ now set the three attributes the new code reads in get_transfer_results(), with the periodic heartbeat disabled; 14 tests failed with AttributeError without this. Test: NIXL unit suites on a B300 pod, before vs after the cherry-pick: no new failures, and these 3 tests now pass. The other failures are the pod's environment (no facebook/opt-125m) and fail identically before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Stu Cao <stu@primeintellect.ai>
mergify
Bot
added
nixl
ovidiusm
marked this pull request as ready for review
November 4, 2026 19:31
ovidiusm
requested review from
ApostaC,
NickLucche,
ivanium,
orozery,
wzhao18 and
xuechendi
as code owners
November 4, 2026 19:31
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
In pull mode, the decoder does not release a remote prefill's NIXL state on failed heartbeat (which is detectable from UCX disconnection error, but current code does not check it).
When a prefill node crashes, its rkeys stay loaded on the decode nodes it recently communicated with, which results in MNNVL stale mappings. When the prefill node attempts to restart, it is unable to remap memory.
The current implementation of heartbeat could detect this, but they are only sent by decoders when they need to extend KV leases for pending requests.
This patch adds periodic sending of heartbeats, with detection and cleanup of disconnected peers.
Addresses #58063
Claims
Validation
On a GB200 NVL72 system, we ran vLLM v0.31.0 with this patch and NIXL 1.5.0, with --enable-cumem-allocator, on every instance, serving with two prefill and two decode instances under continuous traffic. We killed one prefill instance six times and each time started a replacement on the same GPU, restarting one decode instance along the way. In every cycle the decode instances released the killed prefill within about 10 seconds (about 25 seconds with the two keepalive variables left at their defaults), its GPU memory dropped to zero, the replacement started with its full memory budget and served all requests from both decode instances, and the other prefill instance kept serving, with no errors on the decode side.
Caveat
Two workarounds are needed until upstream changes land; two more to shorten detection.
UCX_CUDA_IPC_CACHE=nUCP_EP_PARAMS_FLAGS_RKEY_BOUND_LIFETIMEUCX_RC_GDA_EP_ALLOC_MODE=directrc_gdaQP pool returns QPs still connected to the old peer, and the next connection failsINIT2RTR(syndrome0x5d668c)UCX_KEEPALIVE_INTERVAL=2sUCX_UD_TIMEOUT=5sDetails
Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.