[Bugfix][Rust Frontend] Tolerate NaN-corrupted logprobs in engine-core by almersawi · Pull Request #51026 · vllm-project/vllm

almersawi · 2026-08-04T13:26:00Z

Purpose

Fix a total-outage bug in the Rust frontend's engine-core client: a single request asking for logprobs/prompt_logprobs while the model produces NaN logits permanently kills the client — /health returns 503 and every subsequent request (including ones with no logprobs) fails with engine-core error: messagepack ext value decode failed: token_ranks must be >= 1 for decoded engine-core logprobs until process restart.

Rank 0 is an engine output that must be accepted: the sampler computes the rank as (logprobs >= sampled_logprob).sum(-1) (vllm/v1/sample/sampler.py), which is 0 when the sampled logprob is NaN because every >= comparison against NaN is false. The decode failure is frame-scoped (one bad row fails the whole batched EngineCoreOutputs payload), happens on the shared output-reader path, and is recorded as a sticky health error, so a per-request data anomaly becomes a server-wide outage.

Fix: in PositionLogprobs::from_decoded_row, accept and preserve sampled_rank = 0 along with the engine's original logprob values. This keeps decoding valid for the other requests in the batch. The existing Rust OpenAI serving helper converts NaN and negative infinity to -9999.0 when constructing HTTP responses; engine-core decoding preserves the numerical results for all consumers.

Why this does not duplicate existing PRs:

[Bugfix][Frontend] Handle None/NaN/Inf logprob values when using FP8 quantization #48585 fixes the same class of defect (non-finite logprobs reaching the serving layer) but in the Python frontend, and for logprob values, not ranks. This PR addresses the Rust engine-client rank-zero decode failure; the Rust HTTP helper already handles NaN and negative infinity.
[Bug]: System crash, Rust front end panic (process abort) from an empty stop string via the gRPC Generate API #50725 is a different code path (process abort from an empty stop string); it only shares the blast-radius pattern.
[Core] Zero KV cache when NaN logits detected #50002 proposed aborting requests on optional NaN detection and was superseded by [Core] Add --enable-nan-fault-tolerance for NaN detection and request abort #50283; this PR handles rank-zero payloads that reach the frontend.
[Bugfix]Fix out-of-vocabulary recovered token on all-NaN logits (root cause of empty spec-decode output) #45060 establishes all-NaN logits as a real engine state vLLM already handles elsewhere.

Healthy payloads retain their existing behavior. Model evaluation N/A: this changes wire-payload decode tolerance while preserving the engine's numerical output.

AI assistance was used for the original fix and the follow-up changes and tests. The follow-up preserves rank 0 and raw engine values and adds coverage for the existing HTTP conversion.

Test Plan

Two regression tests cover the layer boundary:

zero_sampled_rank_does_not_fail_frame_mates decodes a shared EngineCoreOutputs batch, preserving rank 0, NaN, and negative infinity in both sample and prompt logprobs while keeping the healthy request intact.
chat_logprobs_clamps_nan_and_negative_infinity_for_json verifies that the existing OpenAI response conversion emits -9999.0 for NaN and negative infinity while preserving a finite logprob.

Test Result

cargo nextest run --locked -p vllm-engine-core-client -p vllm-server -E 'package(vllm-engine-core-client) | test(routes::openai::utils::logprobs::tests)': 118 passed.
cargo clippy --locked -p vllm-engine-core-client -p vllm-server --all-targets --all-features -- -D warnings: passed.
cargo fmt --all -- --check, git diff --check, and commit-time pre-commit hooks: passed.

Commands ran from rust/ with a fresh short TMPDIR for tests and an existing Cargo target directory. No GPU was required.

Model evaluation N/A: the change preserves the engine's numerical output and adjusts frontend decode tolerance.

Essential Elements of an Effective PR Description Checklist

The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
The test plan, such as providing test command.
The test results, such as pasting the results comparison before and after, or e2e results
(Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. (N/A — internal decode-tolerance fix, no docs affected)

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.

BugenZhao · 2026-08-04T23:38:29Z

@codex review

chatgpt-codex-connector · 2026-08-04T23:42:48Z

Codex Review: Didn't find any major issues. 👍

Reviewed commit: 257301a689

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

almersawi · 2026-08-12T09:43:32Z

@BugenZhao Did you have any chance to check this?

…e client The sampler computes a sampled token's vocab rank as `(logprobs >= sampled_logprob).sum(-1)`, which is 0 when the sampled logprob is NaN because every IEEE-754 comparison against NaN is false. The engine-core client treated rank 0 as a corrupt wire payload and bailed during decode of the *batched* EngineCoreOutputs frame, which tripped the shared output-reader path and made the client record a permanent health error - killing every in-flight and subsequent request until process restart. Clamp rank 0 to the lowest valid rank, and map non-finite logprob values to the same -9999.0 sentinel the Python frontend uses (vllm-project#48585), so a NaN state degrades to one request's model-quality problem instead of a server outage. Tests: `cargo test -p vllm-engine-core-client` (100 passed), including new cases pinning rank-0 clamping, NaN-sentinel mapping, and survival of co-scheduled requests sharing the bad row's frame. Co-authored-by: Kimi (AI coding assistant) Signed-off-by: Islam <islam.almersawi@openinnovation.ai>

BugenZhao

Rest LGTM. Thanks!

BugenZhao · 2026-09-16T02:36:09Z

+            // Map non-finite logprobs (from corrupt logits) to the same -9999.0
+            // sentinel the Python frontend uses, so the serving layer can still
+            // serialize the response.
+            let logprob = if logprob.is_finite() {
+                logprob
+            } else {
+                -9999.0
+            };


I think this is not necessary? We can preserve the original value within the frontend pipeline, but only clamp it when it's going to be serialized in the serving layer, which we already have:

vllm/rust/src/server/src/routes/openai/utils/logprobs.rs

Lines 294 to 296 in b0898a4

pub fn clamp_logprob(logprob: f32) -> f32 {

logprob.max(-9999.0)

}

Keep rank 0 and non-finite engine logprobs intact, leaving sentinel conversion at the OpenAI response boundary. Cover sample and prompt logprobs in a shared output batch and verify the existing HTTP NaN/negative-infinity conversion. Co-authored-by: Codex Signed-off-by: Bugen Zhao <i@bugenzhao.com>

BugenZhao · 2026-09-16T02:41:05Z

/ci run

github-actions · 2026-09-16T02:41:22Z

✅ Triggered Buildkite CI #89228 for commit 3344086da2c0.

github-actions · 2026-09-16T02:48:37Z

✅ @almersawi, CI is now available for this PR.

/ci run starts upstream CI; /amd-ci run starts AMD CI only.
Your branch must contain every commit currently on its upstream target branch. Merge or rebase onto the latest target branch, then rerun the command. Append --allow-stale to a run command to test an outdated branch at your own risk.
/ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
/amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
/ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

vllm-project#51026) Co-authored-by: Bugen Zhao <i@bugenzhao.com> Signed-off-by: Islam <islam.almersawi@openinnovation.ai> Signed-off-by: Bugen Zhao <i@bugenzhao.com>

almersawi requested review from BugenZhao and njhill as code owners August 4, 2026 13:26

claude Bot reviewed Aug 4, 2026

View reviewed changes

mergify Bot added rust bug Something isn't working labels Aug 4, 2026

almersawi changed the title ~~[Bugfix][Rust Frontend] Tolerate NaN-corrupted logprobs in engine-cor…~~ [Bugfix][Rust Frontend] Tolerate NaN-corrupted logprobs in engine-core Aug 4, 2026

chelsealong mentioned this pull request Aug 8, 2026

[Bugfix][Rust Frontend] Return an error instead of panicking on empty-prompt logprobs #51474

Open

dvellon mentioned this pull request Aug 9, 2026

[Bug]: [ROCm] v0.26.0 release image: NaN logits from AITER fused-MoE path on gfx942 (Qwen3.5-397B-A17B-FP8) — fixed on main, requesting 0.26.x backport #51580

Open

1 task done

aoshen02 mentioned this pull request Aug 20, 2026

[RFC]: Logprobs/Logits Semantics and Determinism Across the vLLM Ecosystem #42259

Open

BugenZhao approved these changes Sep 16, 2026

View reviewed changes

BugenZhao force-pushed the fix/rust-frontend-nan-logprobs-rank-0 branch from 257301a to 3344086 Compare September 16, 2026 02:40

BugenZhao added the ready ONLY add when PR is ready to merge/full CI is needed label Sep 16, 2026

BugenZhao enabled auto-merge (squash) September 16, 2026 02:48

BugenZhao merged commit c16bb60 into vllm-project:main Sep 16, 2026
37 of 38 checks passed

Navigation Menu

Repository navigation

[Bugfix][Rust Frontend] Tolerate NaN-corrupted logprobs in engine-core - #51026

[Bugfix][Rust Frontend] Tolerate NaN-corrupted logprobs in engine-core#51026
BugenZhao merged 2 commits into
vllm-project:mainfrom
almersawi:fix/rust-frontend-nan-logprobs-rank-0

almersawi commented Aug 4, 2026 •

edited by BugenZhao

Loading

Uh oh!

claude Bot left a comment

Uh oh!

BugenZhao commented Aug 4, 2026

Uh oh!

chatgpt-codex-connector Bot commented Aug 4, 2026

Uh oh!

almersawi commented Aug 12, 2026

Uh oh!

BugenZhao left a comment

Uh oh!

BugenZhao Sep 16, 2026 •

edited

Loading

Uh oh!

BugenZhao commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	pub fn clamp_logprob(logprob: f32) -> f32 {
	logprob.max(-9999.0)
	}

Conversation

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

Uh oh!

Purpose

Test Plan

Test Result

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Claude Code Review

Uh oh!

BugenZhao commented Aug 4, 2026

Uh oh!

chatgpt-codex-connector Bot commented Aug 4, 2026

Uh oh!

almersawi commented Aug 12, 2026

Uh oh!

BugenZhao left a comment

Choose a reason for hiding this comment

Uh oh!

BugenZhao Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

BugenZhao commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

almersawi commented Aug 4, 2026 •

edited by BugenZhao

Loading

BugenZhao Sep 16, 2026 •

edited

Loading