[https://nvbugs/5615248][fix] Beam history copies only on terminal steps#13799

brb-nv · 2026-05-06T05:06:53Z

Description

This addresses https://nvbugspro.nvidia.com/bug/5615248. Two host-side optimizations for the beam-search decode path:

_prepare_beam_history - only copy beam history on terminal steps.
- The per-step D2H copies are skipped entirely on non-terminal steps via a small predictor entirely using information available on host.
- On potentially-terminal steps they run on a private side stream so they never block the main stream. The side-stream completion is recorded into SamplerEvent and awaited host-side in update_requests.
- A predictor-miss falls back to a synchronous .cpu() so correctness is preserved for any unforeseen finish pattern.
seq_slots / seq_lens H2D - cast once per step at the top of _process_requests and reused. Eliminates 2 redundant H2D launches/step on the typical concurrency=1 / single-strategy beam-search path. Multi-strategy keeps the existing per-group fall-back.

Workload: TinyLlama-1.1B-Chat-v1.0 on L40S, ISL=100 / OSL=20, beam_width=10, max_batch_size=1, piecewise CUDA graphs, 5 runs × 16 requests pooled per side (n = 80). Sign convention: negative Δ = improvement.

metric	base median (ms)	exp median (ms)	Δmedian (ms)
TTFT	10.895	10.803	−0.092
E2E	83.918	83.470	−0.448
ITL	0.367	0.365	−0.0016

Reproduction

A companion validation branch carries the bench harness used to produce the table above (kept off this MR to keep the diff minimal):

Branch: user/brb/nvbug-5615248-beam-host-overhead
Repro doc: nvbugs_5615248/trtllm_bench/REPRO_V6_V7.md

Test Coverage

$ pytest tests/unittest/_torch/sampler/test_beam_search_speculative_d2h.py -s -v
$ pytest tests/unittest/_torch/sampler/test_beam_search.py -s -v
$ pytest tests/unittest/_torch/sampler/test_beam_search_util.py -s -v

PR Checklist

Please review the following before submitting your PR:

PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

Summary by CodeRabbit

Refactor
- Improved beam search sampling efficiency with optimized asynchronous data transfer handling to reduce synchronization overhead during batch processing.

PR NVIDIA#13799 (commits 'Defer per-step beam-history D2H to terminal step' and 'Cast seq_slots/seq_lens to CUDA exactly once per step') introduced synchronous .cpu() calls inside _prepare_beam_history's _builder closure on terminal steps, which blocks the host thread and breaks the overlap scheduler's CPU-GPU pipelining. This commit restores overlap by: * Routing the deferred terminal-step D2H copies through a dedicated side stream so they no longer block the main stream. * Recording the side-stream completion via a SamplerEvent secondary event so the main stream is not gated on it. * Batching all per-request side-stream copies inside a single 'with torch.cuda.stream(side)' context to amortize Python overhead. * Adding a host-side predictor (_predict_beam_history_terminal) that skips D2H copies entirely on definitely-non-terminal steps based on py_max_new_tokens / max_seq_len / py_stop_words / lagged first_finish_reasons. The predictor preserves correctness via a fallback synchronous .cpu() path on predictor misses. Robust pooled per-request results (5 runs x 16 requests, Llama-3.1 8B, beam_width=4, ISL 100/OSL 20): * Median E2E: -0.32 to -0.47 ms (-0.4 to -0.55 percent) across two independent benchmark sessions. * Median ITL: -0.0015 to -0.0017 ms (~ -0.4 percent). * Mann-Whitney U two-sided p < 1e-11 in both runs. * TTFT: no significant effect (correctly so; predictor only affects the decode loop's per-step terminal check). Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

brb-nv · 2026-05-09T02:07:00Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-09T02:12:57Z

PR_Github #47463 [ run ] triggered by Bot. Commit: be062da Link to invocation

tensorrt-cicd · 2026-05-09T07:10:07Z

PR_Github #47463 [ run ] completed with state FAILURE. Commit: be062da
/LLM/main/L0_MergeRequest_PR pipeline #37384 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

brb-nv · 2026-05-09T20:21:47Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-09T20:27:33Z

PR_Github #47538 [ run ] triggered by Bot. Commit: be062da Link to invocation

coderabbitai · 2026-05-09T20:28:53Z

📝 Walkthrough

Walkthrough

This PR optimizes beam-search history finalization by introducing side-stream-based asynchronous device-to-host transfers that avoid synchronization on non-terminal steps. The core insight is predicting terminal steps using a lagged snapshot of finish reasons, then selectively issuing non-blocking D2H copies only when needed. CUDA tensor casting is consolidated to reduce redundant transfers.

Changes

Side-Stream Async Beam-History D2H Optimization

Layer / File(s)	Summary
Public Interface Updates `tensorrt_llm/_torch/pyexecutor/sampler.py`	`SamplerEvent` adds optional `side_stream_event` field; `synchronize()` waits on it in addition to `cuda_event`. `_process_requests` return type expanded to include `seq_slots_cuda` and `seq_lens_cuda`.
Async Side-Stream Infrastructure `tensorrt_llm/_torch/pyexecutor/sampler.py`	`AsyncWorkerMixin` introduces `_d2h_side_stream` with lazy initialization, pending-usage tracking fields, and `_side_stream_d2h_batch()` context manager to amortize stream-context overhead. `_async_d2h_via_side_stream()` performs non-blocking copies; `_record_sampler_event()` captures `side_stream_event`. Import adjusted to include `Iterator`.
Terminal-Step Prediction `tensorrt_llm/_torch/pyexecutor/sampler.py`	`TorchSampler` adds `_prev_first_finish_reasons_host` host-side cache per `seq_slot`. New `_predict_beam_history_terminal()` helper decides whether side-stream D2H should be eagerly issued based on beam-search config and lagged finish-reason snapshot.
Beam-History Preparation with Conditional D2H `tensorrt_llm/_torch/pyexecutor/sampler.py`	`_prepare_beam_history` refactored to always compute cheap device tensor slices; only when predicted terminal does it issue side-stream D2H copies, otherwise defers synchronous `.cpu()` to builder fallback. `_prepare_beam_histories` wraps loop in `_side_stream_d2h_batch()` for batched overhead. `update_requests` snapshots current `first_finish_reasons` into `_prev_first_finish_reasons_host` for next step's prediction.
Metadata Building with Cached Tensors `tensorrt_llm/_torch/pyexecutor/sampler.py`	`_add_metadata_to_grouped_requests` signature extended to accept optional `seq_slots_cuda` and `seq_lens_cuda`. For beam-search metadata, reuses cached CUDA tensors when group covers full batch; otherwise performs subset device transfers.
Sampling Pipeline Integration `tensorrt_llm/_torch/pyexecutor/sampler.py`	`_process_requests` performs CUDA casting once (`seq_slots` to int64, `seq_lens` to device dtype) and returns both alongside host outputs. `sample_async`, `_sample_batched_by_strategy`, and `_add_metadata_to_grouped_requests` updated to accept and pass through cached tensors. Greedy fast-path early return includes CUDA tensor outputs.

🎯 4 (Complex) | ⏱️ ~60 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 35.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically summarizes the main optimization: deferring beam history copies to only terminal steps, which is the primary focus of the changeset.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The PR description clearly explains what is being addressed, why the changes are needed, includes a detailed description of the two optimizations, provides microbenchmark results, references test coverage and reproduction steps, and includes the PR checklist.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

Tip

💬 Introducing Slack Agent: The best way for teams to turn conversations into code.

Slack Agent is built on CodeRabbit's deep understanding of your code, so your team can collaborate across the entire SDLC without losing context.

Generate code and open pull requests
Plan features and break down work
Investigate incidents and troubleshoot customer tickets together
Automate recurring tasks and respond to alerts with triggers
Summarize progress and report instantly

Built for teams:

Shared memory across your entire org—no repeating context
Per-thread sandboxes to safely plan and execute work
Governance built-in—scoped access, auditability, and budget controls

One agent for your entire SDLC. Right inside Slack.

👉 Get started

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

🧹 Nitpick comments (2)

tensorrt_llm/_torch/pyexecutor/sampler.py (2)
3514-3518: ⚡ Quick win

Reset the lagged finish-reason cache on slot reuse.

_predict_beam_history_terminal() keys this state by seq_slot, but the snapshot written here is never cleared when a slot is reassigned. A recycled slot can inherit a finished beam from the previous request and force eager beam-history D2H until the new request reaches its first update_requests(). Clearing the entry during new-request setup, or when the request completes, keeps the predictor aligned with the current occupant of the slot.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tensorrt_llm/_torch/pyexecutor/sampler.py` around lines 3514 - 3518, The
snapshot cache _prev_first_finish_reasons_host is never cleared when a seq_slot
is reused, so update the new-request setup (the code path that assigns
req.py_seq_slot and/or the request completion path) to reset the entry for that
slot (indexed by req.py_seq_slot) before the new request begins or when a
request finishes; this keeps _predict_beam_history_terminal keyed by seq_slot
aligned with the current occupant and prevents inheriting stale finish
reasons—clear the entry (e.g., set to None/zeros/empty array) in the code that
assigns req.py_seq_slot or in the request teardown/complete routine and ensure
update_requests() semantics remain correct.
3428-3429: ⚡ Quick win

Keep the predictor enabled for single-token stop words.

This branch disables the optimization for every request with stop words, even though _prepare_beam_history() only forces unconditional finalization for multi-token stop words via _check_stop_words_length(). Using that narrower check here preserves correctness via the .cpu() miss fallback and avoids per-step eager copies for the single-token case.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tensorrt_llm/_torch/pyexecutor/sampler.py` around lines 3428 - 3429, The
current branch unconditionally disables the predictor whenever
request.py_stop_words_list is present, which is too broad; change the condition
so the predictor is only disabled for multi-token stop words (use the same logic
as _prepare_beam_history() via
_check_stop_words_length(request.py_stop_words_list) or equivalent) so
single-token stop words still use the predictor and rely on the .cpu() miss
fallback instead of forcing eager copies.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@tensorrt_llm/_torch/pyexecutor/sampler.py`:
- Around line 3514-3518: The snapshot cache _prev_first_finish_reasons_host is
never cleared when a seq_slot is reused, so update the new-request setup (the
code path that assigns req.py_seq_slot and/or the request completion path) to
reset the entry for that slot (indexed by req.py_seq_slot) before the new
request begins or when a request finishes; this keeps
_predict_beam_history_terminal keyed by seq_slot aligned with the current
occupant and prevents inheriting stale finish reasons—clear the entry (e.g., set
to None/zeros/empty array) in the code that assigns req.py_seq_slot or in the
request teardown/complete routine and ensure update_requests() semantics remain
correct.
- Around line 3428-3429: The current branch unconditionally disables the
predictor whenever request.py_stop_words_list is present, which is too broad;
change the condition so the predictor is only disabled for multi-token stop
words (use the same logic as _prepare_beam_history() via
_check_stop_words_length(request.py_stop_words_list) or equivalent) so
single-token stop words still use the predictor and rely on the .cpu() miss
fallback instead of forcing eager copies.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: a2bdc6b2-42a6-4935-9a98-7c6b54518700

📥 Commits

Reviewing files that changed from the base of the PR and between 43f4b94 and be062da.

📒 Files selected for processing (1)

tensorrt_llm/_torch/pyexecutor/sampler.py

tensorrt-cicd · 2026-05-09T21:38:03Z

PR_Github #47538 [ run ] completed with state FAILURE. Commit: be062da
/LLM/main/L0_MergeRequest_PR pipeline #37455 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

brb-nv · 2026-05-10T00:20:31Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-10T00:26:34Z

PR_Github #47542 [ run ] triggered by Bot. Commit: b85c8cf Link to invocation

tensorrt-cicd · 2026-05-10T05:22:30Z

PR_Github #47542 [ run ] completed with state SUCCESS. Commit: b85c8cf
/LLM/main/L0_MergeRequest_PR pipeline #37458 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

brb-nv · 2026-05-10T19:59:26Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-10T20:04:56Z

PR_Github #47619 [ run ] triggered by Bot. Commit: b85c8cf Link to invocation

tensorrt-cicd · 2026-05-10T21:55:14Z

PR_Github #47619 [ run ] completed with state SUCCESS. Commit: b85c8cf
/LLM/main/L0_MergeRequest_PR pipeline #37524 completed with status: 'SUCCESS'

CI Report

Link to invocation

ixlmar

The changes make sense, but we need to figure out how to keep the overall sampler code maintainable and how to not regress any existing features.

Since the side-stream D2H copies are supposed to overlap with compute, do we even need the speculative D2H-optout with fallback to .cpu() and resulting potential syncs that need to be covered in unit tests and may alter performance characteristics unexpectedly (for big enough models the risk of breaking overlap scheduling by .cpu() may outweigh any perf benefit)?

Generally, this is a form of speculative execution with the usual performance implications. Although I agree that the .cpu() branch would appear to be hit only rarely, this might be unexpected to users running benchmarks etc. and thus perhaps it would be better to gate this optimization by some configuration option (having "spec" in its name)?

PR NVIDIA#13799 (commits 'Defer per-step beam-history D2H to terminal step' and 'Cast seq_slots/seq_lens to CUDA exactly once per step') introduced synchronous .cpu() calls inside _prepare_beam_history's _builder closure on terminal steps, which blocks the host thread and breaks the overlap scheduler's CPU-GPU pipelining. This commit restores overlap by: * Routing the deferred terminal-step D2H copies through a dedicated side stream so they no longer block the main stream. * Recording the side-stream completion via a SamplerEvent secondary event so the main stream is not gated on it. * Batching all per-request side-stream copies inside a single 'with torch.cuda.stream(side)' context to amortize Python overhead. * Adding a host-side predictor (_predict_beam_history_terminal) that skips D2H copies entirely on definitely-non-terminal steps based on py_max_new_tokens / max_seq_len / py_stop_words / lagged first_finish_reasons. The predictor preserves correctness via a fallback synchronous .cpu() path on predictor misses. Robust pooled per-request results (5 runs x 16 requests, Llama-3.1 8B, beam_width=4, ISL 100/OSL 20): * Median E2E: -0.32 to -0.47 ms (-0.4 to -0.55 percent) across two independent benchmark sessions. * Median ITL: -0.0015 to -0.0017 ms (~ -0.4 percent). * Mann-Whitney U two-sided p < 1e-11 in both runs. * TTFT: no significant effect (correctly so; predictor only affects the decode loop's per-step terminal check). Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

brb-nv · 2026-05-15T20:07:54Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-15T20:14:10Z

PR_Github #48634 [ run ] triggered by Bot. Commit: 1216a56 Link to invocation

ixlmar

Had a superficial look at the tests as well (looking good)

brb-nv · 2026-05-16T02:11:19Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-16T02:16:40Z

PR_Github #48664 [ run ] triggered by Bot. Commit: ebdd632 Link to invocation

tensorrt-cicd · 2026-05-16T02:21:41Z

PR_Github #48634 [ run ] completed with state ABORTED. Commit: 1216a56

Link to invocation

tensorrt-cicd · 2026-05-16T12:55:14Z

PR_Github #48664 [ run ] completed with state SUCCESS. Commit: ebdd632
/LLM/main/L0_MergeRequest_PR pipeline #38444 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

brb-nv · 2026-05-16T18:04:32Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-16T18:09:49Z

PR_Github #48705 [ run ] triggered by Bot. Commit: cf4d7cd Link to invocation

brb-nv · 2026-05-16T18:33:53Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-16T18:39:09Z

PR_Github #48706 [ run ] triggered by Bot. Commit: cf4d7cd Link to invocation

tensorrt-cicd · 2026-05-16T18:42:35Z

PR_Github #48705 [ run ] completed with state ABORTED. Commit: cf4d7cd

Link to invocation

tensorrt-cicd · 2026-05-17T16:28:03Z

PR_Github #48706 [ run ] completed with state SUCCESS. Commit: cf4d7cd
/LLM/main/L0_MergeRequest_PR pipeline #38476 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

brb-nv · 2026-05-17T20:11:55Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-17T20:18:06Z

PR_Github #48784 [ run ] triggered by Bot. Commit: cf4d7cd Link to invocation

tensorrt-cicd · 2026-05-18T02:47:06Z

PR_Github #48784 [ run ] completed with state SUCCESS. Commit: cf4d7cd
/LLM/main/L0_MergeRequest_PR pipeline #38548 completed with status: 'SUCCESS'

CI Report

Link to invocation

…eps (NVIDIA#13799) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

github-actions Bot assigned brb-nv May 6, 2026

ixlmar reviewed May 8, 2026

View reviewed changes

Comment thread tensorrt_llm/_torch/pyexecutor/sampler.py Outdated

brb-nv force-pushed the user/brb/nvbug-5615248-beam-host-overhead-pr branch from 09f8feb to 8999b4b Compare May 9, 2026 00:47

brb-nv marked this pull request as ready for review May 9, 2026 20:23

brb-nv requested a review from a team as a code owner May 9, 2026 20:23

brb-nv requested a review from leslie-fang25 May 9, 2026 20:23

brb-nv changed the title ~~[https://nvbugs/5615248][fix] Reduce per-step beam-search host overhead~~ [https://nvbugs/5615248][fix] Beam history copies only on terminal steps May 9, 2026

coderabbitai Bot reviewed May 9, 2026

View reviewed changes

brb-nv requested a review from a team as a code owner May 10, 2026 00:20

brb-nv requested a review from ziyixiong-nv May 10, 2026 00:20

ixlmar reviewed May 11, 2026

View reviewed changes

ixlmar requested a review from dhansen-nvidia May 11, 2026 14:45

ixlmar reviewed May 11, 2026

View reviewed changes

Comment thread tensorrt_llm/_torch/pyexecutor/sampler.py Outdated

brb-nv force-pushed the user/brb/nvbug-5615248-beam-host-overhead-pr branch from b85c8cf to 0b44394 Compare May 14, 2026 02:47

brb-nv requested review from a team as code owners May 15, 2026 19:45

brb-nv requested review from schetlur-nv and suyoggupta May 15, 2026 19:45

address refactor comments - 1

1216a56

Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

brb-nv force-pushed the user/brb/nvbug-5615248-beam-host-overhead-pr branch from 74da52a to 1216a56 Compare May 15, 2026 19:49

brb-nv added the api-compatible Accepted LLM API contract change that is backwards-compatible label May 15, 2026

brb-nv requested a review from ixlmar May 15, 2026 19:57

ixlmar reviewed May 15, 2026

View reviewed changes

address refactor comments - 2

cf4d7cd

Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

brb-nv force-pushed the user/brb/nvbug-5615248-beam-host-overhead-pr branch from ebdd632 to cf4d7cd Compare May 16, 2026 18:04

brb-nv enabled auto-merge (squash) May 17, 2026 22:14

juney-nvidia approved these changes May 18, 2026

View reviewed changes

brb-nv merged commit 371c126 into NVIDIA:main May 18, 2026
8 checks passed

KleinBlueC pushed a commit to KleinBlueC/TensorRT-LLM that referenced this pull request May 19, 2026

[https://nvbugs/5615248][fix] Beam history copies only on terminal st…

0f013f1

…eps (NVIDIA#13799) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

Uh oh!

Conversation

brb-nv commented May 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Reproduction

Test Coverage

PR Checklist

GitHub Bot Help

Summary by CodeRabbit

Uh oh!

Uh oh!

brb-nv commented May 9, 2026

Uh oh!

tensorrt-cicd commented May 9, 2026

Uh oh!

tensorrt-cicd commented May 9, 2026

Uh oh!

brb-nv commented May 9, 2026

Uh oh!

tensorrt-cicd commented May 9, 2026

Uh oh!

coderabbitai Bot commented May 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

tensorrt-cicd commented May 9, 2026

Uh oh!

brb-nv commented May 10, 2026

Uh oh!

tensorrt-cicd commented May 10, 2026

Uh oh!

tensorrt-cicd commented May 10, 2026

Uh oh!

brb-nv commented May 10, 2026

Uh oh!

tensorrt-cicd commented May 10, 2026

Uh oh!

tensorrt-cicd commented May 10, 2026

Uh oh!

ixlmar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

brb-nv commented May 15, 2026

Uh oh!

tensorrt-cicd commented May 15, 2026

Uh oh!

ixlmar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

brb-nv commented May 16, 2026

Uh oh!

tensorrt-cicd commented May 16, 2026

Uh oh!

brb-nv commented May 6, 2026 •

edited

Loading

coderabbitai Bot commented May 9, 2026 •

edited

Loading