Repository navigation
[raycicmd] add gpu-large-flex runner queue mapping - #509
Merged
Merged
Conversation
Maps instance_type gpu-large-flex to gpu_large_flex_runner_queue_{branch,pr}.
That queue (ray-project/reef#286) is a 4-GPU queue that accepts several
instance types (g6.12xlarge, g5.12xlarge, g6.24xlarge, g6e.12xlarge), so
jobs still get machines when g6.12xlarge is out of capacity in every AZ.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Signed-off-by: Ray CI Test <rayci@ray.io>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request adds a new runner queue configuration for gpu-large-flex in both the branch pipeline configuration (gpu_large_flex_runner_queue_branch) and the pull request pipeline configuration (gpu_large_flex_runner_queue_pr) within raycicmd/builtin_configs.go. There are no review comments, and I have no feedback to provide.
4 tasks done
elliot-barn
marked this pull request as ready for review
October 18, 2026 02:12
elliot-barn
added a commit
to ray-project/ray
that referenced
this pull request
Oct 19, 2026
0.50.0 adds the gpu-large-flex instance_type mapping (ray-project/rayci#509), which the llm gpu steps in this PR need. It also picks up the 0.49.0 changes. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: elliot-barn <elliot.barnwell@anyscale.com>
elliot-barn
added a commit
to ray-project/ray
that referenced
this pull request
Oct 20, 2026
## Why are these changes needed? `llm gpu tests` and `llm custom vllm plugin gpu test` run on `g6-large`, which allows only `g6.12xlarge`. On 2026-10-15 the premerge queue (`g6_large_runner_queue_pr`) had 0 of 30 desired instances. Between 19:22 and 21:07 UTC it logged 99 `InsufficientInstanceCapacity` launch failures, spread evenly over all 4 us-west-2 AZs. The queue was already multi-AZ, so the problem was the single instance type. This PR moves both steps to the new `gpu-large-flex` queue. It's 4 GPUs with 24GB+ each, and the types are tried in order: | Priority | Type | GPUs | |---|---|---| | 1 | g6.12xlarge | 4× L4 (same as today) | | 2 | g5.12xlarge | 4× A10G | | 3 | g6.24xlarge | 4× L4, more CPU/RAM | | 4 | g5.24xlarge | 4× A10G, more CPU/RAM | | 5 | g6e.12xlarge | 4× L40S | | 6 | g6e.24xlarge | 4× L40S, more CPU/RAM | `llm gpu tests (4 GPUs)` stays on `gpu-large` (g4dn.12xlarge) for now. `omni_audio_example.py` and `vlm_video_example.py` (run by `llm gpu tests`) asked for `accelerator_type="L4"` in the config they actually run. That would hang on A10G/L40S hosts, so this PR removes it from `create_omni_audio_config()` and `create_vlm_video_config()`. The doc snippets still show the L4 example. ## Depends on - [x] ray-project/reef#286 (+ #287): creates `gpu_large_flex_runner_queue_{branch,pr}` with 6 instance types. Applied and merged. - [x] ray-project/rayci#509: adds the `gpu-large-flex` instance_type mapping. Released in rayci v0.50.0, and `.rayciversion` is bumped to 0.50.0 in this PR (this also picks up the v0.49.0 changes). ## Test plan - [x] Smoke test on the new queues (reef-test-pipeline build 3): pr got a g5.12xlarge and branch got a g5.24xlarge. Both had 4× A10G and `nvidia-smi` passed. - [x] Both steps pass on premerge ([build 75126](https://buildkite.com/ray-project/premerge/builds/75126)): | Step | Instance | Wait | Run | g6-large, last 14 days (p50 wait / p50 run) | |---|---|---|---|---| | `llm gpu tests` | g5.24xlarge (4× A10G) | 1.8 min | 112.5 min | 675 min / 98.5 min | | `llm custom vllm plugin gpu test` | g5.24xlarge (4× A10G) | 1.7 min | 17.3 min | 717 min / 15.6 min | On A10G the tests run about 11–14% longer. The median wait drops from about 11 hours to under 2 minutes. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Signed-off-by: elliot-barn <elliot.barnwell@anyscale.com> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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.
Summary
Adds
instance_type: gpu-large-flexto the branch and PR built-in configs:gpu-large-flexgpu_large_flex_runner_queue_branchgpu_large_flex_runner_queue_prThe queue is created in ray-project/reef#286. It's a 4-GPU queue that accepts
g6.12xlarge,g5.12xlarge,g6.24xlarge,g6e.12xlarge(On-Demand, tried in that order). The goal is that llm gpu tests still get machines wheng6.12xlargeruns out. On 2026-10-15,g6_large_runner_queue_prhad 0 of 30 desired instances, and launches were failing with InsufficientInstanceCapacity in all 4 us-west-2 AZs.The rest of the PR-config diff is just gofmt re-aligning the map.
Rollout
.rayciversionand move the llm gpu tests toinstance_type: gpu-large-flex.Merging this PR before step 1 is harmless. A step that uses
gpu-large-flexwould just sit waiting for agents until the queue exists.Test plan
go build ./...,go test ./raycicmd/🤖 Generated with Claude Code