-
Notifications
You must be signed in to change notification settings - Fork 434
Select PD workers in same roleset #1409
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Summary of Changes
Hello @varungup90, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request refines the pod selection logic within the pdRouter
to ensure that prefill (P) and decode (D) workers are selected from the same roleset. This change is crucial for maintaining consistency and potentially improving performance or resource utilization in disaggregated P/D architectures by ensuring related tasks are handled by co-located or logically grouped resources.
Highlights
- Introduce RoleSet Constant: A new constant
RoleSet
("roleset-name") has been added topkg/plugins/gateway/algorithms/pd_disaggregation.go
to identify pods belonging to a specific roleset, which is crucial for the new selection logic. - Refined Decode Pod Selection: The
selectDecodePod
function now takes theprefillPod
as an argument and filters the available decode pods to only consider those that share the sameRoleSet
label as theprefillPod
. This ensures that prefill and decode operations occur within the same logical grouping. - Improved Prefill Request Handling: The
doPrefillRequest
function has been updated to return the selectedprefillPod
along with any error. This allows the callingRoute
function to use this specific prefill pod's roleset for subsequent decode pod selection, ensuring consistency. - Enhanced Logging: Added
klog.InfoS
statements to log the names of the selected prefill and decode pods, as well as the IP of the prefill pod upon request completion. This enhances operational visibility and aids in debugging the P/D disaggregation routing. - Error Handling for Decode Pod: A new check has been added in the
Route
method to explicitly handle cases where no suitable decode pod is found after filtering by roleset, returning an error to prevent unexpected behavior.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command>
or @gemini-code-assist <command>
. Below is a summary of the supported commands.
Feature | Command | Description |
---|---|---|
Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/
folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Code Review
This pull request introduces a mechanism to select decode pods from the same RoleSet
as the chosen prefill pod, which is a valuable feature for ensuring affinity between prefill and decode workers. The implementation correctly refactors doPrefillRequest
to return the selected prefill pod and updates selectDecodePod
to filter based on the new RoleSet
label.
However, there are a couple of key areas for improvement. Critically, the new logic is not covered by unit tests, and existing tests are likely to fail due to the new RoleSet
label requirement. Additionally, the selectDecodePod
function could be refactored for better clarity, robustness, and traceability. My review includes a critical comment about the need for updated tests and a medium-severity suggestion to refactor selectDecodePod
.
@@ -90,12 +91,18 @@ func (r pdRouter) Route(ctx *types.RoutingContext, readyPodList types.PodList) ( | |||
return "", err | |||
} | |||
|
|||
if err = r.doPrefillRequest(ctx, prefillPods, getLLMEngine(prefillPods[0], LLMEngineIdentifier, VLLMEngine)); err != nil { | |||
prefillPod, err := r.doPrefillRequest(ctx, prefillPods, getLLMEngine(prefillPods[0], LLMEngineIdentifier, VLLMEngine)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This works but may not be optimal.
-
It selects the prefill role without considering the load on the decode nodes. It's possible that the decode nodes in the selected RoleSet are already heavily loaded.
-
It filters decode nodes on every request, which is slightly inefficient—though this is a minor issue given the current scale.
For the first issue, I believe we should improve it. We can either merge this PR and address it in a follow-up, or update this one directly. WDYT?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
- Yes, scoring will change as it needs to take prefill and decode pod both in consideration at the same time. I will need to think off a design, so will be in follow up PR.
- Filtering is in-memory so not expensive, it is ok for this PR. In follow up PR, as prefill and decode pods both are scorered together, at that time this part can be addressed to pre-built cache index on roleset name.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sounds good. Let's merge this one first.
70c9068
to
bbd67e8
Compare
* Select PD workers in same roleset * nit * update ut --------- Signed-off-by: Varun Gupta <varungup90@gmail.com>
…se-0.4 branch (#1468) * Select PD workers in same roleset (#1409) * Select PD workers in same roleset * nit * update ut --------- Signed-off-by: Varun Gupta <varungup90@gmail.com> * [Bug] fix webhook config output when using make manifests (#1412) fix webhook config output when using make manifests Signed-off-by: googs1025 <googs1025@gmail.com> * [Fix] Fix vLLM NIXL-based P/D samples (#1425) Signed-off-by: Haiyang Shi <haiyang.shi@bytedance.com> Co-authored-by: Haiyang Shi <haiyang.shi@bytedance.com> * [Fix] Disable GGA in NIXL samples (#1436) [Fix] Fix NIXL samples Explicitly set UCX_TLS to let UCX not use GGA (GPU Direct) transport Signed-off-by: Haiyang Shi <haiyang.shi@bytedance.com> Co-authored-by: Haiyang Shi <haiyang.shi@bytedance.com> * Fix P/D disaggregation router to follow Nixl kv_transfer_params (#1429) - Add kv_transfer_params configuration to prefill requests and decode requests Signed-off-by: Jiaxin Shan <seedjeffwan@gmail.com> * [Bug] Corrected naming convention for AIBRIX_MODEL_GPU_PROFILE_CACHING_FLAG (#1427) Corrected naming convention for AIBRIX_MODEL_GPU_PROFILE_CACHING_FLAG Signed-off-by: Jonathon Shea <sheajonathon0@gmail.com> * [Bug] stormservice's headless service not set ownerRef (#1442) * fix: stormservice's headless service not set ownerRef Signed-off-by: dajun.cui <dajun.cui@bytedance.com> * fix: patch ut test for service sync Signed-off-by: dajun.cui <dajun.cui@bytedance.com> --------- Signed-off-by: dajun.cui <dajun.cui@bytedance.com> * [Bug] stormservice's headless service need set PublishNotReadyAddresses (#1441) * fix: stormservice's headless service need set PublishNotReadyAddresses Signed-off-by: dajun.cui <dajun.cui@bytedance.com> * fix: isServiceEqual check PublishNotReadyAddresses Signed-off-by: dajun.cui <dajun.cui@bytedance.com> --------- Signed-off-by: dajun.cui <dajun.cui@bytedance.com> --------- Signed-off-by: Varun Gupta <varungup90@gmail.com> Signed-off-by: googs1025 <googs1025@gmail.com> Signed-off-by: Haiyang Shi <haiyang.shi@bytedance.com> Signed-off-by: Jiaxin Shan <seedjeffwan@gmail.com> Signed-off-by: Jonathon Shea <sheajonathon0@gmail.com> Signed-off-by: dajun.cui <dajun.cui@bytedance.com> Co-authored-by: Varun Gupta <varungup90@gmail.com> Co-authored-by: CYJiang <86391540+googs1025@users.noreply.github.com> Co-authored-by: Haiyang Shi <dwyane.shi@gmail.com> Co-authored-by: Haiyang Shi <haiyang.shi@bytedance.com> Co-authored-by: Jonathon Shea <35823125+JonathonShea@users.noreply.github.com> Co-authored-by: cuidajun <dajun.cui@bytedance.com>
Pull Request Description
[Please provide a clear and concise description of your changes here]
Related Issues
Resolves: #[Insert issue number(s)]
Important: Before submitting, please complete the description above and review the checklist below.
Contribution Guidelines (Expand for Details)
We appreciate your contribution to aibrix! To ensure a smooth review process and maintain high code quality, please adhere to the following guidelines:
Pull Request Title Format
Your PR title should start with one of these prefixes to indicate the nature of the change:
[Bug]
: Corrections to existing functionality[CI]
: Changes to build process or CI pipeline[Docs]
: Updates or additions to documentation[API]
: Modifications to aibrix's API or interface[CLI]
: Changes or additions to the Command Line Interface[Misc]
: For changes not covered above (use sparingly)Note: For changes spanning multiple categories, use multiple prefixes in order of importance.
Submission Checklist
By submitting this PR, you confirm that you've read these guidelines and your changes align with the project's contribution standards.