Fix disaggregated prefill/decode RayService for H200 GPU clusters#28
Open
iankouls-aws wants to merge 3 commits intomainfrom
Open
Fix disaggregated prefill/decode RayService for H200 GPU clusters#28iankouls-aws wants to merge 3 commits intomainfrom
iankouls-aws wants to merge 3 commits intomainfrom
Conversation
Adds a new RayService example demonstrating LLM serving with prefill/decode disaggregation using Ray Serve LLM APIs and vLLM's NIXLConnector for KV cache transfer. Files: - disaggregated_prefill_decode.py: Python deployment script - rayservice.disaggregated_prefill_decode.yaml: KubeRay manifest - disaggregated_prefill_decode_req.py: Test client (chat + streaming) - README.md: Documentation with architecture, config, and usage Also updates rayservice-create.sh to list the new example.
…ead GPUs - Switch model from gated meta-llama/Llama-3.1-8B-Instruct to open Qwen/Qwen2.5-7B-Instruct (no HF token required) - Update accelerator_type from A10G to H200 to match actual cluster hardware - Set num-gpus: 0 on head node to prevent Serve replicas from being scheduled on the head and hitting GPU memory contention
The :latest tag does not exist on rayproject/ray-llm. Use the explicit :latest-py311-cu128 tag instead.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Fixes the disaggregated prefill/decode RayService manifest so it works correctly on H200 GPU clusters (p5en instances).
Changes
Model
meta-llama/Llama-3.1-8B-Instruct(gated, requires HF token) toQwen/Qwen2.5-7B-Instruct(open, no auth required)disaggregated_prefill_decode_req.py) to match the new model IDAccelerator
accelerator_typefromA10GtoH200for both prefill and decode configs to match actual cluster hardwareHead node GPU isolation
num-gpus: '0'toheadGroupSpec.rayStartParamsto prevent Ray from scheduling GPU workloads on the head node, which caused OOM failures when Serve replicas landed thereContainer image
:latest(does not exist) to:latest-py311-cu128for all three containers (head, prefill worker, decode worker)Testing
Deployed and verified on EKS cluster (
shared-eks-cluster-cgk, ap-southeast-3) with p5en.48xlarge nodes. All pods reach Ready state and Serve deployments (Prefill, Decode, PDProxyServer, OpenAiIngress) initialize successfully.