[Draft] Dynamic Load Balancing via Request Interruption and Migration for vLLM Engines #134
Conversation
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
…ancing
- Scheduler now sends target_leftover_cnt instead of specific request IDs
- Worker selects requests to interrupt based on overhead:
* First interrupts all unscheduled (waiting) requests
* Then selects running requests with shortest total sequence length
- Add assertions to validate request existence before interruption
- Calculate interruption count as half of load imbalance to avoid over-correction
- Add 3-second freeze period after interruption to allow load redistribution
This minimizes wasted computation by prioritizing requests that haven't started
and selecting running requests based on actual progress rather than arbitrary order.
- Implement abort_to_target_requests_cnt for v1 engine - Support both v0/v1 engines in VllmStrategy - Prioritize interruption: waiting → (swapped+running by length) - Fix swapped_count missing in total request calculation - Return interrupted request IDs for proper tracking Enables efficient request migration by minimizing computational loss.
# Conflicts: # roll/distributed/scheduler/generate_scheduler.py # roll/distributed/strategy/vllm_strategy.py # roll/pipeline/rlvr/rlvr_pipeline.py # roll/third_party/vllm/vllm_0_8_4/llm.py
…balance sensitive threshold to be 1
|
|
taoluo
changed the title
taoluo
changed the title
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Dynamic Load Balancing via Request Interruption for vLLM Engines
Summary
This PR implements a sophisticated dynamic load balancing system for distributed vLLM engines by introducin request interruption. The system automatically redistributes workload across engines to optimize throughput and reduce latency.
Key Features
Request Interruption
target_leftover_cntallowing workers to select which requests to interrupt based on migration overheadDynamic Load Balancing
Implementation Details
generate_scheduler.pyto track load metrics and trigger rebalancingvllm_strategy.pywith request interruption capabilities