How do you retain reinforcement studying for big reasoning fashions from stalling on a couple of very lengthy, very sluggish rollouts whereas GPUs sit below used? a workforce of researchers from Moonshot AI and Tsinghua College introduce ‘Seer’, a brand new on-line context studying system that targets a particular methods bottleneck in reinforcement studying for big language fashions. In synchronous on coverage setups, the rollout section dominates the price of every iteration. Seer restructures this section and experiences rollout throughput positive factors of 74 p.c to 97 p.c and tail latency reductions of 75 p.c to 93 p.c in contrast with a robust synchronous baseline known as veRL.
Why synchronous rollout is sluggish for reasoning fashions?
Fashionable reasoning RL workloads use lengthy chain of thought fashion outputs. Within the Seer experiments, the researchers apply GRPO to 3 completely different fashions, Moonlight, Qwen2 VL 72B and Kimi K2. These workloads run on 32 compute nodes with 8 H800 GPUs per node. The three duties use 32, 128 and 256 GPUs respectively, with 400, 600 and 800 prompts per iteration and eight or 16 responses per immediate.
Most era size is massive. Moonlight is configured for 65,536 tokens, Qwen2 VL 72B for 40,960 tokens and Kimi K2 for 98,304 tokens. A single lengthy chain of thought request can develop from a couple of hundred megabytes of KVCache to tens of gigabytes as decoding progresses. This reminiscence development forces situations to scale back concurrency or to preempt requests, which triggers costly re decoding.
The analysis workforce defines tail requests because the final 10 p.c of requests to complete in a rollout. For Moonlight and Qwen2 VL 72B, this tail alone can eat as much as 50 p.c of the full rollout time within the baseline system. Rollout already dominates iteration time, so this tail impact immediately slows RL.
Seer structure on prime of Mooncake and vLLM
Seer retains the RL algorithm equivalent to synchronous veRL. Every coaching iteration makes use of solely information from the present rollout iteration, so the system preserves on coverage conduct. The coaching section makes use of Megatron for distributed optimization. The rollout section makes use of an in home implementation of vLLM because the inference engine.
To help aggressive request scheduling, Seer depends on a International KVCache Pool constructed on the Mooncake disaggregated KVCache structure utilized in manufacturing for Kimi. Mooncake offers a two tier DRAM and SSD KV cache retailer shared throughout inference nodes, which permits Seer emigrate requests with out recomputing prefills.
On prime of this substrate, Seer introduces three key mechanisms:
- Divided Rollout
- Context Conscious Scheduling
- Adaptive Grouped Speculative Decoding
These are orchestrated by a Request Buffer, a Context Supervisor and an Inference Engine Pool related to the International KVCache Pool.
Divided Rollout, nice grained scheduling and migration
Typical synchronous rollout assigns entire GRPO teams to inference situations. A gaggle is a set of requests that share one immediate. As soon as assigned, a bunch stays on the identical occasion till all responses end. As a result of massive variance in output lengths, this results in load imbalance and lengthy working stragglers.
Seer breaks teams down in two steps. It first decomposes every group into particular person requests. It then divides every request into a number of chunks based mostly on era size. When the scheduler dispatches a request from the Request Buffer, it units a small max tokens worth similar to 8,000 tokens for that chunk. After every chunk, the request is re enqueued till it reaches an finish of sequence token or its authentic max tokens restrict.
As a result of KVCache is saved within the International KVCache Pool, divided requests can transfer between situations at chunk boundaries with out re working the prefill. The scheduler maintains a concurrency stage that retains reminiscence utilization excessive whereas avoiding preemption. This reduces waste and smooths KVCache utilization throughout the iteration.
Context Conscious Scheduling utilizing group size statistics
The analysis workforce observe that completely different requests in the identical group are inclined to have correlated output lengths. Seer makes use of this construction as on-line context. For every immediate group, it designates one request because the speculative request. The scheduler retains speculative requests in a excessive precedence queue and serves them with a smallest first coverage based mostly on generated tokens to this point. Quick requests full shortly and exit. Lengthy requests stay and establish teams which can be potential tail candidates.
The Context Supervisor maintains a size estimate for every group. It updates this estimate to the utmost generated size amongst accomplished requests within the group. If no request has completed, it makes use of the unique max tokens as a conservative sure. As soon as speculative requests are in flight or accomplished, Seer schedules remaining requests with an approximate longest first coverage at group stage. This design achieves throughput and tail conduct near an oracle scheduler that is aware of all output lengths prematurely.
Adaptive Grouped Speculative Decoding
Seer provides Adaptive Grouped Speculative Decoding on prime of the earlier two parts to speed up decoding, particularly for lengthy requests within the tail. It introduces a Distributed Grouped Draft Server, or DGDS. DGDS maintains a Compressed Suffix Tree for every group and aggregates token sequences from all requests in that group. Situations asynchronously append generated tokens to DGDS, periodically fetch up to date suffix bushes and carry out native speculative decoding based mostly on the shared sample statistics.
The system adjusts draft size and the variety of paths in accordance with mannequin structure, batch dimension and measured acceptance size. For dense and Combination of Specialists fashions, it pre-computes completely different hypothesis thresholds and makes use of them to sure draft depth for every batch. In late tail phases, concurrency is low, so Seer will increase draft depth and permits multi path drafting to boost accepted tokens per step.
Ablation outcomes present that divided rollout yields as much as 35 p.c throughput enchancment over the baseline. Including Context Conscious Scheduling will increase this to as much as 47 p.c over baseline. Enabling grouped speculative decoding raises the full speedup to 77 p.c to 87 p.c over the baseline within the evaluated iteration.
Finish to finish influence on RL coaching
The analysis workforce consider Seer on three RL duties constructed on Moonlight, Qwen2 VL 72B and Kimi K2. They run 10 rollout iterations per process and measure output tokens per second and completion time for every rollout. Seer improves rollout throughput by 74 p.c to 97 p.c throughout these workloads relative to veRL with the identical RL algorithm and vLLM based mostly inference engine.
Tail latency is decreased by 75 p.c to 93 p.c. For reminiscence constrained duties, the baseline system spends as much as half of its time on the final 10 p.c of requests. Seer removes most of this tail by combining divided rollout, Context Conscious Scheduling and Adaptive Grouped Speculative Decoding on prime of the Mooncake based mostly International KVCache Pool.
Key Takeaways
- Rollout bottleneck: Seer targets the rollout section of synchronous RL, which accounts for about 63% to 87% of iteration time and is dominated by lengthy tail requests and KV cache fragmentation.
- Three core mechanisms: Seer combines divided rollout, context conscious scheduling and adaptive grouped speculative decoding to take advantage of output size and sample similarity amongst GRPO responses that share a immediate.
- Effective grained scheduling on a worldwide KV cache: Requests are cut up into chunks and migrated throughout a Mooncake fashion International KVCache Pool, which preserves synchronous on coverage RL whereas conserving GPU reminiscence utilization excessive and lowering preemptions.
- On-line context for tail latency discount: Group stage size statistics from speculative requests drive context conscious scheduling that approximates an oracle longest first scheduler and sharply reduces the time spent on the final 10 p.c of requests.
- Measured finish to finish positive factors: On manufacturing grade RL workloads with Moonlight, Qwen2 VL 72B and Kimi K2, Seer improves rollout throughput by 74% to 97% and reduces lengthy tail latency by 75% to 93% relative to a state-of-the-art synchronous vLLM based mostly baseline.
Editorial Feedback
Seer is a vital methods contribution as a result of it optimizes the rollout section in synchronous RL with out altering the underlying GRPO algorithm, so it preserves on coverage ensures and reproducibility whereas fixing an actual infrastructure bottleneck. The mixture of divided rollout, context conscious scheduling and adaptive grouped speculative decoding presents a sensible template for different RL stacks that depend on lengthy chain of thought reasoning fashions and huge KVCache footprints. General, Seer exhibits that on-line context studying on the methods stage is now as important as mannequin structure for scaling reasoning RL effectively.
Try the Paper right here. Be at liberty to take a look at our GitHub Web page for Tutorials, Codes and Notebooks. Additionally, be at liberty to comply with us on Twitter and don’t overlook to affix our 100k+ ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you possibly can be a part of us on telegram as effectively.
The submit Moonshot AI Researchers Introduce Seer: An On-line Context Studying System for Quick Synchronous Reinforcement Studying RL Rollouts appeared first on MarkTechPost.
