{"paper":{"title":"MoE-Prefill: Zero Redundancy Overheads in MoE Prefill Serving","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"MoE prefill serving eliminates redundant overheads by asynchronously gathering expert weights during compute-bound phases.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aurick Qiao, Juncheng Yang, Karthik Ganesan, Olatunji Ruwase, Samyam Rajbhandari, Yue Cheng, Yuxiong He, Zhaoyuan Su","submitted_at":"2026-05-03T03:10:24Z","abstract_excerpt":"Production LLM workloads increasingly serve discriminative tasks, such as classification, recommendation, and verification, whose answers are read from the logits of a single prefill pass with no autoregressive decoding. Serving these prefill-only workloads on mixture-of-experts (MoE) models is bottlenecked not by compute but by the distributed execution required to fit the model: existing parallel strategies (tensor, expert, and pipeline parallelism) trade memory pressure for redundant computation, communication, and synchronization, severely degrading MoE prefill serving efficiency. We obser"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On Qwen3-235B-A22B across four hardware/precision configurations, MoE-Prefill delivers 1.35-1.37x throughput over the strongest distributed baseline on real-world workloads and up to 1.59x on long-context synthetic workloads, sustaining 29.8-36.2% per-GPU model FLOPs utilization.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The long, compute-bound forward passes of large-batch prefill open a per-layer window wide enough to stream expert weights in the background, replacing per-layer activation AllToAll with asynchronous weight AllGather fully overlapped with computation without new bottlenecks or accuracy loss.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MoE-Prefill achieves 1.35-1.59x higher throughput for prefill-only MoE serving by using asynchronous expert parallelism to overlap weight AllGather with computation and prefix-aware routing with true-FLOPs tracking.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"MoE prefill serving eliminates redundant overheads by asynchronously gathering expert weights during compute-bound phases.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"74b8371a6a6832046d254f75f767c3ebd917729fcbae67b252d77215f264f953"},"source":{"id":"2605.02960","kind":"arxiv","version":2},"verdict":{"id":"83754aca-5224-4f93-b30f-082eafe82638","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:38:54.747167Z","strongest_claim":"On Qwen3-235B-A22B across four hardware/precision configurations, MoE-Prefill delivers 1.35-1.37x throughput over the strongest distributed baseline on real-world workloads and up to 1.59x on long-context synthetic workloads, sustaining 29.8-36.2% per-GPU model FLOPs utilization.","one_line_summary":"MoE-Prefill achieves 1.35-1.59x higher throughput for prefill-only MoE serving by using asynchronous expert parallelism to overlap weight AllGather with computation and prefix-aware routing with true-FLOPs tracking.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The long, compute-bound forward passes of large-batch prefill open a per-layer window wide enough to stream expert weights in the background, replacing per-layer activation AllToAll with asynchronous weight AllGather fully overlapped with computation without new bottlenecks or accuracy loss.","pith_extraction_headline":"MoE prefill serving eliminates redundant overheads by asynchronously gathering expert weights during compute-bound phases."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.02960/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T17:01:31.756960Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"a1878eac2e9beaf64465e0bee886c01093de3f9ce883ca40f498e2da03b942f6"},"references":{"count":75,"sample":[{"doi":"","year":2025,"title":"gpt-oss-120b & gpt-oss-20b Model Card","work_id":"178c1f7e-4f19-4392-a45d-45a6dfa88ead","ref_index":1,"cited_arxiv_id":"2508.10925","is_internal_anchor":true},{"doi":"","year":2023,"title":"SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills","work_id":"3dbdd757-ca01-436f-acfd-12ffcd6f64c6","ref_index":2,"cited_arxiv_id":"2308.16369","is_internal_anchor":true},{"doi":"","year":2022,"title":"Deepspeed-inference: enabling efficient in- ference of transformer models at unprecedented scale","work_id":"461bf7aa-e1b9-4309-a540-e2905d823c31","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"A survey on mixture of experts in large language models.IEEE Transactions on Knowledge and Data Engineering, 2025","work_id":"c76fa220-c8af-40e4-a69f-ee8ff33e469b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Moe-lightning: High-throughput moe inference on memory-constrained gpus","work_id":"c7c2193d-2846-4a86-8642-116ade28549b","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":75,"snapshot_sha256":"e667420ecd0ad681b43542be402886428e8f63d7873ba26aa2adae8ae08a04b0","internal_anchors":16},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}