{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:SAQM2GNO3AI7BZRWKFA4PYE6YW","short_pith_number":"pith:SAQM2GNO","schema_version":"1.0","canonical_sha256":"9020cd19aed811f0e6365141c7e09ec58bf6a87f4494e2bf2da62bb4d848ec5d","source":{"kind":"arxiv","id":"2502.19811","version":3},"attestation_state":"computed","paper":{"title":"Comet: Fine-grained Computation-communication Overlapping for Mixture-of-Experts","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.DC","authors_text":"Chengquan Jiang, Haibin Lin, Li-wen Chang, Ningxin Zheng, Qi Hou, Quan Chen, Shulai Zhang, Size Zheng, Weihao Cui, Wenlei Bao, Xin Liu, Ziheng Jiang","submitted_at":"2025-02-27T06:36:45Z","abstract_excerpt":"Mixture-of-experts (MoE) has been extensively employed to scale large language models to trillion-plus parameters while maintaining a fixed computational cost. The development of large MoE models in the distributed scenario encounters the problem of large communication overhead. The inter-device communication of a MoE layer can occupy 47% time of the entire model execution with popular models and frameworks. Therefore, existing methods suggest the communication in a MoE layer to be pipelined with the computation for overlapping. However, these coarse grained overlapping schemes introduce a not"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2502.19811","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.DC","submitted_at":"2025-02-27T06:36:45Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"b0b02fe73ab5c30a537b60ceb9e5ef10d2b9c38d5c278e12acc096aee63b10ab","abstract_canon_sha256":"8d8a26bcd55a81154c38ad946ad075fecc8eeb9cbcb92de3f30411507054710a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:24:02.986906Z","signature_b64":"DgXef19yoth2PNSWcKhrqD9UpIaNBs64CCxkJvLbhvTRNlaWpxtkVxdwB2H8vFRPM7+ZYZRiVbSh01QX4+tdDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9020cd19aed811f0e6365141c7e09ec58bf6a87f4494e2bf2da62bb4d848ec5d","last_reissued_at":"2026-07-05T10:24:02.986235Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:24:02.986235Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Comet: Fine-grained Computation-communication Overlapping for Mixture-of-Experts","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.DC","authors_text":"Chengquan Jiang, Haibin Lin, Li-wen Chang, Ningxin Zheng, Qi Hou, Quan Chen, Shulai Zhang, Size Zheng, Weihao Cui, Wenlei Bao, Xin Liu, Ziheng Jiang","submitted_at":"2025-02-27T06:36:45Z","abstract_excerpt":"Mixture-of-experts (MoE) has been extensively employed to scale large language models to trillion-plus parameters while maintaining a fixed computational cost. The development of large MoE models in the distributed scenario encounters the problem of large communication overhead. The inter-device communication of a MoE layer can occupy 47% time of the entire model execution with popular models and frameworks. Therefore, existing methods suggest the communication in a MoE layer to be pipelined with the computation for overlapping. However, these coarse grained overlapping schemes introduce a not"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.19811","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2502.19811/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2502.19811","created_at":"2026-07-05T10:24:02.986311+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.19811v3","created_at":"2026-07-05T10:24:02.986311+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.19811","created_at":"2026-07-05T10:24:02.986311+00:00"},{"alias_kind":"pith_short_12","alias_value":"SAQM2GNO3AI7","created_at":"2026-07-05T10:24:02.986311+00:00"},{"alias_kind":"pith_short_16","alias_value":"SAQM2GNO3AI7BZRW","created_at":"2026-07-05T10:24:02.986311+00:00"},{"alias_kind":"pith_short_8","alias_value":"SAQM2GNO","created_at":"2026-07-05T10:24:02.986311+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":15,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2607.06202","citing_title":"UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2607.06202","citing_title":"UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2606.09200","citing_title":"Resource-aware Computation-Communication Overlap for multi-GPU ML Workloads","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00735","citing_title":"ViBE: Co-Optimizing Workload Skew and Hardware Variability for MoE Serving","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21603","citing_title":"DynaFlow: Transparent and Flexible Intra-Device Parallelism via Programmable Operator Scheduling","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15617","citing_title":"A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2511.06605","citing_title":"DMA-Latte: Expanding the Reach of DMA Offloads to Latency-bound ML Communication","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2601.20595","citing_title":"Syncopate: Efficient Multi-GPU AI Kernels via Automatic Chunk-Centric Compute-Communication Overlap","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11005","citing_title":"DisagMoE: Computation-Communication overlapped MoE Training via Disaggregated AF-Pipe Parallelism","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08292","citing_title":"Hierarchical Mixture-of-Experts with Two-Stage Optimization","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05607","citing_title":"Accelerating MoE with Dynamic In-Switch Computing on Multi-GPUs","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06206","citing_title":"Federation of Experts: Communication Efficient Distributed Inference for Large Language Models","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2505.07062","citing_title":"Seed1.5-VL Technical Report","ref_index":174,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02953","citing_title":"DITRON: Distributed Multi-level Tiling Compiler for Parallel Tensor Programs","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19241","citing_title":"UniEP: Unified Expert-Parallel MoE MegaKernel for LLM Training","ref_index":48,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW","json":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW.json","graph_json":"https://pith.science/api/pith-number/SAQM2GNO3AI7BZRWKFA4PYE6YW/graph.json","events_json":"https://pith.science/api/pith-number/SAQM2GNO3AI7BZRWKFA4PYE6YW/events.json","paper":"https://pith.science/paper/SAQM2GNO"},"agent_actions":{"view_html":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW","download_json":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW.json","view_paper":"https://pith.science/paper/SAQM2GNO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.19811&json=true","fetch_graph":"https://pith.science/api/pith-number/SAQM2GNO3AI7BZRWKFA4PYE6YW/graph.json","fetch_events":"https://pith.science/api/pith-number/SAQM2GNO3AI7BZRWKFA4PYE6YW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW/action/storage_attestation","attest_author":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW/action/author_attestation","sign_citation":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW/action/citation_signature","submit_replication":"https://pith.science/pith/SAQM2GNO3AI7BZRWKFA4PYE6YW/action/replication_record"}},"created_at":"2026-07-05T10:24:02.986311+00:00","updated_at":"2026-07-05T10:24:02.986311+00:00"}