{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:ZZDC4WM2OJ5CREJ7PXMULCOD7N","short_pith_number":"pith:ZZDC4WM2","schema_version":"1.0","canonical_sha256":"ce462e599a727a28913f7dd94589c3fb7f61f0106a94358dfa5a1fd9409a6fef","source":{"kind":"arxiv","id":"2401.08671","version":1},"attestation_state":"computed","paper":{"title":"DeepSpeed-FastGen: High-throughput Text Generation for LLMs via MII and DeepSpeed-Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.PF","authors_text":"Ammar Ahmad Awan, Arash Bakhtiari, Connor Holmes, Heyang Qin, Jeff Rasley, Lev Kurilenko, Masahiro Tanaka, Michael Wyatt, Reza Yazdani Aminabadi, Samyam Rajbhandari, Yuxiong He","submitted_at":"2024-01-09T06:49:40Z","abstract_excerpt":"The deployment and scaling of large language models (LLMs) have become critical as they permeate various applications, demanding high-throughput and low-latency serving systems. Existing frameworks struggle to balance these requirements, especially for workloads with long prompts. This paper introduces DeepSpeed-FastGen, a system that employs Dynamic SplitFuse, a novel prompt and generation composition strategy, to deliver up to 2.3x higher effective throughput, 2x lower latency on average, and up to 3.7x lower (token-level) tail latency, compared to state-of-the-art systems like vLLM. We leve"},"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":"2401.08671","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.PF","submitted_at":"2024-01-09T06:49:40Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"347faf36cb17aa9af70084b1d6cca28356920dfe4da0252edfa1b9ad0d39d597","abstract_canon_sha256":"14e7d91f6d462597c307ed42930ee697f646ec8270161329b646f6908dc7e777"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:34:19.790390Z","signature_b64":"iTd/HP9gUGdO9ZFgNcEW6ujpQgWNLrs8j05vpv9AOJ8QtCAblHbnzMfSuyADBjZS+GNhmgdPAoC3WFZp/JWvDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ce462e599a727a28913f7dd94589c3fb7f61f0106a94358dfa5a1fd9409a6fef","last_reissued_at":"2026-07-05T07:34:19.790002Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:34:19.790002Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"DeepSpeed-FastGen: High-throughput Text Generation for LLMs via MII and DeepSpeed-Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.PF","authors_text":"Ammar Ahmad Awan, Arash Bakhtiari, Connor Holmes, Heyang Qin, Jeff Rasley, Lev Kurilenko, Masahiro Tanaka, Michael Wyatt, Reza Yazdani Aminabadi, Samyam Rajbhandari, Yuxiong He","submitted_at":"2024-01-09T06:49:40Z","abstract_excerpt":"The deployment and scaling of large language models (LLMs) have become critical as they permeate various applications, demanding high-throughput and low-latency serving systems. Existing frameworks struggle to balance these requirements, especially for workloads with long prompts. This paper introduces DeepSpeed-FastGen, a system that employs Dynamic SplitFuse, a novel prompt and generation composition strategy, to deliver up to 2.3x higher effective throughput, 2x lower latency on average, and up to 3.7x lower (token-level) tail latency, compared to state-of-the-art systems like vLLM. We leve"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2401.08671","kind":"arxiv","version":1},"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/2401.08671/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":"2401.08671","created_at":"2026-07-05T07:34:19.790064+00:00"},{"alias_kind":"arxiv_version","alias_value":"2401.08671v1","created_at":"2026-07-05T07:34:19.790064+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2401.08671","created_at":"2026-07-05T07:34:19.790064+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZZDC4WM2OJ5C","created_at":"2026-07-05T07:34:19.790064+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZZDC4WM2OJ5CREJ7","created_at":"2026-07-05T07:34:19.790064+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZZDC4WM2","created_at":"2026-07-05T07:34:19.790064+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":25,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22983","citing_title":"LiveServe: Interaction-Aware Serving for Real-Time Omni-Modal LLMs","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.18741","citing_title":"ReMP: Low-Downtime Runtime Model-Parallelism Reconfiguration for LLM Serving","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2607.00151","citing_title":"SmoothAgent: Efficient Long-Horizon LLM-Based Agent Serving with Lookahead Context Engineering","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05933","citing_title":"Beyond Greedy Chunking: SLO-Aware Sliding-Window Scheduling for LLM Inference","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06453","citing_title":"Vortex: Efficient and Programmable Sparse Attention Serving for AI Agents","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02964","citing_title":"Multi-Segment Attention: Enabling Efficient KV-Cache Management for Faster Large Language Model Serving","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01839","citing_title":"Observation, Not Prediction: Conversation-Level Disaggregated Scheduling for Agentic Serving","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23389","citing_title":"AlignedServe: Orchestrating Prefix-aware Batching to Build a High-throughput and Computing-efficient LLM Serving System","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2412.03594","citing_title":"BatchLLM: Optimizing Large Batched LLM Inference with Global Prefix Sharing and Throughput-oriented Token Batching","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2412.14590","citing_title":"MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2504.09775","citing_title":"MIST: A Co-Design Framework for Heterogeneous, Multi-Stage LLM Inference","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2505.09999","citing_title":"ServeGen: Workload Characterization and Generation of Large Language Model Serving in Production","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2512.09472","citing_title":"WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2603.10726","citing_title":"PrefixWall: Mitigating Prefix Caching Side Channels in Shared LLM Systems","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2603.19199","citing_title":"FASTER: Rethinking Real-Time Flow VLAs","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18346","citing_title":"Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09611","citing_title":"Characterizing Performance-Energy Trade-offs of Large Language Models in Multi-Request Workflows","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2603.19199","citing_title":"FASTER: Rethinking Real-Time Flow VLAs","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2404.14294","citing_title":"A Survey on Efficient Inference for Large Language Models","ref_index":279,"is_internal_anchor":false},{"citing_arxiv_id":"2603.27960","citing_title":"Towards Efficient Large Vision-Language Models: A Comprehensive Survey on Inference Strategies","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22906","citing_title":"Network Edge Inference for Large Language Models: Principles, Techniques, and Opportunities","ref_index":64,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06206","citing_title":"Federation of Experts: Communication Efficient Distributed Inference for Large Language Models","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2409.19256","citing_title":"HybridFlow: A Flexible and Efficient RLHF Framework","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.26963","citing_title":"MARS: Efficient, Adaptive Co-Scheduling for Heterogeneous Agentic Systems","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15186","citing_title":"Scepsy: Serving Agentic Workflows Using Aggregate LLM Pipelines","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N","json":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N.json","graph_json":"https://pith.science/api/pith-number/ZZDC4WM2OJ5CREJ7PXMULCOD7N/graph.json","events_json":"https://pith.science/api/pith-number/ZZDC4WM2OJ5CREJ7PXMULCOD7N/events.json","paper":"https://pith.science/paper/ZZDC4WM2"},"agent_actions":{"view_html":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N","download_json":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N.json","view_paper":"https://pith.science/paper/ZZDC4WM2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2401.08671&json=true","fetch_graph":"https://pith.science/api/pith-number/ZZDC4WM2OJ5CREJ7PXMULCOD7N/graph.json","fetch_events":"https://pith.science/api/pith-number/ZZDC4WM2OJ5CREJ7PXMULCOD7N/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N/action/storage_attestation","attest_author":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N/action/author_attestation","sign_citation":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N/action/citation_signature","submit_replication":"https://pith.science/pith/ZZDC4WM2OJ5CREJ7PXMULCOD7N/action/replication_record"}},"created_at":"2026-07-05T07:34:19.790064+00:00","updated_at":"2026-07-05T07:34:19.790064+00:00"}