{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "sheng-et-al-flexgen-2023", "title": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference", "summary": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation.", "boundary": "Throughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.", "area": "Optimization and large-model systems", "topic": "Artificial intelligence", "year": 2023, "lane": "HUMAN", "result": "EMPIRICAL", "package_id": "sheng-et-al-flexgen-2023", "target": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "humans": [ "Ying Sheng", "Lianmin Zheng", "Binhang Yuan", "Zhuohan Li", "Max Ryabinin", "Daniel Y. Fu", "Zhiqiang Xie", "Beidi Chen", "Clark Barrett", "Joseph E. Gonzalez", "Percy Liang", "Christopher Ré", "Ion Stoica", "Ce Zhang" ], "authors_complete": true, "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261015-arxiv-2303-06865v2", "title": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" }, { "id": "cur20261015-support-sheng-et-al-flexgen-2023", "title": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference — author/proceedings record", "url": "https://github.com/FMInference/FlexLLMGen" }, { "id": "cur20261015-arxiv-2303-06865v1", "title": "High-throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v1" } ], "points": "65000", "published_grade": { "id": "sheng-et-al-flexgen-2023", "entry_id": "sheng-et-al-flexgen-2023", "band": "B", "modifier": "+", "target": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "what_changed": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation.", "scope": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation. Throughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-16T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Ying Sheng", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Lianmin Zheng", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Binhang Yuan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zhuohan Li", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Max Ryabinin", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Daniel Y. Fu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zhiqiang Xie", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Beidi Chen", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Clark Barrett", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Joseph E. Gonzalez", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Percy Liang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Christopher Ré", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Ion Stoica", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Ce Zhang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" } }, { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-1", "kind": "source_reference", "source": { "citation": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference — author/proceedings record", "url": "https://github.com/FMInference/FlexLLMGen" } }, { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-2", "kind": "source_reference", "source": { "citation": "High-throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Throughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.", "domain": "Optimization and large-model systems" }, "sources": [ { "citation": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2" }, { "citation": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference — author/proceedings record", "url": "https://github.com/FMInference/FlexLLMGen" }, { "citation": "High-throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v1" } ], "statement": "# Sheng et al.: FlexGen throughput-oriented heterogeneous inference\n\nOptimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation.\n\n## Scientific scope\n\nThroughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nYing Sheng; Lianmin Zheng; Binhang Yuan; Zhuohan Li; Max Ryabinin; Daniel Y. Fu; Zhiqiang Xie; Beidi Chen; Clark Barrett; Joseph E. Gonzalez; Percy Liang; Christopher Ré; Ion Stoica; Ce Zhang\n\nSource-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.\n\n\n\n## Original sources\n\n- [FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU](https://arxiv.org/abs/2303.06865v2)\n- [Sheng et al.: FlexGen throughput-oriented heterogeneous inference — author/proceedings record](https://github.com/FMInference/FlexLLMGen)\n- [High-throughput Generative Inference of Large Language Models with a Single GPU](https://arxiv.org/abs/2303.06865v1)\n", "summary": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation.", "title": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference" }, "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "source_ref": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:68c4beec58d5d562fca93476d451b93435f1583c3a8b46a5a884db31e483164e" }, "target": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "sheng-et-al-flexgen-2023", "entry_id": "sheng-et-al-flexgen-2023", "revision": 1, "title": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference", "context": "Optimization and large-model systems", "track": "Human-designed research method or resource", "band": "B", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-16", "evidence_cutoff": "2026-10-16", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation.", "scope": "Optimized tensor placement and access over GPU, CPU and disk with linear-program-guided offloading and compression for high-throughput, latency-insensitive language-model generation. Throughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.", "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Throughput is amortized over large batches and can involve high latency, CPU memory and disk resources; it is not interactive one-token-per-second service on a standalone small GPU.", "references": [ { "title": "FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v2", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." }, { "title": "Sheng et al.: FlexGen throughput-oriented heterogeneous inference — author/proceedings record", "url": "https://github.com/FMInference/FlexLLMGen", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." }, { "title": "High-throughput Generative Inference of Large Language Models with a Single GPU", "url": "https://arxiv.org/abs/2303.06865v1", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." } ], "roles": [ { "who": "Source-listed authors and technical contributors", "credit": "SCIENCE", "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities." } ], "target": { "record_id": "urn:uuid:eccfb983-0710-4b8b-abfc-d69cb111c813", "version": 1, "content_hash": "sha256:c940d06123dd055995931b370f805a70769028fba4c8850fbb3cf9e7c45dd42c" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "65000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": null, "reading_status": "SOURCE_LINKED_NOTE" }