{"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00237552","sets":["1164:1867:11465:11705"]},"path":["11705"],"owner":"44499","recid":"237552","title":["GPU Direct Storageを用いた大規模言語モデルのロード時間短縮手法"],"pubdate":{"attribute_name":"公開日","attribute_value":"2024-08-01"},"_buckets":{"deposit":"d268bf4e-ebae-420f-a5ec-74869851dfd1"},"_deposit":{"id":"237552","pid":{"type":"depid","value":"237552","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"GPU Direct Storageを用いた大規模言語モデルのロード時間短縮手法","author_link":["650454","650452","650451","650453"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"GPU Direct Storageを用いた大規模言語モデルのロード時間短縮手法"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"ストレージ・ネットワーク","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2024-08-01","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"IBM東京基礎研究所"},{"subitem_text_value":"IBM東京基礎研究所"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/237552/files/IPSJ-OS24164007.pdf","label":"IPSJ-OS24164007.pdf"},"date":[{"dateType":"Available","dateValue":"2026-08-01"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-OS24164007.pdf","filesize":[{"value":"1.0 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"11"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"b63575a6-c628-4f96-bdff-c5594ca556db","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2024 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"吉村, 剛"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"千葉, 立寛"}],"nameIdentifiers":[{}]}]},"item_4_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Takeshi, Yoshimura","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Tatsuhiro, Chiba","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10444176","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8795","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"大規模言語モデルは様々な自然言語処理タスクを可能にし，ミリ秒単位のトークン生成の高速化が進んだ結果，様々なコンピュータサービスへの応用が広まっている．しかし，タスク処理能力に直結するモデルの大きさが 100 億パラメータ以上に増加した結果，GPU メモリへロードする時間が分単位近くまで増大してしまっている．その結果，スケーリング時や障害回復処理でのダウンタイム増加や開発時の待機時間の増加など，様々な悪影響が懸念される．本研究は現状の safetensors ファイル向けのモデルローダの非効率的な箇所を分析し，特にホストメモリをバウンスバッファとしてのみ利用している点を解決する．具体的には GPU Direct Storage を用いることで，ファイルシステムから GPU メモリへのロード部分にかかるオーバヘッドを削減する．本論文では実験結果として 400 億パラメータの Falcon および 1760 億パラメータの Bloom においてそれぞれ 9.4 倍と 3.3 倍の性能改善を得られたことを示す．","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"8","bibliographic_titles":[{"bibliographic_title":"研究報告システムソフトウェアとオペレーティング・システム（OS）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2024-08-01","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"7","bibliographicVolumeNumber":"2024-OS-164"}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"id":237552,"updated":"2025-01-19T08:50:52.046979+00:00","links":{},"created":"2025-01-19T01:40:14.567656+00:00"}