{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2024:NVJH5QSPSPVAW4G6BZG4E7XNYP","short_pith_number":"pith:NVJH5QSP","canonical_record":{"source":{"id":"2412.05496","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-07T01:46:38Z","cross_cats_sorted":["cs.PF","cs.PL"],"title_canon_sha256":"02f1b2dde7720112b903632d127dea6e98b8b208365c21be0578a2ebed6860e2","abstract_canon_sha256":"99a5fe2120d519b9735c5507c356d5ed96c76645c2947dde9e523ff66eda917d"},"schema_version":"1.0"},"canonical_sha256":"6d527ec24f93ea0b70de0e4dc27eedc3d7df546776c970c8e09f577f1a23ee90","source":{"kind":"arxiv","id":"2412.05496","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2412.05496","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"arxiv_version","alias_value":"2412.05496v1","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.05496","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"pith_short_12","alias_value":"NVJH5QSPSPVA","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"NVJH5QSPSPVAW4G6","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"NVJH5QSP","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2024:NVJH5QSPSPVAW4G6BZG4E7XNYP","target":"record","payload":{"canonical_record":{"source":{"id":"2412.05496","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-07T01:46:38Z","cross_cats_sorted":["cs.PF","cs.PL"],"title_canon_sha256":"02f1b2dde7720112b903632d127dea6e98b8b208365c21be0578a2ebed6860e2","abstract_canon_sha256":"99a5fe2120d519b9735c5507c356d5ed96c76645c2947dde9e523ff66eda917d"},"schema_version":"1.0"},"canonical_sha256":"6d527ec24f93ea0b70de0e4dc27eedc3d7df546776c970c8e09f577f1a23ee90","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:13.035102Z","signature_b64":"Ssmnl6CBeJjqwRlAo8e+mf+OVbXMRCOrT3xXZvHu7Yt8vHpiXFDlrHEKtCJae2lJph66nupVMSRgnlNtkHY9Aw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6d527ec24f93ea0b70de0e4dc27eedc3d7df546776c970c8e09f577f1a23ee90","last_reissued_at":"2026-05-17T23:38:13.034428Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:13.034428Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2412.05496","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:13Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"GcNpYW88QdxseRGEIwbQhYRg/2X/lY/Fn68SGtjAJZpvW/LKRkWXMlU5vlH+buj7lzFdm6SKzEzFyrONbJKdDw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-02T00:05:16.498713Z"},"content_sha256":"ae163e415cdb30e18d02be923b0067217dd5013a43ba43d546f1b1d2e0fd11f4","schema_version":"1.0","event_id":"sha256:ae163e415cdb30e18d02be923b0067217dd5013a43ba43d546f1b1d2e0fd11f4"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2024:NVJH5QSPSPVAW4G6BZG4E7XNYP","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Flex Attention: A Programming Model for Generating Optimized Attention Kernels","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels.","cross_cats":["cs.PF","cs.PL"],"primary_cat":"cs.LG","authors_text":"Boyuan Feng, Driss Guessous, Horace He, Juechu Dong, Yanbo Liang","submitted_at":"2024-12-07T01:46:38Z","abstract_excerpt":"Over the past 7 years, attention has become one of the most important primitives in deep learning. The primary approach to optimize attention is FlashAttention, which fuses the operation together, drastically improving both the runtime and the memory consumption. However, the importance of FlashAttention combined with its monolithic nature poses a problem for researchers aiming to try new attention variants -- a \"software lottery\". This problem is exacerbated by the difficulty of writing efficient fused attention kernels, resisting traditional compiler-based approaches. We introduce FlexAttent"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The compiler can automatically produce kernels whose performance remains competitive with hand-written code for the majority of attention variants without requiring additional low-level tuning or special-case handling.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FlexAttention supplies a compiler-driven interface that expresses common attention variants in a few lines of PyTorch and emits optimized kernels whose speed matches hand-written implementations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f0db127564b2b583cd42d86866e814cf377bb82a0785044194f2daf93347a94e"},"source":{"id":"2412.05496","kind":"arxiv","version":1},"verdict":{"id":"8e0407a5-d996-42c3-a0f8-4694d46435a8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T21:20:37.421568Z","strongest_claim":"We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels.","one_line_summary":"FlexAttention supplies a compiler-driven interface that expresses common attention variants in a few lines of PyTorch and emits optimized kernels whose speed matches hand-written implementations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The compiler can automatically produce kernels whose performance remains competitive with hand-written code for the majority of attention variants without requiring additional low-level tuning or special-case handling.","pith_extraction_headline":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels."},"references":{"count":45,"sample":[{"doi":"","year":null,"title":"and Ermon, Stefano and Rudra, Atri and R","work_id":"1eb960cf-6552-4e31-ad77-51c3dd0b8fe2","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"International Conference on Learning Representations (ICLR) , year=","work_id":"2685b620-f779-48e2-a07d-1129359ad7d2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"2023 , url =","work_id":"0eb54142-386c-4cec-a539-07ff17367e3c","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=","work_id":"9a3c3fdb-ec54-4e64-9779-63c57fad506c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints , author=. 2023 , eprint=","work_id":"adea1e48-7643-40a8-afd5-bccc76a5c4fc","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":45,"snapshot_sha256":"6e2c9661b126ff3d1788a0e97e655bd3ff56a987a949e1cc2e1ac93bb0d3488f","internal_anchors":6},"formal_canon":{"evidence_count":2,"snapshot_sha256":"576b840979e19a8c874511e7cd833e62478689fb034a2a5c673b33b5f1b6621e"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"8e0407a5-d996-42c3-a0f8-4694d46435a8"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:13Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"UtNgviatEhwHwdR0SoZx6ZsbAxYhs9sbFkqTIXbG+6BzzHPCH2/t1telMsV6M0GHVEq4LGqdH44Zpw209mE0Aw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-02T00:05:16.499252Z"},"content_sha256":"d1c9bb7b114c6814a0ec4df7805ec77009f2eaa418aecff2e04faa15f83bdc13","schema_version":"1.0","event_id":"sha256:d1c9bb7b114c6814a0ec4df7805ec77009f2eaa418aecff2e04faa15f83bdc13"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/bundle.json","state_url":"https://pith.science/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-02T00:05:16Z","links":{"resolver":"https://pith.science/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP","bundle":"https://pith.science/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/bundle.json","state":"https://pith.science/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/state.json","well_known_bundle":"https://pith.science/.well-known/pith/NVJH5QSPSPVAW4G6BZG4E7XNYP/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:NVJH5QSPSPVAW4G6BZG4E7XNYP","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"99a5fe2120d519b9735c5507c356d5ed96c76645c2947dde9e523ff66eda917d","cross_cats_sorted":["cs.PF","cs.PL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-07T01:46:38Z","title_canon_sha256":"02f1b2dde7720112b903632d127dea6e98b8b208365c21be0578a2ebed6860e2"},"schema_version":"1.0","source":{"id":"2412.05496","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2412.05496","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"arxiv_version","alias_value":"2412.05496v1","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.05496","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"pith_short_12","alias_value":"NVJH5QSPSPVA","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"NVJH5QSPSPVAW4G6","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"NVJH5QSP","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:d1c9bb7b114c6814a0ec4df7805ec77009f2eaa418aecff2e04faa15f83bdc13","target":"graph","created_at":"2026-05-17T23:38:13Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The compiler can automatically produce kernels whose performance remains competitive with hand-written code for the majority of attention variants without requiring additional low-level tuning or special-case handling."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"FlexAttention supplies a compiler-driven interface that expresses common attention variants in a few lines of PyTorch and emits optimized kernels whose speed matches hand-written implementations."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels."}],"snapshot_sha256":"f0db127564b2b583cd42d86866e814cf377bb82a0785044194f2daf93347a94e"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"576b840979e19a8c874511e7cd833e62478689fb034a2a5c673b33b5f1b6621e"},"paper":{"abstract_excerpt":"Over the past 7 years, attention has become one of the most important primitives in deep learning. The primary approach to optimize attention is FlashAttention, which fuses the operation together, drastically improving both the runtime and the memory consumption. However, the importance of FlashAttention combined with its monolithic nature poses a problem for researchers aiming to try new attention variants -- a \"software lottery\". This problem is exacerbated by the difficulty of writing efficient fused attention kernels, resisting traditional compiler-based approaches. We introduce FlexAttent","authors_text":"Boyuan Feng, Driss Guessous, Horace He, Juechu Dong, Yanbo Liang","cross_cats":["cs.PF","cs.PL"],"headline":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-07T01:46:38Z","title":"Flex Attention: A Programming Model for Generating Optimized Attention Kernels"},"references":{"count":45,"internal_anchors":6,"resolved_work":45,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"and Ermon, Stefano and Rudra, Atri and R","work_id":"1eb960cf-6552-4e31-ad77-51c3dd0b8fe2","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"International Conference on Learning Representations (ICLR) , year=","work_id":"2685b620-f779-48e2-a07d-1129359ad7d2","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"2023 , url =","work_id":"0eb54142-386c-4cec-a539-07ff17367e3c","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=","work_id":"9a3c3fdb-ec54-4e64-9779-63c57fad506c","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints , author=. 2023 , eprint=","work_id":"adea1e48-7643-40a8-afd5-bccc76a5c4fc","year":2023}],"snapshot_sha256":"6e2c9661b126ff3d1788a0e97e655bd3ff56a987a949e1cc2e1ac93bb0d3488f"},"source":{"id":"2412.05496","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-17T21:20:37.421568Z","id":"8e0407a5-d996-42c3-a0f8-4694d46435a8","model_set":{"reader":"grok-4.3"},"one_line_summary":"FlexAttention supplies a compiler-driven interface that expresses common attention variants in a few lines of PyTorch and emits optimized kernels whose speed matches hand-written implementations.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"FlexAttention is a compiler-driven programming model that allows implementing attention variants in a few lines of PyTorch code while generating competitive performance kernels.","strongest_claim":"We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels.","weakest_assumption":"The compiler can automatically produce kernels whose performance remains competitive with hand-written code for the majority of attention variants without requiring additional low-level tuning or special-case handling."}},"verdict_id":"8e0407a5-d996-42c3-a0f8-4694d46435a8"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:ae163e415cdb30e18d02be923b0067217dd5013a43ba43d546f1b1d2e0fd11f4","target":"record","created_at":"2026-05-17T23:38:13Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"99a5fe2120d519b9735c5507c356d5ed96c76645c2947dde9e523ff66eda917d","cross_cats_sorted":["cs.PF","cs.PL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-07T01:46:38Z","title_canon_sha256":"02f1b2dde7720112b903632d127dea6e98b8b208365c21be0578a2ebed6860e2"},"schema_version":"1.0","source":{"id":"2412.05496","kind":"arxiv","version":1}},"canonical_sha256":"6d527ec24f93ea0b70de0e4dc27eedc3d7df546776c970c8e09f577f1a23ee90","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"6d527ec24f93ea0b70de0e4dc27eedc3d7df546776c970c8e09f577f1a23ee90","first_computed_at":"2026-05-17T23:38:13.034428Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:13.034428Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Ssmnl6CBeJjqwRlAo8e+mf+OVbXMRCOrT3xXZvHu7Yt8vHpiXFDlrHEKtCJae2lJph66nupVMSRgnlNtkHY9Aw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:13.035102Z","signed_message":"canonical_sha256_bytes"},"source_id":"2412.05496","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:ae163e415cdb30e18d02be923b0067217dd5013a43ba43d546f1b1d2e0fd11f4","sha256:d1c9bb7b114c6814a0ec4df7805ec77009f2eaa418aecff2e04faa15f83bdc13"],"state_sha256":"bf82903ca1b1c0e4c64bc3eb20540eecbf778cd792151ed64c13844126ad3513"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"9fcbipzr/cDn759BeqsYNyRlkFlHfi0fRk0rLYXon54d2BreQhpA6grQi4VPXJ5x00MA+z/FmSm6R+Ifo4nhAQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-02T00:05:16.501721Z","bundle_sha256":"0d600063142598da860eba5bcae901b399845856c5d36d43e62b4c147bc3d370"}}