{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:FPQMF4VHSJHQAMD44IOV5LIO2M","short_pith_number":"pith:FPQMF4VH","schema_version":"1.0","canonical_sha256":"2be0c2f2a7924f00307ce21d5ead0ed306ad55f53187b78539bc4fe8fbd22289","source":{"kind":"arxiv","id":"2406.01733","version":2},"attestation_state":"computed","paper":{"title":"Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CV"],"primary_cat":"cs.LG","authors_text":"Gongfan Fang, Michael Bi Mi, Xinchao Wang, Xinyin Ma","submitted_at":"2024-06-03T18:49:57Z","abstract_excerpt":"Diffusion Transformers have recently demonstrated unprecedented generative capabilities for various tasks. The encouraging results, however, come with the cost of slow inference, since each denoising step requires inference on a transformer model with a large scale of parameters. In this study, we make an interesting and somehow surprising observation: the computation of a large proportion of layers in the diffusion transformer, through introducing a caching mechanism, can be readily removed even without updating the model parameters. In the case of U-ViT-H/2, for example, we may remove up to "},"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":"2406.01733","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-06-03T18:49:57Z","cross_cats_sorted":["cs.CV"],"title_canon_sha256":"5928cc602c5cc654998d7cdf768e76a4b028adf2fc450e81477c0d28189a68b5","abstract_canon_sha256":"6ce239597b379b20825eb06884cbcc5a5924386116745d6579768f68cceae696"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:36:12.031037Z","signature_b64":"69GG+I9K89JeHc//CXs87c6xSW0DjOuRPE7sy+p0nHTBeNOJzwLz8KfXW6KaPOw3sx/8kSTW2RRUlIOFqPjyDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2be0c2f2a7924f00307ce21d5ead0ed306ad55f53187b78539bc4fe8fbd22289","last_reissued_at":"2026-07-05T09:36:12.030549Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:36:12.030549Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CV"],"primary_cat":"cs.LG","authors_text":"Gongfan Fang, Michael Bi Mi, Xinchao Wang, Xinyin Ma","submitted_at":"2024-06-03T18:49:57Z","abstract_excerpt":"Diffusion Transformers have recently demonstrated unprecedented generative capabilities for various tasks. The encouraging results, however, come with the cost of slow inference, since each denoising step requires inference on a transformer model with a large scale of parameters. In this study, we make an interesting and somehow surprising observation: the computation of a large proportion of layers in the diffusion transformer, through introducing a caching mechanism, can be readily removed even without updating the model parameters. In the case of U-ViT-H/2, for example, we may remove up to "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.01733","kind":"arxiv","version":2},"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/2406.01733/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":"2406.01733","created_at":"2026-07-05T09:36:12.030611+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.01733v2","created_at":"2026-07-05T09:36:12.030611+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.01733","created_at":"2026-07-05T09:36:12.030611+00:00"},{"alias_kind":"pith_short_12","alias_value":"FPQMF4VHSJHQ","created_at":"2026-07-05T09:36:12.030611+00:00"},{"alias_kind":"pith_short_16","alias_value":"FPQMF4VHSJHQAMD4","created_at":"2026-07-05T09:36:12.030611+00:00"},{"alias_kind":"pith_short_8","alias_value":"FPQMF4VH","created_at":"2026-07-05T09:36:12.030611+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2405.14430","citing_title":"PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02340","citing_title":"Not All Denoising Steps Are Equal: Model Scheduling for Faster Masked Diffusion Language Models","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2602.05449","citing_title":"DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2602.13357","citing_title":"AdaCorrection: Adaptive Offset Cache Correction for Accurate Diffusion Transformers","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2602.22575","citing_title":"S2O: Early Stopping for Sparse Attention via Online Permutation","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14561","citing_title":"CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism","ref_index":20,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M","json":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M.json","graph_json":"https://pith.science/api/pith-number/FPQMF4VHSJHQAMD44IOV5LIO2M/graph.json","events_json":"https://pith.science/api/pith-number/FPQMF4VHSJHQAMD44IOV5LIO2M/events.json","paper":"https://pith.science/paper/FPQMF4VH"},"agent_actions":{"view_html":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M","download_json":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M.json","view_paper":"https://pith.science/paper/FPQMF4VH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.01733&json=true","fetch_graph":"https://pith.science/api/pith-number/FPQMF4VHSJHQAMD44IOV5LIO2M/graph.json","fetch_events":"https://pith.science/api/pith-number/FPQMF4VHSJHQAMD44IOV5LIO2M/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M/action/storage_attestation","attest_author":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M/action/author_attestation","sign_citation":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M/action/citation_signature","submit_replication":"https://pith.science/pith/FPQMF4VHSJHQAMD44IOV5LIO2M/action/replication_record"}},"created_at":"2026-07-05T09:36:12.030611+00:00","updated_at":"2026-07-05T09:36:12.030611+00:00"}