{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:GJMQ5GP2U5QTVEPIHYPZLCNJVY","short_pith_number":"pith:GJMQ5GP2","schema_version":"1.0","canonical_sha256":"32590e99faa7613a91e83e1f9589a9ae22ad3403f1c302db6ca8d00efa39e730","source":{"kind":"arxiv","id":"2405.18710","version":2},"attestation_state":"computed","paper":{"title":"To FP8 and Back Again: Quantifying Reduced Precision Effects on LLM Training Stability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Byeongwook Kim, Dongsoo Lee, Jeongin Bae, Joonhyung Lee, Se Jung Kwon","submitted_at":"2024-05-29T02:42:23Z","abstract_excerpt":"The massive computational costs associated with large language model (LLM) pretraining have spurred great interest in reduced-precision floating-point representations to accelerate the process. As a result, the BrainFloat16 (BF16) precision has become the de facto standard for LLM training, with hardware support included in recent generations of accelerators. This trend has gone even further in the latest processors, where FP8 has recently been introduced. However, prior experience with FP16, which was found to be less stable than BF16, raises concerns as to whether FP8, with even fewer bits t"},"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":"2405.18710","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-05-29T02:42:23Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"40edecbc098c84edb25b2c58a1eaf2924475511ca40a1f111d2f6dfb9a3a33fe","abstract_canon_sha256":"a027307c17817dd5b3516a3e08c79639cae4b4b890f20a99fdcaea5cef275d07"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:38:33.596170Z","signature_b64":"QB+T/dJiKKg5zKJPCtYSyNkKdS/PdGvi8LDCG6UYpOyMdZ4coE25783yuqZvb3uxQ2LXp2CUUUDDzPkcrYI9AQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"32590e99faa7613a91e83e1f9589a9ae22ad3403f1c302db6ca8d00efa39e730","last_reissued_at":"2026-07-05T10:38:33.595692Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:38:33.595692Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"To FP8 and Back Again: Quantifying Reduced Precision Effects on LLM Training Stability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Byeongwook Kim, Dongsoo Lee, Jeongin Bae, Joonhyung Lee, Se Jung Kwon","submitted_at":"2024-05-29T02:42:23Z","abstract_excerpt":"The massive computational costs associated with large language model (LLM) pretraining have spurred great interest in reduced-precision floating-point representations to accelerate the process. As a result, the BrainFloat16 (BF16) precision has become the de facto standard for LLM training, with hardware support included in recent generations of accelerators. This trend has gone even further in the latest processors, where FP8 has recently been introduced. However, prior experience with FP16, which was found to be less stable than BF16, raises concerns as to whether FP8, with even fewer bits t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.18710","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/2405.18710/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":"2405.18710","created_at":"2026-07-05T10:38:33.595752+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.18710v2","created_at":"2026-07-05T10:38:33.595752+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.18710","created_at":"2026-07-05T10:38:33.595752+00:00"},{"alias_kind":"pith_short_12","alias_value":"GJMQ5GP2U5QT","created_at":"2026-07-05T10:38:33.595752+00:00"},{"alias_kind":"pith_short_16","alias_value":"GJMQ5GP2U5QTVEPI","created_at":"2026-07-05T10:38:33.595752+00:00"},{"alias_kind":"pith_short_8","alias_value":"GJMQ5GP2","created_at":"2026-07-05T10:38:33.595752+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.09370","citing_title":"From Detection to Recovery: Operational Analysis on LLM Pre-training with 504 GPUs","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25704","citing_title":"PowLU: An Activation Function for Stable Pre-Training of LLMs","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00539","citing_title":"GNMR: Runtime Stability Control for Low-Precision Large Language Model Training","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2604.03298","citing_title":"ENEC: A Lossless AI Model Compression Method Enabling Fast Inference on Ascend NPUs","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09370","citing_title":"From Detection to Recovery: Operational Analysis on LLM Pre-training with 504 GPUs","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15416","citing_title":"StoSignSGD: Unbiased Structural Stochasticity Fixes SignSGD for Training Large Language Models","ref_index":23,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY","json":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY.json","graph_json":"https://pith.science/api/pith-number/GJMQ5GP2U5QTVEPIHYPZLCNJVY/graph.json","events_json":"https://pith.science/api/pith-number/GJMQ5GP2U5QTVEPIHYPZLCNJVY/events.json","paper":"https://pith.science/paper/GJMQ5GP2"},"agent_actions":{"view_html":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY","download_json":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY.json","view_paper":"https://pith.science/paper/GJMQ5GP2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.18710&json=true","fetch_graph":"https://pith.science/api/pith-number/GJMQ5GP2U5QTVEPIHYPZLCNJVY/graph.json","fetch_events":"https://pith.science/api/pith-number/GJMQ5GP2U5QTVEPIHYPZLCNJVY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY/action/storage_attestation","attest_author":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY/action/author_attestation","sign_citation":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY/action/citation_signature","submit_replication":"https://pith.science/pith/GJMQ5GP2U5QTVEPIHYPZLCNJVY/action/replication_record"}},"created_at":"2026-07-05T10:38:33.595752+00:00","updated_at":"2026-07-05T10:38:33.595752+00:00"}