{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:UMJCWGZ4765FJ3Z5LDUWZH3JQ3","short_pith_number":"pith:UMJCWGZ4","schema_version":"1.0","canonical_sha256":"a3122b1b3cffba54ef3d58e96c9f6986d235ff42a5e4427aa0d27b74163566cc","source":{"kind":"arxiv","id":"2606.06521","version":1},"attestation_state":"computed","paper":{"title":"P-Cast Precision in FP8 Attention: Sink-Induced Collapse and the Optimality of S=2^8","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC","cs.LG","cs.PF"],"primary_cat":"cs.AR","authors_text":"Reed Lau","submitted_at":"2026-06-02T17:29:15Z","abstract_excerpt":"FP8 (E4M3) acceleration for attention computation offers significant throughput gains, but the 3-bit mantissa introduces precision challenges when the softmax probability matrix P is cast to FP8 before the P*V matrix multiplication. We analyze two implementation choices that affect output precision under the Attention Sink phenomenon: (1) the KV block iteration order, and (2) the static scaling factor applied to P before casting. We show that forward KV iteration causes \"P-collapse\" -- to leading order, a fraction Phi(Delta + delta_k - 6.93 - ln S) of non-sink P values underflow to zero, where"},"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":"2606.06521","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AR","submitted_at":"2026-06-02T17:29:15Z","cross_cats_sorted":["cs.AI","cs.DC","cs.LG","cs.PF"],"title_canon_sha256":"3c6c47a901d310b7ee5493b36d7697addaa00b60a48d2ca277e085184ac84ff0","abstract_canon_sha256":"80a3f31470d45f3554423f02487d98d784dff25d02868ff4b31281ef20f34f4a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-08T00:03:40.666997Z","signature_b64":"vGMxtKyMpzd/eXfa7NWQV8Qzekb5eAFrbUzomNRxhVl8AyLefyVnOOsJqwDhuAT48Db0JnrEmbtxnVRlQ5OiCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a3122b1b3cffba54ef3d58e96c9f6986d235ff42a5e4427aa0d27b74163566cc","last_reissued_at":"2026-06-08T00:03:40.666241Z","signature_status":"signed_v1","first_computed_at":"2026-06-08T00:03:40.666241Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"P-Cast Precision in FP8 Attention: Sink-Induced Collapse and the Optimality of S=2^8","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC","cs.LG","cs.PF"],"primary_cat":"cs.AR","authors_text":"Reed Lau","submitted_at":"2026-06-02T17:29:15Z","abstract_excerpt":"FP8 (E4M3) acceleration for attention computation offers significant throughput gains, but the 3-bit mantissa introduces precision challenges when the softmax probability matrix P is cast to FP8 before the P*V matrix multiplication. We analyze two implementation choices that affect output precision under the Attention Sink phenomenon: (1) the KV block iteration order, and (2) the static scaling factor applied to P before casting. We show that forward KV iteration causes \"P-collapse\" -- to leading order, a fraction Phi(Delta + delta_k - 6.93 - ln S) of non-sink P values underflow to zero, where"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.06521","kind":"arxiv","version":1},"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/2606.06521/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":"2606.06521","created_at":"2026-06-08T00:03:40.666365+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.06521v1","created_at":"2026-06-08T00:03:40.666365+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.06521","created_at":"2026-06-08T00:03:40.666365+00:00"},{"alias_kind":"pith_short_12","alias_value":"UMJCWGZ4765F","created_at":"2026-06-08T00:03:40.666365+00:00"},{"alias_kind":"pith_short_16","alias_value":"UMJCWGZ4765FJ3Z5","created_at":"2026-06-08T00:03:40.666365+00:00"},{"alias_kind":"pith_short_8","alias_value":"UMJCWGZ4","created_at":"2026-06-08T00:03:40.666365+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3","json":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3.json","graph_json":"https://pith.science/api/pith-number/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/graph.json","events_json":"https://pith.science/api/pith-number/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/events.json","paper":"https://pith.science/paper/UMJCWGZ4"},"agent_actions":{"view_html":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3","download_json":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3.json","view_paper":"https://pith.science/paper/UMJCWGZ4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.06521&json=true","fetch_graph":"https://pith.science/api/pith-number/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/graph.json","fetch_events":"https://pith.science/api/pith-number/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/action/storage_attestation","attest_author":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/action/author_attestation","sign_citation":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/action/citation_signature","submit_replication":"https://pith.science/pith/UMJCWGZ4765FJ3Z5LDUWZH3JQ3/action/replication_record"}},"created_at":"2026-06-08T00:03:40.666365+00:00","updated_at":"2026-06-08T00:03:40.666365+00:00"}