{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:I5E5PQ66SW7RIJDMUFUJNP2H6R","short_pith_number":"pith:I5E5PQ66","schema_version":"1.0","canonical_sha256":"4749d7c3de95bf14246ca16896bf47f4583c39990d8db0ffa45510093ec80e1c","source":{"kind":"arxiv","id":"2506.08027","version":2},"attestation_state":"computed","paper":{"title":"Recipes for Pre-training LLMs with MXFP8","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC"],"primary_cat":"cs.LG","authors_text":"Asit Mishra, Dusan Stosic, Paulius Micikevicius, Simon Layton","submitted_at":"2025-05-30T21:08:15Z","abstract_excerpt":"Using fewer bits to represent model parameters and related tensors during pre-training has become a required technique for improving GPU efficiency without sacrificing accuracy. Microscaling (MX) formats introduced in NVIDIA Blackwell generation of GPUs represent a major advancement of this technique, making it practical to combine narrow floating-point data types with finer granularity per-block scaling factors. In turn, this enables both quantization of more tensors than previous approaches and more efficient execution of operations on those tensors.\n  Effective use of MX-formats requires ca"},"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":"2506.08027","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-30T21:08:15Z","cross_cats_sorted":["cs.AI","cs.DC"],"title_canon_sha256":"74590c8d473a412368bac8d036cb4a6ae731e4d0f167e7a2694f25c591d8684c","abstract_canon_sha256":"07ff60e8b4fabba9049c540472c5632ce385e05b01c21e55be9f3b87902554b1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:55:40.669070Z","signature_b64":"LJ49yACazz0h5/UdBoj185RQJIabBxfzcdGObraEbW3XQp4LEABRrC5DF5Zs1MABbsDejohqs9BBX6zoCwGSDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4749d7c3de95bf14246ca16896bf47f4583c39990d8db0ffa45510093ec80e1c","last_reissued_at":"2026-07-05T11:55:40.668611Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:55:40.668611Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Recipes for Pre-training LLMs with MXFP8","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC"],"primary_cat":"cs.LG","authors_text":"Asit Mishra, Dusan Stosic, Paulius Micikevicius, Simon Layton","submitted_at":"2025-05-30T21:08:15Z","abstract_excerpt":"Using fewer bits to represent model parameters and related tensors during pre-training has become a required technique for improving GPU efficiency without sacrificing accuracy. Microscaling (MX) formats introduced in NVIDIA Blackwell generation of GPUs represent a major advancement of this technique, making it practical to combine narrow floating-point data types with finer granularity per-block scaling factors. In turn, this enables both quantization of more tensors than previous approaches and more efficient execution of operations on those tensors.\n  Effective use of MX-formats requires ca"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.08027","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/2506.08027/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":"2506.08027","created_at":"2026-07-05T11:55:40.668668+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.08027v2","created_at":"2026-07-05T11:55:40.668668+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.08027","created_at":"2026-07-05T11:55:40.668668+00:00"},{"alias_kind":"pith_short_12","alias_value":"I5E5PQ66SW7R","created_at":"2026-07-05T11:55:40.668668+00:00"},{"alias_kind":"pith_short_16","alias_value":"I5E5PQ66SW7RIJDM","created_at":"2026-07-05T11:55:40.668668+00:00"},{"alias_kind":"pith_short_8","alias_value":"I5E5PQ66","created_at":"2026-07-05T11:55:40.668668+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.00312","citing_title":"Stochastic Rounding Increases Small Singular Values","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28691","citing_title":"OSP-Next: Efficient High-Quality Video Generation with Sparse Sequence Parallelism, HiF8 Quantization, and Reinforcement Learning","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2512.02010","citing_title":"Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15416","citing_title":"StoSignSGD: Unbiased Structural Stochasticity Fixes SignSGD for Training Large Language Models","ref_index":27,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R","json":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R.json","graph_json":"https://pith.science/api/pith-number/I5E5PQ66SW7RIJDMUFUJNP2H6R/graph.json","events_json":"https://pith.science/api/pith-number/I5E5PQ66SW7RIJDMUFUJNP2H6R/events.json","paper":"https://pith.science/paper/I5E5PQ66"},"agent_actions":{"view_html":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R","download_json":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R.json","view_paper":"https://pith.science/paper/I5E5PQ66","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.08027&json=true","fetch_graph":"https://pith.science/api/pith-number/I5E5PQ66SW7RIJDMUFUJNP2H6R/graph.json","fetch_events":"https://pith.science/api/pith-number/I5E5PQ66SW7RIJDMUFUJNP2H6R/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R/action/timestamp_anchor","attest_storage":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R/action/storage_attestation","attest_author":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R/action/author_attestation","sign_citation":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R/action/citation_signature","submit_replication":"https://pith.science/pith/I5E5PQ66SW7RIJDMUFUJNP2H6R/action/replication_record"}},"created_at":"2026-07-05T11:55:40.668668+00:00","updated_at":"2026-07-05T11:55:40.668668+00:00"}