{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:Z5RRMBIO4YDXLMFQXN6UG3AHKV","short_pith_number":"pith:Z5RRMBIO","schema_version":"1.0","canonical_sha256":"cf6316050ee60775b0b0bb7d436c0755694ed693ebc776d5e15e0cb4e23cc20d","source":{"kind":"arxiv","id":"2503.08040","version":3},"attestation_state":"computed","paper":{"title":"Accurate INT8 Training Through Dynamic Block-Level Fallback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianfei Chen, Jia Wei, Jintao Zhang, Jun Zhu, Pengle Zhang","submitted_at":"2025-03-11T04:45:48Z","abstract_excerpt":"Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls bac"},"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":"2503.08040","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-03-11T04:45:48Z","cross_cats_sorted":[],"title_canon_sha256":"5c10b77bc7a06d4868084dadb6921907234c8596c1bbd2e27509a9b3f39ee669","abstract_canon_sha256":"986244ae345a84cae06cf2aeb4c4da7834cc952c7240eb808add609937578627"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:18:19.236235Z","signature_b64":"nevIVGjHOP2PUGIigHQLlvAdCrTApClCDh7bhELWypcNFT/5EwpL15j7i8WVh2Y3k20KoFWeFpNqm0EKn2EoBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cf6316050ee60775b0b0bb7d436c0755694ed693ebc776d5e15e0cb4e23cc20d","last_reissued_at":"2026-07-05T11:18:19.235664Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:18:19.235664Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Accurate INT8 Training Through Dynamic Block-Level Fallback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianfei Chen, Jia Wei, Jintao Zhang, Jun Zhu, Pengle Zhang","submitted_at":"2025-03-11T04:45:48Z","abstract_excerpt":"Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls bac"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.08040","kind":"arxiv","version":3},"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/2503.08040/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":"2503.08040","created_at":"2026-07-05T11:18:19.235738+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.08040v3","created_at":"2026-07-05T11:18:19.235738+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.08040","created_at":"2026-07-05T11:18:19.235738+00:00"},{"alias_kind":"pith_short_12","alias_value":"Z5RRMBIO4YDX","created_at":"2026-07-05T11:18:19.235738+00:00"},{"alias_kind":"pith_short_16","alias_value":"Z5RRMBIO4YDXLMFQ","created_at":"2026-07-05T11:18:19.235738+00:00"},{"alias_kind":"pith_short_8","alias_value":"Z5RRMBIO","created_at":"2026-07-05T11:18:19.235738+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.20402","citing_title":"Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20402","citing_title":"Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20402","citing_title":"Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor","ref_index":43,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV","json":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV.json","graph_json":"https://pith.science/api/pith-number/Z5RRMBIO4YDXLMFQXN6UG3AHKV/graph.json","events_json":"https://pith.science/api/pith-number/Z5RRMBIO4YDXLMFQXN6UG3AHKV/events.json","paper":"https://pith.science/paper/Z5RRMBIO"},"agent_actions":{"view_html":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV","download_json":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV.json","view_paper":"https://pith.science/paper/Z5RRMBIO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.08040&json=true","fetch_graph":"https://pith.science/api/pith-number/Z5RRMBIO4YDXLMFQXN6UG3AHKV/graph.json","fetch_events":"https://pith.science/api/pith-number/Z5RRMBIO4YDXLMFQXN6UG3AHKV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV/action/storage_attestation","attest_author":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV/action/author_attestation","sign_citation":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV/action/citation_signature","submit_replication":"https://pith.science/pith/Z5RRMBIO4YDXLMFQXN6UG3AHKV/action/replication_record"}},"created_at":"2026-07-05T11:18:19.235738+00:00","updated_at":"2026-07-05T11:18:19.235738+00:00"}