{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:D6SPNUM5JFQB6LD2GAWYECBEIT","short_pith_number":"pith:D6SPNUM5","schema_version":"1.0","canonical_sha256":"1fa4f6d19d49601f2c7a302d82082444fdd298497d971a7d64245b2610ef3020","source":{"kind":"arxiv","id":"2404.03605","version":2},"attestation_state":"computed","paper":{"title":"Mitigating the Impact of Outlier Channels for Language Model Quantization with Activation Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Aniruddha Nrusimha, Dan Alistarh, Mayank Mishra, Naigang Wang, Rameswar Panda, Yoon Kim","submitted_at":"2024-04-04T17:25:30Z","abstract_excerpt":"We consider the problem of accurate quantization for language models, where both the weights and activations are uniformly quantized to 4 bits per parameter, the lowest bitwidth format natively supported by GPU hardware. In this context, the key challenge is activation quantization: it is known that language models contain outlier channels whose values on average are orders of magnitude higher than than other channels, which prevents accurate low-bitwidth quantization with known techniques. We systematically study this phenomena and find that these outlier channels emerge early in training, an"},"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":"2404.03605","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-04T17:25:30Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"dc416c678190359daf698cc53b48bcd5d28180ddcfa87d23fdb2255db3106e2d","abstract_canon_sha256":"2fa1839858707ae98eb86221f5484fd3f8a2808ddf2ee9934fdc579e996ec237"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:59:27.021685Z","signature_b64":"Ne5HNcgHiCYevvq9kpHclzOo5k3XeASxHJD/M46jfbhE0rBcLVSSb91H+fL3c1T+6vLHb5a9/ehugeWdW/wRDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1fa4f6d19d49601f2c7a302d82082444fdd298497d971a7d64245b2610ef3020","last_reissued_at":"2026-07-05T08:59:27.021193Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:59:27.021193Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mitigating the Impact of Outlier Channels for Language Model Quantization with Activation Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Aniruddha Nrusimha, Dan Alistarh, Mayank Mishra, Naigang Wang, Rameswar Panda, Yoon Kim","submitted_at":"2024-04-04T17:25:30Z","abstract_excerpt":"We consider the problem of accurate quantization for language models, where both the weights and activations are uniformly quantized to 4 bits per parameter, the lowest bitwidth format natively supported by GPU hardware. In this context, the key challenge is activation quantization: it is known that language models contain outlier channels whose values on average are orders of magnitude higher than than other channels, which prevents accurate low-bitwidth quantization with known techniques. We systematically study this phenomena and find that these outlier channels emerge early in training, an"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.03605","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/2404.03605/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":"2404.03605","created_at":"2026-07-05T08:59:27.021251+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.03605v2","created_at":"2026-07-05T08:59:27.021251+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.03605","created_at":"2026-07-05T08:59:27.021251+00:00"},{"alias_kind":"pith_short_12","alias_value":"D6SPNUM5JFQB","created_at":"2026-07-05T08:59:27.021251+00:00"},{"alias_kind":"pith_short_16","alias_value":"D6SPNUM5JFQB6LD2","created_at":"2026-07-05T08:59:27.021251+00:00"},{"alias_kind":"pith_short_8","alias_value":"D6SPNUM5","created_at":"2026-07-05T08:59:27.021251+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2512.02010","citing_title":"Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01330","citing_title":"Colinearity Decay: Training Quantization-Friendly ViTs with Outlier Decay","ref_index":22,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT","json":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT.json","graph_json":"https://pith.science/api/pith-number/D6SPNUM5JFQB6LD2GAWYECBEIT/graph.json","events_json":"https://pith.science/api/pith-number/D6SPNUM5JFQB6LD2GAWYECBEIT/events.json","paper":"https://pith.science/paper/D6SPNUM5"},"agent_actions":{"view_html":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT","download_json":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT.json","view_paper":"https://pith.science/paper/D6SPNUM5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.03605&json=true","fetch_graph":"https://pith.science/api/pith-number/D6SPNUM5JFQB6LD2GAWYECBEIT/graph.json","fetch_events":"https://pith.science/api/pith-number/D6SPNUM5JFQB6LD2GAWYECBEIT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT/action/storage_attestation","attest_author":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT/action/author_attestation","sign_citation":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT/action/citation_signature","submit_replication":"https://pith.science/pith/D6SPNUM5JFQB6LD2GAWYECBEIT/action/replication_record"}},"created_at":"2026-07-05T08:59:27.021251+00:00","updated_at":"2026-07-05T08:59:27.021251+00:00"}