{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:YKUE67AD6VUSXZ3BOHXVIOSZXB","short_pith_number":"pith:YKUE67AD","schema_version":"1.0","canonical_sha256":"c2a84f7c03f5692be76171ef543a59b867cf0f1f61de72607382056539f910ed","source":{"kind":"arxiv","id":"2607.17644","version":1},"attestation_state":"computed","paper":{"title":"A Training-Memory Regression in MLA Sequence Parallelism: Why Megatron-Core Forbids Absorption, and LAGA -- a Communication-Efficient Fix","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Changzheng Ma","submitted_at":"2026-07-20T07:57:34Z","abstract_excerpt":"Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form -- which slashes collective communication by gathering only the compressed latent -- that is fully implemented but hard-asserted out of training (the forward opens with \"assert not (self.training and self.cache_mla_latents)\"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap -- its intermediates live in n_h x d_kv dimensions per toke"},"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":"2607.17644","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.DC","submitted_at":"2026-07-20T07:57:34Z","cross_cats_sorted":[],"title_canon_sha256":"dc371da43d7bcf12d62cd3e75667e8a06c1257af02b88e29d4f17399c85d0342","abstract_canon_sha256":"998e7d6fc4ec61e701d48cae1b3adbda0d3f59e97e83ef7002eba2969d30a6e7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-21T01:21:56.872778Z","signature_b64":"Vtpido0pshg/8DOs8rPJLg4Om6nZBXGHM3v5G/4E1wfhienRLmeoi9chJrLz0mS5mEJ2+3F/knMQiorV4640Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c2a84f7c03f5692be76171ef543a59b867cf0f1f61de72607382056539f910ed","last_reissued_at":"2026-07-21T01:21:56.872011Z","signature_status":"signed_v1","first_computed_at":"2026-07-21T01:21:56.872011Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Training-Memory Regression in MLA Sequence Parallelism: Why Megatron-Core Forbids Absorption, and LAGA -- a Communication-Efficient Fix","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Changzheng Ma","submitted_at":"2026-07-20T07:57:34Z","abstract_excerpt":"Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form -- which slashes collective communication by gathering only the compressed latent -- that is fully implemented but hard-asserted out of training (the forward opens with \"assert not (self.training and self.cache_mla_latents)\"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap -- its intermediates live in n_h x d_kv dimensions per toke"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.17644","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/2607.17644/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":"2607.17644","created_at":"2026-07-21T01:21:56.872417+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.17644v1","created_at":"2026-07-21T01:21:56.872417+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.17644","created_at":"2026-07-21T01:21:56.872417+00:00"},{"alias_kind":"pith_short_12","alias_value":"YKUE67AD6VUS","created_at":"2026-07-21T01:21:56.872417+00:00"},{"alias_kind":"pith_short_16","alias_value":"YKUE67AD6VUSXZ3B","created_at":"2026-07-21T01:21:56.872417+00:00"},{"alias_kind":"pith_short_8","alias_value":"YKUE67AD","created_at":"2026-07-21T01:21:56.872417+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/YKUE67AD6VUSXZ3BOHXVIOSZXB","json":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB.json","graph_json":"https://pith.science/api/pith-number/YKUE67AD6VUSXZ3BOHXVIOSZXB/graph.json","events_json":"https://pith.science/api/pith-number/YKUE67AD6VUSXZ3BOHXVIOSZXB/events.json","paper":"https://pith.science/paper/YKUE67AD"},"agent_actions":{"view_html":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB","download_json":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB.json","view_paper":"https://pith.science/paper/YKUE67AD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.17644&json=true","fetch_graph":"https://pith.science/api/pith-number/YKUE67AD6VUSXZ3BOHXVIOSZXB/graph.json","fetch_events":"https://pith.science/api/pith-number/YKUE67AD6VUSXZ3BOHXVIOSZXB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB/action/storage_attestation","attest_author":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB/action/author_attestation","sign_citation":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB/action/citation_signature","submit_replication":"https://pith.science/pith/YKUE67AD6VUSXZ3BOHXVIOSZXB/action/replication_record"}},"created_at":"2026-07-21T01:21:56.872417+00:00","updated_at":"2026-07-21T01:21:56.872417+00:00"}