{"id":"https://openalex.org/W7169868603","doi":"https://doi.org/10.48550/arxiv.2607.17644","title":"A Training-Memory Regression in MLA Sequence Parallelism: Why Megatron-Core Forbids Absorption, and LAGA -- a Communication-Efficient Fix","display_name":"A Training-Memory Regression in MLA Sequence Parallelism: Why Megatron-Core Forbids Absorption, and LAGA -- a Communication-Efficient Fix","publication_year":2026,"publication_date":"2026-07-20","ids":{"openalex":"https://openalex.org/W7169868603","doi":"https://doi.org/10.48550/arxiv.2607.17644"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2607.17644","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.17644","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2607.17644","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5074723125","display_name":"C W","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Ma, Changzheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5074723125"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.44190001487731934,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.44190001487731934,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11948","display_name":"Machine Learning in Materials Science","score":0.1088000014424324,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.08510000258684158,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/sequence","display_name":"Sequence (biology)","score":0.6996999979019165},{"id":"https://openalex.org/keywords/porting","display_name":"Porting","score":0.6007000207901001},{"id":"https://openalex.org/keywords/scale","display_name":"Scale (ratio)","score":0.5332000255584717},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5236999988555908},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.5055000185966492},{"id":"https://openalex.org/keywords/kernel","display_name":"Kernel (algebra)","score":0.4867999851703644},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.4090999960899353}],"concepts":[{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.6996999979019165},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6051999926567078},{"id":"https://openalex.org/C106251023","wikidata":"https://www.wikidata.org/wiki/Q851989","display_name":"Porting","level":3,"score":0.6007000207901001},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.5332000255584717},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5236999988555908},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.5055000185966492},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.4867999851703644},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.4090999960899353},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.40540000796318054},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.3702999949455261},{"id":"https://openalex.org/C26713055","wikidata":"https://www.wikidata.org/wiki/Q245962","display_name":"Implementation","level":2,"score":0.35749998688697815},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.34119999408721924},{"id":"https://openalex.org/C118615104","wikidata":"https://www.wikidata.org/wiki/Q121416","display_name":"Discrete mathematics","level":1,"score":0.3285999894142151},{"id":"https://openalex.org/C45347329","wikidata":"https://www.wikidata.org/wiki/Q5166604","display_name":"Convolution (computer science)","level":3,"score":0.32670000195503235},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.29820001125335693},{"id":"https://openalex.org/C40506919","wikidata":"https://www.wikidata.org/wiki/Q7452469","display_name":"Sequence learning","level":2,"score":0.2903999984264374},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.289900004863739},{"id":"https://openalex.org/C133875982","wikidata":"https://www.wikidata.org/wiki/Q764810","display_name":"Shared memory","level":2,"score":0.28450000286102295},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.27639999985694885},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.26899999380111694},{"id":"https://openalex.org/C2780990831","wikidata":"https://www.wikidata.org/wiki/Q319141","display_name":"Conjecture","level":2,"score":0.2662000060081482},{"id":"https://openalex.org/C176649486","wikidata":"https://www.wikidata.org/wiki/Q2308807","display_name":"Memory management","level":3,"score":0.2612999975681305},{"id":"https://openalex.org/C114614502","wikidata":"https://www.wikidata.org/wiki/Q76592","display_name":"Combinatorics","level":1,"score":0.257999986410141},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.2565999925136566}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2607.17644","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.17644","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2607.17644","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.17644","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/16","display_name":"Peace, Justice and strong institutions","score":0.45379167795181274}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Multi-head":[0],"Latent":[1],"Attention":[2],"(MLA)":[3],"ships":[4],"two":[5,132],"implementations":[6],"in":[7,51,82,135,236],"Megatron-Core:":[8],"an":[9,16],"explicit":[10,199,206],"form":[11,18,73],"used":[12],"for":[13],"training":[14,39,155],"and":[15,47,65,137,139,148,209,216,227],"absorbed":[17,72,167],"--":[19,30,78,96,230],"which":[20],"slashes":[21],"collective":[22,195],"communication":[23,170,196],"by":[24,100],"gathering":[25],"only":[26,50],"the":[27,61,71,91,114,145,166,173,182,237],"compressed":[28],"latent":[29],"that":[31],"is":[32,63,74,203,241],"fully":[33],"implemented":[34],"but":[35,171],"hard-asserted":[36],"out":[37],"of":[38],"(the":[40],"forward":[41],"opens":[42],"with":[43,151],"\"assert":[44],"not":[45],"(self.training":[46],"self.cache_mla_latents)\"),":[48],"allowed":[49],"inference":[52],"decode.":[53],"The":[54],"library":[55],"documents":[56],"no":[57,152],"reason.":[58],"We":[59,157],"show":[60],"restriction":[62,147],"well-founded":[64],"quantify":[66],"why:":[67],"ported":[68],"to":[69,103,117,125,205,211],"training,":[70],"a":[75,121,218],"memory":[76,99,200],"trap":[77],"its":[79],"intermediates":[80],"live":[81],"n_h":[83],"x":[84],"d_kv":[85],"dimensions":[86],"per":[87],"token,":[88],"larger":[89],"than":[90],"per-head":[92,178],"K/V":[93,179],"they":[94],"replace":[95],"inflating":[97],"activation":[98],"20-34%,":[101],"up":[102],"9.2":[104],"GB":[105,119],"at":[106,189,207,214,232],"DeepSeek-V3":[107,191],"scale":[108],"(n_h=128,":[109],"seq=16384,":[110],"SP=8,":[111],"eager":[112],"kernel;":[113],"gap":[115],"widens":[116],"19.2":[118],"under":[120,217],"fused":[122,219],"kernel),":[123],"enough":[124],"change":[126],"device-fit.":[127],"This":[128],"measurement,":[129],"validated":[130],"on":[131,141],"axes":[133],"(linear":[134],"seq":[136],"n_h)":[138],"cross-verified":[140],"NVIDIA":[142],"A100,":[143],"explains":[144],"otherwise-undocumented":[146],"leaves":[149],"practitioners":[150],"low-communication":[153],"MLA":[154,240],"path.":[156],"then":[158],"provide":[159],"one.":[160],"LAGA":[161,193],"(Latent":[162],"All-Gather":[163],"Attention)":[164],"keeps":[165],"form's":[168],"latent-gather":[169],"rejects":[172],"absorb":[174],"reformulation,":[175],"instead":[176],"reconstructing":[177],"locally":[180],"from":[181],"gathered":[183],"latent.":[184],"On":[185],"8x":[186],"Ascend":[187],"910B":[188],"real":[190],"dimensions,":[192],"cuts":[194],"1.98x,":[197],"matches":[198],"within":[201,212],"0.5%,":[202],"bit-identical":[204],"SP=1":[208],"equivalent":[210],"1e-3":[213],"SP=2-8,":[215],"attention":[220],"kernel":[221],"improves":[222],"attention-block":[223],"throughput":[224],"1.04-1.06x":[225],"single-node":[226],"1.07-1.24x":[228],"cross-node":[229,238],"leading":[231],"all":[233],"sequence":[234],"lengths":[235],"regime":[239],"deployed":[242],"for.":[243]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-07-22T00:00:00"}