{"id":"https://openalex.org/W4399115604","doi":"https://doi.org/10.48550/arxiv.2405.16406","title":"SpinQuant: LLM quantization with learned rotations","display_name":"SpinQuant: LLM quantization with learned rotations","publication_year":2024,"publication_date":"2024-05-26","ids":{"openalex":"https://openalex.org/W4399115604","doi":"https://doi.org/10.48550/arxiv.2405.16406"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2405.16406","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.16406","pdf_url":"https://arxiv.org/pdf/2405.16406","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,"listed_in":[],"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2405.16406","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5085241696","display_name":"Zechun Liu","orcid":"https://orcid.org/0000-0001-5856-2506","observed_orcids":["https://orcid.org/0000-0001-5856-2506"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Zechun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5109628225","display_name":"Changsheng Zhao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Changsheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5056063665","display_name":"Igor Fedorov","orcid":"https://orcid.org/0000-0002-7235-4823","observed_orcids":["https://orcid.org/0000-0002-7235-4823"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Fedorov, Igor","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5015226839","display_name":"Bilge Soran","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Soran, Bilge","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5046749324","display_name":"Dhruv Choudhary","orcid":"https://orcid.org/0000-0002-4520-765X","observed_orcids":["https://orcid.org/0000-0002-4520-765X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Choudhary, Dhruv","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5039488725","display_name":"Raghuraman Krishnamoorthi","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Krishnamoorthi, Raghuraman","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5016704219","display_name":"Vikas Chandra","orcid":"https://orcid.org/0009-0005-4996-8455","observed_orcids":["https://orcid.org/0009-0005-4996-8455"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chandra, Vikas","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5106668475","display_name":"Yuandong Tian","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tian, Yuandong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5082671220","display_name":"Tijmen Blankevoort","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Blankevoort, Tijmen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":371,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.9707000255584717,"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"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9707000255584717,"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"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.0044999998062849045,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"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/T10363","display_name":"Low-power high-performance VLSI design","score":0.0026000000070780516,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/post-training-quantization","display_name":"post-training quantization","score":0.9390000104904175},{"id":"https://openalex.org/keywords/outlier-removal","display_name":"outlier removal","score":0.6940000057220459},{"id":"https://openalex.org/keywords/kv-cache-quantization","display_name":"KV cache quantization","score":0.8259999752044678},{"id":"https://openalex.org/keywords/activation-quantization","display_name":"activation quantization","score":0.8169999718666077},{"id":"https://openalex.org/keywords/weight-quantization","display_name":"weight quantization","score":0.8880000114440918},{"id":"https://openalex.org/keywords/zero-shot-reasoning","display_name":"zero-shot reasoning","score":0.8289999961853027},{"id":"https://openalex.org/keywords/4-bit-quantization","display_name":"4-bit quantization","score":0.8569999933242798},{"id":"https://openalex.org/keywords/llama-2","display_name":"Llama 2","score":0.7319999933242798},{"id":"https://openalex.org/keywords/llama-3","display_name":"Llama 3","score":0.9520000219345093}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6334540843963623},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4374794065952301},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.33965691924095154},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.32929906249046326},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.30254966020584106}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2405.16406","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.16406","pdf_url":"https://arxiv.org/pdf/2405.16406","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,"listed_in":[],"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2405.16406","doi":"https://doi.org/10.48550/arxiv.2405.16406","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2405.16406","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,"listed_in":[],"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2405.16406","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.16406","pdf_url":"https://arxiv.org/pdf/2405.16406","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,"listed_in":[],"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4399115604.pdf","grobid_xml":"https://content.openalex.org/works/W4399115604.grobid-xml"},"referenced_works_count":28,"referenced_works":["https://openalex.org/W2525332836","https://openalex.org/W2794325560","https://openalex.org/W2809624076","https://openalex.org/W2890894339","https://openalex.org/W2936262349","https://openalex.org/W2950700477","https://openalex.org/W2953271402","https://openalex.org/W2995455260","https://openalex.org/W4307934016","https://openalex.org/W4322718191","https://openalex.org/W4327810158","https://openalex.org/W4378770729","https://openalex.org/W4379260375","https://openalex.org/W4380714727","https://openalex.org/W4384918448","https://openalex.org/W4386185625","https://openalex.org/W4386231419","https://openalex.org/W4387561528","https://openalex.org/W4387635190","https://openalex.org/W4388093177","https://openalex.org/W4389519530","https://openalex.org/W4389524393","https://openalex.org/W4390041933","https://openalex.org/W4390833798","https://openalex.org/W4392181736","https://openalex.org/W4399061648","https://openalex.org/W7079896301","https://openalex.org/W7133204437"],"related_works":["https://openalex.org/W7128722776","https://openalex.org/W7154352817","https://openalex.org/W4405033264","https://openalex.org/W7147106439","https://openalex.org/W7222993696","https://openalex.org/W7214963470","https://openalex.org/W4412889461","https://openalex.org/W4415796210","https://openalex.org/W7117983409","https://openalex.org/W7155044139"],"abstract_inverted_index":{"Post-training":[0],"quantization":[1,30,69,83,119],"(PTQ)":[2],"techniques":[3],"applied":[4],"to":[5,28,60,80,89,137,167,179,185,190,193],"weights,":[6],"activations,":[7],"and":[8,17,44,123,151],"the":[9,127,142,183],"KV":[10],"cache":[11],"greatly":[12],"reduce":[13],"memory":[14],"usage,":[15],"latency,":[16],"power":[18],"consumption":[19],"of":[20,54,120],"Large":[21],"Language":[22],"Models":[23],"(LLMs),":[24],"but":[25],"may":[26],"lead":[27,59,79],"large":[29],"errors":[31],"when":[32],"outliers":[33,43],"are":[34,177],"present.":[35],"Rotating":[36],"activation":[37],"or":[38],"weight":[39],"matrices":[40,111],"helps":[41],"remove":[42,168],"benefits":[45],"quantization.":[46],"In":[47,71,170],"this":[48],"work,":[49],"we":[50,73,101],"identify":[51],"a":[52,99,104],"collection":[53],"applicable":[55],"rotation":[56,110],"parameterizations":[57],"that":[58,75,107,176],"identical":[61],"outputs":[62],"in":[63,93],"full-precision":[64],"Transformer":[65],"architectures":[66],"while":[67],"enhancing":[68],"accuracy.":[70,116],"addition,":[72],"find":[74],"some":[76],"random":[77,165],"rotations":[78,166],"much":[81],"better":[82],"than":[84],"others,":[85],"with":[86,134],"an":[87],"up":[88,189],"13":[90],"points":[91,140,150],"difference":[92],"downstream":[94],"zero-shot":[95,131],"reasoning":[96,132],"performance.":[97],"As":[98],"result,":[100],"propose":[102],"SpinQuant,":[103],"novel":[105],"approach":[106],"incorporates":[108],"learned":[109],"for":[112,172],"optimal":[113],"quantized":[114],"network":[115],"With":[117],"4-bit":[118],"weight,":[121],"activation,":[122],"KV-cache,":[124],"SpinQuant":[125,157,181],"narrows":[126],"accuracy":[128],"gap":[129,184],"on":[130,141],"tasks":[133],"full":[135,186],"precision":[136,187],"merely":[138],"2.9":[139],"LLaMA-2":[143],"7B":[144],"model,":[145],"surpassing":[146],"LLM-QAT":[147],"by":[148,153,188],"19.1":[149],"SmoothQuant":[152],"25.0":[154],"points.":[155],"Furthermore,":[156],"also":[158],"outperforms":[159],"concurrent":[160],"work":[161],"QuaRot,":[162],"which":[163],"applies":[164],"outliers.":[169],"particular,":[171],"LLaMA-3":[173],"8B":[174],"models":[175],"hard":[178],"quantize,":[180],"reduces":[182],"45.1%":[191],"relative":[192],"QuaRot.":[194],"Code":[195],"is":[196],"available":[197],"at":[198],"https://github.com/facebookresearch/SpinQuant.":[199]},"counts_by_year":[{"year":2026,"cited_by_count":189},{"year":2025,"cited_by_count":155},{"year":2024,"cited_by_count":27}],"updated_date":"2026-11-12T12:28:57.059296","created_date":"2024-05-29T00:00:00"}