{"id":"https://openalex.org/W7162545091","doi":"https://doi.org/10.48550/arxiv.2605.26189","title":"Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training","display_name":"Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training","publication_year":2026,"publication_date":"2026-05-25","ids":{"openalex":"https://openalex.org/W7162545091","doi":"https://doi.org/10.48550/arxiv.2605.26189"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.26189","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26189","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.2605.26189","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137156678","display_name":"Yingying Cheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cheng, Yingying","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137103549","display_name":"Jinquan Shi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shi, Jinquan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137146571","display_name":"Li Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137116556","display_name":"Zhiyang He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"He, Zhiyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022030361","display_name":"Zhaoyi Sun","orcid":"https://orcid.org/0009-0003-8197-1465"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Zhaoyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137088772","display_name":"Fan Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Fan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137176292","display_name":"Jie Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Jie","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":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":0.2867000102996826,"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":0.2867000102996826,"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/T11697","display_name":"Numerical Methods and Algorithms","score":0.11969999969005585,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T11181","display_name":"Advanced Data Storage Technologies","score":0.08009999990463257,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/forgetting","display_name":"Forgetting","score":0.7160000205039978},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.6499000191688538},{"id":"https://openalex.org/keywords/scale","display_name":"Scale (ratio)","score":0.45829999446868896},{"id":"https://openalex.org/keywords/scaling","display_name":"Scaling","score":0.447299987077713},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.41510000824928284},{"id":"https://openalex.org/keywords/estimation","display_name":"Estimation","score":0.3328999876976013}],"concepts":[{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.7160000205039978},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.6499000191688538},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5997999906539917},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5651000142097473},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.45829999446868896},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.447299987077713},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4334999918937683},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.41510000824928284},{"id":"https://openalex.org/C96250715","wikidata":"https://www.wikidata.org/wiki/Q965330","display_name":"Estimation","level":2,"score":0.3328999876976013},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.3163999915122986},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.29350000619888306},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.29019999504089355},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.2660999894142151},{"id":"https://openalex.org/C48677424","wikidata":"https://www.wikidata.org/wiki/Q6888088","display_name":"Mode (computer interface)","level":2,"score":0.26440000534057617},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.2621999979019165},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.26159998774528503},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.2513999938964844}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.26189","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26189","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.2605.26189","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26189","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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Quantization-aware":[0],"training":[1,19,84,142],"(QAT)":[2],"with":[3,91,140],"low-bit":[4],"floating-point":[5],"formats":[6],"enables":[7],"efficient":[8],"LLM":[9],"deployment,":[10],"yet":[11],"introduces":[12],"subtle":[13],"failure":[14,50],"modes":[15],"invisible":[16],"to":[17],"standard":[18],"metrics.":[20],"We":[21,87],"present":[22],"a":[23,92,98,106,136,141],"systematic":[24],"study":[25],"of":[26,35,78,145],"HiF8":[27],"W8A8":[28],"QAT":[29,112],"for":[30],"OpenPangu-Embedded-1B":[31],"through":[32],"the":[33],"lens":[34],"Delayed":[36],"Tensor":[37],"Scaling":[38],"(DTS).":[39],"Across":[40],"eight":[41],"controlled":[42],"experiments,":[43],"we":[44],"identify":[45],"and":[46,65,102,119,131],"disentangle":[47],"two":[48],"orthogonal":[49],"modes:":[51],"(i)amax":[52],"saturation,":[53],"where":[54,68],"delayed":[55],"scale":[56],"estimates":[57],"silently":[58],"corrupt":[59],"knowledge-sensitive":[60],"representations":[61],"via":[62,105],"forward-pass":[63],"clipping,":[64],"(ii)catastrophic":[66],"forgetting,":[67],"an":[69],"aggressive":[70],"learning":[71],"rate":[72],"overwrites":[73],"pretrained":[74],"commonsense":[75],"knowledge":[76],"independently":[77],"quantization.":[79],"Neither":[80],"is":[81],"detectable":[82],"from":[83],"loss":[85,143],"alone.":[86],"address":[88],"amax":[89],"saturation":[90],"conservative":[93],"max-algorithm":[94],"DTS":[95],"strategy":[96],"over":[97,148],"64-step":[99],"history":[100],"window,":[101],"mitigate":[103],"forgetting":[104],"500-step":[107],"BF16":[108,138],"warmup":[109],"followed":[110],"by":[111],"at":[113],"lr=10^{-5}.":[114],"Both":[115],"fixes":[116],"are":[117],"necessary":[118],"sufficient:":[120],"our":[121],"final":[122],"configuration":[123],"achieves":[124],"0.43%":[125],"MMLU":[126],"drop,":[127,130],"0.58%":[128],"HellaSwag":[129],"0.22%":[132],"ARC-Challenge":[133],"drop":[134],"versus":[135],"matched":[137],"baseline,":[139],"APE":[144],"only":[146],"0.11%":[147],"10,000":[149],"steps.":[150]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-28T00:00:00"}