[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-huawei-npus-get-4-bit-attention-without-a-retraining-tax":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},3934,"huawei-npus-get-4-bit-attention-without-a-retraining-tax","Huawei NPUs Get 4-bit Attention Without a Retraining Tax","A new training-free method squeezes FlashAttention down to 4-bit integers on Ascend hardware, cutting accuracy loss on standard benchmarks by more than half.","Researchers have found a way to run the attention mechanism inside large language models at 4-bit precision on Huawei's Ascend NPUs — no retraining required.\n\nThe technique, called HiFA4, targets FlashAttention's two core matrix multiplications and executes them as 4-bit HIF4 Cube operations while keeping the softmax state in 16-bit floating point. It bundles two fixes. Smooth-QK redistributes quantization error between the query and key matrices after rotary positional encoding, avoiding per-tile computation at inference time. P-Reordering fuses the softmax normalizer directly into the second matrix multiply, which the authors show eliminates a systematic probability-mass loss they measured across 3.6 million attention tiles in a Qwen3-8B trace. On Qwen3-8B, HiFA4 recovered 37.5% of the accuracy gap introduced by naive 4-bit quantization and cut the rate of predictions that disagreed with full-precision output from 16.3% to 8.2%. On Gemma2-9B, the method stayed within 0.7 percentage points of BF16 baseline. A scheduling analysis projects a 35.4% latency reduction over BF16 by fusing the normalizer into the matrix multiply, though the authors note on-hardware validation remains future work.\n\nMost quantization research targets Nvidia silicon, so a rigorous method benchmarked specifically against Ascend hardware fills a gap that matters as non-Nvidia AI infrastructure becomes harder to ignore. The accuracy-recovery numbers are meaningful: halving disagreement with full-precision output is the difference between a quantized model that is usable and one that is not.\n\nThe projected latency gains are the headline number here, but they come with an asterisk — \"preliminary scheduling analysis\" and \"future work\" are phrases that should temper enthusiasm until the actual hardware numbers arrive.","[\"ai\",\"hardware\",\"llm\",\"quantization\"]","2026-07-07T04:00:00.000Z","2026-07-07T12:35:23.906Z","2026-07-07T12:35:26.886Z","published",null,[],"ai",[24,26,27,28],"hardware","llm","quantization",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04302",0,{"sections":35},[36,40,45,50,55,59,64,69,74,78,83,87,92,97],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":44},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":46,"slug":47,"count":48,"latest_published_at":49},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":56,"slug":26,"count":57,"latest_published_at":58},"Hardware",122,"2026-07-14T19:46:26.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":18},"Dev Tools","dev-tools",59,{"name":79,"slug":80,"count":81,"latest_published_at":82},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":84,"slug":85,"count":81,"latest_published_at":86},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]