{"paper":{"title":"Multi-Scale Dequant: Eliminating Dequantization Bottleneck via Activation Decomposition for Efficient LLM Inference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Decomposing BF16 activations into low-precision scales lets quantized weights multiply directly via native GEMM.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"stat.ML","authors_text":"Chengqiu Hu, Fangzheng Miao, Jun Li, Junyi Fan, Lingchao Zheng, Qichen Liao, Rui Shi, Yuwei Fan","submitted_at":"2026-05-13T09:49:56Z","abstract_excerpt":"Quantization is essential for efficient large language model (LLM) inference, yet the dequantization step-converting low-bit weights back to high-precision for matrix multiplication has become a critical bottleneck on modern AI accelerators. On architectures with decoupled compute units (e.g., Ascend NPUs), dequantization operations can consume more cycles than the matrix multiplication itself, leaving the high-throughput tensor cores underutilized. This paper presents Multi-Scale Dequant (MSD), a quantization framework that removes weight/KV dequantization from the GEMM critical path. Instead"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"For INT8 weights (W4A16), two-pass INT8 decomposition achieves near 16 effective bits. For MXFP4 weights (W4A16), two-pass MXFP4 decomposition yields near 6.6 effective bits with error bound 1/64 per block surpassing single-pass MXFP8 while maintaining the same effective GEMM compute time.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the multi-scale activation decomposition can be implemented with native hardware-accelerated GEMM without introducing pipeline stalls or accuracy loss beyond the derived bounds, and that the closed-form latency models accurately predict real hardware behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MSD eliminates dequantization from the GEMM path by decomposing BF16 activations into multiple low-precision parts that multiply directly with INT8 or MXFP4 weights, achieving near-16 effective bits for INT8 and 6.6 for MXFP4 with reduced HBM traffic.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Decomposing BF16 activations into low-precision scales lets quantized weights multiply directly via native GEMM.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1c51d3ec24f85f8de4ab16604dd1578181b630d655d0ed344186bf5547b845af"},"source":{"id":"2605.13915","kind":"arxiv","version":1},"verdict":{"id":"4b1150f9-d12f-4da5-a483-457377a4b710","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:53:49.087136Z","strongest_claim":"For INT8 weights (W4A16), two-pass INT8 decomposition achieves near 16 effective bits. For MXFP4 weights (W4A16), two-pass MXFP4 decomposition yields near 6.6 effective bits with error bound 1/64 per block surpassing single-pass MXFP8 while maintaining the same effective GEMM compute time.","one_line_summary":"MSD eliminates dequantization from the GEMM path by decomposing BF16 activations into multiple low-precision parts that multiply directly with INT8 or MXFP4 weights, achieving near-16 effective bits for INT8 and 6.6 for MXFP4 with reduced HBM traffic.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the multi-scale activation decomposition can be implemented with native hardware-accelerated GEMM without introducing pipeline stalls or accuracy loss beyond the derived bounds, and that the closed-form latency models accurately predict real hardware behavior.","pith_extraction_headline":"Decomposing BF16 activations into low-precision scales lets quantized weights multiply directly via native GEMM."},"references":{"count":30,"sample":[{"doi":"","year":2024,"title":"DeepSeek.FlashMLA: Efficient MLA for Large Language Models. Technical Report, 2024. https://github.com/deepseek-ai/FlashMLA","work_id":"93f944c5-ef20-4968-bded-2807bc81f255","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Technical Blog, 2025.https://github.com/deepseek-ai/FlashMLA/blob/main/docs/ 20250929-hopper-fp8-sparse-deep-dive.md","work_id":"ab768edf-a287-47b5-ac22-b0b99458446e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers","work_id":"19ed8c44-202a-48f6-8169-637d5a5f2408","ref_index":3,"cited_arxiv_id":"2210.17323","is_internal_anchor":true},{"doi":"","year":2023,"title":"AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration","work_id":"ea9d1d72-db24-4cae-8c89-4ecd83dd87c1","ref_index":4,"cited_arxiv_id":"2306.00978","is_internal_anchor":true},{"doi":"","year":2024,"title":"Castro, Jiale Chen, Torsten Hoefler, and Dan Alistarh","work_id":"19ce2d31-0683-4f81-9db6-b9ca47099921","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":30,"snapshot_sha256":"d39d51c469102fe8b184d2fbc89ed7dc8b3818489b20245e2d902f7caf5c1e76","internal_anchors":6},"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"}