{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/moa-mixture-of-sparse-attention-for-automatic","title":"MoA: Mixture of Sparse Attention for Automatic Large Language Model Compression","arxiv_id":"2406.14909","date":"2024-06-21","proceeding":null,"authors":["Tianyu Fu","Haofeng Huang","Xuefei Ning","Genghan Zhang","Boju Chen","Tianqi Wu","Hongyi Wang","Zixiao Huang","Shiyao Li","Shengen Yan","Guohao Dai","Huazhong Yang","Yu Wang"],"abstract":"Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by $3.9\\times$ with the same average attention span, boosting retrieval accuracy by $1.5-7.1\\times$ over the uniform-attention baseline across Vicuna-{7B,13B}, and Llama3-{8B,70B} models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from $9\\%-36\\%$ to within $5\\%$ across two long-context understanding benchmarks. MoA achieves a $1.2-1.4\\times$ GPU memory reduction, boosting decode throughput by $6.6-8.2\\times$ and $1.7-1.9\\times$ compared to FlashAttention2 and vLLM, with minimal impact on performance. Our code is available at \\url{https://github.com/thu-nics/MoA}.","url_abs":"https://arxiv.org/abs/2406.14909v2","url_pdf":"https://arxiv.org/pdf/2406.14909v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"moa-mixture-of-sparse-attention-for-automatic","repo_url":"https://github.com/thu-nics/moa","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":null,"task_name":"GPU"},{"task_slug":"language-modeling","task_name":"Language Modeling"},{"task_slug":"language-modelling","task_name":"Language Modelling"},{"task_slug":"large-language-model","task_name":"Large Language Model"},{"task_slug":"long-context-understanding","task_name":"Long-Context Understanding"},{"task_slug":"model-compression","task_name":"Model Compression"}],"methods":[{"method_slug":"attention","method_name":"Attention"},{"method_slug":"focus","method_name":"Focus"},{"method_slug":"softmax","method_name":"Softmax"}],"datasets_introduced":[{"slug":"moa","name":"MoA","full_name":"MoA_Long_ModelQA"}],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2406.14909","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2406.14909"}},"developers":"https://syntology.ai/developers","read_at":"2026-09-24T18:15:14+00:00","read_at_is":"when the build read Syntology's graph, not when any sample ran","claim":"Per-sample execution status on synthesized fixtures; not a correctness claim about the paper. Samples come from repositories linked to the paper, official or community; repo_kind says which.","repos":[{"provenance":"external:paperswithcode_snapshot_2025-07-28","url":"https://github.com/thu-nics/moa","reach":{"status":"ok","spdx":"MIT"}}],"summary":{"ran":19,"unverified":3},"by_repo_kind":{"official":{"samples":22,"ran":19,"repositories":1}},"repo_kind_vocabulary":{"official":"The archive marks this repository official for the paper","named_in_paper":"The archive records that the paper mentions this repository; it is not marked official","listed":"In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper","found_in_text":"Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted","community":"Not in the archive's code links for this paper; a community repository Syntology harvested"},"n_pointer_only_for_licence":0,"samples":[{"code_sha256_prefix":"b8258f95889e4952","entry":"Trainer_attention_grad_compute_loss","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/sparse_mask.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/sparse_mask.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"b8258f95889e4952"}},{"code_sha256_prefix":"29414013ce12967c","entry":"count_distinct_elements_per_row","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/density_calculation.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/density_calculation.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"29414013ce12967c"}},{"code_sha256_prefix":"061f1b20e4c107be","entry":"find_subtensor_position","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/dataset/utils.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/dataset/utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"061f1b20e4c107be"}},{"code_sha256_prefix":"149e4b82b7620c4b","entry":"gen_band_pattern","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/pattern.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/pattern.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"149e4b82b7620c4b"}},{"code_sha256_prefix":"6d1389e6b2d85469","entry":"gen_block_sparse_layout_mask","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/sparse_mask.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/sparse_mask.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"6d1389e6b2d85469"}},{"code_sha256_prefix":"7e1d14c04aa0dc2d","entry":"gen_global_pattern","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/pattern.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/pattern.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"7e1d14c04aa0dc2d"}},{"code_sha256_prefix":"2d84082c203b2b52","entry":"head_index_to_split_index","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/kernels/flash_decoding_moa.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/kernels/flash_decoding_moa.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"2d84082c203b2b52"}},{"code_sha256_prefix":"21ce5f5f68d41dad","entry":"hotpotQA_to_multi_round_qa","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/dataset/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/dataset/convert.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"21ce5f5f68d41dad"}},{"code_sha256_prefix":"2866d7c2b62c260d","entry":"layout_to_lut","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/convert.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"2866d7c2b62c260d"}},{"code_sha256_prefix":"e6a764e7c48e4500","entry":"lut_to_permutation","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/permutation_utils.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/permutation_utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"e6a764e7c48e4500"}},{"code_sha256_prefix":"8633f899e4d7407c","entry":"mask_to_layout","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/convert.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"8633f899e4d7407c"}},{"code_sha256_prefix":"0bfc9650b5be48e8","entry":"matrix_reserve_score","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/sparse_mask.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/sparse_mask.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"0bfc9650b5be48e8"}},{"code_sha256_prefix":"1ada47e15f33f461","entry":"moa_config_to_cache_config","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/cache_utils.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/cache_utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"1ada47e15f33f461"}},{"code_sha256_prefix":"2cd7b8febe087a7d","entry":"moa_config_to_permutation","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/permutation_utils.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/permutation_utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"2cd7b8febe087a7d"}},{"code_sha256_prefix":"3dca86fbcf33d8a4","entry":"multiNews_to_multi_round_qa","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/dataset/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/dataset/convert.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"3dca86fbcf33d8a4"}},{"code_sha256_prefix":"77491a1736272578","entry":"permute_lut","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/permutation_utils.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/permutation_utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"77491a1736272578"}},{"code_sha256_prefix":"89ec645a857ef9af","entry":"qasper_to_multi_round_qa","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/dataset/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/dataset/convert.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"89ec645a857ef9af"}},{"code_sha256_prefix":"cd7b0c5e2a16d0cd","entry":"streamingllm_attention_density","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/density_calculation.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/density_calculation.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":true,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"cd7b0c5e2a16d0cd"}},{"code_sha256_prefix":"758a42acae59b91f","entry":"streamingllm_kv_cache_density","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/density_calculation.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/density_calculation.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":true,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"758a42acae59b91f"}},{"code_sha256_prefix":"98558d05575a5d9f","entry":"block_sparse_to_dense","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/convert.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/convert.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"98558d05575a5d9f"}},{"code_sha256_prefix":"ef170fe22d7a924e","entry":"gen_causal_pattern","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/attention/pattern.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/attention/pattern.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"ef170fe22d7a924e"}},{"code_sha256_prefix":"e55fe24f97290938","entry":"update_model_function","repo":"thu-nics/moa","repo_kind":"official","path":"MoA/models/interface.py","file_url":"https://github.com/thu-nics/moa/blob/HEAD/MoA/models/interface.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"e55fe24f97290938"}}]},"arxiv_metadata":null,"syntology_extracted_results":null}