{"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/towards-economical-inference-enabling","title":"Towards Economical Inference: Enabling DeepSeek's Multi-Head Latent Attention in Any Transformer-based LLMs","arxiv_id":"2502.14837","date":"2025-02-20","proceeding":null,"authors":["Tao Ji","Bin Guo","Yuanbin Wu","Qipeng Guo","Lixing Shen","Zhan Chen","Xipeng Qiu","Qi Zhang","Tao Gui"],"abstract":"Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.","url_abs":"https://arxiv.org/abs/2502.14837v1","url_pdf":"https://arxiv.org/pdf/2502.14837v1.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":"towards-economical-inference-enabling","repo_url":"https://github.com/JT-Ushio/MHA2MLA","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[{"task_slug":"quantization","task_name":"Quantization"}],"methods":[{"method_slug":"attention","method_name":"Attention"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"feedforward-network","method_name":"Feedforward Network"},{"method_slug":"grouped-query-attention","method_name":"Grouped-query attention"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"softmax","method_name":"Softmax"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2502.14837","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2502.14837"}},"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/JT-Ushio/MHA2MLA","reach":{"status":"ok","spdx":"Apache-2.0"}}],"summary":{"unverified":15},"by_repo_kind":{"official":{"samples":15,"ran":0,"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":"d1722a47a6e68260","entry":"SVD","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patch_func.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patch_func.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"d1722a47a6e68260"}},{"code_sha256_prefix":"bd8982c35600a2b7","entry":"create_custom_apply_rotary_pos_emb","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_llama.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_llama.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"bd8982c35600a2b7"}},{"code_sha256_prefix":"f73327d915050f1a","entry":"create_custom_apply_rotary_pos_emb","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen2.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen2.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"f73327d915050f1a"}},{"code_sha256_prefix":"4b3aa5813125a737","entry":"create_custom_apply_rotary_pos_emb","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen3.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen3.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"4b3aa5813125a737"}},{"code_sha256_prefix":"e3be3d316dce65ea","entry":"create_custom_apply_rotary_pos_emb","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen3_moe.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen3_moe.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"e3be3d316dce65ea"}},{"code_sha256_prefix":"3e838680cc326966","entry":"create_hook_fn","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/2_norm.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/2_norm.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"3e838680cc326966"}},{"code_sha256_prefix":"19448606f10d10d8","entry":"custom_LlamaAttention_forward","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_llama.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_llama.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"19448606f10d10d8"}},{"code_sha256_prefix":"7ac64d7242a38034","entry":"custom_Qwen2Attention_forward","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen2.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen2.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"7ac64d7242a38034"}},{"code_sha256_prefix":"ccace102b231da7d","entry":"custom_Qwen3Attention_forward","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen3.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen3.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"ccace102b231da7d"}},{"code_sha256_prefix":"88a635fd8ae25685","entry":"custom_Qwen3MoeAttention_forward","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_qwen3_moe.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_qwen3_moe.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"88a635fd8ae25685"}},{"code_sha256_prefix":"c72933732661bbe4","entry":"generate_random_input_ids","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"eval/benchmark_latency_memory.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/eval/benchmark_latency_memory.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"c72933732661bbe4"}},{"code_sha256_prefix":"c836c0fa19b74615","entry":"mha2mla_ref","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/mla_triton_kernel.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/mla_triton_kernel.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"c836c0fa19b74615"}},{"code_sha256_prefix":"818485f3147eb543","entry":"partial_rope_mask","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patch_func.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patch_func.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"818485f3147eb543"}},{"code_sha256_prefix":"937f754d0af3a99a","entry":"reorder_matrix_rows","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patching_model_load.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patching_model_load.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"937f754d0af3a99a"}},{"code_sha256_prefix":"1b387fd9635cb3ad","entry":"svd_low_rank_approx","repo":"JT-Ushio/MHA2MLA","repo_kind":"official","path":"src/mha2mla/patch_func.py","file_url":"https://github.com/JT-Ushio/MHA2MLA/blob/HEAD/src/mha2mla/patch_func.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"Apache-2.0","inline_ok":true,"mcp_get_code":{"code_sha256":"1b387fd9635cb3ad"}}]},"arxiv_metadata":null,"syntology_extracted_results":null}