Multi-Head Low-Rank Attention

Decoding efficiency depends on KV-cache traffic, computation, and parallelism. Multi-head Latent Attention (MLA) offers absorbed and head-expanded forms, but its head-specific projections limit grouped KV sharing. We introduce Group-Query Latent Attention (GQLA), whose group-shared up-projections preserve a joint latent and enable a shardable expanded cache. The same weights support equivalent MQA-absorb and GQA execution using existing kernels. Roofline analysis relates path preference to GPU compute-to-bandwidth ratio and tensor parallelism (TP): absorption trades traffic for computation and replicates the latent across ranks; GQA shards KV groups. Pretraining with eight query heads per group yields terminal training losses close to MLA's across the tested model scales at matched tokens. TransGQLA converts GQA and MLA checkpoints using calibration data. Without gradient updates, converted DeepSeek-V3.1-Base achieves 77.79% eight-task accuracy versus the source's 77.84%, while WikiText-2 perplexity rises from 3.21 to 6.05. On eight H20-3e GPUs, its GQA path improves output throughput over MQA-absorb by 22.42--34.97% at matched resident batches across 8K--64K inputs. Its smaller per-rank cache permits larger batches, giving 50.07% in a separate capacity-tuned comparison. These results support multiple execution paths as an architectural means of adapting decoding to hardware and parallelism.

Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, many practical conversion baselines rely on weight-only low-rank approximations (e.g., SVD-style initializations) and uniform rank allocation. They focus on minimizing the difference between weight matrices rather than on how those weights affect input activations, ignore the covariance structure of activations, and enforce uniform rank across layers, causing activation drift and degraded attention fidelity. To address these issues, we propose CARE, a Covariance-Aware, Rank-Enhanced MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i) activation-preserving factorization, which aligns the approximation with the actual input activations rather than just the weights; (ii) adjusted-rank allocation, which spreads a fixed KV budget across layers by giving more capacity to layers that need it most; and (iii) KV-parity mapping, which reparameterizes the converted K and V to fit the MLA format while keeping the KV-cache size unchanged. Our method outperforms a uniform-rank SVD baseline on Qwen3-4B/30B-A3B-Instruct-2507 and Llama-3.1-8B/70B-Instruct, reducing one-shot perplexity by up to 215x and improving mean accuracy by up to 1.70x at matched KV budgets. With a brief post-SVD healing fine-tune, we fully recover the original model's accuracy.

Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key-value (KV) cache by T. The larger cache limits how many sequences a GPU can decode at once and slows each decoding step, which reads the whole cache. We propose Hybrid Latent Attention (HLA), which keeps exact keys and values within a sliding window of W recent tokens and stores each older token as a compact latent that the query of each loop reads directly, without reconstructing keys and values. We uptrain HLA on Ouro looped models (T=4) with 1.4B and 2.6B parameters, keeping the pretrained weights frozen and training only the added parameters to reproduce the original attention. The cache shrinks by 10.7x per token, fitting 4.0-8.8x as many concurrent sequences per GPU, and decoding throughput improves by 2.5x at 1K-token contexts and by up to 7.4x at 16K. HLA retains over 97% of the original accuracy on math, knowledge and reasoning benchmarks, and 96-100% on long-context retrieval up to 16K tokens. After supervised fine-tuning, it performs on par with the fine-tuned original model on competition-level math.