GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding
For LLM inference on diverse hardware, GQLA provides a single-weight solution that adapts to different compute-bandwidth ratios, enabling efficient decoding on both high-end and commodity GPUs.
GQLA modifies MLA to expose two decoding paths (MQA-absorb and GQA) from the same weights, enabling hardware-adaptive LLM decoding that matches H100 and H20 rooflines without retraining. On LLaMA-3-8B, it compresses KV cache to 28.125% of GQA baseline on the MQA path while preserving GQA traffic on the per-group path.
Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.