Prism Transformer uses a progressive increase in attention heads with depth to form a local-to-global hierarchy, reporting lower validation loss and gains on zero-shot benchmarks versus uniform-head baselines at 124M-757M scales.
super hub Canonical reference
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
Canonical reference. 70% of citing Pith papers cite this work as background.
127
Pith papers citing it
Background
70% of classified citations
abstract
Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.
hub tools
citation-role summary
background 20
method 6
other 1
citation-polarity summary
claims ledger
- abstract Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show t
authors
co-cited works
representative citing papers
Grouped Query Experts applies per-group MoE routing to query heads in GQA, matching baseline accuracy while activating half the query heads on 250M models trained for 30B tokens.
STAR-KV applies differentiable soft thresholding for per-head and per-block adaptive low-rank KV cache compression, combined with hybrid decomposition and low-rank-aware quantization, achieving up to 75% compression and 3.1x throughput gains.
Chronicle is the first model jointly pretrained from scratch on text and time series in a unified transformer that matches a comparable language model on NLU tasks and sets new bars for time series classification and multimodal forecasting.
LLMForge is a NAS framework with Infinite-Head Attention, a Forge-Former surrogate, and Forge-DSE engine that discovers hardware-specific architectures for edge language models, yielding variants with improved accuracy, energy, or latency on different substrates.
Derives μP scalings for GQA via promoted spectral-norm definition of feature learning and a modified norm preserving scaling laws for non-full-rank matrices, with experiments showing learning-rate transfer.
Power capping is illusory in LLM decode as memory-bound operation leaves power headroom untouched on 700 W GPUs, while SM clock locking saves up to 32% energy and three DVFS classes appear across attention types.
3D-ARD+ unifies autoregressive token prediction with diffusion-based 3D latent generation to co-produce indoor scene layouts and object geometries that follow complex text-specified spatial and semantic constraints.
RPA kernel for TPUs achieves 86% MBU in decode and 73% MFU in prefill on Llama 3 8B via tiling for ragged memory, fused pipelines, and specialized compilation for prefill/decode workloads.
ATLAS is the first silicon-validated simulation framework for 3D-DRAM LLM accelerators, achieving under 8.57% error and over 97% correlation with real hardware while supporting design exploration.
MixAtlas uses CLIP-based decomposition and Gaussian process optimization on small proxies to discover data mixtures that improve multimodal benchmark performance by up to 17.6% and transfer to larger models with faster convergence.
This work delivers the first measurements of performance-energy trade-offs across four multi-request LLM workflow patterns on A100 GPUs using vLLM and Parrot.
SnapStream deploys sparse KV attention in a production inference system on dataflow accelerators, delivering 4x on-chip memory savings for DeepSeek-671B at 128k context with up to 1832 tokens/sec and minimal accuracy loss on LongBench-v2, AIME24, and LiveCodeBench.
DELTA partitions layers into full, delta, and sparse groups to select salient tokens via aggregated attention scores, matching full-attention accuracy on AIME and GPQA while cutting attended tokens up to 4.25x and achieving 1.54x speedup.
Dreamer 4 is the first agent to obtain diamonds in Minecraft from only offline data by reinforcement learning inside a scalable world model that accurately predicts game mechanics.
AVA-Bench evaluates vision foundation models by disentangling 14 atomic visual abilities with aligned training-test distributions to reveal precise ability fingerprints.
FastKV decouples prefill context reduction via Token-Selective Propagation from independent KV cache selection, delivering up to 1.82x prefill and 2.87x decoding speedups while matching decoding-only accuracy.
FlashAttention-3 achieves 1.5-2x speedup on H100 GPUs for attention, reaching 740 TFLOPs/s (75% utilization) in FP16 and near 1.2 PFLOPs/s in FP8 while cutting numerical error by 2.6x versus baseline FP8 attention.
Transformers and SSMs are unified through structured state space duality, producing a 2-8X faster Mamba-2 model that remains competitive with Transformers.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
A new shared video-image tokenizer enables large language models to surpass diffusion models on standard visual generation benchmarks.
One-step gradient delay is optimizer-dependent rather than intrinsically unstable, with Muon and error-feedback correction enabling async pipeline parallelism to match synchronous performance on models up to 10B parameters.
HGA uses RoPE-aware chunk summaries for two-level hierarchical routing to approximate dense causal attention at 3% sparsity with 0.01-0.02 nats quality gap, as a drop-in replacement requiring no retraining.
citing papers explorer
-
Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning
Reinforcement learning after SFT conversion narrows the performance gap between sliding-window attention and full self-attention on math reasoning benchmarks while preserving linear complexity.
-
ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning
ArborKV uses search-structure awareness to evict low-reuse KV states in Tree-of-Thoughts inference, delivering up to 4x memory savings with near-full accuracy retention.
-
ZAYA1-8B Technical Report
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
-
Voxtral Realtime
Voxtral Realtime is an end-to-end trained streaming ASR model that achieves Whisper-level transcription quality at 480ms delay after scaling pretraining across 13 languages.
-
SweetSpot: An Analytical Model for Predicting Energy Efficiency of LLM Inference
SweetSpot is an analytical model from Transformer computational and memory complexity that identifies energy minima at short-to-moderate inputs and medium outputs, achieving 1.79% MAPE on H100 GPU measurements across multiple LLMs.