KV cache quantization silently erodes LLM safety alignment via vulnerable low-dimensional subspaces, diagnosed by Per-Channel Reduction into three failure modes and mitigated training-free with up to 97% recovery.
hub Canonical reference
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale
Canonical reference. 89% of citing Pith papers cite this work as background.
53
Pith papers citing it
Background
89% of classified citations
abstract
Large language models have been widely adopted but require significant GPU memory for inference. We develop a procedure for Int8 matrix multiplication for feed-forward and attention projection layers in transformers, which cut the memory needed for inference by half while retaining full precision performance. With our method, a 175B parameter 16/32-bit checkpoint can be loaded, converted to Int8, and used immediately without performance degradation. This is made possible by understanding and working around properties of highly systematic emergent features in transformer language models that dominate attention and transformer predictive performance. To cope with these features, we develop a two-part quantization procedure, LLM.int8(). We first use vector-wise quantization with separate normalization constants for each inner product in the matrix multiplication, to quantize most of the features. However, for the emergent outliers, we also include a new mixed-precision decomposition scheme, which isolates the outlier feature dimensions into a 16-bit matrix multiplication while still more than 99.9% of values are multiplied in 8-bit. Using LLM.int8(), we show empirically it is possible to perform inference in LLMs with up to 175B parameters without any performance degradation. This result makes such models much more accessible, for example making it possible to use OPT-175B/BLOOM on a single server with consumer GPUs. We open-source our software.
hub tools
citation-role summary
background 8
method 1
citation-polarity summary
representative citing papers
Activation outliers quantified by γ = ||μ||/||σ|| cause feature death in SAEs via initialization shifts; mean-centering eliminates it.
A hybrid first-order then zeroth-order optimization approach improves robustness of safety-aligned LLMs while preserving utility, with layer-wise sensitivity estimation for efficiency.
LlamaWeb is a WebGPU backend for llama.cpp that uses static memory planning, tunable kernels, and templated multi-precision support to cut memory use by 29-33% and raise decode throughput by 45-69% versus prior browser frameworks on tested hardware.
The paper introduces an outlier-injection attack that induces targeted weight collapse in LLMs under advanced quantization schemes including AWQ, GPTQ, and GGUF I-quants.
A single fused int4 KV cache kernel on Apple Silicon outperforms fp16 in latency with 3x memory compression and near-zero quality loss on tested models.
ENEC delivers 3.43X higher throughput than DietGPU and 1.12X better compression ratio than nvCOMP for lossless model weight compression on Ascend NPUs, yielding up to 6.3X end-to-end inference speedup.
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.
Chronos pretrains transformer models on tokenized time series to deliver strong zero-shot forecasting across diverse domains.
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.
Speculative sampling accelerates LLM decoding 2-2.5x by letting a draft model propose short sequences that the target model scores in parallel, then applies modified rejection sampling to keep the exact target distribution.
GPTQ quantizes 175B-parameter GPT models to 3-4 bits per weight in one shot using approximate second-order information, achieving negligible accuracy degradation and 3-4x inference speedups.
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
MCWC aligns permutation-symmetric blocks across layers to enable sequential prediction and residual entropy coding, improving rate-accuracy tradeoffs versus quantization and prior codecs on language and vision models.
ThriftAttention recovers 89.1% of the FP16 quality gap versus pure FP4 attention by running only 5% of query-key blocks in FP16 on long-context benchmarks.
TRACE uses cross-layer candidate trajectories inside frozen LLMs to dynamically select and apply one of three correction operators, delivering mean gains of +12.26 MC1 and +8.65 MC2 points across 15 models and 3 benchmarks with no regressions.
LoKA enables practical FP8 use in numerically sensitive large recommendation models via online profiling of activations, reusable model modifications for stability, and dynamic kernel dispatching.
OSAQ suppresses weight outliers in LLMs via a closed-form additive transformation from the Hessian's stable null space, improving 2-bit quantization perplexity by over 40% versus vanilla GPTQ with no inference overhead.
Sharpness-aware pretraining and related flat-minima interventions reduce catastrophic forgetting by up to 80% after post-training across 20M-150M models and by 31-40% at 1B scale.
Fine-tuning 7B code LLMs on a custom multi-file DSL dataset achieves structural fidelity of 1.00, high exact-match accuracy, and practical utility validated by expert survey and execution checks.
GSQ uses Gumbel-Softmax to optimize scalar quantization grids for LLMs, closing most of the accuracy gap to vector methods like QTIP at 2-3 bits per parameter while using symmetric scalar grids compatible with existing kernels.
Redefining residual errors to include compensation-aware discrepancies and realigning calibration to full-precision outputs improves GPTQ and GPTAQ performance on LLMs.
Sol-RL decouples FP4-based candidate exploration from BF16 policy optimization in diffusion RL, delivering up to 4.64x faster convergence with maintained or superior alignment performance on models like FLUX.1 and SD3.5.
citing papers explorer
No citing papers match the current filters.