PapersAnon (@papers

PapersAnon

705 posts

PapersAnon

@papers_anon

Just a fan of acceleration. I read and post interesting papers. Let's all make it through.

SAITAMA

rentry.org/LocalModelsLin…

Joined February 2024

Following

2,368

Followers

Pinned
PapersAnon
@papers_anon
Jun 24, 2024
rentry.org/LocalModelsLin… Various links for ML and local models (not just LLMs) that's kept fairly updated. rentry.org/LocalModelsPap… ML papers I've read that I think are interesting. Also keep a text file at the top of all the abstracts for easy searching.
rentry.org
Local Models Related Links
/lmg/ Accelerate Guides Quick Start Guide Anon's tutorial for getting models running locally SillyTavern Guide Instructions for roleplaying via koboldcpp. Additional GNBF grammar usage LM Tuning...
26K
PapersAnon
@papers_anon
Dec 30, 2024
1.58-bit FLUX Extreme low-bit quantization of text-to-image FLUX.1-dev model. Achieves a 5.1× reduction in inference memory usage and 7.7× reduction in model storage. Calibration dataset remained image-data-free. Also developed a custom kernel optimized for 1.58- bit operations
61K
PapersAnon
@papers_anon
Mar 3, 2025
ByteScale: Efficient Scaling of LLM Training with a 2048K Context Length on More Than 12,000 GPUs From ByteDance. Introduces a novel parallelism strategy, namely Hybrid Data Parallelism, which unifies the inter- and intra-data partitioning with a dynamic mesh design. Evaluated
20K
PapersAnon
@papers_anon
Jan 14, 2025
Transformer^2: Self-adaptive LLMs Introduces a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference uses a dispatch system that identifies the task
14K
PapersAnon
@papers_anon
Apr 4, 2025
Inference-Time Scaling for Generalist Reward Modeling From DeepSeek. Investigates how to improve reward modeling with more inference compute for general queries. Proposes Self-Principled Critique Tuning to foster scalable reward generation behaviors in GRMs through online RL
56K
PapersAnon
@papers_anon
Feb 3, 2025
Efficient Reasoning with Hidden Thinking Proposes an efficient reasoning framework for multimodal LLMs. Uses an encoder to condense each intermediate CoT into a higher-level hidden representation. Uses a decoder to interpret the hidden representations into variable-length
17K
PapersAnon
@papers_anon
Sep 26, 2024
INT-FlashAttention: Enabling Flash Attention for INT8 Quantization Full INT8 activations and GEMM kernels. Useful for Ampere cards that lack FP8. Achieves 72% faster inference speed and 82% smaller quantization error compared to FA with FP16 and FP8 data format. Links below
18K
PapersAnon
@papers_anon
Dec 23, 2024
Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models From Deepmind. Proposes a novel inference-aware finetuning paradigm, in which the model is fine-tuned in a manner that directly optimizes the performance of the inference-time strategy. Improved the Bo32
27K
PapersAnon
@papers_anon
Oct 4, 2024
SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration INT8 attention quantization method that is plug-and-play. Has triton kernels for 3090/4090. About 2.1× and 2.7× faster than FlashAttention2 and xformers. More accurate than FA3's FP8. Links below
18K
PapersAnon
@papers_anon
Dec 30, 2024
Multi-matrix Factorization Attention Novel attention architecture that enhances model capacity by efficiently scaling up both the number and dimension of attention heads through low-rank matrix factorization in the Query-Key circuit. Outperforms MLA and performs comparably to
14K
PapersAnon
@papers_anon
May 19, 2025
MegaScale-MoE: Large-Scale Communication-Efficient Training of Mixture-of-Experts Models in Production From ByteDance. Production system tailored for the efficient training of large-scale mixture-of-experts (MoE) models. 1.88× efficiency improvement compared to Megatron-LM.
18K
PapersAnon
@papers_anon
Oct 31, 2024
TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters Natively scalable architecture that leverages the attention mechanism not only for computations among input tokens but also for interactions between tokens and model parameters. Replaces all the linear
45K
PapersAnon
@papers_anon
Jul 17, 2025
Mixture of Raytraced Experts Stacked MoE architecture that can dynamically select sequences of experts, producing computational graphs of variable width and depth. Allows predictions with increasing accuracy as the computation cycles through the experts' sequence. Links below
12K
PapersAnon
@papers_anon
Mar 10, 2025
Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUs From Ant Group. Ling-Lite 16.8B with 2.75B activated parameters. Ling-Plus has 290B with 28.8B activated parameters. Training focused on reducing costs and they share their findings. Also
20K