srush / do-we-need-attention
โ161Updated last year
Related projects โ
Alternatives and complementary repositories for do-we-need-attention
- Understand and test language model architectures on synthetic tasks.โ161Updated 6 months ago
- A MAD laboratory to improve AI architecture designs ๐งชโ95Updated 6 months ago
- The simplest, fastest repository for training/finetuning medium-sized GPTs.โ83Updated last week
- Language models scale reliably with over-training and on downstream tasksโ94Updated 7 months ago
- Code for the paper "The Impact of Positional Encoding on Length Generalization in Transformers", NeurIPS 2023โ124Updated 6 months ago
- โ76Updated 5 months ago
- โ132Updated last year
- some common Huggingface transformers in maximal update parametrization (ยตP)โ76Updated 2 years ago
- [NeurIPS 2023] Learning Transformer Programsโ157Updated 5 months ago
- Code to reproduce "Transformers Can Do Arithmetic with the Right Embeddings", McLeish et al (NeurIPS 2024)โ176Updated 5 months ago
- Some personal experiments around routing tokens to different autoregressive attention, akin to mixture-of-expertsโ108Updated 3 weeks ago
- Minimal (400 LOC) implementation Maximum (multi-node, FSDP) GPT trainingโ112Updated 6 months ago
- Code for exploring Based models from "Simple linear attention language models balance the recall-throughput tradeoff"โ212Updated 2 months ago
- Some preliminary explorations of Mamba's context scaling.โ190Updated 9 months ago
- Yet another random morning idea to be quickly tried and architecture shared if it works; to allow the transformer to pause for any amountโฆโ49Updated last year
- โ72Updated 4 months ago
- Experiments around a simple idea for inducing multiple hierarchical predictive model within a GPTโ205Updated 2 months ago
- NanoGPT-like codebase for LLM trainingโ73Updated this week
- Experiment of using Tangent to autodiff tritonโ71Updated 9 months ago
- Annotated version of the Mamba paperโ455Updated 8 months ago
- โ53Updated 9 months ago
- โ206Updated 6 months ago
- Implementation of the conditionally routed attention in the CoLT5 architecture, in Pytorchโ224Updated 2 months ago
- Token Omission Via Attentionโ119Updated 3 weeks ago
- Official Repository of Pretraining Without Attention (BiGS), BiGS is the first model to achieve BERT-level transfer learning on the GLUE โฆโ114Updated 7 months ago
- Inference code for LLaMA models in JAXโ112Updated 5 months ago
- โ73Updated 11 months ago
- โ116Updated this week
- Scalable neural net training via automatic normalization in the modular norm.โ118Updated 2 months ago
- โ175Updated this week