AI/ML●●●Banger
TRiP – a complete transformer engine in C built from scratch just by me
From-scratch C transformer engine with training and vision, built by one person.
WizardryDark Horse
carlovalenti
3861mo ago
Back to browse
GitHub Repository
Official PyTorch implementation of NeuroFlow: EMA-Gated Temporal Sequence Compression for Vision Transformers. Achieves up to 55.8x wall-clock speedup for video inference via semantic surprise routing and a training-free Dual-Memory Reconstruction Protocol.
18 starsPython
NeuroFlow 55.8x video inference speedup for Vision Transformers PyTorch
by ynnk·May 26, 2026·8 points·2 comments
AI Analysis
●●SolidBig BrainWizardry
Training-free dual-memory protocol cuts 1792p SigLIP inference from 678ms to 11.9ms.
Strengths
- •Specific 55.8x wall-clock speedup with verifiable benchmark numbers and methodology
- •Training-free Architecture C retains 92.4% of dense accuracy at 84% token sparsity
- •Complete repo with paper, weights, verification scripts, and Hugging Face models
Weaknesses
- •Video ViT optimization is active research space with multiple competing approaches already
- •LLM ablation shows 0% token drift but limited scope beyond Phi-3-mini testing
Category
Target Audience
ML researchers and engineers working on video inference optimization
Similar To
Token Merging (ToMe) · DynamicViT · Sparse ViT approaches
Similar Projects
AI/ML●●Solid
PyTorch on Java
LibTorch bindings bring CUDA and MPS backends to Java with LLaMA-3 inference included.
Niche GemBig Brain
pdsminer
202d ago
AI/ML●●●Banger
We built an LLM inference engine in pure Python – no PyTorch, no Triton
30x faster cold start than vLLM with zero PyTorch dependencies.
WizardryBig BrainZero to One
zyoraclub
2011d ago
Developer Tools●●●Banger
Profine – Profile and rewrite your PyTorch training loop on real GPUs
Automates the painful torch.compile and mixed-precision tuning loop with measured 3x speedups.
Big BrainSolve My Problem
aisinghal
401mo ago
AI/ML●●Solid
Doppler.js – WebGPU inference, faster/simpler than transformer.js
Explicit kernel control over TVM-style black boxes, but benchmarks show mixed wins vs Transformers.js.
Big BrainWizardry
clocksmith
303mo ago
AI/ML●●Solid
BNNR – a closed-loop pipeline for improving vision models
XAI-driven model improvement loop, but Weights & Biases already tracks experiments better.
Big BrainNiche Gem
dominka
102mo ago