AI/ML●●Solid
UATC-Closed-loop VRAM control and dynamic data pruning for LLM training
Kalman filter + PID controller + Smith predictor prevents OOM crashes during edge LLM training.
Big BrainBold Bet
L_u_u_6
208d ago
Back to browse
GitHub Repository
2 starsPython
UATC-Closed-loop VRAM control and dynamic data pruning for LLM training
by L_u_u_6·Jul 13, 2026·1 point·0 comments
AI Analysis
●●●BangerWizardryBig BrainSolve My Problem
Kalman filter and PID loop prevent OOM crashes where DeepSpeed static baselines fail.
Strengths
- •Industrial control theory (PID, Smith predictor) applied to VRAM management is a genuinely novel approach.
- •Recovers from 18 forced OOM events and prunes 86% of redundant samples without fatal crashes.
- •Paradigm-aware adaptation works across full fine-tuning, LoRA, and QLoRA without code changes.
Weaknesses
- •Evaluation limited to a single 1.5B model on a T4; scalability to larger models unproven.
- •High complexity barrier; integrating control theory stacks may deter casual adopters.
Category
Target Audience
ML engineers fine-tuning models on consumer hardware or resource-constrained clusters
Similar To
DeepSpeed · PyTorch FSDP · Accelerate
Similar Projects
Developer Tools●●Solid
UATC – A Closed-Loop Controller to Prevent GPU OOM
Control theory meets GPU memory management with real ablation studies.
Big BrainNiche Gem
L_u_u_6
2011d ago
Developer Tools●●Solid
TruLayer – tracing, evals, and a control loop for production LLMs
Automated rollback on regression is a killer feature LangSmith doesn't have.
Solve My ProblemSlick
trulayer
201mo ago
AI/ML●●●Banger
SCP – A protocol that drops LLM API calls to zero in 60fps physics loop
Subsumption Architecture revival cuts LLM calls with pattern cache misses.
Big BrainWizardry
srk0102
103mo ago
Developer Tools●●Solid
Looplet – a 0-dep agent loop you own
Iterator-first design beats black-box frameworks like LangChain for debugging.
Big BrainNiche Gem
hsaghir
202mo 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
103mo ago