AI/ML●●Solid
SparseLab–real sparse training(CSR+custom kernel) in PyTorch, CPU-first
Custom CPU kernels for sparse training when everyone else chases GPU.
Niche GemBig Brain
DARSHANFOFADIYA
111mo ago
Back to browse
GitHub Repository
metal collective communication library (pytorch DDP)
7 starsC++
MCCL Distributed PyTorch training across MacBooks via Thunderbolt
by sassoshots44·Mar 21, 2026·1 point·0 comments
AI Analysis
●●SolidWizardryNiche GemShip It
Two MacBooks syncing gradients over Thunderbolt — slower than single-GPU but it works.
Strengths
- •Fills genuine gap: PyTorch MPS multi-process collectives didn't exist before this.
- •Honest benchmarks admit 10x slower — rare transparency in ML infrastructure projects.
- •Uses vDSP reductions and Metal for fp16/bf16 with overlapped TCP transport.
Weaknesses
- •Only tested on 2 nodes — no validation for larger clusters or production workloads.
- •Performance is worse than single-GPU, limiting real-world utility.
Category
Target Audience
ML researchers with multiple Macs experimenting with distributed training
Similar To
NCCL · Gloo · PyTorch Distributed
Similar Projects
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
4022d ago
Developer Tools●Mid
easy-torch-tpu – A Flexible Training Pipeline for PyTorch Models on TPU
TPU training wrapper built on torchprime; solves a real problem but torchprime already exists.
Niche Gem
in-silico
103mo ago
AI/ML●●Solid
MLForge – A visual graph editor for building PyTorch models
Infers layer shapes from connections and exports standard PyTorch scripts.
CozyNiche Gem
zaina-ml
101mo ago
AI/ML●●Solid
Neural Abyss – PyTorch multi-agent combat simulator
Per-agent PPO runtime with tensor-first simulation state is genuinely clever architecture.
Big BrainNiche Gem
luthor190397
102mo ago
Infrastructure●●●Banger
Physics-based simulator for distributed LLM training and inference
Estimates LLM training MFU, memory, timeline across 70 models and parallelism strategies—genuinely useful before GPUs commit.
WizardrySolve My ProblemBig Brain
zhebrak
113mo ago