Developer Tools●●●Banger
Trained a 12M transformer on an ML framework we built from scratch
Custom CUDA kernels and Rust backend with a TypeScript API built by students in four months.
WizardryBig BrainShip It
caliandbust
222mo ago
Back to browse
GitHub Repository
Mamba SSM and Mamba-3 SISO in Rust with optional CUDA GPU acceleration. Inference and training (BPTT through SSM state, AdamW), CPU + GPU paths, custom CUDA kernels, CUDA Graph capture, f32 / bf16 / f16. Batch-invariant bf16 inference — per-row output is bit-identical across batch sizes.
10 starsRust
Mamba SSM in Rust – training and inference with custom CUDA kernels
by silvermpx·Mar 23, 2026·1 point·0 comments
AI Analysis
●●SolidWizardryNiche Gem
Custom CUDA kernels for SSM recurrence with zero framework dependencies.
Strengths
- •Full BPTT through recurrent SSM state enables actual training, not just inference.
- •Zero-allocation single-step inference hits ~200μs on CPU without GPU.
- •Standalone design means no PyTorch, Burn, or Candle dependency chain.
Weaknesses
- •Mamba implementations already exist in multiple languages; Rust isn't unique.
- •No benchmark comparisons against official Mamba or other ports.
Category
Target Audience
ML engineers wanting Rust-based SSM implementations
Similar To
mamba-minimal · Candle · Burn
Similar Projects
Developer Tools●●●Banger
cuTile Rust: Safe, data-race-free GPU kernels in Rust
Extends Rust's ownership model across GPU boundary with tile-based partitioning for data-race-free kernels.
WizardryBig BrainNiche Gem
melihelibol
106184d ago
Infrastructure●●Solid
ZSE – Single-file LLM engine with dual INT4 kernels
INT4 inference engine beats llama.cpp on VRAM, but competing against established tools.
WizardryShip It
zyoralabs
103mo 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
2019d ago
AI/ML●●●Banger
Glq LLM quantization using E8 lattice
E8 lattice codebooks beat GPTQ at 2-4 bpw with fused CUDA kernel skipping weight materialization.
WizardryBig Brain
acd
2019d ago
AI/ML●Mid
FlashQwen – A from-scratch CUDA inference engine for Qwen3
Another inference engine when vLLM and llama.cpp already dominate.
Bold BetNiche Gem
langtang1996
505d ago