Search

Few direct matches — filled in with the latest updates.

Tag: #kernels4 results

⚙️

Triton MoE Kernel Beats Megablocks

A pure Triton kernel for fused MoE dispatch outperforms CUDA-optimized Megablocks on Mixtral-8x7B at inference batch sizes (131% faster at 32 tokens). It reduces memory traffic by 35% via fused operations and works across models on NVIDIA and AMD hardware. Code and writeup available on GitHub.

Reddit r/MachineLearningCommunityApr 5#moe#inference#kernels
🔬

SSMs Struggle in 25M Param Training

Empirical findings from OpenAI's Parameter Golf reveal SSMs compress worse than transformers and lose architectural edges at larger vocabs in tiny models (25M params, 16MB, 10min train). Blog details Mamba-3 Triton kernel experiments uncovering fusion slowdowns, quant bugs, and precision fixes.

Reddit r/MachineLearningCommunityMay 4#compression#kernels#parameter-golf
Custom Kernels for All Users

Custom Kernels for All Users

Hugging Face introduces custom kernels powered by Codex and Claude, now available to everyone. This expands access to advanced customization options on the platform. Users can integrate these models seamlessly into their workflows.

Hugging Face BlogOfficialFeb 13#new-feature#hugging-face#codex-claude