Rust LLM training dilemma: Candle vs Burn vs PyTorch for a toy model
Enderlook · reddit · 2026-10-21
Reddit user Enderlook is building and training a toy LLM in Rust/Candle and asks which framework to commit to. Pain points with Candle: flash attention lacks a backward pass, and modern features like GatedDeltaNet2 and Multi-head Latent Attention are missing, forcing manual CUDA kernel porting. Burn offers full training support and, via CubeCL, compiles Rust to CUDA, HIP, Metal, WebGPU and CPU — but runs slower, has limited quantization, and a smaller community. Options on the table: stay on Candle and add CubeCL to reuse Burn kernels, migrate fully to Burn, or jump to Python/PyTorch where every new published feature has a reference implementation. The thread collects real trade-offs of doing LLM training outside the Python ecosystem.
More from coding & agent
- GitHub and Hugging Face Host 'Training Agents IRL' Seminar on RL for Agentic Systems — Hugging Face · 2026-10-21
- AI influencer: skip elaborate agent workflows, most mundane tasks aren't worth automating — EXM7777 · 2026-10-21
- Anthropic's study of 235,000 Claude Code users finds domain expertise beats coding skill — sudoraohacker · 2026-10-21
- System Design for AI Day 3: a visual guide to core building blocks and scaling tradeoffs — Roger_M_Taylor · 2026-10-21
- Claude as the front end, Databricks as the engine: 'most BI tools are fully dead' — teodorio · 2026-10-21
- llampart v2: open-source local LLM chat UI with charts, Mermaid and MCP support — mossy_troll_84 · 2026-10-21