slime v0.3.1 Released: Major Optimizations for Memory and Training Performance
hsu_byron · x · 2026-08-06
THUDM has released slime v0.3.1. Instead of major new features, this update leverages AI agents to improve engineering productivity, focusing on removing previous technical compromises.
Key improvements include:
- Memory & Training: Reduced Megatron offload overhead, support for PYTORCHALLOCCONF, and a new --release-train to free up GPU memory between iterations.
- Faster Weight Updates: Optimized update paths and parallel updatefromdisk.
- Alignment & Algorithms: Introduced Top-p masking for better rollout-training alignment and lower memory PPO logprob/entropy computation.
- Simplification: Removed Megatron Bridge support.
More from Infra
- Goldman Sachs: Hyperscaler Capex to Hit $1.2 Trillion by 2027 — Beth_Kindig · 2026-08-06
- Cloudflare Unveils Kitesurf, a Browser Built for AI Agents — threepointone · 2026-08-06
- Cloudflare Introduces Fully Stateless Next-Gen MCP Architecture — threepointone · 2026-08-06
- FlashInfer Merges Blackwell Kernels, Boosting E2E Inference Speed by 7.6% — YouJiacheng · 2026-08-06
- Making the Rust Compiler 3x Faster Could Save Hundreds of Millions Annually — doodlestein · 2026-08-06
- Alphabet Seeks $25B in New Bonds, Raising Over $150B This Year — firstadopter · 2026-08-06