TensorFold 1.3.3 ships 124 PRs in 24h, boosts GLM decode to 172 tok/s on GB10s
AIFlow_ML · x · 2026-11-22
TensorFold 1.3.3 is out, merging 124 PRs and resolving 34 issues in 24 hours.
- CUDA speedup: GLM-5.3-Flash went from 151.51 to 172.39 tok/s median drafted decode after a 1,048,576-token prompt on four GB10s at TP=4 — 13.8% faster with the same output tokens, thanks to a new kernel from MiaAIlab.
- Metal: GLM-5.3-Flash gets DFlash2 drafting; cold prompts from 2K to 64K on a Mac without tensor units run 5–6% faster.
- Broader support: MOSS-Image generates images on CUDA via /v1/images/generations, GLM-OCR runs on HIP, Idol's image encoder lands on Metal.
- DeepSeek-V4.1 gains structured output in the dsv41-cuda source build.
More from Infra
- Gary Marcus Amplifies Claim That 'Bank of Nvidia' Is Propping Up Token Prices — GaryMarcus · 2026-11-22
- AWS Marketplace adds x402: agents pay per request straight to your AWS bill — MurrLincoln · 2026-11-22
- Power user burns 26.2B tokens in a month for $94 on DeepSeek agents — teortaxesTex · 2026-11-22
- Burkov: open-weight models will soon run as cheap as a cloud VPS as AI demand cools — burkov · 2026-11-22
- Samsung's memory capacity for 2027-2029 is essentially sold out; customers seek supply deals through 2031 — sudoraohacker · 2026-11-22
- Quick math: 1 GW rents for $50B — can Anthropic absorb even half of xAI's 10 new GWs? — toptickcrypto · 2026-11-22