AI Resonance · 阅读最新日报 · AI 入门推荐

2026-11-12 · America/Los_Angeles · 社区动态 · #11

You can read which GPU a model was initialized on from its open weights (PyTorch/cuRAND thread layout)

Interesting forensics writeup: PyTorch fills initial weights with Philox via cuRAND, and the thread count (2,048 per SM) decides where each random number lands. Rows that training never touches keep their initial values, so you can search seed + thread count and recover the SM count of the GPU that generated them. Their readings: Llama 3.1 = 132 SMs (H100, matches Meta's disclosure), Nemotron-3 Ultra = 152 (GB200, matches NVIDIA's report). They say the method agrees with all 24 models whose hardware is publicly disclosed. They also reproduced 445 rows of DeepSeek-V4-Flash bit for bit on a…

You can read which GPU a model was initialized on from its open weights (PyTorch/cuRAND thread layout)

热度 46.0 / 100;排名与评分保留该期记录。