trunk/081f44f105c10ec496471f97634c49227ea9146b: [PyTorch][Inductor] Add GB300 device specifications (#196944)
- PyTorch 近 90 天出现 1019 次
- PyTorch 近 90 天第 1008 次发版
- 上一次:同一天稍早 · trunk/903c6152654a1ae026af919f3d85ebdf5c65ae19
发生了什么
Summary: NVIDIA GB300 is absent from the Inductor static device table, so roofline consumers fall back to the generic Triton hardware model. On a GB300 dev host that fallback reports 0.644 TFLOPS for BF16, versus the published dense peak of 2500 TFLOPS. Add exact-name GB300 device metadata using NVIDIA's published per-GPU specifications: dense throughput values, 8000 GB/s HBM bandwidth, and 279 GB HBM capacity. The…
摘要按规则整理自下方来源原文