← Back to context Comment by zozbot234 4 months ago SOTA models are reportedly MoE, not dense. 3 comments zozbot234 Reply bigyabai 4 months ago A 5T MoE model is still bottlenecked by streaming weights from SSD, in addition to compute bottlenecks during prefill and decode. zozbot234 4 months ago True but a cluster built on pipeline parallelism can naturally stream from multiple SSD's in parallel. That probably makes offload somewhat more effective. And you also have RAM caching available as a natural possibility. bigyabai 4 months ago You won't be RAM caching much of anything with experts that are 220b parameters worth of layers.
bigyabai 4 months ago A 5T MoE model is still bottlenecked by streaming weights from SSD, in addition to compute bottlenecks during prefill and decode. zozbot234 4 months ago True but a cluster built on pipeline parallelism can naturally stream from multiple SSD's in parallel. That probably makes offload somewhat more effective. And you also have RAM caching available as a natural possibility. bigyabai 4 months ago You won't be RAM caching much of anything with experts that are 220b parameters worth of layers.
zozbot234 4 months ago True but a cluster built on pipeline parallelism can naturally stream from multiple SSD's in parallel. That probably makes offload somewhat more effective. And you also have RAM caching available as a natural possibility. bigyabai 4 months ago You won't be RAM caching much of anything with experts that are 220b parameters worth of layers.
bigyabai 4 months ago You won't be RAM caching much of anything with experts that are 220b parameters worth of layers.
A 5T MoE model is still bottlenecked by streaming weights from SSD, in addition to compute bottlenecks during prefill and decode.
True but a cluster built on pipeline parallelism can naturally stream from multiple SSD's in parallel. That probably makes offload somewhat more effective. And you also have RAM caching available as a natural possibility.
You won't be RAM caching much of anything with experts that are 220b parameters worth of layers.