It follows reports of 30% MFU on Ascend hardware.
There are not many 384 clusters. But they'd also fit with DeepSeek's MOPD agenda, if integrated. You can parallelize RL of teacher clones, and eventually merge them into the main branch. I think we'll see this more with 950 servers. V5 report will be a joyful read.
@teortaxesTex i dont really see the direct need for MOPD here U can parallelise the inference across clusters as cursor and we did before For training is the main difference but for RL its unlikely u would reach a point of needing this. But also seems u can do this too without MOPD
@Grad62304977 inference is happening on multiple clusters by default I mean the question of how DeepSeek's training playbook described in V4 paper can fit with this domestic system/method. RL on a small separate server would be an adequate way to produce a marginal teacher.
It follows reports of 30% MFU on Ascend hardware.
There are not many 384 clusters. But they'd also fit with DeepSeek's MOPD agenda, if integrated. You can parallelize RL of teacher clones, and eventually merge them into the main branch. I think we'll see this more with 950 servers. V5 report will be a joyful read.