Easy way to profile model likelihood / gradient evaluation
Easy way to profile model likelihood / gradient evaluation: a task in MiMo-V2.6-RL-oss: Agentic RL Environments (MiMo RL release). When sampling is slow, I want to figure out which part of the log-probability (or its gradient) is the bottleneck. Theano has nice profiling support built in via…
The task
When sampling is slow, I want to figure out which part of the log-probability (or its gradient) is the bottleneck. Theano has nice profiling support built in via theano.function(..., profile=True) plus ProfileStats.summary(), but right now there's no convenient way to use it on…
Part of XiaomiMiMo/MiMo-V2.6-RL-oss.