
Zeeshan Patel
@zeeshanp_ · 10 nov 2026
i wonder how much of an improvement swapping to the c/c++ stack brought in terms of tokens/s/gpu and how fast each pretraining run can be done.
if you use jax, at some point you realize you need to rip out xla and rewrite the comms schedule and compute kernels from scratch
Elon Musk@elonmusk· 10 nov 2026@cradrrat We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to
Elon Musk
@elonmusk
💯
10:23 · 10 de noviembre de 2026 · 14,2 mil visualizaciones
19
10
67