Finetuning 176B
Finetuning 176B at the end of the training might be necessary to ensure exact logits match between Megatron-DS trained model and HF model. For now, there are 2 main bottlenecks that are responsible of not giving 100% logits match between HF model and Megatron model
Diverging bottlenecks
TP merging strategy
See this issue. When merging TP ranks the logits exactness is lost. The idea would be to finetune the 176B model with TP=1
Use torch_softmax instead of fused_softmax
fused_softmax and torch_softmax does not give the same results (ie, torch.testing.assert_allclose(atol=0.0, rtol=0.0) does not pass). The main model could be finetuned with torch_softmax.
See this line