@MayankMish98
We identified an issue with the Mamba-2 ๐ initialization in HuggingFace and FlashLinearAttention repository (dt_bias being incorrectly initialized). This bug is related to 2 main issues: 1. init being incorrect (torch.ones) if Mamba-2 layers are used in isolation without the Mamba2ForCausalLM model class (this has been already fixed: https://t.co/oahfxjIsKb). 2. Skipping initialization due to meta device init for DTensors with FSDP-2 (https://t.co/hLC8nnQFc3 will fix this issue upon merging). The difference is substantial. Mamba-2 seems to be quite sensitive to the initialization. Check out our experiments at the 7B MoE scale: https://t.co/n8iuUICRux Special thanks to @kevinyli_, @bharatrunwal2, @HanGuo97, @tri_dao and @_albertgu ๐ Also thanks to @SonglinYang4 for quickly helping in merging the PR.