Liger-Kernel
                                
                                
                                
                                    Liger-Kernel copied to clipboard
                            
                            
                            
                        Efficient Triton Kernels for LLM Training
### 🐛 Describe the bug Getting `ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)` when doing inference using HF `from_pretrained()` with `device_map="auto"`. ### Error ``` File...
Hello, Regarding inplace modification of PyTorch tensors, there are already multiple (#254, #262, #272) issues. I would also like to point out that according to PyTorch docs for [mark_dirty()](https://pytorch.org/docs/stable/generated/torch.autograd.function.FunctionCtx.mark_dirty.html#torch-autograd-function-functionctx-mark-dirty): ```...
### 🚀 The feature, motivation and pitch Hi, great job on the FusedLinearCrossEntropy kernel! I’ve found it very efficient for model training. However, it seems to lack support for custom...
### 🐛 Describe the bug Instead of only patching the transformers mllama module (`transformers.models.mllama.modeling_mllama`), `apply_liger_kernel_to_mllama` modifies `torch.nn.LayerNorm` globally. The issue is [here](https://github.com/linkedin/Liger-Kernel/commit/6ab3b9febc29f5045e6d2e27ba6bacaa4f041d91#diff-376c16dd1328612cf488158c6ce9805b044773659c8459cdcb2c6ec35dac346bR163). The fix would be to: (1) Not patch...