AITemplate
AITemplate copied to clipboard
AITemplate is a Python framework which renders neural network into high performance CUDA/HIP C++ code. Specialized for FP16 TensorCore (NVIDIA GPU) and MatrixCore (AMD GPU) inference.
Summary: Building a simple AIT cast operator that supports `float16`, `bfloat16`, and `float32`. This first version does not support Vectorization in order to simplify this task Differential Revision: D46603937
# Description Currently `compile_alt.py` generates bad images on T4 GPU (SM75). Related issue - https://github.com/facebookincubator/AITemplate/issues/781 Notebook to reproduce the issue with bad images on T4 GPU - [AIT_alt_bad_image.ipynb](https://colab.research.google.com/drive/1XBupsnP35rJ-3mKheBQVjNfjy1nTcjuO?usp=sharing) I found...
Save all generated images in SD `demo_alt.py` script. This was fixed in `demo.py` already couple days ago.
I tried to use `compile_alt` and `demo_alt` to compile and run stable_diffusion model on T4 GPU (sm 7.5) The generated image is a mess - Example - [bad image 1](https://ibb.co/pb7B0C5)...
## Description This PR allows to compile SD pipeline with variable batch size. H and W are fixed and defined by the model variant. (base - 512 and regular -...
The conv2d profiler can be built well with gcc-12. But, when switching to gcc-13, compilation failed with error listed below. This error may caused by reduced usage of cstdint header...
Summary: During our perf tuning, we are seeing mtml_instagram_model lag behind AIT around 23% on mergenet (525662456/1670) How do we compare the perf between AIT and AOTI? Refering the idea...
Summary: as title, there are still some utilities functions needed from fx2ait for aot_Inductor_lower before we fully moved everything to PT2 full stack Reviewed By: chenyang78 Differential Revision: D56613348
I haven't been able to get grouped ConvTranspose to work correctly. I added a unittest for it here https://github.com/jonpryai/AITemplate_Upstream/commit/47421490c3cec2047328b45d63f5ff3a451efc53 Which fails with a segfault. Any ideas?