AITemplate

AITemplate copied to clipboard

Introduce safe_compile_model with locking

1

Summary: This diff extends D44374161 in the following ways: one criticism with respect to compile_model locking it's build dir is, that this is kind of unexpected behavior and not part...

kadeng

Allow support for additional NVIDIA GPU types

5

This commit will decode the compute capability from `nvidia-smi`, and allow `_detect_cuda` to return an appropriate value for GPUs other than the T4, V100, and A100 (such as the RTX...

space55

Working directory is hardcoded in StableDiffusionAITPipeline and StableDiffusionImg2ImgAITPipeline in examples

4

Issue: StableDiffusionAITPipeline and StableDiffusionImg2ImgAITPipeline in pipeline_stable_diffusion_ait.py and [pipeline_stable_diffusion_img2img_ait.py has workdir hardcoded to "tmp/" which might create problems where the file system locations are either read-only or users have restricted permissions...

dhawalkp

[ask] is it possible to copy weights from PyTorch model to complied AIT model

2

I have a group of fine-tuned stable diffusion models using the same training process. After compiling one model into AIT format, is it possible to reuse the "selected" AIT graph...

YaoAstro

c++ runtime api example

2

Based on what's present, it seems like it's possible to use AI template entirely from native code with no python runtime. Could you add a example which illustrates how to...

mindbeast

usage of nvcc from AI template is a deployment mess

1

It's really not suitable to force users to always provide a nvcc for compilation at runtime. It makes deployments brittle and very complicated, particularly of native code. Nvidia provides nvrtc...

mindbeast

transformer converter

2

Reviewed By: frank-wei Differential Revision: D43677477

wushirong

Adding support for StreamK for GEMM operations (except for group GEMM). As workspace calculations are hard to cap for a particular range of shapes (and max shape calcualation doesn't always...

artkorenev

The result of the StableDiffusionAITPipeline cannot be fixed with a seed value.

8

Even if Seed is given to generator to generate, the result changes slightly each time. ``` def torch_fix_seed(seed=42): # Python random random.seed(seed) # Numpy np.random.seed(seed) # Pytorch torch.manual_seed(seed) torch.cuda.manual_seed(seed) torch.backends.cudnn.deterministic...

ssotabe

AttributeError: module 'cutlass_lib.library' has no attribute 'GroupMode'

1

when I run examples, an error An error occurred： ` File "/opt/conda/lib/python3.8/site-packages/aitemplate/backend/cuda/conv2d/conv2d_bias_relu_few_channels.py", line 47, in conv2d_bias_relu_few_channels_config func_attrs["op_instance"] = cfc.extract_config( File "/opt/conda/lib/python3.8/site-packages/aitemplate/backend/cuda/conv2d/common_conv2d_few_channels.py", line 49, in extract_config return common.extract_config( File "/opt/conda/lib/python3.8/site-packages/aitemplate/backend/cuda/conv2d/common.py", line...

kisisjrlly