-
### 1、Questions
As we Known, SD v1.5 has 1 Billions params , and it's peek GPU memory is about 4G at the precison fp32.
So, the memory of int4 precison (sd_w4a8_chpt.pth) will be about 4G/8 = 500…
-
Hi, I'm working on applying QAT on a model. I made the necessary modifications. However, when I looked into one of the saved checkpoint `.pth` files, I observed that none of the weights were actually …
-
Currently torchao QAT has two APIs, [tensor subclasses](https://github.com/pytorch/ao/blob/a4221df5e10ff8c33854f964fe6b4e00abfbe542/torchao/quantization/prototype/qat/api.py#L41) and [module swap](htt…
-
It can train the ViT model from the Hugging Face transformer,
but when converting to tflite model it appear an error message that I can't solve it.
The following are the tinynn setting and the error…
-
### 🐛 Describe the bug
from torch.ao.quantization.quantizer import (
XNNPACKQuantizer,
get_symmetric_quantization_config,
)
the code abve report error:
ImportError: cannot import name 'X…
-
I want to apply this algorithm to a Jetson AGX Orin development board.
However, many difficulties were encountered, such as failed installation of libraries such as mmcv and mmdet. Pyquaternion inst…
-
**Describe the bug**
非常痛苦 动态shape根本转不出来
**To Reproduce**
```python
import nncase
import numpy as np
import onnx
import onnxsim
# from nncase_base_func import model_simplify, read_model_fil…
-
Add one example to repo and DLM
-
### 🐛 Describe the bug
The current default qconfig for qnnpack is per-tensor quantization. Can we update the default qnnpack qconfig to per-channel quantization? I heard that per-channel has been s…
-
The following problem is solved almost instantaneously in both SageMath and Matlab, but it takes much longer in sympy. The problem has 8 solutions. If I use `manual=True`, solve is very fast, but it f…