-
### 🐛 Describe the bug
Can not use `del` and `gc` to release memory after trace/freeze.
To track memory malloc/release
```
diff --git a/c10/core/impl/alloc_cpu.cpp b/c10/core/impl/alloc_cpu.cpp
…
-
I have turned lustre_exp_ost=true,and after that I can see data in Write/Read Throughtput per Client in Lustre OST panel,but in Lustre STatistics panel the Write/Read Throughtput per Client panel is …
ldd91 updated
4 years ago
-
## 🐛 Describe the bug
There is a bug in PyTorch where data on CPUs is not automatically synchronized before subsequent operations when using the `torch.Tensor.to()` method with `non_blocking=True`.…
-
### 🐛 Describe the bug
`bfloat16` not supported with on `grid_sampler_3d_cuda`.
```python
resampled_roi = F.grid_sample(input=tensor[b_idx: b_idx + 1, :, :, :, :],
…
-
### Describe the bug
The following benchmark on Max Series with xpu backend
```ipex
import intel_extension_for_pytorch as ipex
import torch
device = "xpu"
n_samples = 50_000_000
dim = 10
…
-
### 🐛 Describe the bug
The following code can trigger a Floating Point Exception in `torch.index_select`:
````python
import torch
input = torch.rand([], dtype=torch.float32).cuda()
dim = 0
index…
-
### 🐛 Describe the bug
When I add test case to our device backend. The backward of nearest_upsample op confused me.
For a example, when a tensor of shape {1,3,2,4} resample to {1,3,2,5} using Upsa…
-
### 🐛 Describe the bug
cpu log1p for bfloat16 gives inf on big number.
```
>>> import torch
>>> x = torch.tensor(1.821e+38).bfloat16()
>>> x
tensor(1.8210e+38, dtype=torch.bfloat16)
>>> x.log…
-
```
What steps will reproduce the problem?
1. Setup xuggler to decode and encode a H264 stream (FLV)
2. Stream source being youtube
3. Watch video in real time
4. Frames missing from end of playback
…
-
### 🐛 Describe the bug
When a edge case is given, a **segment fault** can be triggered in `torch.max_pool1d_with_indices`:
````python
import torch
input = torch.rand([12, 2, 2, 2, 0, 4634247419717…