Problem: The memory will slowly increase with the drawing until restarting.
Observation: GC analysis shows that no occupation has occurred, so it is suspected to be a problem with the underlying allocator.
Reason: Under Linux, glibc is used to allocate memory. glibc uses brk and mmap to allocate memory, and the memory allocated by brk cannot be released until the high-address memory is released. That is to say, if you apply for two pieces of memory A and B through brk, it is impossible to release A before B is released, and it is still occupied by the process. Check the suspected "memory leak" through TOP.
So I replaced TCMalloc, but found that libtcmalloc_minimal could not find ptthread_Key_Create. After analysis, it was found that pthread was not entered during compilation.
Fixes failing dependency checks for extensions having a different package name and import name (for example ffmpeg-python / ffmpeg), which currently is causing the unneeded reinstall of packages at runtime.
In fact with current code, the same string is used when installing a package and when checking for its presence, as you can see in the following example:
> launch_utils.run_pip("install ffmpeg-python", "required package")
[ Installing required package: "ffmpeg-python" ... ]
[ Installed ]
> launch_utils.is_installed("ffmpeg-python")
False
... which would actually return true with:
> launch_utils.is_installed("ffmpeg")
True
This tries to execute interpolate with FP32 if it failed.
Background is that
on some environment such as Mx chip MacOS devices, we get error as follows:
```
"torch/nn/functional.py", line 3931, in interpolate
return torch._C._nn.upsample_nearest2d(input, output_size, scale_factors)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: "upsample_nearest2d_channels_last" not implemented for 'Half'
```
In this case, ```--no-half``` doesn't help to solve. Therefore this commits add the FP32 fallback execution to solve it.
Note that the ```upsample_nearest2d``` is called from ```torch.nn.functional.interpolate```.
And the fallback for torch.nn.functional.interpolate is necessary at
```modules/sd_vae_approx.py``` 's ```VAEApprox.forward```
```repositories/stable-diffusion-stability-ai/ldm/modules/diffusionmodules/openaimodel.py``` 's ```Upsample.forward```
