Adam Oster Claude Opus 4.8 (1M context) commited on
Commit ·
d30bc58
1
Parent(s): a1ccfc0
remove torch pin; use ZeroGPU-provided torch
Browse filesThe NVML_SUCCESS / nvmlInit assert in CUDACachingAllocator fires on the
first heavy VAE-decode allocation at ALL resolutions (incl. 360x640), so
it is NVML init failing in the ZeroGPU forked GPU process, not OOM and not
app logic. Both pinned PyPI torch builds (2.9.1, 2.11.0) crash identically.
ZeroGPU ships its own patched torch and discourages pinning, so drop the
pin and let the platform supply a torch matched to the Blackwell runtime.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- requirements.txt +1 -2
requirements.txt
CHANGED
|
@@ -12,5 +12,4 @@ imageio-ffmpeg
|
|
| 12 |
einops
|
| 13 |
timm
|
| 14 |
av
|
| 15 |
-
git+https://github.com/huggingface/diffusers.git@main
|
| 16 |
-
torch==2.11.0
|
|
|
|
| 12 |
einops
|
| 13 |
timm
|
| 14 |
av
|
| 15 |
+
git+https://github.com/huggingface/diffusers.git@main
|
|
|