Add requirements.txt with verified-working versions
Browse files- requirements.txt +28 -0
requirements.txt
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Verified-working environment for this model.
|
| 2 |
+
# Reference machine: NVIDIA A100 40GB, CUDA 12.1.
|
| 3 |
+
#
|
| 4 |
+
# IMPORTANT -- read before substituting versions:
|
| 5 |
+
# You must use `monai-generative` (which provides the `generative` package).
|
| 6 |
+
# MONAI core (>=1.4) ships its own DiffusionModelUNet / DDPMScheduler /
|
| 7 |
+
# DiffusionInferer, but those are NOT weight-compatible with this checkpoint:
|
| 8 |
+
# the attention layers are named differently (52 keys mismatch), even though
|
| 9 |
+
# the parameter count is identical. See the "Requirements" section of the
|
| 10 |
+
# model card. Never load these weights with strict=False.
|
| 11 |
+
|
| 12 |
+
monai-generative==0.2.3
|
| 13 |
+
monai==1.5.2
|
| 14 |
+
torch==2.5.1
|
| 15 |
+
|
| 16 |
+
safetensors>=0.4
|
| 17 |
+
huggingface_hub>=0.23
|
| 18 |
+
numpy>=1.24,<2
|
| 19 |
+
|
| 20 |
+
# Notes:
|
| 21 |
+
# - torch is pinned to the reference version. Newer torch usually works;
|
| 22 |
+
# monai-generative 0.2.3 only emits FutureWarnings about
|
| 23 |
+
# torch.cuda.amp.autocast. If `import generative` fails on a torch API,
|
| 24 |
+
# fall back to torch==2.5.1.
|
| 25 |
+
# - `einops` is NOT required for this model via monai-generative. It is only
|
| 26 |
+
# needed by MONAI core's implementation, which you should not use here.
|
| 27 |
+
# - scipy is not required to load or sample, only for the post-processing
|
| 28 |
+
# (connected-component filtering) described in the model card.
|