Update harmonypy to use PyTorch GPU backend

- Switch Docker image to base_pytorch_nvidia:1.0.0
- Pin harmonypy>=0.2.0 and add torch dependency
- Pass device=None to run_harmony for auto GPU detection
- Remove epsilon workaround as harmonypy v0.2.0 handles
  numerical stability internally via clamping and L2 normalisation
- Update Nextflow label to lowcpu, gpu
- Update CHANGELOG.md

Author: ghar1821

CHANGELOG.md

@@ -81,6 +81,12 @@
   * iLISI is now computed per biological group. Groups where batch is fully confounded by group are skipped and return NaN.
   * Added `helper.py` with `compute_ilisi_per_group` and `compute_clisi` functions.
 
+* Updated `methods/harmonypy` to use the PyTorch GPU backend (harmonypy>=0.2.0):
+  * Switched Docker image to `openproblems/base_pytorch_nvidia:1.0.0`.
+  * Added `device=None` to `run_harmony` for automatic GPU detection (CUDA -> MPS -> CPU).
+  * Removed epsilon workaround as numerical stability is now handled internally by harmonypy.
+  * Updated Nextflow label to `lowcpu, gpu`.
+
 * Updated file schema (PR #18): 
   * Add is_control obs to indicate whether a cell should be used as control when correcting batch effect.
   * Removed donor_id obs from unintegrated censored.

src/methods/harmonypy/config.vsh.yaml

@@ -22,14 +22,15 @@ resources:
 
 engines:
   - type: docker
-    image: openproblems/base_python:1
+    image: openproblems/base_pytorch_nvidia:1.0.0
     setup:
       - type: python
         packages:
-          - harmonypy
+          - harmonypy>=0.2.0
+          - torch
 
 runners:
   - type: executable
   - type: nextflow
     directives:
-      label: [midtime,midmem,midcpu]
+      label: [midtime,midmem,lowcpu,gpu]

src/methods/harmonypy/script.py

@@ -30,14 +30,13 @@
 
 print("Run harmony", flush=True)
 
-# TODO numerical instability in kmeans causing problem with harmony.
-# so adding a very small value to all entries to make sure there are no zeros
-epsilon = 1e-20
-
+# device=None lets harmonypy auto-detect the best available backend:
+# CUDA GPU -> Apple MPS -> CPU
 out = harmonypy.run_harmony(
-    data_mat=adata_to_correct.layers["preprocessed"] + epsilon,
+    data_mat=adata_to_correct.layers["preprocessed"],
     meta_data=adata_to_correct.obs,
     vars_use="batch_str",
+    device=None,
 )
 
 # have to add in the uncorrected markers as well