config.yaml: CIFAR sinkhorn batch 128→32 for T4, pool batches 2500→10000 to compensate

config.yaml

@@ -143,16 +143,21 @@ experiment_cifar10:
     use_scale_shift_norm: true
 
   # Sinkhorn gradient flow (Phase 1)
+  # NOTE: batch_size reduced from paper's 128 to 32 for T4 16GB VRAM.
+  # Sinkhorn on 3072-dim flattened vectors (3x32x32) with tensorized backend
+  # uses O(N^2 * D) memory. 128 samples OOMs on T4; 32 fits comfortably.
+  # Compensate by increasing pool batches (32 * 10000 = 320K ≈ 128 * 2500).
   sinkhorn:
     blur: 1.0
     scaling: 0.85
     eta: 1.0
     num_steps: 5
-    batch_size: 128
+    batch_size: 32
 
-  # Trajectory pool (Appendix E.2: 128 batch * 2500 batches * 5 steps ~ 45GB)
+  # Trajectory pool — adjusted for smaller Sinkhorn batch
+  # 32 batch * 10000 batches * 5 steps = 1.6M entries (same order as paper)
   pool:
-    num_batches: 2500
+    num_batches: 10000
     storage_limit_gb: 45
 
   # Velocity field matching training (NSGF model)