Fix: Only free modules > 1GB (text encoder), keep feat_mlp

The previous 100MB threshold was too aggressive and deleted feat_mlp
(the vision projection head needed for inference). The text encoder
is ~7.5GB while feat_mlp is < 1GB, so a 1GB threshold cleanly
separates them. Also logs all module sizes for debugging.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

radio_backbone.py

@@ -94,39 +94,31 @@ def load_model(device: str = "cuda", vitdet_window_size: int = 8):
 
 
 def _free_text_encoder(adaptor):
-    """Delete large sub-modules from the SigLIP2 adaptor to free RAM.
+    """Delete the SigLIP2 text encoder from the adaptor to free ~7.5GB RAM.
 
-    After pre-computing text embeddings, we no longer need the text
-    encoder, tokenizer model weights, or any module > 100MB.
+    Only targets modules > 1GB — the text encoder is ~7.5GB while vision
+    projection heads (feat_mlp, summary_mlp, etc.) are < 1GB.
     """
     freed = 0
 
-    # Check all direct children of the adaptor
-    for name in list(vars(adaptor).keys()):
-        obj = getattr(adaptor, name, None)
-        if obj is None:
-            continue
-        if hasattr(obj, "parameters"):
-            param_bytes = sum(
-                p.numel() * p.element_size() for p in obj.parameters()
-            )
-            if param_bytes > 100_000_000:  # > 100MB
-                size_gb = param_bytes / 1e9
-                print(f"  Freeing adaptor.{name} ({size_gb:.1f} GB)")
-                try:
-                    delattr(adaptor, name)
-                    freed += param_bytes
-                except Exception:
-                    pass
-
-    # Also check nn.Module named children
-    for name, module in list(adaptor.named_children()):
+    # Log all modules so we can see what's there
+    print("  Adaptor modules:")
+    for name, module in adaptor.named_children():
         param_bytes = sum(
             p.numel() * p.element_size() for p in module.parameters()
         )
-        if param_bytes > 100_000_000:
+        print(f"    {name}: {param_bytes / 1e6:.0f} MB")
+
+    # Only delete modules > 1GB (the text encoder is ~7.5GB,
+    # vision projection heads like feat_mlp are < 1GB)
+    for name in list(dict(adaptor.named_children()).keys()):
+        module = getattr(adaptor, name)
+        param_bytes = sum(
+            p.numel() * p.element_size() for p in module.parameters()
+        )
+        if param_bytes > 1_000_000_000:  # > 1GB
             size_gb = param_bytes / 1e9
-            print(f"  Freeing adaptor child '{name}' ({size_gb:.1f} GB)")
+            print(f"  Freeing adaptor.{name} ({size_gb:.1f} GB)")
             try:
                 delattr(adaptor, name)
                 freed += param_bytes
@@ -140,7 +132,7 @@ def _free_text_encoder(adaptor):
     if freed > 0:
         print(f"  Total freed: {freed / 1e9:.1f} GB")
     else:
-        print("  Warning: could not identify text encoder to free")
+        print("  Warning: no module > 1GB found — text encoder may still be in RAM")
 
 
 def get_model():