cpu memory cleanup

interfaces/cap.py

@@ -11,7 +11,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import CAP_NUM_DICT, CAP_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -98,7 +98,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     output_pred = {f"[{CAP_NUM_DICT[i]}] {CAP_LABEL_NAMES[CAP_NUM_DICT[i]]}": probs[i] for i in np.argsort(probs)[::-1]}

interfaces/cap_media_demo.py

@@ -10,7 +10,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import CAP_MEDIA_NUM_DICT, CAP_MEDIA_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -47,7 +47,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     output_pred = {f"[{CAP_MEDIA_NUM_DICT[i]}] {CAP_MEDIA_LABEL_NAMES[CAP_MEDIA_NUM_DICT[i]]}": probs[i] for i in np.argsort(probs)[::-1]}

interfaces/cap_minor.py

@@ -10,7 +10,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import CAP_MIN_NUM_DICT, CAP_MIN_LABEL_NAMES, CAP_LABEL_NAMES 
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 from itertools import islice
 
 def take(n, iterable):
@@ -79,7 +79,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     output_pred = {f"[{'999' if str(CAP_MIN_NUM_DICT[i]) == '999' else str(CAP_MIN_NUM_DICT[i])[:-2]}]{convert_minor_to_major(CAP_MIN_NUM_DICT[i])} [{CAP_MIN_NUM_DICT[i]}]{CAP_MIN_LABEL_NAMES[CAP_MIN_NUM_DICT[i]]}": probs[i] for i in np.argsort(probs)[::-1]}

interfaces/cap_minor_media.py

@@ -15,7 +15,7 @@ from label_dicts import (CAP_MEDIA_NUM_DICT, CAP_MEDIA_LABEL_NAMES,
                         CAP_MIN_NUM_DICT, CAP_MIN_LABEL_NAMES,
                         CAP_MIN_MEDIA_NUM_DICT)
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -116,8 +116,8 @@ def predict(text, major_model_id, minor_model_id, tokenizer_id, HF_TOKEN=None):
             minor_logits = minor_model(**inputs).logits
             minor_probs = F.softmax(minor_logits, dim=-1)
             
-        free_gpu_memory(major_model, major_model_id)
-        free_gpu_memory(minor_model, minor_model_id)
+        release_model(major_model, major_model_id)
+        release_model(minor_model, minor_model_id)
         
         print(minor_probs) # debug
         # Restrict to valid minor codes
@@ -162,7 +162,7 @@ def predict_flat(text, model_id, tokenizer_id, HF_TOKEN=None):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     top_indices = np.argsort(probs)[::-1][:10]

interfaces/emotion.py

@@ -9,7 +9,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import MANIFESTO_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -40,7 +40,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     output_pred = {model.config.id2label[i]: probs[i] for i in np.argsort(probs)[::-1]}

interfaces/emotion9.py

@@ -9,7 +9,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import EMOTION9_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -38,7 +38,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
 

interfaces/illframes.py

@@ -10,7 +10,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import ILLFRAMES_MIGRATION_LABEL_NAMES, ILLFRAMES_COVID_LABEL_NAMES, ILLFRAMES_WAR_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -80,7 +80,7 @@ def predict(text, model_id, tokenizer_id, label_names):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
 

interfaces/manifesto.py

@@ -9,7 +9,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import MANIFESTO_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -38,7 +38,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     output_pred = {f"[{model.config.id2label[i]}] {MANIFESTO_LABEL_NAMES[int(model.config.id2label[i])]}": probs[i] for i in np.argsort(probs)[::-1]}

interfaces/ontolisst.py

@@ -15,7 +15,7 @@ languages = [
 
 from label_dicts import ONTOLISST_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 # --- DEBUG ---
 import shutil
@@ -67,7 +67,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     predicted_class_id = probs.argmax()

interfaces/sentiment.py

@@ -9,7 +9,7 @@ from huggingface_hub import HfApi
 
 from label_dicts import MANIFESTO_LABEL_NAMES
 
-from .utils import is_disk_full, free_gpu_memory
+from .utils import is_disk_full, release_model
 
 HF_TOKEN = os.environ["hf_read"]
 
@@ -43,7 +43,7 @@ def predict(text, model_id, tokenizer_id):
 
     with torch.no_grad():
         logits = model(**inputs).logits
-    free_gpu_memory(model, model_id)
+    release_model(model, model_id)
 
     probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
     predicted_class_id = probs.argmax()

interfaces/utils.py

@@ -1,5 +1,6 @@
 import shutil
 import torch
+import psutil
 import gc
 
 
@@ -15,25 +16,52 @@ def is_disk_full(min_free_space_in_GB=10):
         return True
         
         
-def free_gpu_memory(obj=None, label='Object'):
+def release_model(model=None, label='Model'):
     """
-    Frees GPU memory by deleting the passed object, collecting garbage,
-    and clearing PyTorch's CUDA cache.
+    Releases CPU and GPU memory used by a model or pipeline.
 
     Args:
-        obj (optional): The model or pipeline to delete.
-        label (str): Label for logging (e.g., 'model', 'pipeline').
+        model: The object to delete (e.g., model, pipeline).
+        label: String label for log output.
     """
-    if torch.cuda.is_available():
-        print(f"\n[{label}] Before deletion: {torch.cuda.memory_allocated() / 1e6:.2f} MB")
-    
-    if obj is not None:
-        del obj
+    using_cuda = torch.cuda.is_available()
+    was_cuda = False
+
+    # CPU memory before
+    process = psutil.Process(os.getpid())
+    mem_cpu_before = process.memory_info().rss / 1e6  # MB
+
+    if using_cuda:
+        mem_gpu_before = torch.cuda.memory_allocated()
+        print(f"\n[{label}] GPU memory before release: {mem_gpu_before:.2f} MB")
+
+    print(f"[{label}] CPU memory before release: {mem_cpu_before:.2f} MB")
+
+    # Try to detect if model was on CUDA
+    if model is not None:
+        try:
+            if hasattr(model, 'parameters'):
+                was_cuda = any(p.is_cuda for p in model.parameters())
+        except Exception as e:
+            print(f"[{label}] Could not check device: {e}")
+        del model
 
+    # Garbage collection and cache clearing
     gc.collect()
-    torch.cuda.empty_cache()
+    if using_cuda:
+        if was_cuda:
+            torch.cuda.empty_cache()
+        else:
+            print(f"[{label}] ⚠️ Model was not using CUDA, but CUDA is available.")
+
+    # CPU memory after
+    mem_cpu_after = process.memory_info().rss / 1e6  # MB
+    print(f"[{label}] CPU memory after release:  {mem_cpu_after:.2f} MB")
+
+    if using_cuda:
+        mem_gpu_after = torch.cuda.memory_allocated()
+        print(f"[{label}] GPU memory after release: {mem_gpu_after:.2f} MB\n")
+    else:
+        print(f"[{label}] CUDA not available — GPU memory not tracked.\n")
 
-    if torch.cuda.is_available():
-        print(f"[{label}] After deletion:  {torch.cuda.memory_allocated() / 1e6:.2f} MB\n")
-