Attempt at fixing 4-bit with CPU offload

modules/GPTQ_loader.py

@@ -1,3 +1,4 @@
+import re
 import sys
 from pathlib import Path
 
@@ -56,16 +57,20 @@ def load_quantized(model_name):
 
     # Multiple GPUs or GPU+CPU
     if shared.args.gpu_memory:
+        memory_map = list(map(lambda x : x.strip(), shared.args.gpu_memory))
+        max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
         max_memory = {}
-        for i in range(len(shared.args.gpu_memory)):
-            max_memory[i] = f"{shared.args.gpu_memory[i]}GiB"
-        max_memory['cpu'] = f"{shared.args.cpu_memory or '99'}GiB"
+        for i in range(len(memory_map)):
+            max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
+        max_memory['cpu'] = max_cpu_memory
 
         device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
-        model = accelerate.dispatch_model(model, device_map=device_map)
+        print("Using the following device map for the 4-bit model:", device_map)
+        # https://huggingface.co/docs/accelerate/package_reference/big_modeling#accelerate.dispatch_model
+        model = accelerate.dispatch_model(model, device_map=device_map, offload_buffers=True)
 
     # Single GPU
-    else:
+    elif not shared.args.cpu:
         model = model.to(torch.device('cuda:0'))
 
     return model