diff --git a/ethosu/vela/tensor_allocation.py b/ethosu/vela/tensor_allocation.py
index 1e5eb85..0a7da5d 100644
--- a/ethosu/vela/tensor_allocation.py
+++ b/ethosu/vela/tensor_allocation.py
@@ -202,12 +202,6 @@
             else:
                 sg.memory_used_per_type[mem_type] += total_sz
 
-        if mem_area == arch.fast_storage_mem_area:
-            for tens in lrs.ranges:
-                if tens.purpose == TensorPurpose.Weights:
-                    nng.total_compressed_weights += tens.storage_size()
-                    nng.total_original_weights += tens.elements() * tens.element_size()
-
         print_allocation(lrs, mem_area, mem_type_set, sg, verbose_allocation)
 
         if mem_area == MemArea.Sram:
@@ -217,10 +211,9 @@
 
     if sg == nng.get_root_subgraph():
         nng.memory_used = sg.memory_used
-        if mem_area == arch.fast_storage_mem_area:
-            try:
-                nng.weights_compression_ratio = nng.total_compressed_weights / nng.total_original_weights
-            except ZeroDivisionError:
-                nng.weights_compression_ratio = 0.0
+        try:
+            nng.weights_compression_ratio = nng.total_compressed_weights / nng.total_original_weights
+        except ZeroDivisionError:
+            nng.weights_compression_ratio = 0.0
 
     return True
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 188b16a..60ca846 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -521,6 +521,9 @@
                 compress_weights(
                     arch, nng, tens, op.type.npu_block_type, ps.block_config[-1], ofm_depth_step, op.get_dilation_h_w()
                 )
+                nng.total_compressed_weights += tens.weight_compressed_offsets[-1]
+                nng.total_original_weights += tens.elements() * tens.element_size()
+
                 # Update source tensor
                 if needs_dma:
                     src_tens = tens.get_dma_src_tensor()