diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
index 42cb3d974b0e..8a89fc0e6841 100644
--- a/src/transformers/modeling_utils.py
+++ b/src/transformers/modeling_utils.py
@@ -475,7 +475,7 @@ def remove_tied_weights_from_state_dict(
             # In offloaded cases, there may be meta tensors in the state_dict.
             # For these cases, key by the pointer of the original tensor object
             # (state_dict tensors are detached and therefore no longer shared)
-            tensor = model.get_parameter(name)
+            tensor = model.get_parameter_or_buffer(name)
             ptrs[id(tensor)].append(name)
 
         else:
@@ -3538,8 +3538,11 @@ def save_pretrained(
         # Remove tied weights as safetensors do not handle them
         state_dict = remove_tied_weights_from_state_dict(state_dict, model_to_save)
 
-        # Revert all renaming and/or weight operations
-        if save_original_format and not _hf_peft_config_loaded:
+        # Revert all renaming and/or weight operations. In general, due to potential many-weights-to-one conversion patterns,
+        # we need to revert the whole state_dict at once to make sure all weights are available. For offloaded models though,
+        # some weights are on meta device, so we need to first load them back into cpu then convert (and we do it later inside a
+        # given shard, to avoid blowing cpu memory since offloading means constrained resources)
+        if save_original_format and not is_offloaded and not _hf_peft_config_loaded:
             state_dict = revert_weight_conversion(model_to_save, state_dict)
 
         # Shard the model if it is too big.
@@ -3553,13 +3556,6 @@ def save_pretrained(
         state_dict_split = split_torch_state_dict_into_shards(
             state_dict, filename_pattern=filename_pattern, max_shard_size=max_shard_size
         )
-        # Save index if sharded
-        index = None
-        if state_dict_split.is_sharded:
-            index = {
-                "metadata": {"total_parameters": self.num_parameters(), **state_dict_split.metadata},
-                "weight_map": state_dict_split.tensor_to_filename,
-            }
 
         # Clean the folder from a previous save
         for filename in os.listdir(save_directory):
@@ -3581,6 +3577,17 @@ def save_pretrained(
             ):
                 os.remove(full_filename)
 
+        # The weight_map may change compared to state_dict_split.tensor_to_filename due to revert weight conversions
+        weight_map = None
+        if state_dict_split.is_sharded:
+            # For offloaded weights, we will convert each shard later, so the weight names will change and we will fill
+            # the weight_map as we get them
+            weight_map = (
+                state_dict_split.tensor_to_filename
+                if not (is_offloaded and save_original_format and not _hf_peft_config_loaded)
+                else {}
+            )
+
         # Save the model
         for shard_file, tensor_names in logging.tqdm(
             state_dict_split.filename_to_tensors.items(), desc="Writing model shards"
@@ -3600,6 +3607,23 @@ def save_pretrained(
                 # only do contiguous after it's permuted correctly in case of TP
                 shard_state_dict[tensor_name] = tensor.contiguous()
 
+            # As explained above, for offloaded scenarios, weight format could not be reverted before due to meta weights,
+            # so do it now after they were loaded onto cpu. For one-weight-to-many operations, it may be an issue, but usually the shards
+            # contain all the necessary params, except if we are quite unlucky on the sharding. The failure surface is (very few models
+            # with one-weight-to-many + offloading to disk + unlucky sharding), so it will almost never happen
+            if is_offloaded and save_original_format and not _hf_peft_config_loaded:
+                try:
+                    shard_state_dict = revert_weight_conversion(model_to_save, shard_state_dict)
+                    # Save the weight_map, since some names etc may have changed due to conversion compared to initial `state_dict_split`
+                    if state_dict_split.is_sharded:
+                        weight_map.update({k: os.path.basename(shard_file)} for k in shard_state_dict.keys())  # ty: ignore[unresolved-attribute]
+                except Exception:
+                    raise RuntimeError(
+                        "We could not revert some weight conversions because of offlading, and several weights needed for a single "
+                        "conversion operation living in different shard files. Try reducing `max_shard_size` a bit, or worst case "
+                        "set `save_original_format=False`."
+                    )
+
             # TODO: it would be very nice to do the writing concurrently, but safetensors never releases the GIL,
             # so it's not possible for now....
             # Write the shard to disk
@@ -3607,6 +3631,14 @@ def save_pretrained(
             # Cleanup the data before next loop (important with offloading, so we don't blowup cpu RAM)
             del shard_state_dict
 
+        # Save index if sharded
+        index = None
+        if state_dict_split.is_sharded:
+            index = {
+                "metadata": {"total_parameters": self.num_parameters(), **state_dict_split.metadata},
+                "weight_map": weight_map,
+            }
+
         if index is None:
             path_to_weights = os.path.join(save_directory, weights_name)
             logger.info(f"Model weights saved in {path_to_weights}")