diff --git a/src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py b/src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py
index c6c3be62c4..c8a069784d 100644
--- a/src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py
+++ b/src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py
@@ -191,20 +191,21 @@ def convert_bloom_checkpoint_to_pytorch(
                     tensors[key] = tensors[key] / pretraining_tp
 
             other_keys = model.load_state_dict(tensors, strict=False)
-            assert not other_keys.unexpected_keys
+            assert not other_keys.unexpected_keys, f"The keys {other_keys.unexpected_keys} are unexpected"
             if missing_keys is None:
                 missing_keys = set(other_keys.missing_keys)
             else:
                 missing_keys = missing_keys.intersection(set(other_keys.missing_keys))
 
-        assert not missing_keys
+        assert not missing_keys, f"The keys {missing_keys} are missing"
 
         # Save pytorch-model
         os.makedirs(pytorch_dump_folder_path, exist_ok=True)
         pytorch_weights_dump_path = pytorch_dump_folder_path + "/" + WEIGHTS_NAME
         pytorch_config_dump_path = pytorch_dump_folder_path + "/" + CONFIG_NAME
         print(f"Save PyTorch model to {pytorch_weights_dump_path} with dtype {config.torch_dtype}")
-        model = model.to(config.torch_dtype)
+        if config.torch_dtype is not None:
+            model = model.to(config.torch_dtype)
         torch.save(model.state_dict(), pytorch_weights_dump_path)
         print(f"Save configuration file to {pytorch_config_dump_path}")
         with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
diff --git a/src/transformers/models/gpt2/CONVERSION.md b/src/transformers/models/gpt2/CONVERSION.md
new file mode 100644
index 0000000000..d42ea1db9c
--- /dev/null
+++ b/src/transformers/models/gpt2/CONVERSION.md
@@ -0,0 +1,9 @@
+Here is how to convert a GPT2 model generated outside of `transformers`
+
+* [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)-generated model:
+
+Use [convert_megatron_gpt2_checkpoint.py](../megatron_gpt2/convert_megatron_gpt2_checkpoint.py)
+
+* [big-science fork of Megatron-Deepspeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed/)-generated model:
+
+Use the instructions [here](https://github.com/bigscience-workshop/bigscience/tree/aa872e754106f6678e8a9dac8c6962404ba39a6d/train/tr1-13B-base#checkpoint-conversion-and-upload). This approach uses a set of scripts that require the use of this particular fork of Megatron-Deepspeed.