diff --git a/speculator/train_speculator_utils.py b/speculator/train_speculator_utils.py
index 5cf31f9..1403732 100644
--- a/speculator/train_speculator_utils.py
+++ b/speculator/train_speculator_utils.py
@@ -554,15 +554,18 @@ def factory(**kwargs):
 register_model(
     "embedgpt_bigcode", "20b", _gpt_bigcode_factory_factory(_gpt_bigcode_20b_config)
 )
-serialization.register_adapter("embedgpt_bigcode", "hf", _gptbigcode_hf_sd_to_fms_sd)
+serialization.register_adapter_step("embedgpt_bigcode", "hf_to_fms", _gptbigcode_hf_sd_to_fms_sd)
+serialization.register_adapter("embedgpt_bigcode", "hf", ["hf_to_fms"])
 
 register_model(
     "embedllama", "7b", _llama_factory_factory(get_model_config("llama2_7b"))
 )
 register_model(
     "embedllama", "8b", _llama_factory_factory(get_model_config("llama3_8b"))
-)
-serialization.register_adapter("embedllama", "hf", _llama_hf_sd_to_fms_sd)
+    )
+serialization.register_adapter_step("embedllama", "hf_to_fms", _llama_hf_sd_to_fms_sd)
+serialization.register_adapter("embedllama", "hf", ["hf_to_fms"])
 
 register_model("embedmixtral", "8x7b", _mixtral_factory_factory(MixtralConfig()))
-serialization.register_adapter("embedmixtral", "hf", _mixtral_hf_sd_to_fms_sd)
+serialization.register_adapter_step("embedmixtral", "hf_to_fms", _mixtral_hf_sd_to_fms_sd)
+serialization.register_adapter("embedmixtral", "hf", ["hf_to_fms"])