diff --git a/requirements.txt b/requirements.txt
index 8bb8d804..530e1dfb 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -9,7 +9,8 @@ datasets>=2.15.0
 numba
 numpy
 rich
-dolomite-engine @ git+https://github.com/ibm-granite/dolomite-engine.git@main
+instructlab-dolomite @ git+https://github.com/instructlab/GPTDolomite.git@initial
+
 trl==0.9.4
 peft
 pydantic>=2.7.0
diff --git a/src/instructlab/training/main_ds.py b/src/instructlab/training/main_ds.py
index eeb0c077..4d18ba7d 100644
--- a/src/instructlab/training/main_ds.py
+++ b/src/instructlab/training/main_ds.py
@@ -43,6 +43,9 @@
     setup_logger,
 )
 import instructlab.training.data_process as dp
+from instructlab.dolomite.hf_models import GPTDolomiteForCausalLM
+from instructlab.dolomite.enums import GradientCheckpointingMethod
+from instructlab.dolomite.gradient_checkpointing import apply_gradient_checkpointing
 
 
 def get_ds_config(world_size, samples_per_gpu, grad_accum, opts: DeepSpeedOptions):
@@ -88,8 +91,6 @@ def setup_model(args, tokenizer, train_loader, grad_accum):
         )
 
     if args.is_granite:
-        # Third Party
-        from dolomite_engine.hf_models.models import GPTDolomiteForCausalLM
 
         model = GPTDolomiteForCausalLM.from_pretrained(
             args.model_name_or_path,
@@ -201,9 +202,6 @@ def setup_model(args, tokenizer, train_loader, grad_accum):
     # granite gradient checkpointing is handled uniformly
     # for both lora and full here
     if args.is_granite:
-        # Third Party
-        from dolomite_engine.enums import GradientCheckpointingMethod
-        from dolomite_engine.gradient_checkpointing import apply_gradient_checkpointing
 
         block_name = model._no_split_modules[0]
         apply_gradient_checkpointing(
diff --git a/src/instructlab/training/utils.py b/src/instructlab/training/utils.py
index 6feaa548..a52fe9a0 100644
--- a/src/instructlab/training/utils.py
+++ b/src/instructlab/training/utils.py
@@ -13,6 +13,7 @@
 import warnings
 
 # Third Party
+from instructlab.dolomite.hf_models import export_to_huggingface
 from rich.logging import RichHandler
 from torch import distributed as dist
 from torch.distributed import get_rank, is_initialized
@@ -539,7 +540,6 @@ def save_hf_format_ds(args, model, tokenizer, samples_seen, convert_granite=True
             from tempfile import TemporaryDirectory
 
             # Third Party
-            from dolomite_engine.hf_models import export_to_huggingface
             from safetensors.torch import save_file
 
             with TemporaryDirectory("w") as tmpdir: