diff --git a/QEfficient/cloud/finetune.py b/QEfficient/cloud/finetune.py
index c440e73c0..ea36ca93f 100644
--- a/QEfficient/cloud/finetune.py
+++ b/QEfficient/cloud/finetune.py
@@ -9,7 +9,6 @@
 import warnings
 from typing import Any, Dict, Optional, Union
 
-import fire
 import numpy as np
 import torch
 import torch.distributed as dist
@@ -31,6 +30,7 @@
     get_custom_data_collator,
     get_preprocessed_dataset,
 )
+from QEfficient.finetune.utils.parser import get_finetune_parser
 from QEfficient.finetune.utils.train_utils import get_longest_seq_length, print_model_size, train
 from QEfficient.utils._utils import login_and_download_hf_lm
 
@@ -354,4 +354,6 @@ def main(peft_config_file: str = None, **kwargs) -> None:
 
 
 if __name__ == "__main__":
-    fire.Fire(main)
+    parser = get_finetune_parser()
+    args = parser.parse_args()
+    main(**args.__dict__)
diff --git a/QEfficient/finetune/configs/training.py b/QEfficient/finetune/configs/training.py
index 69b083b6a..32889954d 100644
--- a/QEfficient/finetune/configs/training.py
+++ b/QEfficient/finetune/configs/training.py
@@ -16,7 +16,8 @@ class TrainConfig:
         model_name (str): Name of the pre-trained model to fine-tune (default: "meta-llama/Llama-3.2-1B").
         tokenizer_name (str): Name of the tokenizer (defaults to model_name if None).
         run_validation (bool): Whether to run validation during training (default: True).
-        batch_size_training (int): Batch size for training (default: 1).
+        train_batch_size (int): Batch size for training (default: 1).
+        val_batch_size (int): Batch size for validation (default: 1).
         context_length (Optional[int]): Maximum sequence length for inputs (default: None).
         gradient_accumulation_steps (int): Steps for gradient accumulation (default: 4).
         gradient checkpointing (bool): Enable gradient checkpointing to save the memory by compromising the speed. (default: False).
@@ -29,9 +30,10 @@ class TrainConfig:
         weight_decay (float): Weight decay for optimizer (default: 0.0).
         gamma (float): Learning rate decay factor (default: 0.85).
         seed (int): Random seed for reproducibility (default: 42).
+        use_autocast (bool): Use autocast for mixed precision (default: True).
+        dataset (str): Dataset name for training (default: "samsum_dataset").
         use_fp16 (bool): Use mixed precision training (default: True).
         use_autocast (bool): Use autocast for mixed precision (default: True).
-        val_batch_size (int): Batch size for validation (default: 1).
         dataset (str): Dataset name for training (default: "samsum_dataset").
         task_type (str): Type of task for which the finetuning is to be done. Options: "generation" and "seq_classification". (default: "generation")
         peft_method (str): Parameter-efficient fine-tuning method (default: "lora").
@@ -39,7 +41,6 @@ class TrainConfig:
         from_peft_checkpoint (str): Path to PEFT checkpoint (default: "").
         output_dir (str): Directory to save outputs (default: "meta-llama-samsum").
         num_freeze_layers (int): Number of layers to freeze (default: 1).
-        one_qaic (bool): Use single QAIC device (default: False).
         save_model (bool): Save the trained model (default: True).
         save_metrics (bool): Save training metrics (default: True).
         intermediate_step_save (int): Steps between intermediate saves (default: 1000).
@@ -58,7 +59,8 @@ class TrainConfig:
     model_name: str = "meta-llama/Llama-3.2-1B"
     tokenizer_name: str = None  # if not passed as an argument, it uses the value of model_name
     run_validation: bool = True
-    batch_size_training: int = 1
+    train_batch_size: int = 1
+    val_batch_size: int = 1
     context_length: int = None
     gradient_accumulation_steps: int = 4
     gradient_checkpointing: bool = False
@@ -71,9 +73,7 @@ class TrainConfig:
     weight_decay: float = 0.0
     gamma: float = 0.85  # multiplicatively decay the learning rate by gamma after each epoch
     seed: int = 42
-    use_fp16: bool = True
     use_autocast: bool = True
-    val_batch_size: int = 1
     dataset = "samsum_dataset"
     task_type = "generation"  # "generation" / "seq_classification"
     peft_method: str = "lora"
@@ -81,7 +81,6 @@ class TrainConfig:
     from_peft_checkpoint: str = ""  # if not empty and use_peft=True, will load the peft checkpoint and resume the fine-tuning on that checkpoint
     output_dir: str = "meta-llama-samsum"
     num_freeze_layers: int = 1
-    one_qaic: bool = False
     save_model: bool = True
     save_metrics: bool = True  # saves training metrics to a json file for later plotting
     intermediate_step_save: int = 1000
diff --git a/QEfficient/finetune/utils/config_utils.py b/QEfficient/finetune/utils/config_utils.py
index c5c7fe615..6eb9984f2 100644
--- a/QEfficient/finetune/utils/config_utils.py
+++ b/QEfficient/finetune/utils/config_utils.py
@@ -117,7 +117,7 @@ def generate_dataset_config(dataset_name: str) -> Any:
 
 def get_dataloader_kwargs(train_config, dataset, dataset_processer, mode):
     kwargs = {}
-    batch_size = train_config.batch_size_training if mode == "train" else train_config.val_batch_size
+    batch_size = train_config.train_batch_size if mode == "train" else train_config.val_batch_size
     if train_config.enable_ddp:
         if train_config.enable_sorting_for_ddp:
             if train_config.context_length:
diff --git a/QEfficient/finetune/utils/dataset_utils.py b/QEfficient/finetune/utils/dataset_utils.py
index b3447e0db..c1c110380 100644
--- a/QEfficient/finetune/utils/dataset_utils.py
+++ b/QEfficient/finetune/utils/dataset_utils.py
@@ -9,7 +9,6 @@
 
 # from QEfficient.finetune.data.concatenator import ConcatDataset
 from QEfficient.finetune.dataset.dataset_config import DATALOADER_COLLATE_FUNC, DATASET_PREPROC
-from QEfficient.finetune.utils.config_utils import get_dataloader_kwargs
 
 
 def get_preprocessed_dataset(
@@ -29,20 +28,3 @@ def get_custom_data_collator(dataset_processer, dataset_config) -> torch.utils.d
         return None
 
     return DATALOADER_COLLATE_FUNC[dataset_config.dataset](dataset_processer, dataset_config)
-
-
-def get_dataloader(tokenizer, dataset_config, train_config, split: str = "train"):
-    dataset = get_preprocessed_dataset(tokenizer, dataset_config, split)
-    dl_kwargs = get_dataloader_kwargs(train_config, dataset, tokenizer, split)
-
-    # if split == "train" and train_config.batching_strategy == "packing":
-    #    dataset = ConcatDataset(dataset, chunk_size=train_config.context_length)
-
-    # Create data loader
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        num_workers=train_config.num_workers_dataloader,
-        pin_memory=True,
-        **dl_kwargs,
-    )
-    return dataloader
diff --git a/QEfficient/finetune/utils/parser.py b/QEfficient/finetune/utils/parser.py
new file mode 100644
index 000000000..5690adbb8
--- /dev/null
+++ b/QEfficient/finetune/utils/parser.py
@@ -0,0 +1,242 @@
+import argparse
+
+
+def get_finetune_parser():
+    parser = argparse.ArgumentParser(
+        description="Finetune command, the model will be downloaded from Huggingface and finetuned on Cloud AI 100 and weights are saved."
+    )
+    parser.add_argument(
+        "--model-name",
+        "--model_name",
+        required=False,
+        type=str,
+        default="meta-llama/Llama-3.2-1B",
+        help="Name of the pre-trained model to fine-tune",
+    )
+    parser.add_argument(
+        "--tokenizer-name",
+        "--tokenizer_name",
+        required=False,
+        type=str,
+        default=None,
+        help="Name of the tokenizer,if not passed as an argument, it uses the value of model_name",
+    )
+    parser.add_argument(
+        "--run-validation",
+        "--run_validation",
+        required=False,
+        type=bool,
+        default=True,
+        help="To run validation during training",
+    )
+    parser.add_argument(
+        "--train-batch-size", "--train_batch_size", required=False, type=int, default=1, help="Batch size for training"
+    )
+    parser.add_argument(
+        "--val-batch-size", "--val_batch_size", required=False, type=int, default=1, help="Batch size for validation"
+    )
+    parser.add_argument(
+        "--context-length",
+        "--context_length",
+        required=False,
+        type=int,
+        default=None,
+        help="Maximum sequence length for inputs",
+    )
+    parser.add_argument(
+        "--gradient-accumulation-steps",
+        "--gradient_accumulation_steps",
+        required=False,
+        type=int,
+        default=4,
+        help="Steps for gradient accumulation",
+    )
+    parser.add_argument(
+        "--gradient-checkpointing",
+        "--gradient_checkpointing",
+        required=False,
+        type=bool,
+        default=False,
+        help="Use gradient checkpointing",
+    )
+    parser.add_argument(
+        "--num-epochs", "--num_epochs", required=False, type=int, default=1, help="Number of training epochs"
+    )
+    parser.add_argument(
+        "--max-train-step",
+        "--max_train_step",
+        required=False,
+        type=int,
+        default=0,
+        help="Maximum training steps, unlimited if 0",
+    )
+    parser.add_argument(
+        "--max-eval-step",
+        "--max_eval_step",
+        required=False,
+        type=int,
+        default=0,
+        help="Maximum evaluation steps, unlimited if 0",
+    )
+    parser.add_argument("--device", required=False, type=str, default="qaic", help="Device to train on")
+    parser.add_argument(
+        "--num-workers-dataloader",
+        "--num_workers_dataloader",
+        required=False,
+        type=int,
+        default=1,
+        help="Number of workers for dataloader",
+    )
+    parser.add_argument("--lr", required=False, type=float, default=3e-4, help="Learning rate ")
+    parser.add_argument(
+        "--weight-decay", "--weight_decay", required=False, type=float, default=0.0, help="Weight decay for optimizer"
+    )
+    parser.add_argument(
+        "--gamma",
+        required=False,
+        type=float,
+        default=0.85,
+        help="Learning rate decay factor, multiplicatively decays the learning rate by gamma after each epoch",
+    )
+    parser.add_argument("--seed", required=False, type=int, default=42, help="Random seed for reproducibility")
+    parser.add_argument(
+        "--use-autocast",
+        "--use_autocast",
+        required=False,
+        type=bool,
+        default=True,
+        help="Use autocast for mixed precision",
+    )
+
+    parser.add_argument(
+        "--dataset", required=False, type=str, default="samsum_dataset", help="Dataset name for finetuning"
+    )
+    parser.add_argument(
+        "--task-type",
+        "--task_type",
+        required=False,
+        type=str,
+        default="generation",
+        help="generation/seq_classification",
+    )
+    parser.add_argument(
+        "--peft-method",
+        "--peft_method",
+        required=False,
+        type=str,
+        default="lora",
+        help="Parameter-efficient fine-tuning method",
+    )
+    parser.add_argument(
+        "--use-peft",
+        "--use_peft",
+        required=False,
+        type=bool,
+        default=True,
+        help="Whether to use PEFT(parameter efficient fine tuning)",
+    )
+    parser.add_argument(
+        "--from-peft-checkpoint",
+        "--from_peft_checkpoint",
+        required=False,
+        type=str,
+        default="",
+        help="Path to load PEFT checkpoint and resume the fine-tuning on that checkpoint",
+    )
+    parser.add_argument(
+        "--output-dir",
+        "--output_dir",
+        required=False,
+        type=str,
+        default="meta-llama-samsum",
+        help="Directory to save outputs",
+    )
+    parser.add_argument(
+        "--num-freeze-layers",
+        "--num_freeze_layers",
+        required=False,
+        type=int,
+        default=1,
+        help="Number of layers to freeze",
+    )
+    parser.add_argument(
+        "--save-model", "--save_model", required=False, type=bool, default=True, help="Save the trained model"
+    )
+    parser.add_argument(
+        "--save-metrics",
+        "--save_metrics",
+        required=False,
+        type=bool,
+        default=True,
+        help="Save training metrics to a json file for later plotting",
+    )
+    parser.add_argument(
+        "--intermediate-step-save",
+        "--intermediate_step_save",
+        required=False,
+        type=int,
+        default=1000,
+        help="Steps between intermediate saves",
+    )
+    parser.add_argument(
+        "--batching-strategy",
+        "--batching_strategy",
+        required=False,
+        type=str,
+        default="packing",
+        help="Batching strategy",
+    )
+    parser.add_argument(
+        "--enable-sorting-for-ddp",
+        "--enable_sorting_for_ddp",
+        required=False,
+        type=bool,
+        default=True,
+        help="Sort data for DDP",
+    )
+    parser.add_argument(
+        "--convergence-counter",
+        "--convergence_counter",
+        required=False,
+        type=int,
+        default=5,
+        help="Steps to check convergence, its value should be >= 1, stop fine tuning when loss <= convergence_loss (defined below) for #convergence_counter steps",
+    )
+    parser.add_argument(
+        "--convergence-loss",
+        "--convergence_loss",
+        required=False,
+        type=float,
+        default=1e-4,
+        help="Loss threshold for convergence, if loss value is <= convergence_loss for #convergence_counter consecutive steps, fine tuning stops",
+    )
+    parser.add_argument(
+        "--use-profiler", "--use_profiler", required=False, type=bool, default=False, help="Enable profiling"
+    )
+    parser.add_argument(
+        "--enable-ddp", "--enable_ddp", required=False, type=int, default=1000, help="Enable distributed data parallel"
+    )
+    parser.add_argument(
+        "--dist-backend",
+        "--dist_backend",
+        required=False,
+        type=str,
+        default="cpu:gloo,qaic:qccl,cuda:gloo",
+        help="Backend for distributed training",
+    )
+    parser.add_argument(
+        "--grad-scaler", "--grad_scaler", required=False, type=bool, default=True, help="Use gradient scaler"
+    )
+    parser.add_argument(
+        "--dump-root-dir",
+        "--dump_root_dir",
+        required=False,
+        type=str,
+        default="meta-llama-samsum-mismatches/step_",
+        help="Directory for mismatch dumps",
+    )
+    parser.add_argument(
+        "--opByOpVerifier", required=False, type=bool, default=True, help="Enable operation-by-operation verification"
+    )
+
+    return parser
diff --git a/QEfficient/finetune/utils/train_utils.py b/QEfficient/finetune/utils/train_utils.py
index 8693ae32d..02035b388 100644
--- a/QEfficient/finetune/utils/train_utils.py
+++ b/QEfficient/finetune/utils/train_utils.py
@@ -469,14 +469,6 @@ def get_longest_seq_length(data: List[Dict]) -> Tuple[int, int]:
     return longest_seq_length, longest_seq_ix
 
 
-def get_parameter_dtypes(model):
-    """Get the data types of model parameters"""
-    parameter_dtypes = {}
-    for name, parameter in model.named_parameters():
-        parameter_dtypes[name] = parameter.dtype
-    return parameter_dtypes
-
-
 def print_model_size(model, config) -> None:
     """
     Print model name, the number of trainable parameters and initialization time.
diff --git a/docs/source/finetune.md b/docs/source/finetune.md
index 40df4401c..e5a8bc475 100644
--- a/docs/source/finetune.md
+++ b/docs/source/finetune.md
@@ -13,6 +13,7 @@ For torch_qaic, assuming QEfficient is already installed,
 ```bash
 pip install /opt/qti-aic/integrations/torch_qaic/py310/torch_qaic-0.1.0-cp310-cp310-linux_x86_64.whl
 ```
+If qeff-env inside docker is used then torch_qaic and accelerate packages are already installed.
 
 ## Finetuning