Improve READMEs (#3)

Align with open-source repo
Add SQFT bibtex and Remove abstract
Updated device_map configuration to auto
refactor: move datasets to global scope
Update environment in notebooks
feat: Add support for automatic generation of NNCF Config files
refactor: Replace ${...} with <...> to improve readability
Add device_map in check_sparsity.py
Refactor type annotations for compatibility (Python versions below 3.9)
Add missing 'fi' to if-else statements in script
---------

Signed-off-by: Yuan0320 <jinjie.yuan@intel.com>
Co-authored-by: J. Pablo Muñoz <pablo.munoz@intel.com>

Author: Yuan0320

LoNAS/README.md

@@ -170,13 +170,25 @@ Please refer to `running_commands` for all commands related to reproducing the p
 # Citation
 
 ```bibtex
-@inproceedings{
-munoz2024lonas,
-title={LoNAS: Elastic Low-Rank Adapters for Efficient Large Language Models},
-author={J. Pablo Muñoz and Jinjie Yuan and Yi Zheng and Nilesh Jain},
-booktitle={The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation},
-year={2024},
-url={https://aclanthology.org/2024.lrec-main.940/}
+@inproceedings{munoz-etal-2024-lonas,
+    title = "{L}o{NAS}: Elastic Low-Rank Adapters for Efficient Large Language Models",
+    author = "Munoz, Juan Pablo  and
+      Yuan, Jinjie  and
+      Zheng, Yi  and
+      Jain, Nilesh",
+    editor = "Calzolari, Nicoletta  and
+      Kan, Min-Yen  and
+      Hoste, Veronique  and
+      Lenci, Alessandro  and
+      Sakti, Sakriani  and
+      Xue, Nianwen",
+    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
+    month = may,
+    year = "2024",
+    address = "Torino, Italia",
+    publisher = "ELRA and ICCL",
+    url = "https://aclanthology.org/2024.lrec-main.940",
+    pages = "10760--10776",
 }
 ```
 

LoNAS/run_commonsense.py

@@ -224,7 +224,7 @@ def main():
         model_args.model_name_or_path,
         load_in_8bit=False,
         torch_dtype=torch.float16,
-        device_map={"": 0},
+        device_map="auto",
         trust_remote_code=True,
         cache_dir=model_args.cache_dir,
     )
@@ -243,7 +243,7 @@ def main():
         model.print_trainable_parameters()
     elif training_args.lora:
         logger.info("Loading LoRA modules...")
-        model = PeftModel.from_pretrained(model, model_args.lora_weights, torch_dtype=torch.float16, device_map={"": 0})
+        model = PeftModel.from_pretrained(model, model_args.lora_weights, torch_dtype=torch.float16, device_map="auto")
 
     nncf_config = None
     if training_args.nncf_config is not None:

LoNAS/run_glue.py

@@ -558,7 +558,7 @@ def get_inputs(self, dataloader_output):
         model.print_trainable_parameters()
     elif training_args.lora:
         logger.info("Loading LoRA modules...")
-        model = PeftModel.from_pretrained(model, model_args.lora_weights, device_map={"": 0})
+        model = PeftModel.from_pretrained(model, model_args.lora_weights, device_map="auto")
 
     compression_ctrl = None
     if nncf_config is not None:

LoNAS/run_math.py

@@ -244,7 +244,7 @@ def main():
         model_args.model_name_or_path,
         load_in_8bit=False,
         torch_dtype=torch.float16,
-        device_map={"": 0},
+        device_map="auto",
         trust_remote_code=True,
         cache_dir=model_args.cache_dir,
     )
@@ -263,7 +263,7 @@ def main():
         model.print_trainable_parameters()
     elif training_args.lora:
         logger.info("Loading LoRA modules...")
-        model = PeftModel.from_pretrained(model, model_args.lora_weights, torch_dtype=torch.float16, device_map={"": 0})
+        model = PeftModel.from_pretrained(model, model_args.lora_weights, torch_dtype=torch.float16, device_map="auto")
 
     nncf_config = None
     if training_args.nncf_config is not None:

SQFT/README.md

@@ -11,6 +11,8 @@
 from peft import PeftModel
 
 
+DATASETS = ["mawps", "SVAMP", "gsm8k"]
+
 def extract_answer_number(dataset_name: str, sentence: str) -> float:
     """
     Extracts the numerical answer from a given sentence based on the dataset type.
@@ -194,9 +196,7 @@ def main():
     dtype = args.dtype
     adapter_model_path = args.adapter_model_path
     output_dir = args.output_dir
-
     os.makedirs(output_dir, exist_ok=True)
-    datasets = ["mawps", "SVAMP", "gsm8k"]
 
     tokenizer = AutoTokenizer.from_pretrained(
         base_model_path if tokenizer_path is None else tokenizer_path,
@@ -205,13 +205,13 @@ def main():
 
     model = AutoModelForCausalLM.from_pretrained(
         base_model_path,
-        device_map={"": 0},
+        device_map="auto",
         trust_remote_code=True,
         torch_dtype=dtype,
     )
 
     if adapter_model_path is not None:
-        model = PeftModel.from_pretrained(model, adapter_model_path, torch_dtype=dtype, device_map={"": 0})
+        model = PeftModel.from_pretrained(model, adapter_model_path, torch_dtype=dtype, device_map="auto")
 
     model.eval()
 
@@ -222,7 +222,7 @@ def main():
     else:
         dataset_to_accuracy = {}
 
-    for dataset_name in datasets:
+    for dataset_name in DATASETS:
         if dataset_name in dataset_to_accuracy:
             continue
         print(f"*** Evaluation on {dataset_name} ***")
@@ -231,7 +231,7 @@ def main():
         dataset_to_accuracy[dataset_name] = accuracy
         print(f"{dataset_name} - Accuracy: {accuracy}")
 
-    accuracies = [acc for dataset_name, acc in dataset_to_accuracy.items() if dataset_name in datasets]
+    accuracies = [acc for dataset_name, acc in dataset_to_accuracy.items() if dataset_name in DATASETS]
     average_accuracy = sum(accuracies) / len(accuracies)
     dataset_to_accuracy["average"] = average_accuracy
     with open(result_file_path, "w") as file:

SQFT/eval/evaluate_math.py

@@ -9,19 +9,19 @@ mkdir third_party && cd third_party
 
 # transformers
 git clone https://github.com/huggingface/transformers.git
-cd transformers && git checkout v4.44.2 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/transformers-modifications-for-sqft-usage.patch && pip install -e . && cd ..
+cd transformers && git checkout v4.44.2 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/transformers-v4.44.2.patch && pip install -e . && cd ..
 
 # peft
 git clone https://github.com/huggingface/peft.git
-cd peft && git checkout v0.10.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/peft-modifications-for-sqft-usage.patch && pip install -e . && cd ..
+cd peft && git checkout v0.10.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/peft-v0.10.0.patch && pip install -e . && cd ..
 
 pip install datasets accelerate sentencepiece protobuf
 pip install optimum==1.18.0 --no-deps
 pip install git+https://github.com/AutoGPTQ/AutoGPTQ@866b4c8
 
 # nncf
 git clone https://github.com/openvinotoolkit/nncf.git
-cd nncf && git checkout v2.12.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/nncf-modifications-for-sqft-usage.patch && pip install -e . && cd ..
+cd nncf && git checkout v2.12.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/nncf-v2.12.0.patch && pip install -e . && cd ..
 
 # lm-eval-harness
 pip install lm-eval==0.4.2

SQFT/install.sh

@@ -5,7 +5,7 @@ set -x
 pip install 'numpy<2.0.0' setuptools==69.5.1 wheel
 
 # transformers
-pip install transformers==v4.44.2
+pip install transformers==4.44.2
 pip install datasets accelerate sentencepiece protobuf
 pip install optimum==1.18.0 --no-deps
 pip install git+https://github.com/AutoGPTQ/AutoGPTQ@866b4c8
@@ -14,7 +14,7 @@ pip install git+https://github.com/AutoGPTQ/AutoGPTQ@866b4c8
 SQFT_PATH=$PWD
 mkdir third_party_inference && cd third_party_inference
 git clone https://github.com/huggingface/peft.git
-cd peft && git checkout v0.10.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/peft-modifications-for-sqft-usage.patch && pip install -e . && cd ..
+cd peft && git checkout v0.10.0 && git apply --ignore-space-change --ignore-whitespace ${SQFT_PATH}/patches/peft-v0.10.0.patch && pip install -e . && cd ..
 
 # lm-eval-harness (for evaluation)
 pip install lm-eval==0.4.2

SQFT/install_inference.sh

@@ -1,4 +1,4 @@
-from typing import Optional
+from typing import List, Optional
 
 import numpy as np
 import torch
@@ -197,11 +197,11 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
 
         return result
 
-    def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None:
+    def merge(self, safe_merge: bool = False, adapter_names: Optional[List[str]] = None) -> None:
         """
         Merges the adapter weights into the base weight.
 
-        Referenced from:
+        Reference:
         https://github.com/AutoGPTQ/AutoGPTQ/blob/866b4c8c2cbb893f1156cb6c114625bba2e4d7c5/auto_gptq/nn_modules/qlinear/qlinear_cuda_old.py#L135-L140
         """
         base_weight = self.base_weight