add grpo

examples/train_lora/llama3_preprocess.yaml

+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+tokenized_path: saves/llama3-8b/dataset/sft
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+overwrite_output_dir: true

examples/train_lora/llama4_lora_sft_ds3.yaml

+# pip install git+https://github.com/hiyouga/transformers.git@llama4_train
+
+### model
+model_name_or_path: meta-llama/Llama-4-Scout-17B-16E-Instruct
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+deepspeed: examples/deepspeed/ds_z3_config.json  # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]
+
+### dataset
+dataset: mllm_demo,identity,alpaca_en_demo
+template: llama4
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama4-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 2
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+
+### eval
+# eval_dataset: alpaca_en_demo
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_lora/qwen2_5vl_lora_dpo.yaml

+### model
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
+image_max_pixels: 262144
+video_max_pixels: 16384
+trust_remote_code: true
+
+### method
+stage: dpo
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+pref_beta: 0.1
+pref_loss: sigmoid  # choices: [sigmoid (dpo), orpo, simpo]
+
+### dataset
+dataset: rlhf_v
+template: qwen2_vl
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/qwen2_5vl-7b/lora/dpo
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 5.0e-6
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_lora/qwen2_5vl_lora_sft.yaml

+### model
+model_name_or_path: Qwen/Qwen2.5-VL-7B-Instruct
+image_max_pixels: 262144
+video_max_pixels: 16384
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: mllm_demo,identity,alpaca_en_demo  # video: mllm_video_demo
+template: qwen2_vl
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/qwen2_5vl-7b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_qlora/llama3_lora_sft_aqlm.yaml

+### model
+model_name_or_path: ISTA-DASLab/Meta-Llama-3-8B-Instruct-AQLM-2Bit-1x16
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_qlora/llama3_lora_sft_awq.yaml

+### model
+model_name_or_path: TechxGenus/Meta-Llama-3-8B-Instruct-AWQ
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_qlora/llama3_lora_sft_bnb_npu.yaml

+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+quantization_bit: 4
+quantization_method: bnb
+double_quantization: false
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_qlora/llama3_lora_sft_gptq.yaml

+### model
+model_name_or_path: TechxGenus/Meta-Llama-3-8B-Instruct-GPTQ
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

examples/train_qlora/llama3_lora_sft_otfq.yaml

+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+quantization_bit: 4  # choices: [8 (bnb/hqq/eetq), 4 (bnb/hqq), 3 (hqq), 2 (hqq)]
+quantization_method: bnb  # choices: [bnb, hqq, eetq]
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 8
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500

pyproject.toml

+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "llamafactory"
+dynamic = [
+    "version",
+    "dependencies",
+    "optional-dependencies",
+    "requires-python",
+    "scripts",
+    "authors",
+    "description",
+    "readme",
+    "license",
+    "keywords",
+    "classifiers"
+]
+
+[tool.ruff]
+target-version = "py39"
+line-length = 119
+indent-width = 4
+
+[tool.ruff.lint]
+ignore = [
+    "C408", # collection
+    "C901", # complex
+    "E501", # line too long
+    "E731", # lambda function
+    "E741", # ambiguous var name
+    "D100", # no doc public module
+    "D101", # no doc public class
+    "D102", # no doc public method
+    "D103", # no doc public function
+    "D104", # no doc public package
+    "D105", # no doc magic method
+    "D107", # no doc __init__
+]
+extend-select = [
+    "C",      # complexity
+    "E",      # error
+    "F",      # pyflakes
+    "I",      # isort
+    "W",      # warning
+    "UP",     # pyupgrade
+    "D",      # pydocstyle
+    "PT009",  # pytest assert
+    "RUF022", # sort __all__
+]
+
+[tool.ruff.lint.isort]
+lines-after-imports = 2
+known-first-party = ["llamafactory"]
+known-third-party = [
+    "accelerate",
+    "datasets",
+    "gradio",
+    "numpy",
+    "peft",
+    "torch",
+    "transformers",
+    "trl",
+]
+
+[tool.ruff.lint.pydocstyle]
+convention = "google"
+
+[tool.ruff.format]
+quote-style = "double"
+indent-style = "space"
+docstring-code-format = true
+skip-magic-trailing-comma = false
+line-ending = "auto"
+
+[tool.uv]
+conflicts = [
+    [
+        { extra = "torch-npu" },
+        { extra = "aqlm" },
+    ],
+    [
+        { extra = "torch-npu" },
+        { extra = "liger-kernel" },
+    ],
+    [
+        { extra = "torch-npu" },
+        { extra = "vllm" },
+    ],
+    [
+        { extra = "sglang" },
+        { extra = "minicpm_v" },
+    ],
+]

requirements.txt

+# core deps
+transformers>=4.49.0,<=4.52.4,!=4.52.0; sys_platform != 'darwin'
+transformers>=4.49.0,<=4.51.3,!=4.52.0; sys_platform == 'darwin'
+datasets>=2.16.0,<=3.6.0
+accelerate>=1.3.0,<=1.7.0
+peft>=0.14.0,<=0.15.2
+trl>=0.8.6,<=0.9.6
+tokenizers>=0.19.0,<=0.21.1
+# gui
+gradio>=4.38.0,<=5.31.0
+matplotlib>=3.7.0
+tyro<0.9.0
+# ops
+einops
+numpy<2.0.0
+pandas>=2.0.0
+scipy
+# model and tokenizer
+sentencepiece
+tiktoken
+modelscope>=1.14.0
+hf-transfer
+# python
+fire
+omegaconf
+packaging
+protobuf
+pyyaml
+pydantic<=2.10.6
+# api
+uvicorn
+fastapi
+sse-starlette
+# media
+av
+librosa

scripts/api_example/test_image.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+from openai import OpenAI
+from transformers.utils.versions import require_version
+
+
+require_version("openai>=1.5.0", "To fix: pip install openai>=1.5.0")
+
+
+def main():
+    client = OpenAI(
+        api_key="{}".format(os.getenv("API_KEY", "0")),
+        base_url="http://localhost:{}/v1".format(os.getenv("API_PORT", 8000)),
+    )
+    messages = []
+    messages.append(
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Output the color and number of each box."},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-VL/boxes.png"},
+                },
+            ],
+        }
+    )
+    result = client.chat.completions.create(messages=messages, model="test")
+    messages.append(result.choices[0].message)
+    print("Round 1:", result.choices[0].message.content)
+    # The image shows a pyramid of colored blocks with numbers on them. Here are the colors and numbers of ...
+    messages.append(
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What kind of flower is this?"},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-VL/flowers.jpg"},
+                },
+            ],
+        }
+    )
+    result = client.chat.completions.create(messages=messages, model="test")
+    messages.append(result.choices[0].message)
+    print("Round 2:", result.choices[0].message.content)
+    # The image shows a cluster of forget-me-not flowers. Forget-me-nots are small ...
+
+
+if __name__ == "__main__":
+    main()

scripts/api_example/test_toolcall.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+
+from openai import OpenAI
+from transformers.utils.versions import require_version
+
+
+require_version("openai>=1.5.0", "To fix: pip install openai>=1.5.0")
+
+
+def calculate_gpa(grades: list[str], hours: list[int]) -> float:
+    grade_to_score = {"A": 4, "B": 3, "C": 2}
+    total_score, total_hour = 0, 0
+    for grade, hour in zip(grades, hours):
+        total_score += grade_to_score[grade] * hour
+        total_hour += hour
+    return round(total_score / total_hour, 2)
+
+
+def main():
+    client = OpenAI(
+        api_key="{}".format(os.getenv("API_KEY", "0")),
+        base_url="http://localhost:{}/v1".format(os.getenv("API_PORT", 8000)),
+    )
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "calculate_gpa",
+                "description": "Calculate the Grade Point Average (GPA) based on grades and credit hours",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "grades": {"type": "array", "items": {"type": "string"}, "description": "The grades"},
+                        "hours": {"type": "array", "items": {"type": "integer"}, "description": "The credit hours"},
+                    },
+                    "required": ["grades", "hours"],
+                },
+            },
+        }
+    ]
+    tool_map = {"calculate_gpa": calculate_gpa}
+
+    messages = []
+    messages.append({"role": "user", "content": "My grades are A, A, B, and C. The credit hours are 3, 4, 3, and 2."})
+    result = client.chat.completions.create(messages=messages, model="test", tools=tools)
+    if result.choices[0].message.tool_calls is None:
+        raise ValueError("Cannot retrieve function call from the response.")
+
+    messages.append(result.choices[0].message)
+    tool_call = result.choices[0].message.tool_calls[0].function
+    print(tool_call)
+    # Function(arguments='{"grades": ["A", "A", "B", "C"], "hours": [3, 4, 3, 2]}', name='calculate_gpa')
+    name, arguments = tool_call.name, json.loads(tool_call.arguments)
+    tool_result = tool_map[name](**arguments)
+    messages.append({"role": "tool", "content": json.dumps({"gpa": tool_result}, ensure_ascii=False)})
+    result = client.chat.completions.create(messages=messages, model="test", tools=tools)
+    print(result.choices[0].message.content)
+    # Based on the grades and credit hours you provided, your Grade Point Average (GPA) is 3.42.
+
+
+if __name__ == "__main__":
+    main()

scripts/convert_ckpt/llamafy_baichuan2.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from collections import OrderedDict
+from typing import Any
+
+import fire
+import torch
+from huggingface_hub import split_torch_state_dict_into_shards
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers.modeling_utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, WEIGHTS_NAME
+
+
+CONFIG_NAME = "config.json"
+
+
+def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool):
+    baichuan2_state_dict: dict[str, torch.Tensor] = OrderedDict()
+    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
+        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".bin"):
+            shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu", weights_only=True)
+            baichuan2_state_dict.update(shard_weight)
+
+    llama_state_dict: dict[str, torch.Tensor] = OrderedDict()
+    for key, value in tqdm(baichuan2_state_dict.items(), desc="Convert format"):
+        if "W_pack" in key:
+            proj_size = value.size(0) // 3
+            llama_state_dict[key.replace("W_pack", "q_proj")] = value[:proj_size, :]
+            llama_state_dict[key.replace("W_pack", "k_proj")] = value[proj_size : 2 * proj_size, :]
+            llama_state_dict[key.replace("W_pack", "v_proj")] = value[2 * proj_size :, :]
+        elif "lm_head" in key:
+            llama_state_dict[key] = torch.nn.functional.normalize(value)
+        else:
+            llama_state_dict[key] = value
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+    state_dict_split = split_torch_state_dict_into_shards(
+        llama_state_dict, filename_pattern=filename_pattern, max_shard_size=shard_size
+    )
+    for shard_file, tensors in tqdm(state_dict_split.filename_to_tensors.items(), desc="Save weights"):
+        shard = {tensor: llama_state_dict[tensor].contiguous() for tensor in tensors}
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if not state_dict_split.is_sharded:
+        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}.")
+    else:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+
+        print(f"Model weights saved in {output_dir}.")
+
+
+def save_config(input_dir: str, output_dir: str):
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
+        llama2_config_dict: dict[str, Any] = json.load(f)
+
+    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
+    llama2_config_dict.pop("auto_map", None)
+    llama2_config_dict.pop("tokenizer_class", None)
+    llama2_config_dict["model_type"] = "llama"
+
+    with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
+        json.dump(llama2_config_dict, f, indent=2)
+
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
+
+
+def llamafy_baichuan2(
+    input_dir: str,
+    output_dir: str,
+    shard_size: str = "2GB",
+    save_safetensors: bool = True,
+):
+    r"""Convert the Baichuan2-7B model in the same format as LLaMA2-7B.
+
+    Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
+    """
+    try:
+        os.makedirs(output_dir, exist_ok=False)
+    except Exception as e:
+        raise print("Output dir already exists", e)
+
+    save_weight(input_dir, output_dir, shard_size, save_safetensors)
+    save_config(input_dir, output_dir)
+
+
+if __name__ == "__main__":
+    fire.Fire(llamafy_baichuan2)

scripts/convert_ckpt/llamafy_qwen.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from collections import OrderedDict
+from typing import Any
+
+import fire
+import torch
+from huggingface_hub import split_torch_state_dict_into_shards
+from safetensors import safe_open
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers.modeling_utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, WEIGHTS_NAME
+from transformers.utils import check_min_version
+
+
+try:
+    check_min_version("4.34.0")
+except Exception:
+    raise ValueError("Please upgrade `transformers` to 4.34.0")
+
+
+CONFIG_NAME = "config.json"
+
+
+def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool) -> str:
+    qwen_state_dict: dict[str, torch.Tensor] = OrderedDict()
+    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
+        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".safetensors"):
+            with safe_open(os.path.join(input_dir, filepath), framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    qwen_state_dict[key] = f.get_tensor(key)
+
+    llama_state_dict: dict[str, torch.Tensor] = OrderedDict()
+    torch_dtype = None
+    for key, value in tqdm(qwen_state_dict.items(), desc="Convert format"):
+        if torch_dtype is None:
+            torch_dtype = value.dtype
+        if "wte" in key:
+            llama_state_dict["model.embed_tokens.weight"] = value
+        elif "ln_f" in key:
+            llama_state_dict["model.norm.weight"] = value
+        else:
+            key = key.replace("transformer.h", "model.layers")
+            if "attn.c_attn" in key:
+                proj_size = value.size(0) // 3
+                llama_state_dict[key.replace("attn.c_attn", "self_attn.q_proj")] = value[:proj_size, ...]
+                llama_state_dict[key.replace("attn.c_attn", "self_attn.k_proj")] = value[
+                    proj_size : 2 * proj_size, ...
+                ]
+                llama_state_dict[key.replace("attn.c_attn", "self_attn.v_proj")] = value[2 * proj_size :, ...]
+            elif "attn.c_proj" in key:
+                llama_state_dict[key.replace("attn.c_proj", "self_attn.o_proj")] = value
+                llama_state_dict[key.replace("attn.c_proj.weight", "self_attn.o_proj.bias")] = torch.zeros_like(
+                    value[:, 0]
+                ).squeeze()
+            elif "ln_1" in key:
+                llama_state_dict[key.replace("ln_1", "input_layernorm")] = value
+            elif "ln_2" in key:
+                llama_state_dict[key.replace("ln_2", "post_attention_layernorm")] = value
+            elif "mlp.w1" in key:
+                llama_state_dict[key.replace("mlp.w1", "mlp.up_proj")] = value
+            elif "mlp.w2" in key:
+                llama_state_dict[key.replace("mlp.w2", "mlp.gate_proj")] = value
+            elif "mlp.c_proj" in key:
+                llama_state_dict[key.replace("mlp.c_proj", "mlp.down_proj")] = value
+            elif "lm_head" in key:
+                llama_state_dict[key] = value
+            else:
+                raise KeyError(f"Unable to process key {key}")
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+    state_dict_split = split_torch_state_dict_into_shards(
+        llama_state_dict, filename_pattern=filename_pattern, max_shard_size=shard_size
+    )
+    for shard_file, tensors in tqdm(state_dict_split.filename_to_tensors.items(), desc="Save weights"):
+        shard = {tensor: llama_state_dict[tensor].contiguous() for tensor in tensors}
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if not state_dict_split.is_sharded:
+        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}.")
+    else:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+
+        print(f"Model weights saved in {output_dir}.")
+
+    return str(torch_dtype).replace("torch.", "")
+
+
+def save_config(input_dir: str, output_dir: str, torch_dtype: str):
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
+        qwen_config_dict: dict[str, Any] = json.load(f)
+
+    llama2_config_dict: dict[str, Any] = OrderedDict()
+    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
+    llama2_config_dict["hidden_act"] = "silu"
+    llama2_config_dict["hidden_size"] = qwen_config_dict["hidden_size"]
+    llama2_config_dict["initializer_range"] = qwen_config_dict["initializer_range"]
+    llama2_config_dict["intermediate_size"] = qwen_config_dict["intermediate_size"] // 2
+    llama2_config_dict["max_position_embeddings"] = qwen_config_dict["max_position_embeddings"]
+    llama2_config_dict["model_type"] = "llama"
+    llama2_config_dict["num_attention_heads"] = qwen_config_dict["num_attention_heads"]
+    llama2_config_dict["num_hidden_layers"] = qwen_config_dict["num_hidden_layers"]
+    llama2_config_dict["num_key_value_heads"] = qwen_config_dict["hidden_size"] // qwen_config_dict["kv_channels"]
+    llama2_config_dict["pretraining_tp"] = 1
+    llama2_config_dict["rms_norm_eps"] = qwen_config_dict["layer_norm_epsilon"]
+    llama2_config_dict["rope_scaling"] = None
+    llama2_config_dict["tie_word_embeddings"] = qwen_config_dict["tie_word_embeddings"]
+    llama2_config_dict["torch_dtype"] = torch_dtype
+    llama2_config_dict["transformers_version"] = "4.34.0"
+    llama2_config_dict["use_cache"] = True
+    llama2_config_dict["vocab_size"] = qwen_config_dict["vocab_size"]
+    llama2_config_dict["attention_bias"] = True
+
+    with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
+        json.dump(llama2_config_dict, f, indent=2)
+
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
+
+
+def llamafy_qwen(
+    input_dir: str,
+    output_dir: str,
+    shard_size: str = "2GB",
+    save_safetensors: bool = False,
+):
+    r"""Convert the Qwen models in the same format as LLaMA2.
+
+    Usage: python llamafy_qwen.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
+    """
+    try:
+        os.makedirs(output_dir, exist_ok=False)
+    except Exception as e:
+        raise print("Output dir already exists", e)
+
+    torch_dtype = save_weight(input_dir, output_dir, shard_size, save_safetensors)
+    save_config(input_dir, output_dir, torch_dtype)
+
+
+if __name__ == "__main__":
+    fire.Fire(llamafy_qwen)

scripts/convert_ckpt/tiny_llama4.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from transformers import Llama4Config, Llama4ForConditionalGeneration, Llama4TextConfig, Llama4VisionConfig
+
+
+if __name__ == "__main__":
+    vision_config = Llama4VisionConfig(
+        hidden_size=1408,
+        image_size=336,
+        intermediate_size=5632,
+        num_attention_heads=16,
+        num_hidden_layers=4,
+        vision_output_dim=4096,
+    )
+    text_config = Llama4TextConfig(
+        hidden_size=512,
+        intermediate_size=1024,
+        intermediate_size_mlp=1024,
+        num_hidden_layers=4,
+        num_attention_heads=8,
+        num_key_value_heads=2,
+        head_dim=512 // 8,
+        num_local_experts=2,
+    )
+    config = Llama4Config(vision_config=vision_config, text_config=text_config)
+    model = Llama4ForConditionalGeneration._from_config(config)
+    model.save_pretrained("tiny-llama4")

scripts/eval_bleu_rouge.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import logging
+import time
+
+import fire
+from datasets import load_dataset
+
+
+try:
+    import jieba  # type: ignore
+    from nltk.translate.bleu_score import SmoothingFunction, sentence_bleu  # type: ignore
+    from rouge_chinese import Rouge  # type: ignore
+
+    jieba.setLogLevel(logging.CRITICAL)
+    jieba.initialize()
+except ImportError:
+    print("Please install llamafactory with `pip install -e .[metrics]`.")
+    raise
+
+
+def compute_metrics(sample):
+    hypothesis = list(jieba.cut(sample["predict"]))
+    reference = list(jieba.cut(sample["label"]))
+
+    bleu_score = sentence_bleu(
+        [list(sample["label"])],
+        list(sample["predict"]),
+        smoothing_function=SmoothingFunction().method3,
+    )
+
+    if len(" ".join(hypothesis).split()) == 0 or len(" ".join(reference).split()) == 0:
+        result = {"rouge-1": {"f": 0.0}, "rouge-2": {"f": 0.0}, "rouge-l": {"f": 0.0}}
+    else:
+        rouge = Rouge()
+        scores = rouge.get_scores(" ".join(hypothesis), " ".join(reference))
+        result = scores[0]
+
+    metric_result = {}
+    for k, v in result.items():
+        metric_result[k] = round(v["f"] * 100, 4)
+
+    metric_result["bleu-4"] = round(bleu_score * 100, 4)
+
+    return metric_result
+
+
+def main(filename: str):
+    start_time = time.time()
+    dataset = load_dataset("json", data_files=filename, split="train")
+    dataset = dataset.map(compute_metrics, num_proc=8, remove_columns=dataset.column_names)
+    score_dict = dataset.to_dict()
+
+    average_score = {}
+    for task, scores in sorted(score_dict.items(), key=lambda x: x[0]):
+        print(f"{task}: {sum(scores) / len(scores):.4f}")
+        average_score[task] = sum(scores) / len(scores)
+
+    with open("predictions_score.json", "w", encoding="utf-8") as f:
+        json.dump(average_score, f, indent=4)
+
+    print(f"\nDone in {time.time() - start_time:.3f}s.\nScore file saved to predictions_score.json")
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

scripts/llama_pro.py

+# Copyright 2025 Tencent Inc. and the LlamaFactory team.
+#
+# This code is inspired by the Tencent's LLaMA-Pro library.
+# https://github.com/TencentARC/LLaMA-Pro/blob/main/scripts/block_expansion.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING
+
+import fire
+import torch
+from huggingface_hub import split_torch_state_dict_into_shards
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, PreTrainedModel
+from transformers.modeling_utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, WEIGHTS_NAME
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig
+
+
+def change_name(name: str, old_index: int, new_index: int) -> str:
+    return name.replace(f".{old_index:d}.", f".{new_index:d}.")
+
+
+def block_expansion(
+    model_name_or_path: str,
+    output_dir: str,
+    num_expand: int,
+    shard_size: str = "5GB",
+    save_safetensors: bool = True,
+):
+    r"""Perform block expansion for LLaMA, Mistral, Qwen2 or Yi models.
+
+    Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
+    """
+    config: PretrainedConfig = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
+    num_layers = getattr(config, "num_hidden_layers")
+    if num_layers % num_expand != 0:
+        raise ValueError(f"`num_layers` {num_layers} should be divisible by `num_expand` {num_expand}.")
+
+    setattr(config, "num_hidden_layers", num_layers + num_expand)
+    config.save_pretrained(output_dir)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
+    tokenizer.save_pretrained(output_dir)
+
+    print(f"Expanding model of {num_layers} layers to {num_layers + num_expand} layers.")
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name_or_path, torch_dtype="auto", device_map="cpu", trust_remote_code=True, low_cpu_mem_usage=True
+    )
+    assert isinstance(model, PreTrainedModel)  # type hint
+    if save_safetensors and getattr(model.config, "tie_word_embeddings", False):
+        del model.lm_head  # safetensors does not allow shared weights
+
+    split = num_layers // num_expand
+    layer_cnt = 0
+    state_dict = model.state_dict()
+    output_state_dict: dict[str, torch.Tensor] = OrderedDict()
+    for i in range(num_layers):
+        for key, value in state_dict.items():
+            if f".{i:d}." in key:
+                output_state_dict[change_name(key, i, layer_cnt)] = value
+
+        print(f"Add layer {layer_cnt} copied from layer {i}.")
+        layer_cnt += 1
+        if (i + 1) % split == 0:
+            for key, value in state_dict.items():
+                if f".{i:d}." in key:
+                    if "down_proj" in key or "o_proj" in key:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.zeros_like(value)
+                    else:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.clone(value)
+
+            print(f"Add layer {layer_cnt} expanded from layer {i}.")
+            layer_cnt += 1
+
+    for key, value in state_dict.items():
+        if key not in output_state_dict:
+            output_state_dict[key] = value
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+    state_dict_split = split_torch_state_dict_into_shards(
+        output_state_dict, filename_pattern=filename_pattern, max_shard_size=shard_size
+    )
+    for shard_file, tensors in tqdm(state_dict_split.filename_to_tensors.items(), desc="Save weights"):
+        shard = {tensor: output_state_dict[tensor].contiguous() for tensor in tensors}
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if not state_dict_split.is_sharded:
+        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}.")
+    else:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+
+        print(f"Model weights saved in {output_dir}.")
+
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
+    print("finetuning_type: freeze")
+    print(f"freeze_trainable_layers: {num_expand}")
+    print("use_llama_pro: true")
+
+
+if __name__ == "__main__":
+    fire.Fire(block_expansion)

scripts/loftq_init.py

+# Copyright 2025 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.10.0/examples/loftq_finetuning/quantize_save_load.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import TYPE_CHECKING
+
+import fire
+from peft import LoftQConfig, LoraConfig, TaskType, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+def quantize_loftq(
+    model_name_or_path: str,
+    output_dir: str,
+    loftq_bits: int = 4,
+    loftq_iter: int = 4,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
+):
+    r"""Initialize LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ).
+
+    Usage: python loftq_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
+    loftq_config = LoftQConfig(loftq_bits=loftq_bits, loftq_iter=loftq_iter)
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        inference_mode=True,
+        r=lora_rank,
+        lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
+        init_lora_weights="loftq",
+        loftq_config=loftq_config,
+    )
+
+    # Init LoftQ model
+    print("Initializing LoftQ weights, it may be take several minutes, wait patiently.")
+    peft_model = get_peft_model(model, lora_config)
+    loftq_dir = os.path.join(output_dir, "loftq_init")
+
+    # Save LoftQ model
+    setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply loftq again
+    peft_model.save_pretrained(loftq_dir, safe_serialization=save_safetensors)
+    print(f"Adapter weights saved in {loftq_dir}")
+
+    # Save base model
+    base_model: PreTrainedModel = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print(f"Model weights saved in {output_dir}")
+
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {loftq_dir}")
+    print("finetuning_type: lora")
+    print(f"quantization_bit: {loftq_bits}")
+
+
+if __name__ == "__main__":
+    fire.Fire(quantize_loftq)

scripts/pissa_init.py

+# Copyright 2025 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.11.0/examples/pissa_finetuning/preprocess.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import TYPE_CHECKING
+
+import fire
+from peft import LoraConfig, TaskType, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+def quantize_pissa(
+    model_name_or_path: str,
+    output_dir: str,
+    pissa_iter: int = 16,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
+):
+    r"""Initialize LoRA weights with Principal Singular values and Singular vectors Adaptation (PiSSA).
+
+    Usage: python pissa_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        r=lora_rank,
+        lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
+        init_lora_weights="pissa" if pissa_iter == -1 else f"pissa_niter_{pissa_iter}",
+    )
+
+    # Init PiSSA model
+    peft_model = get_peft_model(model, lora_config)
+    pissa_dir = os.path.join(output_dir, "pissa_init")
+
+    # Save PiSSA model
+    setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply pissa again
+    peft_model.save_pretrained(pissa_dir, safe_serialization=save_safetensors)
+    print(f"Adapter weights saved in {pissa_dir}")
+
+    # Save base model
+    base_model: PreTrainedModel = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print(f"Model weights saved in {output_dir}")
+
+    print("- Fine-tune this model with:")
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {pissa_dir}")
+    print("finetuning_type: lora")
+    print("pissa_init: false")
+    print("pissa_convert: true")
+    print("- and optionally with:")
+    print("quantization_bit: 4")
+
+
+if __name__ == "__main__":
+    fire.Fire(quantize_pissa)