convert_hf_checkpoint.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import sys
sys.path.append("..")
import json
import re
import shutil
import sys
from pathlib import Path
from typing import Optional
from safetensors.torch import load_file as load_safetensors_file
import torch

# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))

from MagicDec.Engine.SnapKV.model import ModelArgs


@torch.inference_mode()
def convert_hf_checkpoint(
    *,
    checkpoint_dir: Path = Path("checkpoints/meta-Transformer/Transformer-2-7b-chat-hf"),
    model_name: Optional[str] = None,
) -> None:
    if model_name is None:
        model_name = checkpoint_dir.name

    config = ModelArgs.from_name(model_name)
    print(f"Model config {config.__dict__}")

    # Load the json file containing weight mapping
    model_map_json_safetensors = checkpoint_dir / 'model.safetensors.index.json'
    model_map_json_pytorch = checkpoint_dir / "pytorch_model.bin.index.json"
    model_safetensors = checkpoint_dir / "model.safetensors"
    model_pytorch = checkpoint_dir / "pytorch_model.bin"
    model_map_json = None
    model_file = None
   
    try:
      assert model_map_json_safetensors.is_file()
      model_map_json = model_map_json_safetensors
      print(f"Found safetensors index at {model_map_json_safetensors}")
    except AssertionError:
      print(f"{model_map_json_safetensors} not found")
    if model_map_json is None:
      try:
        assert model_map_json_pytorch.is_file()
        model_map_json = model_map_json_pytorch
        print(f"Found pytorch index at {model_map_json_pytorch}")
      except AssertionError:
        print(f"{model_map_json_pytorch} not found")
   
    if model_map_json is None:
       try:
           assert model_safetensors.is_file()
           model_file = model_safetensors
           print(f"Found safetensors weights at {model_safetensors}")
       except AssertionError:
           print(f"{model_safetensors} not found")
    
    if model_map_json is None and model_file is None:
        try:
           assert model_safetensors.is_file()
           model_file = model_pytorch
           print(f"Found pytorch weights at {model_pytorch}")
        except AssertionError:
           print(f"{model_pytorch} not found, can't find any weights or index.")
           exit()

    if model_map_json != None:
        with open(model_map_json) as json_map:
            bin_index = json.load(json_map)

    weight_map = {
        "model.embed_tokens.weight": "tok_embeddings.weight",
        "model.layers.{}.self_attn.q_proj.weight": "layers.{}.attention.wq.weight",
        "model.layers.{}.self_attn.k_proj.weight": "layers.{}.attention.wk.weight",
        "model.layers.{}.self_attn.v_proj.weight": "layers.{}.attention.wv.weight",
        "model.layers.{}.self_attn.o_proj.weight": "layers.{}.attention.wo.weight",
        'model.layers.{}.self_attn.rotary_emb.inv_freq': None,
        'model.layers.{}.mlp.gate_proj.weight': 'layers.{}.feed_forward.w1.weight',
        "model.layers.{}.mlp.up_proj.weight": "layers.{}.feed_forward.w3.weight",
        "model.layers.{}.mlp.down_proj.weight": "layers.{}.feed_forward.w2.weight",
        "model.layers.{}.input_layernorm.weight": "layers.{}.attention_norm.weight",
        "model.layers.{}.post_attention_layernorm.weight": "layers.{}.ffn_norm.weight",
        "model.norm.weight": "norm.weight",
        "lm_head.weight": "output.weight",
    }
    if "qwen" in model_name.lower():
        weight_map.update({
            "model.layers.{}.self_attn.q_proj.bias": "layers.{}.attention.wq.bias",
            "model.layers.{}.self_attn.k_proj.bias": "layers.{}.attention.wk.bias",
            "model.layers.{}.self_attn.v_proj.bias": "layers.{}.attention.wv.bias",
        })
    if model_map_json != None:
        bin_files = {checkpoint_dir / bin for bin in bin_index["weight_map"].values()}

    def permute(w, n_head):
        if len(w.shape) == 2:
            # weight term
            dim = config.dim
            return (
                w.view(n_head, 2, config.head_dim // 2, dim)
                .transpose(1, 2)
                .reshape(config.head_dim * n_head, dim)
            )
        else:
            # bias term
            return w.view(n_head, 2, config.head_dim // 2).transpose(1, 2).reshape(config.head_dim * n_head)

    merged_result = {}
    if model_map_json != None:
        for file in sorted(bin_files):
            if "safetensors" in str(file):
                state_dict = load_safetensors_file(str(file), device="cpu")
                merged_result.update(state_dict)
            else:
                state_dict = torch.load(str(file), map_location="cpu", mmap=True, weights_only=True)
                merged_result.update(state_dict)
    else:
        # state_dict = state_dict = torch.load(str(bin_file), map_location="cpu", mmap=True, weights_only=True)
        if "safetensors" in str(model_file):
            state_dict = load_safetensors_file(str(model_file), device="cpu")
            merged_result.update(state_dict)
        else:
            state_dict = state_dict = torch.load(str(model_file), map_location="cpu", mmap=True, weights_only=True)
            merged_result.update(state_dict)
 
    final_result = {}
    for key, value in merged_result.items():
        if "layers" in key:
            abstract_key = re.sub(r'(\d+)', '{}', key)
            layer_num = re.search(r'\d+', key).group(0)
            new_key = weight_map[abstract_key]
            if new_key is None:
                continue
            new_key = new_key.format(layer_num)
        else:
            new_key = weight_map[key]

        final_result[new_key] = value
    if 'output.weight' not in final_result.keys():
        final_result['output.weight'] = merged_result["model.embed_tokens.weight"]
        print("Use input embedding as the output lm_head")

    for key in tuple(final_result.keys()):
        if "wq" in key:
            q = final_result[key]
            k = final_result[key.replace("wq", "wk")]
            v = final_result[key.replace("wq", "wv")]
            q = permute(q, config.n_head)
            k = permute(k, config.n_local_heads)
            final_result[key.replace("wq", "wqkv")] = torch.cat([q, k, v])
            del final_result[key]
            del final_result[key.replace("wq", "wk")]
            del final_result[key.replace("wq", "wv")]
    print(f"Saving checkpoint to {checkpoint_dir / 'model.pth'}")
    torch.save(final_result, checkpoint_dir / "model.pth")
    if 'llama-3' in model_name.lower():
        original_dir = checkpoint_dir / "original"
        tokenizer_model = original_dir / "tokenizer.model"
        tokenizer_model_tiktoken = checkpoint_dir / "tokenizer.model"
        print(f"Copying {tokenizer_model} to {tokenizer_model_tiktoken}")
        shutil.copy(tokenizer_model, tokenizer_model_tiktoken)

if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser(description='Convert HuggingFace checkpoint.')
    parser.add_argument('--checkpoint_dir', type=Path, default=Path("/scratch/models/meta-llama/Llama-3.2-1B"))
    parser.add_argument('--model_name', type=str, default=None)

    args = parser.parse_args()
    convert_hf_checkpoint(
        checkpoint_dir=args.checkpoint_dir,
        model_name=args.model_name,
    )


# # Copyright (c) Meta Platforms, Inc. and affiliates.
# # All rights reserved.

# # This source code is licensed under the license found in the
# # LICENSE file in the root directory of this source tree.
# import sys
# sys.path.append("..")
# import json
# import re
# import shutil
# import sys
# import copy
# from pathlib import Path
# from typing import Optional
# from safetensors.torch import load_file as load_safetensors_file
# import torch

# # support running without installing as a package
# wd = Path(__file__).parent.parent.resolve()
# sys.path.append(str(wd))

# from MagicDec.Engine.model import ModelArgs


# @torch.inference_mode()
# def convert_hf_checkpoint(
#     *,
#     checkpoint_dir: Path = Path("checkpoints/meta-Transformer/Transformer-2-7b-chat-hf"),
#     model_name: Optional[str] = None,
# ) -> None:
#     if model_name is None:
#         model_name = checkpoint_dir.name

#     config = ModelArgs.from_name(model_name)
#     print(f"Model config {config.__dict__}")

#     # Load the json file containing weight mapping
#     model_map_json_safetensors = checkpoint_dir / 'model.safetensors.index.json'
#     model_map_json_pytorch = checkpoint_dir / "pytorch_model.bin.index.json"
#     model_map_json = None
   
#     try:
#       assert model_map_json_safetensors.is_file()
#       model_map_json = model_map_json_safetensors
#       print(f"Found safetensors index at {model_map_json_safetensors}")
#     except AssertionError:
#       print(f"{model_map_json_safetensors} not found")
#     if model_map_json is None:
#       try:
#         assert model_map_json_pytorch.is_file()
#         model_map_json = model_map_json_pytorch
#         print(f"Found pytorch index at {model_map_json_pytorch}")
#       except AssertionError:
#         print(f"{model_map_json_pytorch} not found")
   
#     if model_map_json is None: raise Exception("No model map found!")

#     with open(model_map_json) as json_map:
#         bin_index = json.load(json_map)

#     weight_map = {
#         "model.embed_tokens.weight": "tok_embeddings.weight",
#         "model.layers.{}.self_attn.q_proj.weight": "layers.{}.attention.wq.weight",
#         "model.layers.{}.self_attn.k_proj.weight": "layers.{}.attention.wk.weight",
#         "model.layers.{}.self_attn.v_proj.weight": "layers.{}.attention.wv.weight",
#         "model.layers.{}.self_attn.o_proj.weight": "layers.{}.attention.wo.weight",
#         'model.layers.{}.self_attn.rotary_emb.inv_freq': None,
#         'model.layers.{}.mlp.gate_proj.weight': 'layers.{}.feed_forward.w1.weight',
#         "model.layers.{}.mlp.up_proj.weight": "layers.{}.feed_forward.w3.weight",
#         "model.layers.{}.mlp.down_proj.weight": "layers.{}.feed_forward.w2.weight",
#         "model.layers.{}.input_layernorm.weight": "layers.{}.attention_norm.weight",
#         "model.layers.{}.post_attention_layernorm.weight": "layers.{}.ffn_norm.weight",
#         "model.norm.weight": "norm.weight",
#         "lm_head.weight": "output.weight"
#     }
#     bin_files = {checkpoint_dir / bin for bin in bin_index["weight_map"].values()}

#     def permute(w, n_head):
#         dim = config.dim
#         return (
#             w.view(n_head, 2, config.head_dim // 2, dim)
#             .transpose(1, 2)
#             .reshape(config.head_dim * n_head, dim)
#         )

#     merged_result = {}
#     for file in sorted(bin_files):
#        if "safetensors" in str(file):
#            state_dict = load_safetensors_file(str(file), device="cpu")
#            merged_result.update(state_dict)
#        else:
#            state_dict = torch.load(str(file), map_location="cpu", mmap=True, weights_only=True)
#            merged_result.update(state_dict)
#     if "lm_head.weight" not in merged_result.keys():
#         merged_result["lm_head.weight"] = copy.deepcopy(merged_result["model.embed_tokens.weight"])
#     final_result = {}
#     for key, value in merged_result.items():
#         if "layers" in key:
#             abstract_key = re.sub(r'(\d+)', '{}', key)
#             layer_num = re.search(r'\d+', key).group(0)
#             new_key = weight_map[abstract_key]
#             if new_key is None:
#                 continue
#             new_key = new_key.format(layer_num)
#         else:
#             new_key = weight_map[key]

#         final_result[new_key] = value

#     for key in tuple(final_result.keys()):
#         if "wq" in key:
#             q = final_result[key]
#             k = final_result[key.replace("wq", "wk")]
#             v = final_result[key.replace("wq", "wv")]
#             q = permute(q, config.n_head)
#             k = permute(k, config.n_local_heads)
#             final_result[key.replace("wq", "wqkv")] = torch.cat([q, k, v])
#             del final_result[key]
#             del final_result[key.replace("wq", "wk")]
#             del final_result[key.replace("wq", "wv")]
   
#     print(f"Saving checkpoint to {checkpoint_dir / 'model.pth'}")
#     torch.save(final_result, checkpoint_dir / "model.pth")
#     if 'llama-3-' in model_name.lower() or 'llama-3.' in model_name.lower():
#         if 'llama-3.1-405b' in model_name.lower():
#             original_dir = checkpoint_dir / "original" / "mp16"
#         else:
#             original_dir = checkpoint_dir / "original"
#         tokenizer_model = original_dir / "tokenizer.model"
#         tokenizer_model_tiktoken = checkpoint_dir / "tokenizer.model"
#         print(f"Copying {tokenizer_model} to {tokenizer_model_tiktoken}")
#         shutil.copy(tokenizer_model, tokenizer_model_tiktoken)

# if __name__ == '__main__':
#     import argparse
#     parser = argparse.ArgumentParser(description='Convert HuggingFace checkpoint.')
#     parser.add_argument('--checkpoint_dir', type=Path, default=Path("checkpoints/meta-llama/llama-2-7b-chat-hf"))
#     parser.add_argument('--model_name', type=str, default=None)

#     args = parser.parse_args()
#     convert_hf_checkpoint(
#         checkpoint_dir=args.checkpoint_dir,
#         model_name=args.model_name,
#     )