|
| 1 | +import os |
| 2 | +import re |
| 3 | +from pathlib import Path |
| 4 | +from typing import Optional |
| 5 | +from collections import OrderedDict |
| 6 | + |
| 7 | +import torch |
| 8 | +from tqdm.auto import tqdm |
| 9 | +from transformers import LlamaForCausalLM, AutoTokenizer |
| 10 | + |
| 11 | + |
| 12 | +scale2emb = { |
| 13 | + '7B': 4096, |
| 14 | + '13B': 5120, |
| 15 | + '30B': 6656, |
| 16 | + '65B': 8192, |
| 17 | + '70B': 8192, |
| 18 | +} |
| 19 | + |
| 20 | + |
| 21 | +key_to_dim = { |
| 22 | + "w1": 0, |
| 23 | + "w2": -1, |
| 24 | + "w3": 0, |
| 25 | + "wo": -1, |
| 26 | + "wq": 0, |
| 27 | + "wk": 0, |
| 28 | + "wv": 0, |
| 29 | + "output": 0, |
| 30 | + "tok_embeddings": -1, |
| 31 | + "ffn_norm": None, |
| 32 | + "attention_norm": None, |
| 33 | + "norm": None, |
| 34 | + "rope": None, |
| 35 | +} |
| 36 | + |
| 37 | + |
| 38 | +def init_merged_ckpt(pth_00, num_pth=8, emb_dim=8192): |
| 39 | + merged_ckpt = OrderedDict() |
| 40 | + for parameter_name, parameter in pth_00.items(): |
| 41 | + short_name = parameter_name.split(".")[-2] |
| 42 | + if key_to_dim[short_name] is None: |
| 43 | + merged_ckpt[parameter_name] = parameter |
| 44 | + del parameter |
| 45 | + elif key_to_dim[short_name] == 0: |
| 46 | + size = parameter.shape[0] |
| 47 | + merged_param_shape = [ parameter.shape[0] * num_pth, parameter.shape[1] ] |
| 48 | + merged_ckpt[parameter_name] = torch.zeros(merged_param_shape) |
| 49 | + merged_ckpt[parameter_name][0 : size, :] = parameter |
| 50 | + del parameter |
| 51 | + elif key_to_dim[short_name] == -1: |
| 52 | + size = parameter.shape[-1] |
| 53 | + merged_param_shape = [ parameter.shape[0], parameter.shape[1] * num_pth] |
| 54 | + merged_ckpt[parameter_name] = torch.zeros(merged_param_shape) |
| 55 | + merged_ckpt[parameter_name][:, 0 : size] = parameter |
| 56 | + del parameter |
| 57 | + return merged_ckpt |
| 58 | + |
| 59 | + |
| 60 | +def merge_meta_llama(size: int, root_dir: Path): |
| 61 | + paths = sorted(path for path in root_dir.iterdir() |
| 62 | + if re.match(r"^consolidated\.[0-9]+\.pth$", path.name)) |
| 63 | + if len(paths) == 1: # no sharded checkpoints, return everything |
| 64 | + return torch.load(paths[0], map_location=torch.device("cpu")) |
| 65 | + |
| 66 | + num_pth = len(paths) |
| 67 | + for i, ckpt_path in enumerate(tqdm(paths, desc="Merging llama")): |
| 68 | + llama_config = torch.load(ckpt_path, map_location=torch.device('cpu')) |
| 69 | + if i == 0: |
| 70 | + merged_ckpt = init_merged_ckpt(llama_config, num_pth=num_pth, |
| 71 | + emb_dim=scale2emb[f"{size}B"]) |
| 72 | + else: |
| 73 | + for parameter_name, parameter in llama_config.items(): |
| 74 | + short_name = parameter_name.split(".")[-2] |
| 75 | + if key_to_dim[short_name] == 0: |
| 76 | + size = parameter.shape[0] |
| 77 | + merged_param_shape = [ parameter.shape[0] * num_pth, parameter.shape[1] ] |
| 78 | + merged_ckpt[parameter_name][size * i : size * (i + 1), :] = parameter |
| 79 | + del parameter |
| 80 | + if key_to_dim[short_name] == -1: |
| 81 | + size = parameter.shape[-1] |
| 82 | + merged_param_shape = [ parameter.shape[0], parameter.shape[1] * num_pth] |
| 83 | + merged_ckpt[parameter_name][:, size * i : size * (i + 1)] = parameter |
| 84 | + del parameter |
| 85 | + del llama_config |
| 86 | + return merged_ckpt |
| 87 | + |
| 88 | + |
| 89 | +def merge_hf_llama(size: int, version: int, cache_dir: Optional[Path] = None, model_path=None, tokenizer_len=32000): |
| 90 | + assert version == 2, "Only llama v2 available using huggingface" |
| 91 | + print(cache_dir) |
| 92 | + model = LlamaForCausalLM.from_pretrained(cache_dir, cache_dir=cache_dir, local_files_only=True, use_safetensors=False) |
| 93 | + # resize token embeddings size according saved tokenizer for model extend token size. |
| 94 | + # model.resize_token_embeddings(tokenizer_len) |
| 95 | + weights = model.state_dict() |
| 96 | + weights["tok_embeddings.weight"] = weights.pop("model.embed_tokens.weight") |
| 97 | + weights["norm.weight"] = weights.pop("model.norm.weight") |
| 98 | + weights["output.weight"] = weights.pop("lm_head.weight") |
| 99 | + for key in list(weights.keys()): |
| 100 | + if rmatch := re.match(r"^model\.(layers\.[0-9]+\.)(.+)(\.weight)$", key): |
| 101 | + new_key = { |
| 102 | + "self_attn.q_proj": "attention.wq", |
| 103 | + "self_attn.k_proj": "attention.wk", |
| 104 | + "self_attn.v_proj": "attention.wv", |
| 105 | + "self_attn.o_proj": "attention.wo", |
| 106 | + "mlp.gate_proj": "feed_forward.w1", |
| 107 | + "mlp.down_proj": "feed_forward.w2", |
| 108 | + "mlp.up_proj": "feed_forward.w3", |
| 109 | + "input_layernorm": "attention_norm", |
| 110 | + "post_attention_layernorm": "ffn_norm" |
| 111 | + }[rmatch.group(2)] |
| 112 | + weights[rmatch.group(1) + new_key + rmatch.group(3)] = weights.pop(key) |
| 113 | + return weights |
| 114 | + |
| 115 | + |
| 116 | +def merge_llama(size: int, version: int, root_dir: Optional[Path] = None, tokenizer_len: Optional[int] = 32000): |
| 117 | + if root_dir is not None and (root_dir/"consolidated.00.pth").exists(): |
| 118 | + return merge_meta_llama(size, root_dir), "meta" |
| 119 | + print(f"Weights at {root_dir} do not look like a meta checkpoint, assuming " |
| 120 | + "huggingface cache_dir instead") |
| 121 | + return merge_hf_llama(size, version, root_dir, tokenizer_len), "hf" |
0 commit comments