|
| 1 | +#!/usr/bin/env python3 |
| 2 | +# HF CodeShell --> gguf conversion |
| 3 | + |
| 4 | +from __future__ import annotations |
| 5 | + |
| 6 | +import argparse |
| 7 | +import json |
| 8 | +import os |
| 9 | +import struct |
| 10 | +import sys |
| 11 | +from pathlib import Path |
| 12 | +from typing import Any |
| 13 | + |
| 14 | +import numpy as np |
| 15 | +import torch |
| 16 | +from transformers import AutoTokenizer # type: ignore[import] |
| 17 | + |
| 18 | +if 'NO_LOCAL_GGUF' not in os.environ: |
| 19 | + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) |
| 20 | +import gguf |
| 21 | + |
| 22 | +def count_model_parts(dir_model: Path) -> int: |
| 23 | + num_parts = 0 |
| 24 | + for filename in os.listdir(dir_model): |
| 25 | + if filename.startswith("pytorch_model-"): |
| 26 | + num_parts += 1 |
| 27 | + |
| 28 | + if num_parts > 0: |
| 29 | + print("gguf: found " + str(num_parts) + " model parts") |
| 30 | + return num_parts |
| 31 | + |
| 32 | + |
| 33 | +def parse_args() -> argparse.Namespace: |
| 34 | + parser = argparse.ArgumentParser(description="Convert a CodeShell model to a GGML compatible file") |
| 35 | + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") |
| 36 | + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") |
| 37 | + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") |
| 38 | + parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) |
| 39 | + return parser.parse_args() |
| 40 | + |
| 41 | +args = parse_args() |
| 42 | + |
| 43 | +dir_model = args.model |
| 44 | +ftype = args.ftype |
| 45 | +if not dir_model.is_dir(): |
| 46 | + print(f'Error: {args.model} is not a directory', file = sys.stderr) |
| 47 | + sys.exit(1) |
| 48 | + |
| 49 | +# possible tensor data types |
| 50 | +# ftype == 0 -> float32 |
| 51 | +# ftype == 1 -> float16 |
| 52 | + |
| 53 | +# map from ftype to string |
| 54 | +ftype_str = ["f32", "f16"] |
| 55 | + |
| 56 | +if args.outfile is not None: |
| 57 | + fname_out = args.outfile |
| 58 | +else: |
| 59 | + # output in the same directory as the model by default |
| 60 | + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' |
| 61 | + |
| 62 | +print("gguf: loading model "+dir_model.name) |
| 63 | + |
| 64 | +with open(dir_model / "config.json", "r", encoding="utf-8") as f: |
| 65 | + hparams = json.load(f) |
| 66 | + |
| 67 | +if hparams["architectures"][0] != "CodeShellForCausalLM": |
| 68 | + print("Model architecture not supported: " + hparams["architectures"][0]) |
| 69 | + sys.exit(1) |
| 70 | + |
| 71 | +# get number of model parts |
| 72 | +num_parts = count_model_parts(dir_model) |
| 73 | + |
| 74 | +ARCH = gguf.MODEL_ARCH.CODESHELL |
| 75 | +gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) |
| 76 | + |
| 77 | +print("gguf: get model metadata") |
| 78 | + |
| 79 | +block_count = hparams["n_layer"] |
| 80 | + |
| 81 | +gguf_writer.add_name("CodeShell") |
| 82 | +gguf_writer.add_context_length(hparams["n_positions"]) |
| 83 | +gguf_writer.add_embedding_length(hparams["n_embd"]) |
| 84 | +gguf_writer.add_feed_forward_length(4 * hparams["n_embd"]) |
| 85 | +gguf_writer.add_block_count(block_count) |
| 86 | +gguf_writer.add_head_count(hparams["n_head"]) |
| 87 | +gguf_writer.add_head_count_kv(hparams["num_query_groups"]) |
| 88 | +gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) |
| 89 | +gguf_writer.add_file_type(ftype) |
| 90 | +gguf_writer.add_rope_freq_base(10000.0) |
| 91 | +gguf_writer.add_rope_scale_linear(1.0) |
| 92 | + |
| 93 | +# TOKENIZATION |
| 94 | + |
| 95 | +print("gguf: get tokenizer metadata") |
| 96 | + |
| 97 | +tokens: list[bytearray] = [] |
| 98 | +scores: list[float] = [] |
| 99 | +toktypes: list[int] = [] |
| 100 | + |
| 101 | +# gpt2 tokenizer |
| 102 | +gguf_writer.add_tokenizer_model("gpt2") |
| 103 | + |
| 104 | +print("gguf: get gpt2 tokenizer vocab") |
| 105 | + |
| 106 | +# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py |
| 107 | +tokenizer = AutoTokenizer.from_pretrained(dir_model) |
| 108 | + |
| 109 | +# The number of tokens in tokenizer.json can differ from the expected vocab size. |
| 110 | +# This causes downstream issues with mismatched tensor sizes when running the inference |
| 111 | +vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) |
| 112 | +assert max(tokenizer.vocab.values()) < vocab_size |
| 113 | + |
| 114 | +reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} |
| 115 | + |
| 116 | +for i in range(vocab_size): |
| 117 | + tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]") |
| 118 | + scores.append(0.0) # dummy |
| 119 | + toktypes.append(gguf.TokenType.NORMAL) |
| 120 | + |
| 121 | +gguf_writer.add_token_list(tokens) |
| 122 | +gguf_writer.add_token_scores(scores) |
| 123 | +gguf_writer.add_token_types(toktypes) |
| 124 | + |
| 125 | +special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) |
| 126 | +special_vocab.add_to_gguf(gguf_writer) |
| 127 | + |
| 128 | +# TENSORS |
| 129 | + |
| 130 | +tensor_map = gguf.get_tensor_name_map(ARCH,block_count) |
| 131 | + |
| 132 | +# params for qkv transform |
| 133 | +n_head = hparams["n_head"] |
| 134 | +n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1 |
| 135 | + |
| 136 | +head_dim = hparams["n_embd"] // n_head |
| 137 | + |
| 138 | +# tensor info |
| 139 | +print("gguf: get tensor metadata") |
| 140 | + |
| 141 | +if num_parts == 0: |
| 142 | + part_names = iter(("pytorch_model.bin",)) |
| 143 | +else: |
| 144 | + part_names = ( |
| 145 | + f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) |
| 146 | + ) |
| 147 | + |
| 148 | +for part_name in part_names: |
| 149 | + if args.vocab_only: |
| 150 | + break |
| 151 | + print("gguf: loading model part '" + part_name + "'") |
| 152 | + model_part = torch.load(dir_model / part_name, map_location="cpu") |
| 153 | + |
| 154 | + for name in model_part.keys(): |
| 155 | + data = model_part[name] |
| 156 | + |
| 157 | + old_dtype = data.dtype |
| 158 | + |
| 159 | + # convert any unsupported data types to float32 |
| 160 | + if data.dtype != torch.float16 and data.dtype != torch.float32: |
| 161 | + data = data.to(torch.float32) |
| 162 | + |
| 163 | + data = data.squeeze().numpy() |
| 164 | + |
| 165 | + # map tensor names |
| 166 | + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) |
| 167 | + if new_name is None: |
| 168 | + print("Can not map tensor '" + name + "'") |
| 169 | + sys.exit() |
| 170 | + |
| 171 | + if "rotary_emb.inv_freq" in name: |
| 172 | + print(f"skipping tensor {new_name}") |
| 173 | + continue |
| 174 | + |
| 175 | + n_dims = len(data.shape) |
| 176 | + data_dtype = data.dtype |
| 177 | + |
| 178 | + # if f32 desired, convert any float16 to float32 |
| 179 | + if ftype == 0 and data_dtype == np.float16: |
| 180 | + data = data.astype(np.float32) |
| 181 | + |
| 182 | + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 |
| 183 | + if ftype == 1 and data_dtype == np.float16 and n_dims == 1: |
| 184 | + data = data.astype(np.float32) |
| 185 | + |
| 186 | + # if f16 desired, convert any float32 2-dim weight tensors to float16 |
| 187 | + if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: |
| 188 | + data = data.astype(np.float16) |
| 189 | + |
| 190 | + print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype)) |
| 191 | + |
| 192 | + gguf_writer.add_tensor(new_name, data) |
| 193 | + |
| 194 | + |
| 195 | +print("gguf: write header") |
| 196 | +gguf_writer.write_header_to_file() |
| 197 | +print("gguf: write metadata") |
| 198 | +gguf_writer.write_kv_data_to_file() |
| 199 | +if not args.vocab_only: |
| 200 | + print("gguf: write tensors") |
| 201 | + gguf_writer.write_tensors_to_file() |
| 202 | + |
| 203 | +gguf_writer.close() |
| 204 | + |
| 205 | +print(f"gguf: model successfully exported to '{fname_out}'") |
| 206 | +print("") |
0 commit comments