File size: 8,455 Bytes
48ecd01 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 | #!/usr/bin/env python3
"""Fix GGUF newline crash by adding byte-fallback tokens to the tokenizer.
Problem: The SentencePiece Unigram tokenizer was trained without byte_fallback=True,
so characters like \n have no token representation. llama.cpp crashes when it
encounters these characters because there's no byte-fallback.
Fix:
1. Add 256 byte-fallback tokens (<0x00> .. <0xFF>) to tokenizer.json
2. Resize model embeddings from 64000 -> 64256
3. Update config.json vocab_size
4. Copy tokenizer.model for proper GGUF conversion
Usage:
python scripts/fix_tokenizer_byte_fallback.py \
--input outputs/hf_checkpoint-best \
--output outputs/hf_checkpoint-best-fixed \
--sp_model tokenizer/korean_sp/tokenizer.model
"""
import argparse
import json
import shutil
from pathlib import Path
import torch
from safetensors.torch import load_file, save_file
BYTE_FALLBACK_COUNT = 256
BYTE_TOKEN_TEMPLATE = "<0x{:02X}>"
def fix_tokenizer_json(input_path: Path, output_path: Path):
"""Add byte_fallback=True and 256 byte tokens to tokenizer.json."""
with open(input_path) as f:
tok = json.load(f)
model = tok["model"]
vocab = model["vocab"] # list of [piece, score]
original_size = len(vocab)
# Enable byte_fallback
model["byte_fallback"] = True
# Add 256 byte tokens with very low score (they're fallback only)
for i in range(BYTE_FALLBACK_COUNT):
byte_token = BYTE_TOKEN_TEMPLATE.format(i)
vocab.append([byte_token, 0.0])
new_size = len(vocab)
print(f" Vocab: {original_size} -> {new_size} (+{BYTE_FALLBACK_COUNT} byte tokens)")
print(f" byte_fallback: False -> True")
# Also add byte tokens to added_tokens list
added = tok.get("added_tokens", [])
for i in range(BYTE_FALLBACK_COUNT):
byte_token = BYTE_TOKEN_TEMPLATE.format(i)
added.append({
"id": original_size + i,
"content": byte_token,
"single_word": False,
"lstrip": False,
"rstrip": False,
"normalized": False,
"special": True,
})
tok["added_tokens"] = added
with open(output_path, "w") as f:
json.dump(tok, f, ensure_ascii=False, indent=2)
return original_size, new_size
def fix_config_json(input_path: Path, output_path: Path, new_vocab_size: int):
"""Update vocab_size in config.json."""
with open(input_path) as f:
config = json.load(f)
old_size = config["vocab_size"]
config["vocab_size"] = new_vocab_size
print(f" config.json vocab_size: {old_size} -> {new_vocab_size}")
with open(output_path, "w") as f:
json.dump(config, f, indent=2)
def resize_embeddings(input_path: Path, output_path: Path,
old_vocab: int, new_vocab: int, tie_embeddings: bool):
"""Resize embedding and lm_head weights to accommodate new tokens."""
print(f" Loading model weights from {input_path} ...")
state_dict = load_file(str(input_path))
embed_key = "model.embed_tokens.weight"
lm_head_key = "lm_head.weight"
if embed_key not in state_dict:
raise KeyError(f"{embed_key} not found in state_dict. Keys: {list(state_dict.keys())[:10]}")
embed = state_dict[embed_key]
print(f" embed_tokens shape: {embed.shape}")
hidden_size = embed.shape[1]
extra = new_vocab - old_vocab
# Initialize new embeddings as mean of existing (better than random for byte tokens)
mean_embed = embed.mean(dim=0, keepdim=True)
# Add small noise to avoid identical embeddings
noise = torch.randn(extra, hidden_size, dtype=embed.dtype) * 0.01
new_rows = mean_embed.expand(extra, -1) + noise
new_embed = torch.cat([embed, new_rows], dim=0)
state_dict[embed_key] = new_embed
print(f" embed_tokens resized: {embed.shape} -> {new_embed.shape}")
if tie_embeddings:
# When tie_word_embeddings=True, lm_head shares embed_tokens
# Remove lm_head if present (it will be tied automatically)
if lm_head_key in state_dict:
del state_dict[lm_head_key]
print(f" lm_head removed (tie_word_embeddings=True)")
else:
if lm_head_key in state_dict:
lm_head = state_dict[lm_head_key]
mean_lm = lm_head.mean(dim=0, keepdim=True)
noise_lm = torch.randn(extra, hidden_size, dtype=lm_head.dtype) * 0.01
new_lm = torch.cat([lm_head, mean_lm.expand(extra, -1) + noise_lm], dim=0)
state_dict[lm_head_key] = new_lm
print(f" lm_head resized: {lm_head.shape} -> {new_lm.shape}")
print(f" Saving to {output_path} ...")
save_file(state_dict, str(output_path))
def main():
parser = argparse.ArgumentParser(description="Fix tokenizer byte-fallback for GGUF")
parser.add_argument("--input", type=Path, required=True, help="Input HF checkpoint dir")
parser.add_argument("--output", type=Path, required=True, help="Output fixed HF checkpoint dir")
parser.add_argument("--sp_model", type=Path, default=None,
help="SentencePiece .model file to copy (for GGUF conversion)")
args = parser.parse_args()
input_dir = args.input
output_dir = args.output
if not input_dir.exists():
print(f"ERROR: Input directory not found: {input_dir}")
return 1
output_dir.mkdir(parents=True, exist_ok=True)
# Load config to check tie_word_embeddings
with open(input_dir / "config.json") as f:
config = json.load(f)
old_vocab = config["vocab_size"]
new_vocab = old_vocab + BYTE_FALLBACK_COUNT
tie_embeddings = config.get("tie_word_embeddings", False)
print(f"=== Byte-Fallback Fix ===")
print(f"Input: {input_dir}")
print(f"Output: {output_dir}")
print(f"Old vocab: {old_vocab}, New vocab: {new_vocab}")
print(f"tie_word_embeddings: {tie_embeddings}")
print()
# 1. Fix tokenizer.json
print("[1/4] Fixing tokenizer.json ...")
fix_tokenizer_json(
input_dir / "tokenizer.json",
output_dir / "tokenizer.json",
)
# 2. Fix config.json
print("[2/4] Fixing config.json ...")
fix_config_json(
input_dir / "config.json",
output_dir / "config.json",
new_vocab,
)
# 3. Resize model weights
print("[3/4] Resizing embeddings ...")
resize_embeddings(
input_dir / "model.safetensors",
output_dir / "model.safetensors",
old_vocab, new_vocab, tie_embeddings,
)
# 4. Copy other files
print("[4/4] Copying remaining files ...")
for fname in ["tokenizer_config.json", "generation_config.json"]:
src = input_dir / fname
if src.exists():
shutil.copy2(src, output_dir / fname)
print(f" Copied {fname}")
# Copy SentencePiece model if provided (needed for GGUF conversion)
if args.sp_model and args.sp_model.exists():
shutil.copy2(args.sp_model, output_dir / "tokenizer.model")
print(f" Copied tokenizer.model from {args.sp_model}")
elif (input_dir / "tokenizer.model").exists():
shutil.copy2(input_dir / "tokenizer.model", output_dir / "tokenizer.model")
print(f" Copied tokenizer.model from input dir")
# Update tokenizer_config.json to add added_tokens_decoder for byte tokens
tc_path = output_dir / "tokenizer_config.json"
if tc_path.exists():
with open(tc_path) as f:
tc = json.load(f)
added_tokens_decoder = tc.get("added_tokens_decoder", {})
for i in range(BYTE_FALLBACK_COUNT):
token_id = old_vocab + i
byte_token = BYTE_TOKEN_TEMPLATE.format(i)
added_tokens_decoder[str(token_id)] = {
"content": byte_token,
"lstrip": False,
"normalized": False,
"rstrip": False,
"single_word": False,
"special": True,
}
tc["added_tokens_decoder"] = added_tokens_decoder
with open(tc_path, "w") as f:
json.dump(tc, f, indent=2)
print(f" Updated tokenizer_config.json with {BYTE_FALLBACK_COUNT} byte tokens")
print()
print(f"=== Done! Fixed checkpoint at: {output_dir} ===")
print(f"Next: python outputs/llama.cpp/convert_hf_to_gguf.py {output_dir} --outfile outputs/gguf/frankenstallm-3b-f16.gguf --outtype f16")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|