llama.cpp/gguf-py/gguf/gguf_writer_split.py

260 lines
10 KiB
Python

from __future__ import annotations
import os
import logging
from enum import IntEnum
from typing import TYPE_CHECKING, Any, Sequence
from argparse import Namespace
from collections import deque
from dataclasses import dataclass
from pathlib import Path
import numpy as np
if TYPE_CHECKING:
from typing_extensions import TypeAlias
from .constants import (
GGMLQuantizationType,
GGUFEndian,
GGUFValueType
)
from .gguf_writer import GGUFWriter, WriterState
from .constants import Keys
logger = logging.getLogger(__name__)
SHARD_NAME_FORMAT = "{:s}-{:05d}-of-{:05d}.gguf"
METADATA_ONLY_INDICATOR = -1
KVTempData: TypeAlias = dict[str, tuple[Any, GGUFValueType | None]] # {key: (value, type)}
TensorTempData: TypeAlias = tuple[str, np.ndarray[Any, Any], GGMLQuantizationType | None] # (tensor name, tensor data, tensor dtype)
@dataclass
class Shard:
path: Path
tensor_count: int
size: int
tensors: deque[TensorTempData]
class SplitStyle(IntEnum):
NONE = 0
TENSORS = 1
SIZE = 2
class SplitArguments:
def __init__(self, args: Namespace) -> None:
self.split = args.split
self.split_max_tensors = args.split_max_tensors if args.split else 0
self.split_max_size = GGUFWriterSplit.split_str_to_n_bytes(args.split_max_size) if args.split and args.split_max_size else 0
self.split_style = SplitStyle.NONE if not self.split \
else SplitStyle.TENSORS if self.split_max_tensors \
else SplitStyle.SIZE
self.dry_run = args.dry_run
self.small_first_shard = args.small_first_shard
class GGUFWriterSplit(GGUFWriter):
kv_data: KVTempData
split_arguments: SplitArguments
shards: list[Shard]
shard_writers: list[GGUFWriter]
def __init__(self, path: os.PathLike[str] | str, arch: str, split_arguments: SplitArguments,
use_temp_file: bool = True, endianess: GGUFEndian = GGUFEndian.LITTLE
) -> None:
# we intentionally don't call superclass constructor
self.arch = arch
self.path = Path(path)
self.endianess = endianess
self.kv_data = {}
self.shards = []
self.shard_writers = []
self.total_tensors = 0
self.use_temp_file = use_temp_file
self.split_arguments = split_arguments
self.recent_key = None
self.state = WriterState.EMPTY
if self.split_arguments.small_first_shard:
self.shards.append(Shard(Path(), 0, METADATA_ONLY_INDICATOR, deque()))
def init_shards(self) -> None:
self.total_tensors = sum(shard.tensor_count for shard in self.shards)
total_size = sum(shard.size for shard in self.shards)
# check if we need to split
if self.split_arguments.split_max_tensors and self.total_tensors < self.split_arguments.split_max_tensors:
logger.warning("Model has fewer tensors than the split threshold, not splitting")
self.split_style = SplitStyle.NONE
if self.split_arguments.split_max_size and total_size < self.split_arguments.split_max_size:
logger.warning("Model has smaller size than the split threshold, not splitting")
self.split_style = SplitStyle.NONE
# no shards are created when writing vocab so make one
if not self.shards:
self.shards.append(Shard(Path(), 0, METADATA_ONLY_INDICATOR, deque()))
# format shard names
if len(self.shards) == 1:
self.shards[0].path = self.path
else:
for i in range(len(self.shards)):
self.shards[i].path = self.path.with_name(SHARD_NAME_FORMAT.format(self.path.stem, i + 1, len(self.shards)))
# print shard info
logger.info("Writing the following files:")
for shard in self.shards:
logger.info(f" {shard.path}: n_tensors = {shard.tensor_count}, total_size = {GGUFWriterSplit.format_n_bytes_to_str(shard.size)}")
if self.split_arguments.dry_run:
logger.info("Dry run, not writing files")
exit()
# we don't want to initialize GGUFWriters until now because they create files
for i, shard in enumerate(self.shards):
# add_architecture is used for consistency - examples/gguf_split doesn't add arch to all shards
writer = GGUFWriter(shard.path, self.arch, use_temp_file=self.use_temp_file,
endianess=self.endianess, add_architecture=(i == 0))
# only the first shard needs all the KV data
if i == 0:
for key, (value, etype) in self.kv_data.items():
writer.add_key(key)
writer.add_val(value, etype)
# add split metadata unless it's one file - small first shard splits even with SplitStyle.NONE
if self.split_arguments.split_style != SplitStyle.NONE or self.split_arguments.small_first_shard:
writer.add_uint16(Keys.Split.LLM_KV_SPLIT_NO, i)
writer.add_uint16(Keys.Split.LLM_KV_SPLIT_COUNT, len(self.shards))
writer.add_int32(Keys.Split.LLM_KV_SPLIT_TENSORS_COUNT, self.total_tensors)
# add tensors, deque popleft() ensures references to eager tensors are not kept
while True:
try:
(name, tensor, dtype) = shard.tensors.popleft()
writer.add_tensor(name, tensor, raw_dtype=dtype)
except IndexError:
break
self.shard_writers.append(writer)
def write_header_to_file(self) -> None:
if self.state is not WriterState.EMPTY:
raise ValueError(f'Expected GGUFWriterSplit state to be EMPTY, got {self.state}')
for writer in self.shard_writers:
writer.write_header_to_file()
self.state = WriterState.HEADER
def write_kv_data_to_file(self) -> None:
if self.state is not WriterState.HEADER:
raise ValueError(f'Expected GGUFWriterSplit state to be HEADER, got {self.state}')
for writer in self.shard_writers:
writer.write_kv_data_to_file()
self.state = WriterState.KV_DATA
def write_tensors_to_file(self, *, progress: bool = False) -> None:
if self.state is not WriterState.KV_DATA:
raise ValueError(f'Expected GGUFWriterSplit state to be KV_DATA, got {self.state}')
running_total = self.total_tensors
for i in range(len(self.shard_writers)):
writer = self.shard_writers[i]
is_metadata = writer.ti_data_count == 0
if is_metadata:
logger.info(f"Writing to shard {i + 1}/{len(self.shards)} with metadata only")
else:
logger.info(f"Writing to shard {i + 1}/{len(self.shards)} with {writer.ti_data_count}/{running_total} remaining tensors (of {self.total_tensors} total)")
running_total -= writer.ti_data_count
writer.write_tensors_to_file(progress=(progress and not is_metadata))
del writer
self.state = WriterState.TI_DATA
# override add_key, add_val to handle kv data separately
def add_key(self, key: str) -> None:
self.recent_key = key
def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True) -> None:
if self.recent_key is None:
raise ValueError("No key set for value")
self.kv_data[self.recent_key] = (val, vtype)
# need to handle arrays separately
def add_array(self, key: str, val: Sequence[Any]) -> None:
if not isinstance(val, Sequence):
raise ValueError(f'Expected a sequence for {key}, got {type(val)}')
self.kv_data[key] = (val, GGUFValueType.ARRAY)
def add_tensor(
self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
raw_dtype: GGMLQuantizationType | None = None,
) -> None:
# we build splits as tensors are added so we need logic to figure out when to split
# logic is all in the conditional because it short-circuits, otherwise accessing self.shards[-1] would throw an error
# create a first shard to start it off
if (len(self.shards) == self.split_arguments.small_first_shard \
# or split when over tensor limit
or (self.split_arguments.split_style == SplitStyle.TENSORS \
and self.shards[-1].tensor_count >= self.split_arguments.split_max_tensors) \
# or split when over size limit
or (self.split_arguments.split_style == SplitStyle.SIZE \
and self.shards[-1].size + GGUFWriterSplit.get_tensor_size(tensor) > self.split_arguments.split_max_size)):
# we fill in the name later when we know how many shards there are
self.shards.append(Shard(Path(), 1, GGUFWriterSplit.get_tensor_size(tensor), deque([(name, tensor, raw_dtype)])))
else:
self.shards[-1].tensor_count += 1
self.shards[-1].size += GGUFWriterSplit.get_tensor_size(tensor)
self.shards[-1].tensors.append((name, tensor, raw_dtype))
def close(self) -> None:
for writer in self.shard_writers:
writer.close()
@staticmethod
def get_tensor_size(tensor) -> int:
try:
return tensor.data_type.elements_to_bytes(np.prod(tensor.shape))
except AttributeError: # numpy ndarray[Any, Any]
return tensor.nbytes
@staticmethod
def split_str_to_n_bytes(split_str: str) -> int:
if split_str.endswith("K"):
n = int(split_str[:-1]) * 1000
elif split_str.endswith("M"):
n = int(split_str[:-1]) * 1000 * 1000
elif split_str.endswith("G"):
n = int(split_str[:-1]) * 1000 * 1000 * 1000
elif split_str.isnumeric():
n = int(split_str)
else:
raise ValueError(f"Invalid split size: {split_str}, must be a number, optionally followed by K, M, or G")
if n <= 0:
raise ValueError(f"Invalid split size: {split_str}, must be positive")
return n
@staticmethod
def format_n_bytes_to_str(num: int) -> str:
if num == METADATA_ONLY_INDICATOR:
return "negligible - metadata only"
fnum = float(num)
for unit in ("", "K", "M", "G"):
if abs(fnum) < 1000.0:
return f"{fnum:3.1f}{unit}"
fnum /= 1000.0
return f"{fnum:.1f}T - over 1TB, --split recommended"