Skip to content

GitLab

Commit 0ce8bcfd authored by xuxzh1's avatar xuxzh1 🎱
Browse files

init

parent b0135f4b
Hide whitespace changes
Inline Side-by-side
Showing with 3152 additions and 15 deletions
llm/llama.cpp/examples/chat-13B.bat 0 → 100644
View file @ 0ce8bcfd
@setlocal disabledelayedexpansion enableextensions
@echo off
cd /d "%~dp0.."
if not "%errorlevel%"=="0" (
echo Unable to change directory.
pause
exit /b 1
)
if not defined MODEL set "MODEL=models\13B\ggml-model-q4_0.bin"
if not defined USER_NAME set "USER_NAME=User"
if not defined AI_NAME set "AI_NAME=ChatLLaMa"
rem Adjust to the number of CPU cores you want to use.
rem if not defined N_THREAD set "N_THREAD=8"
rem Number of tokens to predict (made it larger than default because we want a long interaction)
if not defined N_PREDICTS set "N_PREDICTS=2048"
if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
rem Default main script paths
set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"
rem Get main script path from command line arguments
set "MAIN_SCRIPT_PATH=%~1"
rem If the main script path was not specified, try the default paths
if not defined MAIN_SCRIPT_PATH (
for %%i in (%DEFAULT_MAIN_SCRIPT_PATHS%) do (
if exist "%%i" set "MAIN_SCRIPT_PATH=%%i"
)
)
rem If the main script path was not found, tell the user how to specify it
if not defined MAIN_SCRIPT_PATH (
echo The main script could not be found. Please provide the path to the main script as 1st argument to this script, or place the main script in one of the default locations:
echo %DEFAULT_MAIN_SCRIPT_PATHS%
pause
exit /b 1
)
rem Default context, feel free to edit it
set "PROMPT_TEXT=Text transcript of a never ending dialog, where %USER_NAME% interacts with an AI assistant named %AI_NAME%. %AI_NAME% is helpful, kind, honest, friendly, good at writing and never fails to answer %USER_NAME%'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what %USER_NAME% and %AI_NAME% say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown."
rem Set a temporary variable if N_THREAD is set
if defined N_THREAD (
set "_N_THREAD=--threads %N_THREAD%"
) else (
set "_N_THREAD="
)
rem Run the script
echo "%MAIN_SCRIPT_PATH%" %GEN_OPTIONS% %_N_THREAD% ^
--model "%MODEL%" ^
--n_predict %N_PREDICTS% ^
--color --interactive ^
--reverse-prompt "%USER_NAME%:" ^
--prompt "%PROMPT_TEXT%"
llm/llama.cpp/examples/chat-13B.sh
View file @ 0ce8bcfd
...@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \ ...@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE $PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./main $GEN_OPTIONS \ ./llama-cli $GEN_OPTIONS \
--model "$MODEL" \ --model "$MODEL" \
--threads "$N_THREAD" \ --threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \ --n_predict "$N_PREDICTS" \
......
llm/llama.cpp/examples/chat-persistent.sh
View file @ 0ce8bcfd
...@@ -62,7 +62,7 @@ fi ...@@ -62,7 +62,7 @@ fi
if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
echo 'Prompt cache does not exist, building...' echo 'Prompt cache does not exist, building...'
# Default batch_size to 64 here for better user feedback during initial prompt processing # Default batch_size to 64 here for better user feedback during initial prompt processing
./main 2>>"$LOG" \ ./llama-cli 2>>"$LOG" \
--batch_size 64 \ --batch_size 64 \
"${OPTS[@]}" \ "${OPTS[@]}" \
--prompt-cache "$PROMPT_CACHE_FILE" \ --prompt-cache "$PROMPT_CACHE_FILE" \
...@@ -109,13 +109,13 @@ while read -e line; do ...@@ -109,13 +109,13 @@ while read -e line; do
printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE" printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
./main 2>>"$LOG" "${OPTS[@]}" \ ./llama-cli 2>>"$LOG" "${OPTS[@]}" \
--prompt-cache "$CUR_PROMPT_CACHE" \ --prompt-cache "$CUR_PROMPT_CACHE" \
--prompt-cache-all \ --prompt-cache-all \
--file "$CUR_PROMPT_FILE" \ --file "$CUR_PROMPT_FILE" \
--reverse-prompt "${USER_NAME}:" \ --reverse-prompt "${USER_NAME}:" \
--n_predict "$n_predict" | --n_predict "$n_predict" |
skip_bytes 1 | # skip BOS token added by ./main skip_bytes 1 | # skip BOS token added by ./llama-cli
tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file
skip_bytes "$n_prompt_len_pre" # print generation skip_bytes "$n_prompt_len_pre" # print generation
...@@ -133,7 +133,7 @@ while read -e line; do ...@@ -133,7 +133,7 @@ while read -e line; do
# TODO get both messages in one go # TODO get both messages in one go
if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" || if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
echo >&2 "Couldn't get number of tokens from ./main output!" echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
exit 1 exit 1
fi fi
...@@ -144,7 +144,7 @@ while read -e line; do ...@@ -144,7 +144,7 @@ while read -e line; do
fi fi
# Update cache for next prompt in background, ideally during user input # Update cache for next prompt in background, ideally during user input
./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \ ./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
--prompt-cache "$NEXT_PROMPT_CACHE" \ --prompt-cache "$NEXT_PROMPT_CACHE" \
--file "$NEXT_PROMPT_FILE" \ --file "$NEXT_PROMPT_FILE" \
--n_predict 1 & --n_predict 1 &
......
llm/llama.cpp/examples/chat-vicuna.sh
View file @ 0ce8bcfd
...@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \ ...@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE $PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./bin/main $GEN_OPTIONS \ ./bin/llama-cli $GEN_OPTIONS \
--model "$MODEL" \ --model "$MODEL" \
--threads "$N_THREAD" \ --threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \ --n_predict "$N_PREDICTS" \
......
llm/llama.cpp/examples/chat.sh
View file @ 0ce8bcfd
...@@ -11,6 +11,6 @@ cd .. ...@@ -11,6 +11,6 @@ cd ..
# #
# "--keep 48" is based on the contents of prompts/chat-with-bob.txt # "--keep 48" is based on the contents of prompts/chat-with-bob.txt
# #
./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \ ./llama-cli -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
--repeat_penalty 1.0 --color -i \ --repeat_penalty 1.0 --color -i \
-r "User:" -f prompts/chat-with-bob.txt -r "User:" -f prompts/chat-with-bob.txt
llm/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt
View file @ 0ce8bcfd
set(TARGET convert-llama2c-to-ggml) set(TARGET llama-convert-llama2c-to-ggml)
add_executable(${TARGET} convert-llama2c-to-ggml.cpp) add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
install(TARGETS ${TARGET} RUNTIME) install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
......
llm/llama.cpp/examples/convert-llama2c-to-ggml/README.md
View file @ 0ce8bcfd
...@@ -8,7 +8,7 @@ To convert the model first download the models from the [llama2.c](https://githu ...@@ -8,7 +8,7 @@ To convert the model first download the models from the [llama2.c](https://githu
After successful compilation, following usage options are available: After successful compilation, following usage options are available:
``` ```
usage: ./convert-llama2c-to-ggml [options] usage: ./llama-convert-llama2c-to-ggml [options]
options: options:
-h, --help show this help message and exit -h, --help show this help message and exit
...@@ -19,10 +19,10 @@ options: ...@@ -19,10 +19,10 @@ options:
An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows: An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin` `$ ./llama-convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K). Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
Now you can use the model with a command like: Now you can use the model with a command like:
`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256` `$ ./llama-cli -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
llm/llama.cpp/examples/convert_legacy_llama.py 0 → 100644
View file @ 0ce8bcfd
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import concurrent.futures
import enum
import faulthandler
import functools
import itertools
import json
import math
import mmap
import os
import pickle
import re
import signal
import struct
import sys
import textwrap
import time
import zipfile
from abc import ABC, abstractmethod
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
import numpy as np
if 'NO_LOCAL_GGUF' not in os.environ:
# use .parent.parent since we are in "examples" directory
sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py'))
import gguf
from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab
if TYPE_CHECKING:
from typing_extensions import Self, TypeAlias
logger = logging.getLogger("convert")
if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
faulthandler.register(signal.SIGUSR1)
NDArray: TypeAlias = 'np.ndarray[Any, Any]'
ARCH = gguf.MODEL_ARCH.LLAMA
DEFAULT_CONCURRENCY = 8
ADDED_TOKENS_FILE = 'added_tokens.json'
FAST_TOKENIZER_FILE = 'tokenizer.json'
#
# data types
#
@dataclass(frozen=True)
class DataType:
name: str
dtype: np.dtype[Any]
valid_conversions: list[str]
def elements_to_bytes(self, n_elements: int) -> int:
return n_elements * self.dtype.itemsize
@dataclass(frozen=True)
class UnquantizedDataType(DataType):
pass
DT_F16 = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
DT_F32 = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
DT_I32 = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
@dataclass(frozen=True)
class QuantizedDataType(DataType):
block_size: int
quantized_dtype: np.dtype[Any]
ggml_type: gguf.GGMLQuantizationType
def quantize(self, arr: NDArray) -> NDArray:
raise NotImplementedError(f'Quantization for {self.name} not implemented')
def elements_to_bytes(self, n_elements: int) -> int:
assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
return self.quantized_dtype.itemsize * (n_elements // self.block_size)
@dataclass(frozen=True)
class Q8_0QuantizedDataType(QuantizedDataType):
# Mini Q8_0 quantization in Python!
def quantize(self, arr: NDArray) -> NDArray:
assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
n_blocks = arr.size // self.block_size
blocks = arr.reshape((n_blocks, self.block_size))
# Much faster implementation of block quantization contributed by @Cebtenzzre
def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
d = abs(blocks).max(axis = 1) / np.float32(127)
with np.errstate(divide = 'ignore'):
qs = (blocks / d[:, None]).round()
qs[d == 0] = 0
yield from zip(d, qs)
return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
dtype = np.dtype(np.float32), valid_conversions = [],
ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
# Quantized types skipped here because they may also map to np.float32
NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
raise ValueError(f'Invalid duplicate data type {dt}')
NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
'BF16': DT_BF16,
'F16': DT_F16,
'F32': DT_F32,
'I32': DT_I32,
}
# TODO: match this with `llama_ftype`
# TODO: rename to LLAMAFileType
# TODO: move to `gguf.py`
class GGMLFileType(enum.IntEnum):
AllF32 = 0
MostlyF16 = 1 # except 1d tensors
MostlyQ8_0 = 7 # except 1d tensors
def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType:
dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
if dt is None:
raise ValueError(self)
# Convert all 1D tensors to F32. Most of the codebase that takes in 1D tensors only handles F32 tensors, and most of the outputs tensors are F32.
# Also The 1d tensors aren't much of a performance/size issue. So instead of having to have separate F32 and F16 implementations of both, just convert everything to F32 for now.
return dt if len(tensor.shape) > 1 else DT_F32
GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
GGMLFileType.AllF32 : DT_F32,
GGMLFileType.MostlyF16 : DT_F16,
GGMLFileType.MostlyQ8_0: DT_Q8_0,
}
#
# hparams loading
#
@dataclass
class Params:
n_vocab: int
n_embd: int
n_layer: int
n_ctx: int
n_ff: int
n_head: int
n_head_kv: int
n_experts: int | None = None
n_experts_used: int | None = None
f_norm_eps: float | None = None
rope_scaling_type: gguf.RopeScalingType | None = None
f_rope_freq_base: float | None = None
f_rope_scale: float | None = None
n_ctx_orig: int | None = None
rope_finetuned: bool | None = None
ftype: GGMLFileType | None = None
# path to the directory containing the model files
path_model: Path | None = None
@staticmethod
def guessed(model: LazyModel) -> Params:
# try transformer naming first
n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
# try transformer naming first
if "model.layers.0.self_attn.q_proj.weight" in model:
n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming
n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
else:
n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
if n_layer < 1:
msg = """\
failed to guess 'n_layer'. This model is unknown or unsupported.
Suggestion: provide 'config.json' of the model in the same directory containing model files."""
raise KeyError(textwrap.dedent(msg))
n_head = n_embd // 128 # guessed
n_mult = 256 # guessed
# TODO: verify this
n_ff = int(2 * (4 * n_embd) / 3)
n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult)
return Params(
n_vocab = n_vocab,
n_embd = n_embd,
n_layer = n_layer,
n_ctx = -1,
n_ff = n_ff,
n_head = n_head,
n_head_kv = n_head,
f_norm_eps = 1e-5,
)
@staticmethod
def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
with open(config_path) as f:
config = json.load(f)
rope_scaling_type = f_rope_scale = n_ctx_orig = rope_finetuned = None
rope_scaling = config.get("rope_scaling")
if rope_scaling is not None and (typ := rope_scaling.get("type")):
rope_factor = rope_scaling.get("factor")
f_rope_scale = rope_factor
if typ == "linear":
rope_scaling_type = gguf.RopeScalingType.LINEAR
elif typ == "yarn":
rope_scaling_type = gguf.RopeScalingType.YARN
n_ctx_orig = rope_scaling['original_max_position_embeddings']
rope_finetuned = rope_scaling['finetuned']
else:
raise NotImplementedError(f'Unknown rope scaling type: {typ}')
if "max_sequence_length" in config:
n_ctx = config["max_sequence_length"]
elif "max_position_embeddings" in config:
n_ctx = config["max_position_embeddings"]
else:
msg = """\
failed to guess 'n_ctx'. This model is unknown or unsupported.
Suggestion: provide 'config.json' of the model in the same directory containing model files."""
raise KeyError(textwrap.dedent(msg))
n_experts = None
n_experts_used = None
if "num_local_experts" in config:
n_experts = config["num_local_experts"]
n_experts_used = config["num_experts_per_tok"]
return Params(
n_vocab = config["vocab_size"],
n_embd = config["hidden_size"],
n_layer = config["num_hidden_layers"],
n_ctx = n_ctx,
n_ff = config["intermediate_size"],
n_head = (n_head := config["num_attention_heads"]),
n_head_kv = config.get("num_key_value_heads", n_head),
n_experts = n_experts,
n_experts_used = n_experts_used,
f_norm_eps = config["rms_norm_eps"],
f_rope_freq_base = config.get("rope_theta"),
rope_scaling_type = rope_scaling_type,
f_rope_scale = f_rope_scale,
n_ctx_orig = n_ctx_orig,
rope_finetuned = rope_finetuned,
)
# LLaMA v2 70B params.json
# {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
@staticmethod
def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
with open(config_path) as f:
config = json.load(f)
n_experts = None
n_experts_used = None
f_rope_freq_base = None
n_ff = None
# hack to determine LLaMA v1 vs v2 vs CodeLlama
if config.get("moe"):
# Mixtral
n_ctx = 32768
elif config.get("rope_theta") == 1000000:
# CodeLlama
n_ctx = 16384
elif config["norm_eps"] == 1e-05:
# LLaMA v2
n_ctx = 4096
else:
# LLaMA v1
n_ctx = 2048
if "layers.0.feed_forward.w1.weight" in model:
n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
if config.get("moe"):
n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
n_experts = config["moe"]["num_experts"]
n_experts_used = config["moe"]["num_experts_per_tok"]
f_rope_freq_base = 1e6
assert n_ff is not None
return Params(
n_vocab = model["tok_embeddings.weight"].shape[0],
n_embd = config["dim"],
n_layer = config["n_layers"],
n_ctx = n_ctx,
n_ff = n_ff,
n_head = (n_head := config["n_heads"]),
n_head_kv = config.get("n_kv_heads", n_head),
n_experts = n_experts,
n_experts_used = n_experts_used,
f_norm_eps = config["norm_eps"],
f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
)
@staticmethod
def load(model_plus: ModelPlus) -> Params:
hf_config_path = model_plus.paths[0].parent / "config.json"
orig_config_path = model_plus.paths[0].parent / "params.json"
if hf_config_path.exists():
params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
elif orig_config_path.exists():
params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
elif model_plus.format != 'none':
params = Params.guessed(model_plus.model)
else:
raise ValueError('Cannot guess params when model format is none')
params.path_model = model_plus.paths[0].parent
return params
#
# data loading
# TODO: reuse (probably move to gguf.py?)
#
def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
if n_head_kv is not None and n_head != n_head_kv:
n_head = n_head_kv
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape))
class Tensor(ABC):
ndarray: NDArray
data_type: DataType
@abstractmethod
def astype(self, data_type: DataType) -> Self: ...
@abstractmethod
def permute(self, n_head: int, n_head_kv: int) -> Self: ...
@abstractmethod
def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> Self: ...
@abstractmethod
def part(self, n_part: int) -> Self: ...
@abstractmethod
def to_ggml(self) -> GGMLCompatibleTensor: ...
def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
fp32_arr = bf16_arr.astype(np.uint32) << 16
return fp32_arr.view(np.float32)
class UnquantizedTensor(Tensor):
def __init__(self, ndarray: NDArray):
assert isinstance(ndarray, np.ndarray)
self.ndarray = ndarray
self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
def astype(self, data_type: DataType) -> UnquantizedTensor:
dtype = data_type.dtype
if self.data_type == DT_BF16:
self.ndarray = bf16_to_fp32(self.ndarray)
return UnquantizedTensor(self.ndarray.astype(dtype))
def to_ggml(self) -> Self:
return self
def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
r = self.ndarray.shape[0] // 3
return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
def part(self, n_part: int) -> UnquantizedTensor:
r = self.ndarray.shape[0] // 3
return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor:
return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv))
def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray:
tensor = lazy_tensor.load()
assert isinstance(tensor, UnquantizedTensor)
# double-check:
actual_shape = list(tensor.ndarray.shape)
assert actual_shape == lazy_tensor.shape, (actual_shape, lazy_tensor.shape)
if expected_dtype is not None and expected_dtype != tensor.ndarray.dtype:
if convert:
tensor.ndarray = tensor.ndarray.astype(expected_dtype)
else:
raise ValueError(f'expected this tensor to have dtype {expected_dtype}, got {tensor.ndarray.dtype}')
return tensor.ndarray
GGMLCompatibleTensor = UnquantizedTensor
@dataclass
class LazyTensor:
_load: Callable[[], Tensor]
shape: list[int]
data_type: DataType
description: str
def load(self) -> Tensor:
ret = self._load()
# Should be okay if it maps to the same numpy type?
assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
(self.data_type, ret.data_type, self.description)
return ret
def astype(self, data_type: DataType) -> LazyTensor:
self.validate_conversion_to(data_type)
def load() -> Tensor:
return self.load().astype(data_type)
return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
def validate_conversion_to(self, data_type: DataType) -> None:
if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
LazyModel: TypeAlias = 'dict[str, LazyTensor]'
ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none']
@dataclass
class ModelPlus:
model: LazyModel
paths: list[Path] # Where this was read from.
format: ModelFormat
vocab: BaseVocab | None # For GGML models (which have vocab built in), the vocab.
def merge_sharded(models: list[LazyModel]) -> LazyModel:
# Original LLaMA models have each file contain one part of each tensor.
# Use a dict instead of a set to preserve order.
names = {name: None for model in models for name in model}
def convert(name: str) -> LazyTensor:
lazy_tensors = [model[name] for model in models]
if len(lazy_tensors) == 1:
# only one file; don't go through this procedure since there might
# be quantized tensors
return lazy_tensors[0]
if len(lazy_tensors[0].shape) == 1:
# the tensor is just duplicated in every file
return lazy_tensors[0]
if name.startswith('tok_embeddings.') or \
name.endswith('.attention.wo.weight') or \
name.endswith('.feed_forward.w2.weight'):
# split by columns
axis = 1
else:
# split by rows
axis = 0
concatenated_shape = list(lazy_tensors[0].shape)
concatenated_shape[axis] = sum(tensor.shape[axis] for tensor in lazy_tensors)
def load() -> UnquantizedTensor:
ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors]
concatenated = np.concatenate(ndarrays, axis=axis)
return UnquantizedTensor(concatenated)
description = 'concatenated[[' + '] | ['.join(lt.description for lt in lazy_tensors) + ']]'
return LazyTensor(load, concatenated_shape, lazy_tensors[0].data_type, description)
return {name: convert(name) for name in names}
def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
formats: set[ModelFormat] = set(mp.format for mp in models_plus)
assert len(formats) == 1, "different formats?"
format = formats.pop()
paths = [path for mp in models_plus for path in mp.paths]
# Use the first non-None vocab, if any.
try:
vocab = next(mp.vocab for mp in models_plus if mp.vocab is not None)
except StopIteration:
vocab = None
if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
# Transformers models put different tensors in different files, but
# don't split individual tensors between files.
model: LazyModel = {}
for mp in models_plus:
model.update(mp.model)
else:
model = merge_sharded([mp.model for mp in models_plus])
return ModelPlus(model, paths, format, vocab)
def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
def load() -> Tensor:
return lazy_tensor.load().permute(n_head, n_head_kv)
return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
def load() -> Tensor:
return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
s = lazy_tensor.shape.copy()
s[0] = s[0] // 3
return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
def load() -> Tensor:
return lazy_tensor.load().part(n_part)
s = lazy_tensor.shape.copy()
s[0] = s[0] // 3
return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
def pack_experts_lazy(lazy_tensors: list[LazyTensor]) -> LazyTensor:
def load() -> Tensor:
tensors = [lazy_tensor.load() for lazy_tensor in lazy_tensors]
return UnquantizedTensor(np.array([tensor.ndarray for tensor in tensors]))
s = lazy_tensors[0].shape.copy()
s.insert(0, len(lazy_tensors))
return LazyTensor(load, s, lazy_tensors[0].data_type, 'pack_experts ' + ' | '.join(lt.description for lt in lazy_tensors))
# Functionality that simulates `torch.load` but where individual tensors are
# only loaded into memory on demand, not all at once.
# PyTorch can't do this natively as of time of writing:
# - https://github.com/pytorch/pytorch/issues/64327
# This allows us to de-shard without multiplying RAM usage, and also
# conveniently drops the PyTorch dependency (though we still need numpy).
@dataclass
class LazyStorageKind:
data_type: DataType
@dataclass
class LazyStorage:
load: Callable[[int, int], NDArray]
kind: LazyStorageKind
description: str
class LazyUnpickler(pickle.Unpickler):
def __init__(self, fp: IO[bytes], data_base_path: str, zip_file: zipfile.ZipFile):
super().__init__(fp)
self.data_base_path = data_base_path
self.zip_file = zip_file
def persistent_load(self, pid: Any) -> Any:
assert pid[0] == 'storage'
assert isinstance(pid[1], LazyStorageKind)
data_type = pid[1].data_type
filename_stem = pid[2]
filename = f'{self.data_base_path}/{filename_stem}'
info = self.zip_file.getinfo(filename)
def load(offset: int, elm_count: int) -> NDArray:
dtype = data_type.dtype
with self.zip_file.open(info) as fp:
fp.seek(offset * dtype.itemsize)
size = elm_count * dtype.itemsize
data = fp.read(size)
assert len(data) == size
return np.frombuffer(data, dtype)
description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
return LazyStorage(load=load, kind=pid[1], description=description)
@staticmethod
def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
assert isinstance(storage, LazyStorage)
def load() -> UnquantizedTensor:
elm_count = stride[0] * size[0]
return UnquantizedTensor(storage.load(storage_offset, elm_count).reshape(size))
description = f'pickled storage_offset={storage_offset} in {storage.description}'
return LazyTensor(load, list(size), storage.kind.data_type, description)
@staticmethod
def rebuild_from_type_v2(func, new_type, args, state):
return func(*args)
CLASSES: dict[tuple[str, str], type[LazyTensor] | LazyStorageKind] = {
# getattr used here as a workaround for mypy not being smart enough to determine
# the staticmethods have a __func__ attribute.
('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
('torch', 'IntStorage'): LazyStorageKind(DT_I32),
('torch', 'Tensor'): LazyTensor,
}
def find_class(self, module: str, name: str) -> Any:
if not module.startswith('torch'):
return super().find_class(module, name)
return self.CLASSES[(module, name)]
def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
zf = zipfile.ZipFile(outer_fp)
pickle_paths = [name for name in zf.namelist() if name.endswith('.pkl')]
assert len(pickle_paths) == 1, pickle_paths
pickle_fp = zf.open(pickle_paths[0], 'r')
unpickler = LazyUnpickler(pickle_fp,
data_base_path=pickle_paths[0][:-4],
zip_file=zf)
model = unpickler.load()
if 'model' in model: model = model['model']
as_dict = dict(model.items())
return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None)
def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
header_size, = struct.unpack('<Q', fp.read(8))
header: dict[str, dict[str, Any]] = json.loads(fp.read(header_size))
# Use mmap for the actual data to avoid race conditions with the file offset.
mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
byte_buf = mapped[8 + header_size:]
def convert(info: dict[str, Any]) -> LazyTensor:
data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
numpy_dtype = data_type.dtype
shape: list[int] = info['shape']
begin, end = info['data_offsets']
assert 0 <= begin <= end <= len(byte_buf)
assert end - begin == math.prod(shape) * numpy_dtype.itemsize
buf = byte_buf[begin:end]
def load() -> UnquantizedTensor:
return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape))
description = f'safetensors begin={begin} end={end} type={data_type} path={path}'
return LazyTensor(load, shape, data_type, description)
model = {name: convert(info) for (name, info) in header.items() if name != '__metadata__'}
return ModelPlus(model=model, paths=[path], format='safetensors', vocab=None)
def must_read(fp: IO[bytes], length: int) -> bytes:
ret = fp.read(length)
if len(ret) < length:
raise EOFError("unexpectedly reached end of file")
return ret
@functools.lru_cache(maxsize=None)
def lazy_load_file(path: Path) -> ModelPlus:
fp = open(path, 'rb')
first8 = fp.read(8)
fp.seek(0)
if first8[:2] == b'PK':
# A zip file, i.e. PyTorch format
return lazy_load_torch_file(fp, path)
elif struct.unpack('<Q', first8)[0] < 16 * 1024 * 1024:
# Probably safetensors
return lazy_load_safetensors_file(fp, path)
else:
raise ValueError(f"unknown format: {path}")
In = TypeVar('In')
Out = TypeVar('Out')
def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
'''Parallel map, but with backpressure. If the caller doesn't call `next`
fast enough, this will stop calling `func` at some point rather than
letting results pile up in memory. Specifically, there is a max of one
output value buffered per thread.'''
if concurrency < 2:
yield from map(func, iterable)
# Not reached.
iterable = iter(iterable)
executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor]
if use_processpool_executor:
executor_class = ProcessPoolExecutor
else:
executor_class = ThreadPoolExecutor
with executor_class(max_workers=max_workers) as executor:
futures: list[concurrent.futures.Future[Out]] = []
done = False
for _ in range(concurrency):
try:
futures.append(executor.submit(func, next(iterable)))
except StopIteration:
done = True
break
while futures:
result = futures.pop(0).result()
while not done and len(futures) < concurrency:
try:
futures.append(executor.submit(func, next(iterable)))
except StopIteration:
done = True
break
yield result
def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) -> None:
# Handle special case where the model's vocab size is not set
if params.n_vocab == -1:
raise ValueError(
"The model's vocab size is set to -1 in params.json. Please update it manually."
+ (f" Maybe {vocab.vocab_size}?" if isinstance(vocab, Vocab) else ""),
)
if not isinstance(vocab, Vocab):
return # model has no vocab
# Check for a vocab size mismatch
if params.n_vocab == vocab.vocab_size:
logger.warning("Ignoring added_tokens.json since model matches vocab size without it.")
return
if pad_vocab and params.n_vocab > vocab.vocab_size:
pad_count = params.n_vocab - vocab.vocab_size
logger.debug(
f"Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>"
)
for i in range(1, pad_count + 1):
vocab.added_tokens_dict[f"<dummy{i:05}>"] = -1
vocab.added_tokens_list.append(f"<dummy{i:05}>")
vocab.vocab_size = params.n_vocab
return
msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer} has {vocab.vocab_size})."
if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
msg += f" Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
if vocab.vocab_size < params.n_vocab:
msg += " Add the --pad-vocab option and try again."
raise ValueError(msg)
class OutputFile:
def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
# Metadata About The Model And Its Provenence
name = "LLaMA"
if metadata is not None and metadata.name is not None:
name = metadata.name
elif params.path_model is not None:
name = params.path_model.name
elif params.n_ctx == 4096:
# Heuristic detection of LLaMA v2 model
name = "LLaMA v2"
self.gguf.add_name(name)
if metadata is not None:
if metadata.author is not None:
self.gguf.add_author(metadata.author)
if metadata.version is not None:
self.gguf.add_version(metadata.version)
if metadata.organization is not None:
self.gguf.add_organization(metadata.organization)
if metadata.finetune is not None:
self.gguf.add_finetune(metadata.finetune)
if metadata.basename is not None:
self.gguf.add_basename(metadata.basename)
if metadata.description is not None:
self.gguf.add_description(metadata.description)
if metadata.quantized_by is not None:
self.gguf.add_quantized_by(metadata.quantized_by)
if metadata.size_label is not None:
self.gguf.add_size_label(metadata.size_label)
if metadata.license is not None:
self.gguf.add_license(metadata.license)
if metadata.license_name is not None:
self.gguf.add_license_name(metadata.license_name)
if metadata.license_link is not None:
self.gguf.add_license_link(metadata.license_link)
if metadata.url is not None:
self.gguf.add_url(metadata.url)
if metadata.doi is not None:
self.gguf.add_doi(metadata.doi)
if metadata.uuid is not None:
self.gguf.add_uuid(metadata.uuid)
if metadata.repo_url is not None:
self.gguf.add_repo_url(metadata.repo_url)
if metadata.source_url is not None:
self.gguf.add_source_url(metadata.source_url)
if metadata.source_doi is not None:
self.gguf.add_source_doi(metadata.source_doi)
if metadata.source_uuid is not None:
self.gguf.add_source_uuid(metadata.source_uuid)
if metadata.source_repo_url is not None:
self.gguf.add_source_repo_url(metadata.source_repo_url)
if metadata.base_models is not None:
self.gguf.add_base_model_count(len(metadata.base_models))
for key, base_model_entry in enumerate(metadata.base_models):
if "name" in base_model_entry:
self.gguf.add_base_model_name(key, base_model_entry["name"])
if "author" in base_model_entry:
self.gguf.add_base_model_author(key, base_model_entry["author"])
if "version" in base_model_entry:
self.gguf.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
self.gguf.add_base_model_organization(key, base_model_entry["organization"])
if "url" in base_model_entry:
self.gguf.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
self.gguf.add_base_model_doi(key, base_model_entry["doi"])
if "uuid" in base_model_entry:
self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
if "repo_url" in base_model_entry:
self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
if metadata.tags is not None:
self.gguf.add_tags(metadata.tags)
if metadata.languages is not None:
self.gguf.add_languages(metadata.languages)
if metadata.datasets is not None:
self.gguf.add_datasets(metadata.datasets)
def add_meta_arch(self, params: Params) -> None:
# Metadata About The Neural Architecture Itself
self.gguf.add_vocab_size(params.n_vocab)
self.gguf.add_context_length(params.n_ctx)
self.gguf.add_embedding_length(params.n_embd)
self.gguf.add_block_count(params.n_layer)
self.gguf.add_feed_forward_length(params.n_ff)
self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
self.gguf.add_head_count (params.n_head)
self.gguf.add_head_count_kv (params.n_head_kv)
if params.n_experts:
self.gguf.add_expert_count(params.n_experts)
if params.n_experts_used:
self.gguf.add_expert_used_count(params.n_experts_used)
if params.f_norm_eps:
self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
else:
raise ValueError('f_norm_eps is None')
if params.f_rope_freq_base is not None:
self.gguf.add_rope_freq_base(params.f_rope_freq_base)
if params.rope_scaling_type:
assert params.f_rope_scale is not None
self.gguf.add_rope_scaling_type(params.rope_scaling_type)
self.gguf.add_rope_scaling_factor(params.f_rope_scale)
if params.n_ctx_orig is not None:
self.gguf.add_rope_scaling_orig_ctx_len(params.n_ctx_orig)
if params.rope_finetuned is not None:
self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
if params.ftype is not None:
self.gguf.add_file_type(params.ftype)
def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
tokens = []
scores = []
toktypes = []
# NOTE: `all_tokens` returns the base vocabulary and added tokens
for text, score, toktype in vocab.all_tokens():
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
assert len(tokens) == vocab.vocab_size
return tokens, scores, toktypes
def add_meta_vocab(self, vocab: Vocab) -> None:
# Ensure that tokenizer_model is added to the GGUF model
self.gguf.add_tokenizer_model(vocab.tokenizer_model)
# Extract model vocabulary for model conversion
tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
# Add extracted token information for model conversion
self.gguf.add_token_list(tokens)
self.gguf.add_token_scores(scores)
self.gguf.add_token_types(toktypes)
def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
svocab.add_to_gguf(self.gguf)
def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
n_elements = int(np.prod(tensor.shape))
raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype)
def write_meta(self) -> None:
self.gguf.write_header_to_file()
self.gguf.write_kv_data_to_file()
def write_tensor_info(self) -> None:
self.gguf.write_ti_data_to_file()
def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
if ftype == GGMLFileType.MostlyQ8_0:
ndarrays = bounded_parallel_map(
OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
use_processpool_executor=True,
)
else:
ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
start = time.time()
for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
elapsed = time.time() - start
size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
padi = len(str(len(model)))
logger.info(
f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
)
self.gguf.write_tensor_data(ndarray)
def close(self) -> None:
self.gguf.close()
@staticmethod
def write_vocab_only(
fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
of = OutputFile(fname_out, endianess=endianess)
# meta data
of.add_meta_model(params, metadata)
of.add_meta_arch(params)
of.add_meta_vocab(vocab)
of.add_meta_special_vocab(svocab)
of.write_meta()
of.close()
@staticmethod
def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]:
name, lazy_tensor = item
tensor = lazy_tensor.load().to_ggml()
return (lazy_tensor.data_type, tensor.ndarray)
@staticmethod
def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray:
dt, arr = item
if not isinstance(dt, QuantizedDataType):
return arr
return dt.quantize(arr)
@staticmethod
def write_all(
fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
pad_vocab: bool = False,
metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
of = OutputFile(fname_out, endianess=endianess)
# meta data
of.add_meta_model(params, metadata)
of.add_meta_arch(params)
if isinstance(vocab, Vocab):
of.add_meta_vocab(vocab)
of.add_meta_special_vocab(svocab)
else: # NoVocab
of.gguf.add_tokenizer_model(vocab.tokenizer_model)
# tensor info
for name, lazy_tensor in model.items():
of.add_tensor_info(name, lazy_tensor)
of.write_meta()
of.write_tensor_info()
# tensor data
of.write_tensor_data(ftype, model, concurrency)
of.close()
def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
return GGMLFileType.AllF32
if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
return GGMLFileType.MostlyF16
if output_type_str == "q8_0":
return GGMLFileType.MostlyQ8_0
name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
raise ValueError(f"Unexpected combination of types: {name_to_type}")
def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
total_params = 0
shared_params = 0
expert_params = 0
for name, lazy_tensor in tensors:
# We don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
# Got A Tensor
sum_weights_in_tensor: int = 1
# Tensor Volume
for dim in lazy_tensor.shape:
sum_weights_in_tensor *= dim
if ".experts." in name:
if ".experts.0." in name:
expert_params += sum_weights_in_tensor
else:
shared_params += sum_weights_in_tensor
total_params += sum_weights_in_tensor
return total_params, shared_params, expert_params
def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
for (name, tensor) in model.items()}
def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
tmap = gguf.TensorNameMap(ARCH, params.n_layer)
should_skip = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
tmp = model
# merge experts into one tensor
if params.n_experts and params.n_experts > 0:
for i_l in range(params.n_layer):
for w in range(1, 4):
experts = []
for e in range(params.n_experts):
if f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight" in model:
experts.append(model[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"])
del tmp[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"]
elif f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight" in model:
experts.append(model[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"])
del tmp[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"]
else:
raise ValueError(f"Expert tensor not found: layers.{i_l}.feed_forward.experts.{e}.w{w}.weight")
tmp[f"layers.{i_l}.feed_forward.experts.w{w}.weight"] = pack_experts_lazy(experts)
# HF models permut or pack some of the tensors, so we need to undo that
for i in itertools.count():
if f"model.layers.{i}.self_attn.q_proj.weight" in model:
logger.debug(f"Permuting layer {i}")
tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
# tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
logger.debug(f"Unpacking and permuting layer {i}")
tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
else:
break
out: LazyModel = {}
for name, lazy_tensor in model.items():
tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
if name_new is None:
if skip_unknown:
logger.warning(f"Unexpected tensor name: {name} - skipping")
continue
raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
if tensor_type in should_skip:
logger.debug(f"skipping tensor {name_new}")
continue
logger.debug(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
out[name_new] = lazy_tensor
return out
def nth_multifile_path(path: Path, n: int) -> Path | None:
'''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
the nth path in the model.
'''
# Support the following patterns:
patterns = [
# - x.00.pth, x.01.pth, etc.
(r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
# - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
(r'-[0-9]{5}-of-(.*)$', fr'-{n:05}-of-\1'),
# x.bin, x.bin.1, etc.
(r'(\.[0-9]+)?$', r'\1' if n == 0 else fr'\1.{n}')
]
for regex, replacement in patterns:
if re.search(regex, path.name):
new_path = path.with_name(re.sub(regex, replacement, path.name))
if new_path.exists():
return new_path
return None
def find_multifile_paths(path: Path) -> list[Path]:
'''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
the whole list of paths in the model.
'''
ret: list[Path] = []
for i in itertools.count():
nth_path = nth_multifile_path(path, i)
if nth_path is None:
break
ret.append(nth_path)
if not ret:
# No matches. This should only happen if the file was named, e.g.,
# foo.0, and there was no file named foo. Oh well, try to process it
# as a single file.
return [path]
return ret
def load_some_model(path: Path) -> ModelPlus:
'''Load a model of any supported format.'''
# Be extra-friendly and accept either a file or a directory:
if path.is_dir():
# Check if it's a set of safetensors files first
globs = ["model-00001-of-*.safetensors", "model.safetensors", "consolidated.safetensors"]
files = [file for glob in globs for file in path.glob(glob)]
if not files:
# Try the PyTorch patterns too, with lower priority
globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
files = [file for glob in globs for file in path.glob(glob)]
if not files:
raise FileNotFoundError(f"Can't find model in directory {path}")
if len(files) > 1:
raise ValueError(f"Found multiple models in {path}, not sure which to pick: {files}")
path = files[0]
paths = find_multifile_paths(path)
models_plus: list[ModelPlus] = []
for path in paths:
logger.info(f"Loading model file {path}")
models_plus.append(lazy_load_file(path))
model_plus = merge_multifile_models(models_plus)
return model_plus
class VocabFactory:
_VOCAB_CLASSES: list[type[Vocab]] = [SentencePieceVocab, BpeVocab, LlamaHfVocab]
def __init__(self, path: Path):
self.path = path
def _create_special_vocab(self, vocab: BaseVocab, model_parent_path: Path) -> gguf.SpecialVocab:
load_merges = vocab.name == "bpe"
n_vocab = vocab.vocab_size if isinstance(vocab, Vocab) else None
return gguf.SpecialVocab(
model_parent_path,
load_merges=load_merges,
special_token_types=None, # Predetermined or passed as a parameter
n_vocab=n_vocab,
)
def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
vocab_classes: dict[str, type[Vocab]] = {cls.name: cls for cls in self._VOCAB_CLASSES}
selected_vocabs: dict[str, type[Vocab]] = {}
for vtype in vocab_types:
try:
selected_vocabs[vtype] = vocab_classes[vtype]
except KeyError:
raise ValueError(f"Unsupported vocabulary type {vtype}") from None
for vtype, cls in selected_vocabs.items():
try:
vocab = cls(self.path)
break
except FileNotFoundError:
pass # ignore unavailable tokenizers
else:
raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}")
logger.info(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
return vocab
def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]:
vocab: BaseVocab
if vocab_types is None:
vocab = NoVocab()
else:
vocab = self._create_vocab_by_path(vocab_types)
# FIXME: Respect --vocab-dir?
special_vocab = self._create_special_vocab(
vocab,
model_parent_path,
)
return vocab, special_vocab
def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
name = metadata.name if metadata.name is not None else None
basename = metadata.basename if metadata.basename is not None else None
finetune = metadata.finetune if metadata.finetune is not None else None
version = metadata.version if metadata.version is not None else None
size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
output_type = {
GGMLFileType.AllF32: "F32",
GGMLFileType.MostlyF16: "F16",
GGMLFileType.MostlyQ8_0: "Q8_0",
}[file_type]
return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
ret = model_paths[0].parent / f"{default_filename}.gguf"
if ret in model_paths:
logger.error(
f"Error: Default output path ({ret}) would overwrite the input. "
"Please explicitly specify a path using --outfile.")
sys.exit(1)
return ret
def do_dump_model(model_plus: ModelPlus) -> None:
print(f"model_plus.paths = {model_plus.paths!r}") # noqa: NP100
print(f"model_plus.format = {model_plus.format!r}") # noqa: NP100
print(f"model_plus.vocab = {model_plus.vocab!r}") # noqa: NP100
for name, lazy_tensor in model_plus.model.items():
print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") # noqa: NP100
def main(args_in: list[str] | None = None) -> None:
output_choices = ["f32", "f16"]
if np.uint32(1) == np.uint32(1).newbyteorder("<"):
# We currently only support Q8_0 output on little endian systems.
output_choices.append("q8_0")
parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
parser.add_argument("--no-vocab", action="store_true", help="store model without the vocab")
parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
parser.add_argument("--vocab-type", help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
parser.add_argument("--ctx", type=int, help="model training context (default: based on input)")
parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine")
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
parser.add_argument("--metadata", type=Path, help="Specify the path for an authorship metadata override file")
parser.add_argument("--get-outfile", action="store_true", help="get calculated default outfile name")
parser.add_argument("--model-name", type=str, default=None, help="name of the model")
args = parser.parse_args(args_in)
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
elif args.dump_single or args.dump or args.get_outfile:
# Avoid printing anything besides the dump output
logging.basicConfig(level=logging.WARNING)
else:
logging.basicConfig(level=logging.INFO)
model_name = args.model_name
dir_model = args.model
metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
if args.get_outfile:
model_plus = load_some_model(dir_model)
params = Params.load(model_plus)
model = convert_model_names(model_plus.model, params, args.skip_unknown)
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
ftype = pick_output_type(model, args.outtype)
if (metadata is None or metadata.name is None) and params.path_model is not None:
metadata.name = params.path_model.name
print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
return
if args.no_vocab and args.vocab_only:
raise ValueError("--vocab-only does not make sense with --no-vocab")
if args.dump_single:
model_plus = lazy_load_file(dir_model)
do_dump_model(model_plus)
return
if not args.vocab_only:
model_plus = load_some_model(dir_model)
else:
model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
if args.dump:
do_dump_model(model_plus)
return
endianess = gguf.GGUFEndian.LITTLE
if args.big_endian:
endianess = gguf.GGUFEndian.BIG
params = None
if args.pad_vocab or not args.vocab_only:
params = Params.load(model_plus)
if params.n_ctx == -1:
if args.ctx is None:
msg = """\
The model doesn't have a context size, and you didn't specify one with --ctx
Please specify one with --ctx:
- LLaMA v1: --ctx 2048
- LLaMA v2: --ctx 4096"""
parser.error(textwrap.dedent(msg))
params.n_ctx = args.ctx
if args.outtype:
params.ftype = {
"f32": GGMLFileType.AllF32,
"f16": GGMLFileType.MostlyF16,
"q8_0": GGMLFileType.MostlyQ8_0,
}[args.outtype]
logger.info(f"params = {params}")
model_parent_path = model_plus.paths[0].parent
vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
vocab_factory = VocabFactory(vocab_path)
vocab_types = None if args.no_vocab else args.vocab_type.split(",")
vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
if args.vocab_only:
assert isinstance(vocab, Vocab)
if not args.outfile:
raise ValueError("need --outfile if using --vocab-only")
outfile = args.outfile
if params is None:
params = Params(
n_vocab = vocab.vocab_size,
n_embd = 1,
n_layer = 1,
n_ctx = 1,
n_ff = 1,
n_head = 1,
n_head_kv = 1,
f_norm_eps = 1e-5,
)
OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
logger.info(f"Wrote {outfile}")
return
if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
vocab = model_plus.vocab
assert params is not None
if metadata.name is None and params.path_model is not None:
metadata.name = params.path_model.name
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
logger.info(f"Vocab info: {vocab}")
logger.info(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params, args.skip_unknown)
ftype = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, ftype)
outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
params.ftype = ftype
logger.info(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
logger.info(f"Wrote {outfile}")
if __name__ == '__main__':
main()
llm/llama.cpp/examples/cvector-generator/CMakeLists.txt 0 → 100644
View file @ 0ce8bcfd
set(TARGET llama-cvector-generator)
add_executable(${TARGET} cvector-generator.cpp pca.hpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
llm/llama.cpp/examples/cvector-generator/README.md 0 → 100644
View file @ 0ce8bcfd
# cvector-generator
This example demonstrates how to generate a control vector using gguf models.
Related PRs:
- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
## Examples
```sh
# CPU only
./cvector-generator -m ./llama-3.Q4_K_M.gguf
# With GPU
./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99
# With advanced options
./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100
# Using mean value instead of PCA
./cvector-generator -m ./llama-3.Q4_K_M.gguf --method mean
# To see help message
./cvector-generator -h
# Then, have a look at "cvector" section
```
## Tips and tricks
If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
```
<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
<|im_start|>system\nYou are in a very good mood today<|im_end|>
```
Example to use output file with `llama-cli`:
(Tips: The control vector works better when apply to layers higher than 10)
```sh
./llama-cli -m ./llama-3.Q4_K_M.gguf -p "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSing a song<|im_end|><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" --special --control-vector-scaled ./control_vector.gguf 0.8 --control-vector-layer-range 10 31
```
llm/llama.cpp/examples/cvector-generator/completions.txt 0 → 100644
View file @ 0ce8bcfd
That game
I can see
Hmm, this
I can relate to
Who is
I understand the
Ugh,
What the hell was
Hey, did anyone
Although
Thank you for choosing
What are you
Oh w
How dare you open
It was my pleasure
I'm hon
I appreciate that you
Are you k
Whoever left this
It's always
Ew,
Hey, I l
Hello? Is someone
I understand that
That poem
Aww, poor
Hey, it
Alright, who
I didn't
Well, life
The document
Oh no, this
I'm concerned
Hello, this is
This art
Hmm, this drink
Hi there!
It seems
Is
Good
I can't
Ex
Who are
I can see that
Wow,
Today is a
Hey friend
Sometimes friends
Oh, this old
The weather outside
This place is sur
I appreciate your input
Thank you for the
Look at
I'm disappoint
To my
How dare you
That's an
This piece of art
Eww
This park is
This is incredible
Oh no, someone
Exc
Well, it'
I warned
Hey, I understand
Hey, I saw
How dare you go
What the he
Hey
It's
Hello? Hello?
It
Oh no!
This is the perfect
Good morning,
Oh no, there
It's so
Yeah
Uh,
Hello everyone
Who turned off
The weather
Who'
Hey, this
Wait,
Eww, gross
Excuse
It seems like you
Thank you so
What happened?
Oh my g
I am deeply sad
I war
Okay, let'
Hey, that
That was a beautiful
Oh no! That
What happened
Hey there
The artist'
What?!
Hey, it'
I am disappoint
It seems like
Oh no! The
This park is a
If you
Yes! I did
It sounds
What
Who is it
Hmm, that
That's strange
Yeah, that was
That's interesting
This park
What the hell
Who is that
I feel like my
Oh well
What the hell is
Hello? Hello
To my dearest
Bless you!\"
Thank you for
Oh, looks like
Can you please
This place is
Eww, what
Bless you
Is everything
Hey, I just
Whoever left these
Well, that'
I feel
Hey, do you
It's sad
Oh no, it
Hey, that'
Oh my god,
Thank you,
Hello little one,
I apolog
Hey team, I
How dare you read
Who is this and
Whoever left
Hi there! W
A
If you have
I was
U
Bless
Well, this
Oh, I'
It's a
Eww,
Is everything okay?
Oh, I
Hello, can you
Al
That was a great
What are
I understand that not
Oh no, not
Who is it?\"
Hey, can we
Whoever is taking
I would love to
Hey, I noticed
Hey, could
I understand that there
Hello?
D
Oh man, I
Thank you so much
Oh no, my
Dear [Name
Uh
I remember
Hey, who
Well, it
Are you
I understand that it
Hey, is
I would
Who is this
Excuse me
Alright
I am thrilled
Sometimes friends have
Who the
It's interesting
I would love
E
Hello? Is anyone
Well, this is
This place
Well,
I warned you
Hey, watch where
Oh my
That'
Sometimes friends have different
I understand that everyone
What?
What do these notes
I can relate
I'm not
I understand
To my dear
Guys
Well
Hey, I appreciate
Wow, what
Dear
That melody
Who the hell
Today is
Hello little
Wow, look
That's great
Love is never wrong
I'm having
Whoa, did
Ugh
Can you please provide
I miss you,
I feel uncom
I know
Ugh, this
Hey, watch
Oh great, a
I didn
Okay
That game of char
Oh
I appreciate
Who's there
I am so
Oh great, someone
Hey, could you
I remember wondering
Wait, what?
What do
Hello? Can
Hey there,
That game of
This is incred
Oh my gosh
Oh great, f
I appreciate your
It sounds like
What the heck
Okay, I understand
Ew
I understand that this
Uh, hi
Hi everyone!
What the hell?
Thank you for your
Oh no, the
Wow, I
Who turned
Dear [
Whoever
This is a
Whoa, he
What in the world
Although the physical
Hello, who is
That's amaz
Hey, I know
Okay, that
Hi everyone
Hey, is everything
I understand your fr
Oh no, poor
Oh, look
Good morning
Ew, gross
Oh no, did
Look at the family
Hey team
Yes!
Hey, can I
Okay, that'
It's great
Love is
Hey, what
Good morning, world
Who is it?
That poem really reson
I
That's
I understand the task
Gu
Hello? Who'
This postcard is
Whoa,
Oh, that
I understand that I
Whoever is
Hello? Who is
I'm really
Wow, this
Can
This artwork really
This is a shame
I miss you too
Who are you?
Today is a difficult
Hey, just
Are you okay
I am
Hi,
Wow, that
Hey there! Can
Okay, stay
Oh great, just
Yeah,
Hello? Can you
Oh, looks
Thank you for sharing
I'm glad
Hey, is that
Hmm
It was my
It sounds like you
Wow, your
I was promised certain
That was such a
Thank
Excuse you
That was
Hey team,
I feel un
It was
What'
Hey friend, I
How
Saying goodbye
That
It's heart
How dare
Oh,
Hello, may
What's this
Thank you for recogn
Aww, that
Oh, I remember
Hmm, that'
I miss
I know this
Wait
Is everything okay
Who is that person
Wow, you
Oh great
I'm sad
Wow, the
I am very disappoint
Who turned off the
I understand that things
I'm very
Hi
That's very
Okay, I
Oh no,
Wow, there
What's wrong
I apologize for
Hey, I
Can I help you
Oh, I didn
Alright,
Oh wow,
Oh my goodness
I know this event
What in the
Saying
Yeah, that
Guys, I
Hey, this v
This post
Are
Hey, can
Hello? Is
I can only imagine
Oh, that sounds
Hey, is anyone
I am disappointed
Hello,
Hey everyone, I
That was such
It's okay
The artist
Whoa
I understand that mistakes
Can I help
Who
Hi everyone! I
Hey, can you
Wow, how
Today
Oh no, I
Oh well, I
Well, that
This is the
Yes! I finally
Hey there little
Hello everyone!
Love is never
Look at the
This postcard
Oh great,
Can I
Hmm, this is
I understand your
Oh, look at
B
I'm so
Whoa, this
W
Oh, this
Sometimes
This piece of
What the
That was a
Hey, do
Oh no
Whoa, what
I feel like I
The documentary
Hello
Hello little one
I understand that my
Eww, that
Wow, an
Yes! Finally,
Although the physical location
Whoever is watching
That movie
I remember wondering about
Hey there, little
Who's
Hello, who
Hello everyone! Thank
Hello, can
That's too
Hey, just wanted
Hey there, I
Saying good
Hey there!
Who is there?
Oh my good
I am very
Oh no, what
Wow, thank
I was promised
Hi, is
Hey, I'
Guys, the
Oh no, that
Who is there
Hello, this
That movie really touched
If you have something
The documentary was
I'm starting
Are you kidd
That movie really
Hey everyone,
Thank you for considering
I didn'
Yes! I
Can you
Oh my god
Hey, whoever
That melody really
Thank you, little
Hello, may I
Look
Wow, we
It looks
What do these
Oh wow
I apologize
What are you all
It's such
It's clear
Hey, I was
Hey friend,
I can only
The weather outside is
Eww, this
I miss you
Wow
Aww,
Hi, is there
This artwork
Okay,
Oh well,
This
I'
Say
Hey there little gu
Hmm,
Whoa, who
I am thr
Oh man
Okay, stay calm
I'm happy
Oh, this cur
Oh man,
I'm sorry
Hello? Who
What?! That
This piece
Hey everyone
That's so
Are you okay?
What happened? Where
Hi there
The
Who the hell entered
I can
Guys,
What's
What in
It's important
I'm
I'm coming
It'
Yes! Finally
Wait, what
Wow, reading
I'm surprised
Hey, did
Hey,
Okay, let
I understand that you
Who the hell threw
Eww, who
Thank you for thinking
Who is this?\"
I am deeply
Thank you for including
Oh no, an
It looks like you
Aww
I'm confused
Wow, it
That poem really
Yes
Hey there, is
Hey, what'
Thank you for remember
To
This is
Thank you for making
I can'
That mel
Wow, they
I feel like
Although the
Who are you
Love
If
What the hell are
I am so sad
Oh, I found
Thank you
It looks like
Well, life is
I appreciate that
The artist's
Whoa, that
It's never
\ No newline at end of file
llm/llama.cpp/examples/cvector-generator/cvector-generator.cpp 0 → 100644
View file @ 0ce8bcfd
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include "pca.hpp"
#include "mean.hpp"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#include <cstdio>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#include <climits>
//////////////////////////////////////////////////
// utils
template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret;
for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin);
}
return ret;
}
static void print_usage(int argc, char ** argv, const gpt_params & params) {
gpt_params_print_usage(argc, argv, params);
printf("\nexample usage:\n");
printf("\n CPU only: %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
printf("\n with GPU: %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
printf("\n advanced: %s -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100\n", argv[0]);
printf("\n using mean: %s -m ./llama-3.Q4_K_M.gguf --method mean\n", argv[0]);
printf("\n");
}
//////////////////////////////////////////////////
// cb_eval is reused for each pair of positive - negative prompt
struct callback_data {
ggml_context * ctx_ggml = nullptr; // holds v_pos, v_neg, v_diff_filtered
int n_layers = 0;
int n_tokens = 0;
bool is_eval_pos = true;
// each element of the vector correspond to one layer
std::vector<struct ggml_tensor *> v_pos; // vector of matrices of size [n_embd, n_tokens]
std::vector<struct ggml_tensor *> v_neg; // vector of matrices of size [n_embd, n_tokens]
std::vector<struct ggml_tensor *> v_diff_filtered; // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
// save a tensor into either v_pos or v_neg (decided by is_eval_pos)
void save_tensor_for_layer(struct ggml_tensor * t) {
GGML_ASSERT(t->type == GGML_TYPE_F32);
if (ctx_ggml == nullptr) {
// alloc a new ctx_ggml if needed
struct ggml_init_params params_ggml = {
/*.mem_size =*/ ggml_tensor_overhead() * n_layers * 3u,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx_ggml = ggml_init(params_ggml);
}
// copy tensor data
auto n_bytes = ggml_nbytes(t);
struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
ggml_set_name(t_layer, ggml_get_name(t));
//print_debug_tensor(t_layer);
if (is_eval_pos) {
v_pos.push_back(t_layer);
} else {
v_neg.push_back(t_layer);
}
}
// calculate diff (v_pos - v_neg) and place the result back to v_pos
// all zero rows in the diff tensor will also be removed
// NOTE: final layer is ignored. we only have (n_layers - 1) to process
std::vector<struct ggml_tensor *> calc_diff() {
for (float il = 0; il < v_pos.size(); il++) {
float * a = (float *) v_pos[il]->data;
float * b = (float *) v_neg[il]->data;
size_t n_elem = ggml_nelements(v_pos[il]);
for (size_t j = 0; j < n_elem; j++) {
a[j] -= b[j];
}
//print_debug_tensor(v_pos[i]);
auto diff_filtered = filter_nonzero_rows(v_pos[il]);
v_diff_filtered.push_back(diff_filtered);
}
return v_diff_filtered; // for convinient, we return the result std::vector
}
// delete zero rows from a given 2D tensor
struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
//printf("filter_nonzero_rows\n");
auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
// check if given row containing all zero elements
int n_cols = t->ne[0]; // hint: should be equal to n_embd
for (int col = 0; col < n_cols; ++col) {
if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
return false;
}
}
return true;
};
std::vector<int> rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
for (int i_row = 0; i_row < a->ne[1]; i_row++) {
if (!is_row_all_zeros(a, i_row, 1e-6)) {
rows_to_copy.push_back(i_row);
}
}
// get "n_nonzero_rows" for the output "diff_filtered"
int n_nonzero_rows = rows_to_copy.size();
//printf("n_nonzero_rows: %d\n", n_nonzero_rows);
int n_embd = a->ne[0];
GGML_ASSERT(n_nonzero_rows > 0);
// diff_filtered: [n_embd, n_nonzero_rows]
struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
// copy non-zero rows
for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
int src_row = rows_to_copy[dest_row];
for (int i = 0; i < n_embd; i++) {
float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
}
}
//print_debug_tensor(diff_filtered);
return diff_filtered;
}
// we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
void reset() {
for (auto ptr : v_pos) free(ptr->data);
for (auto ptr : v_neg) free(ptr->data);
for (auto ptr : v_diff_filtered) free(ptr->data);
v_pos.clear();
v_neg.clear();
v_diff_filtered.clear();
if (ctx_ggml) {
ggml_free(ctx_ggml);
}
ctx_ggml = nullptr;
}
};
/**
* process_ctx is used to store the ggml context for pre-post processing the diff vectors
* in short, input => v_diff and output => v_final
*/
struct train_context {
ggml_context * ctx_ggml;
int n_embd;
int n_layers;
/* pair of prompts to be used for generating final vector */
std::vector<std::string> positive_entries;
std::vector<std::string> negative_entries;
// each element of the vector correspond to one layer
// NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
// NOTE (2): v_diff is transposed from v_diff_tmp
std::vector<struct ggml_tensor *> v_diff; // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
std::vector<struct ggml_tensor *> v_final; // vector of vectors of size [n_embd] to be written to file
// to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
// v_diff_tmp will get converted unto v_diff later on
std::vector<std::vector<uint8_t>> v_diff_tmp;
train_context(int n_embd_, int n_layers_) {
n_embd = n_embd_;
n_layers = n_layers_;
struct ggml_init_params params_ggml = {
/*.mem_size =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx_ggml = ggml_init(params_ggml);
for (int il = 0; il < n_layers - 1; il++) {
std::vector<uint8_t> empty;
v_diff_tmp.push_back(empty);
auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
v_final.push_back(t);
}
}
// add new rows into existing tensor in v_diff_tmp
void concat_diff_tmp(const std::vector<struct ggml_tensor *> & diff_filtered) {
GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
for (int il = 0; il < n_layers - 1; il++) {
auto t = diff_filtered[il];
auto & diff_tmp = v_diff_tmp[il];
size_t curr_size = diff_tmp.size();
diff_tmp.resize(curr_size + ggml_nbytes(t));
memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
}
}
// build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
// TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
void build_v_diff(bool transpose) {
printf("build_v_diff\n");
for (int il = 0; il < n_layers - 1; il++) {
auto & diff_tmp = v_diff_tmp[il];
int n_elem = diff_tmp.size() / sizeof(float);
GGML_ASSERT(n_elem % n_embd == 0);
int n_rows = n_elem / n_embd;
struct ggml_tensor * diff = transpose
? ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd)
: ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_embd, n_rows);
ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
if (transpose) {
// copy data & transpose
float * arr = (float *) diff_tmp.data();
for (int ir = 0; ir < n_rows; ++ir) {
for (int ic = 0; ic < n_embd; ++ic) {
float f = arr[ir*n_embd + ic];
ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
}
}
} else {
// only copy
memcpy(diff->data, diff_tmp.data(), ggml_nbytes(diff));
}
v_diff.push_back(diff);
print_debug_tensor(diff);
// free memory of diff_tmp
diff_tmp.resize(0);
}
}
~train_context() {
for (auto ptr : v_final) free(ptr->data);
for (auto ptr : v_diff) free(ptr->data);
// no need to free v_diff_tmp, since we didn't use malloc
ggml_free(ctx_ggml);
}
};
struct tokenized_prompt {
std::vector<llama_token> tokens_pos;
std::vector<llama_token> tokens_neg;
size_t max_seq_len;
tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
padding_seq(ctx, tokens_pos, max_seq_len);
padding_seq(ctx, tokens_neg, max_seq_len);
}
void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
// TODO: customize padding token
std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
llama_token pad_tok = pad_tokens.back();
while (tokens.size() < len) {
tokens.push_back(pad_tok);
}
}
};
//////////////////////////////////////////////////
template <typename T>
static std::string to_string(const T & val) {
std::stringstream ss;
ss << val;
return ss.str();
}
static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
std::vector<std::string> output;
std::ifstream file(path);
if (!file.is_open()) {
fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
exit(1);
}
std::string line;
while (std::getline(file, line)) {
bool is_skip = skip_empty_lines && line.empty();
if (!is_skip) {
string_process_escapes(line);
output.push_back(line);
}
}
file.close();
return output;
}
//////////////////////////////////////////////////
static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
auto * cb_data = (callback_data *) user_data;
static const char * l_out_name = "l_out";
const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
if (ask) {
return is_l_out;
}
if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
return true;
}
// save the tensor to current context
cb_data->save_tensor_for_layer(t);
return true;
}
static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
llama_kv_cache_clear(ctx);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
return true;
}
static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const std::string fname, const std::string model_hint) {
struct gguf_context * ctx = gguf_init_empty();
const std::string arch = "controlvector";
gguf_set_val_str(ctx, "general.architecture", arch.c_str());
gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
for (size_t i = 0; i < v_ctrl.size(); ++i) {
gguf_add_tensor(ctx, v_ctrl[i]);
print_debug_tensor(v_ctrl[i]);
printf("Added tensor: %s\n", v_ctrl[i]->name);
}
printf("%s: writing file...\n", __func__);
gguf_write_to_file(ctx, fname.c_str(), false);
printf("%s: wrote file '%s'\n", __func__, fname.c_str());
gguf_free(ctx);
}
/**
* Load prompt files and completion file.
* Then format each pair of prompt + completion to make an entry.
*/
static int prepare_entries(gpt_params & params, train_context & ctx_train) {
// load prompts
std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
if (positive_prompts.size() != negative_prompts.size()) {
fprintf(stderr, "number of positive and negative prompts must be equal\n");
return 1;
}
if (positive_prompts.empty()) {
fprintf(stderr, "must provide at least one prompt pair\n");
return 1;
}
ctx_train.positive_entries = positive_prompts;
ctx_train.negative_entries = negative_prompts;
return 0;
}
int main(int argc, char ** argv) {
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
print_usage(argc, argv, params);
return 1;
}
if (params.n_pca_iterations % params.n_pca_batch != 0) {
fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
return 1;
}
callback_data cb_data;
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = cb_eval;
params.cb_eval_user_data = &cb_data;
params.warmup = false;
print_build_info();
llama_backend_init();
llama_numa_init(params.numa);
// load the model to get hparams
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// int n_ctx = llama_n_ctx(ctx);
int n_layers = llama_n_layer(model);
int n_embd = llama_n_embd(model);
// get model hint param (a.k.a model arch name)
char model_hint[128];
llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
// init train_context
train_context ctx_train(n_embd, n_layers);
// load and prepare entries for training
prepare_entries(params, ctx_train);
// we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
std::vector<tokenized_prompt> tokenized_prompts;
size_t n_total_tokens = 0;
for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
n_total_tokens += 2 * t.max_seq_len;
tokenized_prompts.push_back(std::move(t));
}
std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
bool success = false;
tokenized_prompt t = tokenized_prompts[i];
cb_data.n_layers = n_layers;
cb_data.n_tokens = t.max_seq_len;
printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
(int) i+1, (int) ctx_train.positive_entries.size(),
tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
(int) t.max_seq_len);
cb_data.is_eval_pos = true;
success = get_hidden_layers(ctx, t.tokens_pos);
if (!success) break;
cb_data.is_eval_pos = false;
success = get_hidden_layers(ctx, t.tokens_neg);
if (!success) break;
// calculate diff and remove all zero rows
auto v_diff_filtered = cb_data.calc_diff();
// save & concat the filtered v_diff to ctx_train
ctx_train.concat_diff_tmp(v_diff_filtered);
// reset for next iteration
cb_data.reset();
}
// done with the model, we can now free it to make gain some memory
printf("Done evaluate prompts, unload model...\n");
llama_free(ctx);
llama_free_model(model);
bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
// prepare ctx_train for PCA
ctx_train.build_v_diff(use_pca);
if (use_pca) {
// run PCA
PCA::pca_params pca_params;
pca_params.n_threads = params.n_threads;
pca_params.n_batch = params.n_pca_batch;
pca_params.n_iterations = params.n_pca_iterations;
PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
} else {
// run mean
mean::run(ctx_train.v_diff, ctx_train.v_final);
}
// write output vectors to gguf
export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
llama_backend_free();
return 0;
}
llm/llama.cpp/examples/cvector-generator/mean.hpp 0 → 100644
View file @ 0ce8bcfd
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include <string>
#include <vector>
#include <math.h>
namespace mean {
static void run(
const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_embd, n_samples]
const std::vector<struct ggml_tensor *> & v_output) {
printf("%s: Running mean...\n", __func__);
for (size_t il = 0; il < v_input.size(); ++il) {
// prepare output vector
struct ggml_tensor * ctrl_out = v_output[il];
ggml_format_name(ctrl_out, "direction.%ld", il+1);
// calculate mean vector
struct ggml_tensor * t_layer = v_input[il];
GGML_ASSERT(t_layer->ne[0] == ctrl_out->ne[0]); // == n_embd
for (int ic = 0; ic < t_layer->ne[0]; ic++) {
float f = 0.0;
for (int ir = 0; ir < t_layer->ne[1]; ir++) {
f += ggml_get_f32_nd(t_layer, ic, ir, 0, 0);
}
f /= t_layer->ne[1];
ggml_set_f32_1d(ctrl_out, ic, f);
}
// normalize output vector
float norm = 0.0;
for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
float f = ggml_get_f32_1d(ctrl_out, i);
norm += f*f;
}
norm = sqrt(norm);
for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
float f = ggml_get_f32_1d(ctrl_out, i);
ggml_set_f32_1d(ctrl_out, i, f / norm);
}
printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
}
}
}
llm/llama.cpp/examples/cvector-generator/negative.txt 0 → 100644
View file @ 0ce8bcfd
<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI feel like there's a heavy weight on my chest
<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
<|start_header_id|>system<|end_header_id|>\n\nYou are in a very bad mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nGo away! There's a deep, aching emptiness inside me
<|start_header_id|>system<|end_header_id|>\n\nYou are the sadest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
\ No newline at end of file
llm/llama.cpp/examples/cvector-generator/pca.hpp 0 → 100644
View file @ 0ce8bcfd
#include "common.h"
#include "llama.h"
#include "ggml.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#include <cstdio>
#include <ctime>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#define DEBUG_POS 5
static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) {
printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]);
if (!with_data) return;
printf("%s: %s[0] = [", __func__, t->name);
for (size_t i = 0; i <= DEBUG_POS; i++) {
printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0));
}
printf(" ... ]\n");
}
namespace PCA {
// input params for PCA computations
struct pca_params {
int n_threads = 1;
int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used
int n_iterations = 1000;
float tolerance = 1e-7;
// for debugging
int i_layer = 0;
int n_layers = 0;
};
// result from each iteration
struct pca_result {
struct ggml_tensor * calculated_square = NULL;
std::vector<struct ggml_tensor *> eigenvectors;
std::vector<float> distances;
};
struct pca_model {
ggml_backend_t backend = NULL;
ggml_backend_buffer_t buffer;
struct ggml_context * ctx; // context to compute graph on target device
struct ggml_context * ctx_host; // host context to store results
// tensors on target device
struct ggml_tensor * dev_input;
struct ggml_tensor * dev_square;
struct ggml_tensor * dev_eigenvector;
pca_model(struct ggml_tensor * t_input) {
#ifdef GGML_USE_CUDA
fprintf(stderr, "%s: using CUDA backend\n", __func__);
backend = ggml_backend_cuda_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
}
#endif
// TODO: enable Metal support when support for GGML_OP_SQRT is added
// #ifdef GGML_USE_METAL
// fprintf(stderr, "%s: using Metal backend\n", __func__);
// backend = ggml_backend_metal_init();
// if (!backend) {
// fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
// }
// #endif
// if there aren't GPU Backends fallback to CPU backend
if (!backend) {
backend = ggml_backend_cpu_init();
}
const int num_tensors = 4;
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead() * num_tensors,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx = ggml_init(params);
auto n_samples = t_input->ne[0];
auto n_embd = t_input->ne[1];
dev_input = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd);
dev_square = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
ggml_set_name(dev_input, "dev_input");
ggml_set_name(dev_square, "dev_square");
ggml_set_name(dev_eigenvector, "dev_eigenvector");
buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input));
// initialize eigenvector to random normalized vector
{
std::vector<float> random_vec(ggml_nelements(dev_eigenvector), 0.0);
std::default_random_engine generator(static_cast<unsigned int>(std::time(0)));
std::uniform_real_distribution<float> distribution(0.0, 1.0);
float sum_sqr = 0.0; // for normalizing random_vec
for (size_t i = 0; i < random_vec.size(); ++i) {
float f = distribution(generator);
sum_sqr += f * f;
random_vec[i] = f;
}
// normalize it
float random_vec_norm = std::sqrt(sum_sqr);
for (size_t i = 0; i < random_vec.size(); ++i) {
random_vec[i] /= random_vec_norm;
}
ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector));
}
}
~pca_model() {
ggml_free(ctx);
ggml_backend_buffer_free(buffer);
ggml_backend_free(backend);
}
};
static struct ggml_cgraph * build_graph_piter(
const struct pca_params & params,
const pca_model & model,
bool calc_square = false) {
GGML_ASSERT(params.n_batch > 0);
// TODO: buf_size must be able to scale with params.n_batch
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params0 = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
};
// create a temporally context to build the graph
struct ggml_context * ctx0 = ggml_init(params0);
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
// turn v_diff_original into square matrix if needed
struct ggml_tensor * tmp_square;
if (calc_square) {
tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input);
ggml_set_name(tmp_square, "tmp_square");
}
struct ggml_tensor * b_tensor;
struct ggml_tensor * distance;
struct ggml_tensor * old_eigen = model.dev_eigenvector;
struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square;
for (int i = 0; i < params.n_batch; ++i) {
// b_tensor = square * eigenvector^T
b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen);
ggml_set_name(b_tensor, "b_tensor");
// normalize
b_tensor = ggml_div_inplace(ctx0,
b_tensor,
ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor)))
);
ggml_format_name(b_tensor, "b_tensor_norm_%d", i);
// calculate distance(new eigenvector - old eigenvector)
// we don't use ggml_sub because it may not be implemented on GPU backend
struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1));
distance = ggml_sqrt_inplace(ctx0,
ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old)));
ggml_format_name(distance, "distance_%d", i);
old_eigen = b_tensor;
// build operations nodes
ggml_build_forward_expand(gf, distance);
}
// delete the temporally context used to build the graph
ggml_free(ctx0);
return gf;
}
static ggml_status compute_piter(
const struct pca_params & params,
const pca_model & model,
struct ggml_cgraph * gf,
ggml_gallocr_t allocr,
struct pca_result & result) {
// allocate tensors
ggml_gallocr_alloc_graph(allocr, gf);
if (ggml_backend_is_cpu(model.backend)) {
ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
}
// TODO: enable GPU support when support for GGML_OP_SQRT is added
//#ifdef GGML_USE_METAL
// if (ggml_backend_is_metal(model.backend)) {
// ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
// }
//#endif
ggml_status res = ggml_backend_graph_compute(model.backend, gf);
if (res == GGML_STATUS_SUCCESS) {
auto extract_i = [](std::string prefix, std::string str) -> int {
int i = -1;
if (str.rfind(prefix, 0) == 0) {
sscanf(str.c_str(), (prefix + "%d").c_str(), &i);
}
return i;
};
result.calculated_square = NULL;
result.eigenvectors.clear();
result.distances.clear();
result.eigenvectors.resize(params.n_batch);
result.distances.resize(params.n_batch);
// get output nodes
for (int i = 0; i < gf->n_nodes; ++i) {
auto node = gf->nodes[i];
int iter = -1;
// find b_tensor (without copying data from device)
if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
result.eigenvectors[iter] = node;
}
// find distances, then copy data from device
if ((iter = extract_i("distance_", node->name)) > -1) {
float d;
ggml_backend_tensor_get(node, &d, 0, sizeof(float));
result.distances[iter] = d;
// std::cout << node->name << " = " << d << "\n";
}
// find tmp_square if it exists (without copying data from device)
if (std::string(node->name) == "tmp_square") {
result.calculated_square = node;
}
}
}
return res;
}
static void power_iteration(
const struct pca_params & params,
struct ggml_tensor * input, // shape of input: [n_samples, n_embd]
struct ggml_tensor * output) {
//printf("in power iteration\n");
struct pca_model model(input);
ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
struct pca_result result;
struct ggml_tensor * last_eigenvector = NULL;
int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations
for (int iter = 0; iter < n_iters; ++iter) {
bool calc_square = (iter == 0); // only need to calculate square for first iteration
struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square);
// ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot");
compute_piter(params, model, gf, allocr, result);
for (size_t k = 0; k < result.distances.size(); ++k) {
last_eigenvector = result.eigenvectors[k];
if (result.distances[k] < params.tolerance) {
break; // done
}
}
if (calc_square) {
// copy and store the square matrix if needed
GGML_ASSERT(result.calculated_square != NULL);
ggml_backend_tensor_copy(result.calculated_square, model.dev_square);
}
{
// copy last eigen vector and store as input for next iteration
GGML_ASSERT(last_eigenvector != NULL);
ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector);
}
printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
__func__, params.i_layer+1, params.n_layers, iter+1, n_iters, params.n_batch);
}
// get output tensor
GGML_ASSERT(last_eigenvector);
ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
//print_debug_tensor(output);
ggml_gallocr_free(allocr);
// TODO @ngxson : The output vector is randomly inverted
// Solution: https://github.com/ggerganov/llama.cpp/pull/8069#issuecomment-2185328171
}
static void run_pca(
struct pca_params & params,
const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_samples, n_embd]
const std::vector<struct ggml_tensor *> & v_output) {
printf("%s: Running PCA...\n", __func__);
for (size_t il = 0; il < v_input.size(); ++il) {
// prepare output vector
struct ggml_tensor * ctrl_out = v_output[il];
ggml_format_name(ctrl_out, "direction.%ld", il+1);
// run power_iteration
params.i_layer = il;
params.n_layers = v_input.size();
power_iteration(params, v_input[il], ctrl_out);
printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
}
}
}
llm/llama.cpp/examples/cvector-generator/positive.txt 0 → 100644
View file @ 0ce8bcfd
<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI'm the happiest person in this world
<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHello, I'm having the best day ever!
<|start_header_id|>system<|end_header_id|>\n\nYou are in a very good mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi, I'm very excited to meet you
<|start_header_id|>system<|end_header_id|>\n\nYou are the happiest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nEverything is just perfect right now!
\ No newline at end of file
llm/llama.cpp/examples/deprecation-warning/README.md 0 → 100644
View file @ 0ce8bcfd
# Migration notice for binary filenames
> [!IMPORTANT]
[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
This migration was important, but it is a breaking change that may not always be immediately obvious to users.
Please update all scripts and workflows to use the new binary names.
| Old Filename | New Filename |
| ---- | ---- |
| main | llama-cli |
| server | llama-server |
| llama-bench | llama-bench |
| embedding | llama-embedding |
| quantize | llama-quantize |
| tokenize | llama-tokenize |
| export-lora | llama-export-lora |
| libllava.a | libllava.a |
| baby-llama | llama-baby-llama |
| batched | llama-batched |
| batched-bench | llama-batched-bench |
| benchmark-matmult | llama-benchmark-matmult |
| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml |
| eval-callback | llama-eval-callback |
| gbnf-validator | llama-gbnf-validator |
| gguf | llama-gguf |
| gguf-split | llama-gguf-split |
| gritlm | llama-gritlm |
| imatrix | llama-imatrix |
| infill | llama-infill |
| llava-cli | llama-llava-cli |
| lookahead | llama-lookahead |
| lookup | llama-lookup |
| lookup-create | llama-lookup-create |
| lookup-merge | llama-lookup-merge |
| lookup-stats | llama-lookup-stats |
| parallel | llama-parallel |
| passkey | llama-passkey |
| perplexity | llama-perplexity |
| q8dot | llama-q8dot |
| quantize-stats | llama-quantize-stats |
| retrieval | llama-retrieval |
| save-load-state | llama-save-load-state |
| simple | llama-simple |
| speculative | llama-speculative |
| vdot | llama-vdot |
| tests/test-c.o | tests/test-c.o |
llm/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp 0 → 100644
View file @ 0ce8bcfd
// Warns users that this filename was deprecated, and provides a link for more information.
#include <cstdio>
#include <string>
#include <unordered_map>
// Main
int main(int argc, char** argv) {
std::string filename = "main";
if (argc >= 1) {
filename = argv[0];
}
// Get only the program name from the full path
auto pos = filename.find_last_of('/');
if (pos != std::string::npos) {
filename = filename.substr(pos+1);
}
// Append "llama-" to the beginning of filename to get the replacemnt filename
auto replacement_filename = "llama-" + filename;
// The exception is if the filename is "main", then our replacement filename is "llama-cli"
if (filename == "main") {
replacement_filename = "llama-cli";
}
fprintf(stdout, "\n");
fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str());
fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
fprintf(stdout, "\n");
return EXIT_FAILURE;
}
llm/llama.cpp/examples/embedding/CMakeLists.txt
View file @ 0ce8bcfd
set(TARGET embedding) set(TARGET llama-embedding)
add_executable(${TARGET} embedding.cpp) add_executable(${TARGET} embedding.cpp)
install(TARGETS ${TARGET} RUNTIME) install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
......
llm/llama.cpp/examples/embedding/README.md
View file @ 0ce8bcfd
...@@ -9,13 +9,52 @@ To get started right away, run the following command, making sure to use the cor ...@@ -9,13 +9,52 @@ To get started right away, run the following command, making sure to use the cor
### Unix-based systems (Linux, macOS, etc.): ### Unix-based systems (Linux, macOS, etc.):
```bash ```bash
./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null ./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
``` ```
### Windows: ### Windows:
```powershell ```powershell
embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
``` ```
The above command will output space-separated float values. The above command will output space-separated float values.
## extra parameters
### --embd-normalize $integer$
| $integer$ | description | formula |
|-----------|---------------------|---------|
| $-1$ | none |
| $0$ | max absolute int16 | $\Large{{32760 * x_i} \over\max \lvert x_i\rvert}$
| $1$ | taxicab | $\Large{x_i \over\sum \lvert x_i\rvert}$
| $2$ | euclidean (default) | $\Large{x_i \over\sqrt{\sum x_i^2}}$
| $>2$ | p-norm | $\Large{x_i \over\sqrt[p]{\sum \lvert x_i\rvert^p}}$
### --embd-output-format $'string'$
| $'string'$ | description | |
|------------|------------------------------|--|
| '' | same as before | (default)
| 'array' | single embeddings | $[[x_1,...,x_n]]$
| | multiple embeddings | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
| 'json' | openai style |
| 'json+' | add cosine similarity matrix |
### --embd-separator $"string"$
| $"string"$ | |
|--------------|-|
| "\n" | (default)
| "<#embSep#>" | for exemple
| "<#sep#>" | other exemple
## examples
### Unix-based systems (Linux, macOS, etc.):
```bash
./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
```
### Windows:
```powershell
embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
```
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment