vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'upstream/master' into eval-thread-count

Browse source
This commit is contained in:
ml6 2023-04-16 14:03:18 -07:00
commit c770e0145f
43 changed files with 2683 additions and 1891 deletions
Download patch file Download diff file Expand all files Collapse all files

5
.devops/full.Dockerfile
View file

@ -5,9 +5,10 @@ FROM ubuntu:$UBUNTU_VERSION as build
RUN apt-get update && \
apt-get install -y build-essential python3 python3-pip
COPY requirements.txt requirements.txt
RUN pip install --upgrade pip setuptools wheel \
&& pip install numpy requests sentencepiece tqdm \
&& pip install torch --index-url https://download.pytorch.org/whl/cpu
&& pip install -r requirements.txt
WORKDIR /app

5
.ecrc Normal file
View file

@ -0,0 +1,5 @@
{
"Disable": {
"IndentSize": true
}
}

19
.editorconfig Normal file
View file

@ -0,0 +1,19 @@
# https://EditorConfig.org
# Top-most EditorConfig file
root = true
# Unix-style newlines with a newline ending every file, utf-8 charset
[*]
end_of_line = lf
insert_final_newline = true
trim_trailing_whitespace = true
charset = utf-8
indent_style = space
indent_size = 4
[Makefile]
indent_style = tab
[prompts/*.txt]
insert_final_newline = unset

17
.github/workflows/editorconfig.yml vendored Normal file
View file

@ -0,0 +1,17 @@
name: EditorConfig Checker
on:
push:
branches:
- master
pull_request:
branches:
- master
jobs:
editorconfig:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
- uses: editorconfig-checker/action-editorconfig-checker@main
- run: editorconfig-checker

1
.gitignore vendored
View file

@ -23,6 +23,7 @@ models/*
/result
/perplexity
/embedding
/benchmark-q4_0-matmult
/Pipfile
arm_neon.h

22
CMakeLists.txt
View file

@ -56,6 +56,10 @@ option(LLAMA_AVX "llama: enable AVX"
option(LLAMA_AVX2 "llama: enable AVX2" ON)
option(LLAMA_AVX512 "llama: enable AVX512" OFF)
option(LLAMA_FMA "llama: enable FMA" ON)
# in MSVC F16C is implied with AVX2/AVX512
if (NOT MSVC)
option(LLAMA_F16C "llama: enable F16C" ON)
endif()
# 3rd party libs
option(LLAMA_ACCELERATE "llama: enable Accelerate framework" ON)
@ -116,6 +120,21 @@ if (LLAMA_OPENBLAS)
add_compile_definitions(GGML_USE_OPENBLAS)
add_link_options(${BLAS_LIBRARIES})
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas)
# find header file
set(OPENBLAS_INCLUDE_SEARCH_PATHS
/usr/include
/usr/include/openblas
/usr/include/openblas-base
/usr/local/include
/usr/local/include/openblas
/usr/local/include/openblas-base
/opt/OpenBLAS/include
$ENV{OpenBLAS_HOME}
$ENV{OpenBLAS_HOME}/include
)
find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
add_compile_options(-I${OPENBLAS_INC})
else()
message(WARNING "OpenBLAS not found")
endif()
@ -207,7 +226,9 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
add_compile_options(/arch:AVX)
endif()
else()
if (LLAMA_F16C)
add_compile_options(-mf16c)
endif()
if (LLAMA_FMA)
add_compile_options(-mfma)
endif()
@ -247,7 +268,6 @@ endif()
add_library(llama
llama.cpp
llama.h
llama_internal.h
llama_util.h)
target_include_directories(llama PUBLIC .)

29
Makefile
View file

@ -133,48 +133,53 @@ $(info I CC: $(CCV))
$(info I CXX: $(CXXV))
$(info )
default: main quantize perplexity embedding
default: main quantize quantize-stats perplexity embedding
#
# Build library
#
ggml.o: ggml.c ggml.h
$(CC) $(CFLAGS) -c ggml.c -o ggml.o
$(CC) $(CFLAGS) -c $< -o $@
llama.o: llama.cpp llama.h llama_util.h llama_internal.h
$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
llama.o: llama.cpp ggml.h llama.h llama_util.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common.o: examples/common.cpp examples/common.h
$(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o
$(CXX) $(CXXFLAGS) -c $< -o $@
clean:
rm -vf *.o main quantize quantize-stats perplexity embedding
rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
main: examples/main/main.cpp ggml.o llama.o common.o
$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@echo
@echo '==== Run ./main -h for help. ===='
@echo
quantize: examples/quantize/quantize.cpp ggml.o llama.o
$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp ggml.o llama.o -o quantize $(LDFLAGS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
$(CXX) $(CXXFLAGS) examples/quantize-stats/quantize-stats.cpp ggml.o llama.o -o quantize-stats $(LDFLAGS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
$(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
$(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
libllama.so: llama.o ggml.o
$(CXX) $(CXXFLAGS) -shared -fPIC -o libllama.so llama.o ggml.o $(LDFLAGS)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
#
# Tests
#
benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o
$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
./benchmark-q4_0-matmult
.PHONY: tests
tests:
bash ./tests/run-tests.sh

50
README.md
View file

@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
**Hot topics:**
- [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915)
- [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784)
## Description
@ -48,6 +49,7 @@ New features will probably be added mostly through community contributions.
- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
**UI:**
@ -148,30 +150,52 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
## Usage
Here are the step for the LLaMA-7B model:
Here are the step for the LLaMA-7B model.
### Get the Code
```bash
# build this repo
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make
```
#For Windows and CMake, use the following command instead:
cd <path_to_llama_folder>
mkdir build
cd build
cmake ..
cmake --build . --config Release
### Build
Note: For Windows, CMake or Zig can be used.
1. Use `make`
```bash
make
```
1. Use CMake
```bash
mkdir build
cd build
cmake ..
cmake --build . --config Release
```
1. Use Zig
```bash
zig build -Drelease-fast
```
### Prepare Data & Run
```bash
# obtain the original LLaMA model weights and place them in ./models
ls ./models
65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
# install Python dependencies
python3 -m pip install torch numpy sentencepiece
python3 -m pip install -r requirements.txt
# convert the 7B model to ggml FP16 format
python3 convert-pth-to-ggml.py models/7B/ 1
python3 convert.py models/7B/
# quantize the model to 4-bits (using method 2 = q4_0)
./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin 2
@ -180,8 +204,6 @@ python3 convert-pth-to-ggml.py models/7B/ 1
./main -m ./models/7B/ggml-model-q4_0.bin -n 128
```
Currently, it's best to use Python 3.9 or Python 3.10, as `sentencepiece` has not yet published a wheel for Python 3.11.
When running the larger models, make sure you have enough disk space to store all the intermediate files.
### Memory/Disk Requirements
@ -266,7 +288,7 @@ convert the model from the old format to the new format with [./migrate-ggml-202
- **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.**
- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
- Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
- Please verify the sha256 checksums of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
- Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
- The following command will verify if you have all possible latest files in your self-installed `./models` subdirectory:
`sha256sum --ignore-missing -c SHA256SUMS` on Linux

22
build.zig
View file

@ -1,16 +1,14 @@
const std = @import("std");
pub fn build(b: *std.Build) void {
pub fn build(b: *std.build.Builder) void {
const target = b.standardTargetOptions(.{});
const optimize = b.standardOptimizeOption(.{});
const optimize = b.standardReleaseOptions();
const want_lto = b.option(bool, "lto", "Want -fLTO");
const lib = b.addStaticLibrary(.{
.name = "llama",
.target = target,
.optimize = optimize,
});
const lib = b.addStaticLibrary("llama", null);
lib.want_lto = want_lto;
lib.setTarget(target);
lib.setBuildMode(optimize);
lib.linkLibCpp();
lib.addIncludePath(".");
lib.addIncludePath("examples");
@ -44,16 +42,12 @@ pub fn build(b: *std.Build) void {
fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
const b = args.b;
const lib = args.lib;
const target = args.target;
const optimize = args.optimize;
const want_lto = args.want_lto;
const exe = b.addExecutable(.{
.name = name,
.target = target,
.optimize = optimize,
});
const exe = b.addExecutable(name, null);
exe.want_lto = want_lto;
lib.setTarget(args.target);
lib.setBuildMode(args.optimize);
exe.addIncludePath(".");
exe.addIncludePath("examples");
exe.addCSourceFiles(&.{

299
convert-ggml-to-pth.py
View file

@ -1,299 +0,0 @@
# Author: github.com/ductai199x
import argparse
import os
import struct
import numpy as np
import torch
from numba import njit
from tqdm.auto import tqdm
def read_header(fin):
values = struct.unpack("i" * 9, fin.read(4 * 9))
_, _, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype = values
return {
"vocab_size": vocab_size,
"dim": dim,
"multiple_of": multiple_of,
"n_heads": n_heads,
"n_layers": n_layers,
}, ftype
def read_tokens(fin, vocab_size):
tokens = []
for _ in range(vocab_size):
text_len = struct.unpack("i", fin.read(4))[0]
text_bytes = fin.read(text_len)
try:
text = text_bytes.decode()
except UnicodeDecodeError:
text = text_bytes.decode(errors="replace")
score = struct.unpack("f", fin.read(4))[0]
tokens.append((text, score))
return tokens
@njit
def dequantize_weights_numba(fin_data, n_rows, n_cols):
qk = 32
nb = n_cols // qk
bs = 4 + (qk // 2)
weights = np.zeros((n_rows, n_cols), dtype=np.float32)
data_pos = 0
for row in range(n_rows):
for block in range(nb):
d = np.frombuffer(fin_data[data_pos : data_pos + 4], dtype=np.float32)[0]
data_pos += 4
packed_values = fin_data[data_pos : data_pos + (qk // 2)]
data_pos += qk // 2
for i in range(qk // 2):
packed_value = packed_values[i]
v0 = np.float32((packed_value & 0b00001111) - 8) * d
v1 = np.float32((packed_value >> 4) - 8) * d
weights[row, block * qk + 2 * i] = v0
weights[row, block * qk + 2 * i + 1] = v1
return weights
def dequantize_weights(fin, n_rows, n_cols):
qk = 32
nb = n_cols // qk
data_size = n_rows * n_cols // 2 + n_rows * nb * 4
fin_data = fin.read(data_size)
return dequantize_weights_numba(fin_data, n_rows, n_cols)
def read_variables(fin):
model = {}
pbar = tqdm(total=os.path.getsize(fin.name), unit="B", unit_scale=True, desc="Reading variables")
while True:
start_pos = fin.tell()
try:
n_dims, name_length, ftype_cur = struct.unpack("iii", fin.read(4 * 3))
except struct.error:
break
shape = tuple(struct.unpack("i" * n_dims, fin.read(4 * n_dims)))
shape = shape[::-1]
name = fin.read(name_length).decode()
# ensure tensor data is aligned
tensor_data_offset = fin.tell()
tensor_data_offset = (tensor_data_offset + 31) & -32
fin.seek(tensor_data_offset)
if ftype_cur == 2:
# 4-bit quantized weights
dtype = np.uint8
data = dequantize_weights(fin, shape[0], shape[1])
data = data.reshape(shape)
elif ftype_cur == 0:
dtype = np.float32
data_size = np.prod(shape)
data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
elif ftype_cur == 1:
dtype = np.float16
data_size = np.prod(shape)
data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
model[name] = torch.tensor(data, dtype=torch.float32 if dtype == np.float32 else torch.float16)
pbar.update(fin.tell() - start_pos)
return model
def convert_to_hf_format(model, hparams):
# This works for llama 7B, need to test with other models
n_layers = hparams["n_layers"]
n_heads = hparams["n_heads"]
dim = hparams["dim"]
dims_per_head = dim // n_heads
base = 10000.0
inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
# permute for sliced rotary
def permute(w):
return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim)
state_dict = {}
for layer_i in range(n_layers):
state_dict.update(
{
f"model.layers.{layer_i}.self_attn.q_proj.weight": permute(
model[f"layers.{layer_i}.attention.wq.weight"]
),
f"model.layers.{layer_i}.self_attn.k_proj.weight": permute(
model[f"layers.{layer_i}.attention.wk.weight"]
),
f"model.layers.{layer_i}.self_attn.v_proj.weight": model[
f"layers.{layer_i}.attention.wv.weight"
],
f"model.layers.{layer_i}.self_attn.o_proj.weight": model[
f"layers.{layer_i}.attention.wo.weight"
],
f"model.layers.{layer_i}.mlp.gate_proj.weight": model[
f"layers.{layer_i}.feed_forward.w1.weight"
],
f"model.layers.{layer_i}.mlp.down_proj.weight": model[
f"layers.{layer_i}.feed_forward.w2.weight"
],
f"model.layers.{layer_i}.mlp.up_proj.weight": model[
f"layers.{layer_i}.feed_forward.w3.weight"
],
f"model.layers.{layer_i}.input_layernorm.weight": model[
f"layers.{layer_i}.attention_norm.weight"
],
f"model.layers.{layer_i}.post_attention_layernorm.weight": model[
f"layers.{layer_i}.ffn_norm.weight"
],
}
)
state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
state_dict.update(
{
"model.embed_tokens.weight": model["tok_embeddings.weight"],
"model.norm.weight": model["norm.weight"],
"lm_head.weight": model["output.weight"],
}
)
return state_dict
def chat(model, hparams, llama_dir):
from transformers import (GenerationConfig, LlamaForCausalLM,
LlamaTokenizer, StoppingCriteria,
StoppingCriteriaList)
from transformers.models.llama.configuration_llama import LlamaConfig
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self):
super().__init__()
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops=[]):
print(tokenizer.decode(input_ids[0]), end="", flush=True)
if input_ids[0][-1] == 13:
return True
return False
config = LlamaConfig(
vocab_size=hparams["vocab_size"],
dim=hparams["dim"],
num_hidden_layers=hparams["n_layers"],
num_attention_heads=hparams["n_heads"],
)
llama = LlamaForCausalLM(config=config)
llama.load_state_dict(state_dict=model, strict=True)
tokenizer = LlamaTokenizer.from_pretrained(llama_dir)
device = torch.device("cpu")
llama = llama.to(device)
ctx = """You are AI.
This is a dialog, where User interacts with AI. AI is helpful, kind, obedient, honest, respectful, direct, concise, should try to protect User's privacy, and knows its own limits. Also, AI must answer User and AI cannot stop the conversation by itself.
User: Hello, AI.
AI: Hello! How can I assist you today?
"""
print(ctx.rstrip("\n"))
while True:
print("-" * 60)
prompt = input("User: ")
if ctx != "":
ctx = f"{ctx}User: {prompt}\n"
else:
ctx = f"{prompt}\nAI:"
ctx = (ctx[-1920:]) if len(ctx) >= 2048 else ctx
print("-" * 60)
if len(ctx.strip()) > 0:
input_ids = tokenizer(ctx, return_tensors="pt")["input_ids"].to(device)
generation_config = GenerationConfig(
temperature=0.8,
top_p=0.95,
top_k=50,
repetition_penalty=1.1764,
)
with torch.no_grad():
generation_output = llama.generate(
input_ids=input_ids,
generation_config=generation_config,
return_dict_in_generate=True,
output_scores=True,
max_length=2048,
do_sample=True,
stopping_criteria=StoppingCriteriaList([StoppingCriteriaSub()]),
)
s = generation_output.sequences[0]
decoded = tokenizer.decode(s)
ctx = f"{decoded}\n"
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--input_dir", "-i", type=str, required=True, help="The input directory containing the ggml files."
)
parser.add_argument(
"--prefix",
"-p",
type=str,
required=True,
help="The prefix of the ggml files (ggml-model-f16 or ggml-model-q4_0).",
)
parser.add_argument(
"--hf",
action="store_true",
help="Whether to save the model in the Hugging Face format. (default: False)",
)
parser.add_argument(
"--chat", "-c", action="store_true", help="Whether to open a chat with the model. (default: False)"
)
args = parser.parse_args()
llama_dir = os.path.abspath(f"{args.input_dir}/../")
ggml_files = sorted(
[f"{args.input_dir}/{f}" for f in os.listdir(args.input_dir) if f.startswith(args.prefix)]
)
fin = open(ggml_files[0], "rb")
hparams, ftype = read_header(fin)
tokens = read_tokens(fin, hparams["vocab_size"])
model = read_variables(fin)
for f in tqdm(ggml_files[1:]):
fin = open(f, "rb")
read_header(fin)
read_tokens(fin, hparams["vocab_size"])
model.update(read_variables(fin))
if args.hf:
model = convert_to_hf_format(model, hparams)
pth_ckpt = {
"state_dict": model,
"hparams": hparams,
"tokens": tokens,
}
torch.save(pth_ckpt, f"{args.input_dir}/{args.prefix}-to-torch.pth")
if args.chat:
if not args.hf:
model = convert_to_hf_format(model, hparams)
chat(model, hparams, llama_dir)
if __name__ == "__main__":
main()

107
convert-gpt4all-to-ggml.py
View file

@ -1,107 +0,0 @@
#!/usr/bin/env python3
#
# TODO: deduplicate GPT4All with convert-unversioned-ggml-to-ggml.py
#
# Original by https://github.com/eiz
# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
import argparse
import glob
import os
import struct
import sys
from sentencepiece import SentencePieceProcessor
HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
def parse_args():
parser = argparse.ArgumentParser(description='Upgrade a GPT4All model to the current format')
parser.add_argument('gpt4all_model', help='path to gpt4all-lora-quantized.bin')
parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
return parser.parse_args()
def read_header(f_in):
struct_fmt = "i" * (3 + len(HPARAMS))
struct_size = struct.calcsize(struct_fmt)
buf = f_in.read(struct_size)
return struct.unpack(struct_fmt, buf)
def write_header(f_out, header):
(magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
if magic != 0x67676d6c:
raise Exception('Invalid file magic. Must be an old style ggml file.')
values = [
0x67676d66, # magic: ggml in hex
1, # file version
vocab_size,
dim,
multiple_of,
n_heads,
n_layers,
rot,
ftype
]
f_out.write(struct.pack("i" * len(values), *values))
def write_tokens(fout, tokenizer):
for i in range(tokenizer.vocab_size()):
if tokenizer.is_unknown(i):
text = " \u2047 ".encode()
elif tokenizer.is_control(i):
text = b""
elif tokenizer.is_byte(i):
piece = tokenizer.id_to_piece(i)
if len(piece) != 6:
print(f"Invalid token: {piece}")
sys.exit(1)
byte_value = int(piece[3:-1], 16)
text = struct.pack("B", byte_value)
else:
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
fout.write(struct.pack("i", len(text)))
fout.write(text)
fout.write(struct.pack("f", tokenizer.get_score(i)))
# TODO: GPT4All - add extra <pad> token
text = "<pad>".encode()
fout.write(struct.pack("i", len(text)))
fout.write(text)
fout.write(struct.pack("f", 0.0))
def read_tokens(f_in, tokenizer):
for i in range(tokenizer.vocab_size()):
len_b = f_in.read(4)
(length,) = struct.unpack("i", len_b)
f_in.read(length)
def copy_all_data(f_out, f_in):
while True:
buf = f_in.read(1024 * 1024)
if not buf:
break
f_out.write(buf)
def convert_one_file(path_in, tokenizer):
path_tmp = f"{path_in}.tmp"
path_orig= f"{path_in}.orig"
print(f"converting {path_in}")
with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
write_header(f_out, read_header(f_in))
read_tokens(f_in, tokenizer)
write_tokens(f_out, tokenizer)
copy_all_data(f_out, f_in)
os.rename(path_in, path_orig)
os.rename(path_tmp, path_in)
def main():
args = parse_args()
tokenizer = SentencePieceProcessor(args.tokenizer_model)
convert_one_file(args.gpt4all_model, tokenizer)
if __name__ == "__main__":
main()

172
convert-gptq-to-ggml.py
View file

@ -1,172 +0,0 @@
# Convert a GPTQ quantized LLaMA model to a ggml compatible file
# Based on: https://github.com/qwopqwop200/GPTQ-for-LLaMa
#
import os
import re
import sys
import json
import struct
import numpy as np
import torch
from sentencepiece import SentencePieceProcessor
if len(sys.argv) != 4:
print("Usage: convert-gptq-to-ggml.py llamaXXb-4bit.pt tokenizer.model out.bin\n")
sys.exit(1)
fname_model = sys.argv[1]
fname_tokenizer = sys.argv[2]
dir_out = sys.argv[3]
model = torch.load(fname_model, map_location="cpu")
n_vocab, n_embd = model['model.embed_tokens.weight'].shape
n_layer = 1 + max(int(m.group(1)) for name in model
if (m := re.match(r'model\.layers\.([0-9]+)', name)))
# hardcoded:
n_mult = 256
n_head = {32: 32, 40: 40, 60: 52, 80: 64}[n_layer]
tokenizer = SentencePieceProcessor(fname_tokenizer)
assert tokenizer.vocab_size() == n_vocab
fname_out = sys.argv[3]
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d66)) # magic: ggmf in hex
fout.write(struct.pack("i", 1)) # file version
fout.write(struct.pack("i", n_vocab))
fout.write(struct.pack("i", n_embd))
fout.write(struct.pack("i", n_mult))
fout.write(struct.pack("i", n_head))
fout.write(struct.pack("i", n_layer))
fout.write(struct.pack("i", n_embd // n_head)) # rot (obsolete)
fout.write(struct.pack("i", 4))
# This loop unchanged from convert-pth-to-ggml.py:
for i in range(tokenizer.vocab_size()):
if tokenizer.is_unknown(i):
text = " \u2047 ".encode()
elif tokenizer.is_control(i):
text = b""
elif tokenizer.is_byte(i):
piece = tokenizer.id_to_piece(i)
if len(piece) != 6:
print(f"Invalid token: {piece}")
sys.exit(1)
byte_value = int(piece[3:-1], 16)
text = struct.pack("B", byte_value)
else:
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
fout.write(struct.pack("i", len(text)))
fout.write(text)
fout.write(struct.pack("f", tokenizer.get_score(i)))
def write_header(shape, dst_name, ftype_cur):
sname = dst_name.encode()
fout.write(struct.pack("iii", len(shape), len(sname), ftype_cur))
fout.write(struct.pack("i" * len(shape), *shape[::-1]))
fout.write(sname)
# ensure tensor data is aligned
tensor_data_offset = fout.tell()
tensor_data_offset = (tensor_data_offset + 31) & -32
fout.seek(tensor_data_offset)
def convert_non_q4(src_name, dst_name):
v = model[src_name]
shape = v.shape
print(f"Processing non-Q4 variable: {src_name} with shape: {shape} and type: {v.dtype}")
if len(shape) == 1:
print(" Converting to float32")
v = v.to(torch.float32)
ftype_cur = {torch.float16: 1, torch.float32: 0}[v.dtype]
# header
write_header(shape, dst_name, ftype_cur)
# data
v.numpy().tofile(fout)
def convert_q4(src_name, dst_name, permute=False):
zeros = model[f"{src_name}.zeros"].numpy()
scales = model[f"{src_name}.scales"].numpy()
bias = model[f"{src_name}.bias"].numpy()
qweight = model[f"{src_name}.qweight"].numpy().T # transpose
# Q4_1 does not support bias; good thing the bias is always all zeros.
assert not np.any(bias)
# Each int32 item is actually 8 int4 items packed together, and it's transposed.
shape = (qweight.shape[0], qweight.shape[1] * 8)
print(f"Processing Q4 variable: {src_name} with shape: {shape}")
# The output format has the int4 weights in groups of 32 rather than 8.
# It looks like this:
# For each row:
# For each group of 32 columns:
# - addend (float32, 4 bytes)
# - scale (float32, 4 bytes)
# - weights (int4 * 32, 16 bytes)
# Note that in the input, the scales and addends are shared between all
# the columns in a row, so we end up wasting quite a bit of memory with
# repeated scales and addends.
addends = -zeros # flip sign
# Since the output format is mixed between integers and floats, we have
# to hackily view the floats as int32s just so numpy will let us
# concatenate them.
addends_view = addends.view(dtype=np.int32)
scales_view = scales.view(dtype=np.int32)
# Split into groups of 4 columns (i.e. 32 columns of quantized data):
grouped = qweight.reshape([qweight.shape[0], qweight.shape[1] // 4, 4])
# Repeat addends and scales:
addends_rep = np.atleast_3d(addends_view).repeat(grouped.shape[1], axis=1)
scales_rep = np.atleast_3d(scales_view).repeat(grouped.shape[1], axis=1)
blob = np.concatenate([scales_rep, addends_rep, grouped], axis=2, casting='no')
if permute:
# Permute some rows to undo the permutation done by convert_llama_weights_to_hf.py.
# This can be done after the above conversion because it doesn't affect column order/layout.
blob = (blob.reshape(n_head, 2, shape[0] // n_head // 2, *blob.shape[1:])
.swapaxes(1, 2)
.reshape(blob.shape))
# header
write_header(shape, dst_name, 3) # ftype = Q4_1
# data
blob.tofile(fout)
convert_non_q4("model.embed_tokens.weight", "tok_embeddings.weight")
convert_non_q4("model.norm.weight", "norm.weight")
convert_non_q4("lm_head.weight", "output.weight")
for i in range(n_layer):
convert_q4(f"model.layers.{i}.self_attn.q_proj", f"layers.{i}.attention.wq.weight", permute=True)
convert_q4(f"model.layers.{i}.self_attn.k_proj", f"layers.{i}.attention.wk.weight", permute=True)
convert_q4(f"model.layers.{i}.self_attn.v_proj", f"layers.{i}.attention.wv.weight")
convert_q4(f"model.layers.{i}.self_attn.o_proj", f"layers.{i}.attention.wo.weight")
convert_q4(f"model.layers.{i}.mlp.gate_proj", f"layers.{i}.feed_forward.w1.weight")
convert_q4(f"model.layers.{i}.mlp.down_proj", f"layers.{i}.feed_forward.w2.weight")
convert_q4(f"model.layers.{i}.mlp.up_proj", f"layers.{i}.feed_forward.w3.weight")
convert_non_q4(f"model.layers.{i}.input_layernorm.weight", f"layers.{i}.attention_norm.weight")
convert_non_q4(f"model.layers.{i}.post_attention_layernorm.weight", f"layers.{i}.ffn_norm.weight")
fout.close()
print(f"Done. Output file: {fname_out}")
print()

277
convert-pth-to-ggml.py
View file

@ -1,274 +1,11 @@
# Convert a LLaMA model checkpoint to a ggjt compatible file
#
# Load the model using Torch
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
# - Number of dimensions (int)
# - Name length (int)
# - Dimensions (int[n_dims])
# - Name (char[name_length])
# - Data (float[n_dims])
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#
# Compatibility stub
import argparse
import os
import sys
import json
import struct
import numpy as np
import torch
from sentencepiece import SentencePieceProcessor
import convert
QK = 32
GGML_TYPE_Q4_0 = 0
GGML_TYPE_Q4_1 = 1
GGML_TYPE_I8 = 2
GGML_TYPE_I16 = 3
GGML_TYPE_I32 = 4
GGML_TYPE_F16 = 5
GGML_TYPE_F32 = 6
WTYPES = {
0: GGML_TYPE_F32,
1: GGML_TYPE_F16,
2: GGML_TYPE_Q4_0,
3: GGML_TYPE_Q4_1,
}
GGML_BLCK_SIZE = {
GGML_TYPE_Q4_0: QK,
GGML_TYPE_Q4_1: QK,
GGML_TYPE_I8: 1,
GGML_TYPE_I16: 1,
GGML_TYPE_I32: 1,
GGML_TYPE_F16: 1,
GGML_TYPE_F32: 1,
}
GGML_TYPE_SIZE = {
GGML_TYPE_Q4_0: 4 + QK//2,
GGML_TYPE_Q4_1: 4*2 + QK//2,
GGML_TYPE_I8: 1,
GGML_TYPE_I16: 2,
GGML_TYPE_I32: 4,
GGML_TYPE_F16: 2,
GGML_TYPE_F32: 4,
}
def ggml_nelements(shape):
r = 1
for i in shape:
r *= i
return r
def ggml_nbytes(shape, ftype):
x = ggml_nelements(shape)
t = WTYPES[ftype]
x *= GGML_TYPE_SIZE[t]
x //= GGML_BLCK_SIZE[t]
return x
def parse_args():
parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
parser.add_argument('dir_model', help='directory containing the model checkpoint')
parser.add_argument('ftype', help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
parser.add_argument('vocab_only', help='only write vocab to file', type=int, default=0, nargs='?')
return parser.parse_args()
def get_n_parts(dim):
mappings = {4096: 1, 5120: 2, 6656: 4, 8192: 8}
n_parts = mappings.get(dim)
if n_parts is None:
print(f"Invalid dim: {dim}")
sys.exit(1)
print(f"n_parts = {n_parts}\n")
return n_parts
def load_hparams_and_tokenizer(dir_model):
# `dir_model` is something like `models/7B` or `models/7B/`.
# "tokenizer.model" is expected under model's parent dir.
# When `dir_model` is a symlink, f"{dir_model}/../tokenizer.model" would not be found.
# Let's use the model's parent dir directly.
model_parent_dir = os.path.dirname(os.path.normpath(dir_model))
fname_hparams = f"{dir_model}/params.json"
fname_tokenizer = f"{model_parent_dir}/tokenizer.model"
with open(fname_hparams, "r") as f:
hparams = json.load(f)
print(hparams)
tokenizer = SentencePieceProcessor(fname_tokenizer)
hparams.update({"vocab_size": tokenizer.vocab_size()})
return hparams, tokenizer
def write_header(fout, hparams, ftype):
keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
values = [
0x67676a74, # magic: ggjt in hex
1, # file version
*[hparams[key] for key in keys],
hparams["dim"] // hparams["n_heads"], # rot (obsolete)
ftype
]
fout.write(struct.pack("i" * len(values), *values))
def write_tokens(fout, tokenizer):
for i in range(tokenizer.vocab_size()):
if tokenizer.is_unknown(i):
text = " \u2047 ".encode()
elif tokenizer.is_control(i):
text = b""
elif tokenizer.is_byte(i):
piece = tokenizer.id_to_piece(i)
if len(piece) != 6:
print(f"Invalid token: {piece}")
sys.exit(1)
byte_value = int(piece[3:-1], 16)
text = struct.pack("B", byte_value)
else:
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
fout.write(struct.pack("i", len(text)))
fout.write(text)
fout.write(struct.pack("f", tokenizer.get_score(i)))
def process_and_write_variables(fout, model, ftype, part_id, n_parts):
for name, datao in model.items():
if name.endswith("freqs"):
continue
# remove dimensions with a single element
data = datao.numpy().squeeze()
partshape = data.shape
n_dims = len(data.shape)
assert n_dims in (1, 2)
print(f"Processing variable: {name} with shape: {partshape} and type: {datao.dtype}")
# coerce single-dimensional tensors from float16 to float32
ftype_cur = 1
if ftype == 0 or n_dims == 1:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
blck_size = GGML_BLCK_SIZE[WTYPES[ftype_cur]]
type_size = GGML_TYPE_SIZE[WTYPES[ftype_cur]]
# determine dimension along which multipart tensor is sharded
#
# split_dim 0 regex:
# - output.*
# - layers.*.attention.wq.weight
# - layers.*.attention.wk.weight
# - layers.*.attention.wv.weight
# - layers.*.feed_forward.w1.weight
# - layers.*.feed_forward.w3.weight
#
# split_dim 1 regex:
# - tok_embeddings.*
# - layers.*.attention.wo.weight
# - layers.*.feed_forward.w2.weight
#
if n_dims > 1:
split_dim = 1
if "tok_embeddings" in name:
split_dim = 1
elif "layers" in name:
if "attention.wo.weight" in name:
split_dim = 1
elif "feed_forward.w2.weight" in name:
split_dim = 1
else:
split_dim = 0
elif "output" in name:
split_dim = 0
# output tensor header
fullshape = list(partshape)
if n_dims > 1:
fullshape[split_dim] *= n_parts
sname = name.encode()
fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
for dim in reversed(fullshape):
fout.write(struct.pack("i", dim))
fout.write(sname)
# ensure tensor data is aligned
tensor_data_offset = fout.tell()
while tensor_data_offset % QK != 0:
fout.write(struct.pack("B", 0))
tensor_data_offset += 1
# output unified mappable tensor data
if n_dims == 1 or n_parts == 1:
# copy tensor which we thankfully received in one piece
if part_id == 0:
data.tofile(fout)
elif split_dim == 0:
# reassemble multifile tensor containing some of the rows
rows_per_chunk = partshape[0]
current_row = part_id * rows_per_chunk
bytes_per_row = fullshape[1] // blck_size * type_size
offset = current_row * bytes_per_row
fout.seek(tensor_data_offset + offset)
data.tofile(fout)
elif split_dim == 1:
# reassemble multifile tensor containing some of the cols
cols_per_chunk = partshape[1]
current_col = part_id * cols_per_chunk
bytes_per_row = fullshape[1] // blck_size * type_size
offset_current_col = current_col // blck_size * type_size
for row in range(partshape[0]):
offset_row = row * bytes_per_row
offset = offset_row + offset_current_col
fout.seek(tensor_data_offset + offset)
data[row].tofile(fout)
# advance file position to next tensor
fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype_cur))
def main():
args = parse_args()
dir_model = args.dir_model
ftype = args.ftype
ftype_str = ["f32", "f16"]
hparams, tokenizer = load_hparams_and_tokenizer(dir_model)
print(args)
# if only writing vocab to file
if args.vocab_only:
fname_model = f"{dir_model}/consolidated.00.pth"
fname_out = f"{dir_model}/ggml-vocab.bin"
print(f"Extracting only the vocab from '{fname_model}'\n")
with open(fname_out, "wb") as fout:
write_header(fout, hparams, ftype)
write_tokens(fout, tokenizer)
print(f"Done. Output file: {fname_out}\n")
return
n_parts = get_n_parts(hparams["dim"])
fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin"
# we output a single file for ggml
with open(fname_out, "wb") as fout:
write_header(fout, hparams, ftype)
write_tokens(fout, tokenizer)
offset_of_tensors = fout.tell()
# the tensors we load could be split across multiple files
for part_id in range(n_parts):
fout.seek(offset_of_tensors)
print(f"Processing part {part_id+1} of {n_parts}\n")
fname_model = f"{dir_model}/consolidated.0{part_id}.pth"
model = torch.load(fname_model, map_location="cpu")
process_and_write_variables(fout, model, ftype, part_id, n_parts)
del model
print(f"Done. Output file: {fname_out}\n")
if __name__ == "__main__":
main()
parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
parser.add_argument('dir_model', help='directory containing the model checkpoint')
parser.add_argument('ftype', help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
args = parser.parse_args()
convert.main(['--outtype', 'f16' if args.ftype == 1 else 'f32', '--', args.dir_model])

100
convert-unversioned-ggml-to-ggml.py
View file

@ -1,100 +0,0 @@
#!/usr/bin/env python3
# Original by https://github.com/eiz
# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
import argparse
import glob
import os
import struct
import sys
from sentencepiece import SentencePieceProcessor
HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
def parse_args():
parser = argparse.ArgumentParser(description='Upgrade old ggml model files to the current format')
parser.add_argument('dir_model', help='directory containing ggml .bin files')
parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
return parser.parse_args()
def read_header(f_in):
struct_fmt = "i" * (3 + len(HPARAMS))
struct_size = struct.calcsize(struct_fmt)
buf = f_in.read(struct_size)
return struct.unpack(struct_fmt, buf)
def write_header(f_out, header):
(magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
if magic != 0x67676d6c:
raise Exception('Invalid file magic. Must be an old style ggml file.')
values = [
0x67676d66, # magic: ggml in hex
1, # file version
vocab_size,
dim,
multiple_of,
n_heads,
n_layers,
rot,
ftype
]
f_out.write(struct.pack("i" * len(values), *values))
def write_tokens(fout, tokenizer):
for i in range(tokenizer.vocab_size()):
if tokenizer.is_unknown(i):
text = " \u2047 ".encode()
elif tokenizer.is_control(i):
text = b""
elif tokenizer.is_byte(i):
piece = tokenizer.id_to_piece(i)
if len(piece) != 6:
print(f"Invalid token: {piece}")
sys.exit(1)
byte_value = int(piece[3:-1], 16)
text = struct.pack("B", byte_value)
else:
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
fout.write(struct.pack("i", len(text)))
fout.write(text)
fout.write(struct.pack("f", tokenizer.get_score(i)))
def read_tokens(f_in, tokenizer):
for i in range(tokenizer.vocab_size()):
len_b = f_in.read(4)
(length,) = struct.unpack("i", len_b)
f_in.read(length)
def copy_all_data(f_out, f_in):
while True:
buf = f_in.read(1024 * 1024)
if not buf:
break
f_out.write(buf)
def convert_one_file(path_in, tokenizer):
path_tmp = f"{path_in}.tmp"
path_orig= f"{path_in}.orig"
print(f"converting {path_in}")
with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
write_header(f_out, read_header(f_in))
read_tokens(f_in, tokenizer)
write_tokens(f_out, tokenizer)
copy_all_data(f_out, f_in)
os.rename(path_in, path_orig)
os.rename(path_tmp, path_in)
def main():
args = parse_args()
files = []
files.extend(glob.glob(f"{args.dir_model}/*.bin"))
files.extend(glob.glob(f"{args.dir_model}/*.bin.*"))
tokenizer = SentencePieceProcessor(args.tokenizer_model)
for file in files:
convert_one_file(file, tokenizer)
if __name__ == "__main__":
main()

1148
convert.py Normal file
View file

File diff suppressed because it is too large Load diff

2
examples/alpaca.sh
View file

@ -7,4 +7,4 @@
cd `dirname $0`
cd ..
./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7
./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7

270
examples/benchmark/benchmark-q4_0-matmult.c Normal file
View file

@ -0,0 +1,270 @@
/*
License: MIT License
Changelog:
- 2023-03-31 Initial version by Sebastian Apel (https://github.com/SebastianApel)
*/
#include <locale.h>
#include "ggml.h"
#include <assert.h>
#include <math.h>
#include <cstring>
#include <cstdio>
#include <cinttypes>
#include <unordered_map>
#include <queue>
#include <string.h>
#include <cassert>
#include <fstream>
#include <string>
#include <iterator>
#include <algorithm>
float tensor_sum_elements(struct ggml_tensor * tensor) {
float sum = 0;
if (tensor->type==GGML_TYPE_F32) {
for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) {
sum += ((float *) tensor->data)[j*tensor->ne[0]+k];
}
}
}
return sum;
}
/*
These are mapping to unknown
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
GGML_TYPE_COUNT,
*/
#define TENSOR_TYPE_AS_STR(TYPE) TYPE == GGML_TYPE_F32 ? "FP32" : TYPE == GGML_TYPE_F16 ? "FP16" : TYPE == GGML_TYPE_Q4_0 ? "Q4_0" : TYPE == GGML_TYPE_Q4_1 ? "Q4_1" : "UNKNOWN"
#define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", #TENSOR, \
TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\
TENSOR->ne[0], TENSOR->ne[1], TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \
{ float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); }
struct benchmark_params_struct {
int32_t n_threads = 1;
int32_t n_iterations = 10;
};
void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -i N, --iter N number of iterations to use during computation (default: %d)\n", params.n_iterations);
fprintf(stderr, "\n");
}
int main(int argc, char ** argv) {
struct benchmark_params_struct benchmark_params;
bool invalid_param = false;
std::string arg;
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg == "-t" || arg == "--threads") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_threads = std::stoi(argv[i]);
} else if (arg == "-i" || arg == "--iter") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_iterations = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") {
print_usage(argc, argv, benchmark_params);
exit(0);
}
if (invalid_param) {
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
print_usage(argc, argv, benchmark_params);
exit(1);
}
}
// create the ggml context
printf("Starting Test\n");
struct ggml_context * ctx;
//const int sizex = 4096;
//const int sizey = 11008;
#undef VERBOSE_DEBUGGING
#ifndef VERBOSE_DEBUGGING
const int sizey = 4096;
const int sizex = 11008;
const int sizez = 128;
#else
/* Working - let's increase size */
const int sizey = 1;
const int sizex = (8*32);
const int sizez = 1;
/*const int sizey = 1;
const int sizex = 3*(8*32);
const int sizez = 1;*/
#endif
//printf("Memsize required = %i\n", sizex*sizex);
ggml_type wtype = GGML_TYPE_F32;
size_t ctx_size = 0;
ctx_size += sizex*sizey*ggml_type_sizef(wtype);
ctx_size += sizex*sizey*ggml_type_sizef(wtype);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizeof(float);
ctx_size += 1024*1024*100;
printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024));
struct ggml_init_params params = {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/* no_alloc =*/ 0
};
ctx = ggml_init(params);
if (!ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false;
}
printf("Creating new tensors\n");
// printf("Creating new tensor m1\n");
struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m11, 1.0f);
// printf("Creating new tensor m1\n");
struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m12, 1.5f);
// printf("Creating new tensor m2\n");
struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
ggml_set_f32(m2, 2.0f);
printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
// printf("Creating new tensor m11xm2\n");
struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
// printf("Creating compute graph\n");
struct ggml_cgraph gf = ggml_build_forward(m11xm2);
gf.n_threads=benchmark_params.n_threads;
printf("cgraph->n_threads=%i\n",gf.n_threads);
TENSOR_DUMP(m11);
TENSOR_DUMP(m2);
ggml_graph_compute(ctx, &gf);
TENSOR_DUMP(gf.nodes[0]);
printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
int32_t nelements = sizex*sizey;
int32_t ne[2] = { sizex, sizey };
std::vector<int64_t> hist_cur(1 << 4, 0);
// Set up a the benchmark matrices
// printf("Creating new tensor q11 & Running quantize\n");
struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
// Set up a the compute graph
// printf("Creating new tensor q31\n");
struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
// printf("Creating compute graph\n");
struct ggml_cgraph gf31 = ggml_build_forward(q31);
gf31.n_threads=benchmark_params.n_threads;
// Set up a second graph computation to make sure we override the CPU cache lines
// printf("Creating new tensor q12 & Running quantize\n");
struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
// printf("Creating new tensor q32\n");
struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
//printf("Creating compute graph\n");
struct ggml_cgraph gf32 = ggml_build_forward(q32);
gf32.n_threads=benchmark_params.n_threads;
printf("cgraph->n_threads=%i\n",gf31.n_threads);
const int dimx = sizex;
const int dimy = sizey;
const int dimz = sizez;
long long int flops_per_dot_product = dimy + dimy;
long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
// Let's use the F32 result from above as a reference for the q4_0 multiplication
float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; FLOPS_per_u_Second\n");
printf("==============================================================================================\n");
for (int i=0;i<benchmark_params.n_iterations ;i++) {
long long int start = ggml_time_us();
//printf("Running ggml_graph_compute\n");
ggml_graph_compute(ctx, &gf31);
long long int stop = ggml_time_us();
long long int usec = stop-start;
float sec = usec/1000000;
float flops_per_usec = (1.0f*flops_per_matrix)/usec;
printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%19.2f\n",
i,
gf31.n_threads,
sizex, sizey, sizez, flops_per_matrix,
usec,flops_per_usec);
#ifdef VERBOSE_DEBUGGING
TENSOR_DUMP("res",gf31.nodes[0])
#endif
// Check that the matrix multiplication result is in the right ballpark
// We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; // Let's accept an epsilon of 10^-6
if (delta > allowed_delta) {
printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
sum_of_F32_reference,
sum_of_Q4_result,
delta,
allowed_delta
);
exit(0);
}
// Running a different graph computation to make sure we override the CPU cache lines
ggml_graph_compute(ctx, &gf32);
}
}

6
examples/common.cpp
View file

@ -7,12 +7,6 @@
#include <iterator>
#include <algorithm>
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
#include <alloca.h>
#endif
#if defined (_WIN32)
#include <fcntl.h>
#include <io.h>

2
examples/embedding/embedding.cpp
View file

@ -1,6 +1,8 @@
#include "common.h"
#include "llama.h"
#include <ctime>
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";

2
examples/gpt4all.sh
View file

@ -10,6 +10,6 @@ cd ..
./main --color --instruct --threads 4 \
--model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
--file ./prompts/alpaca.txt \
--batch_size 8 --ctx_size 2048 \
--batch_size 8 --ctx_size 2048 -n -1 \
--repeat_last_n 64 --repeat_penalty 1.3 \
--n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95

6
examples/main/main.cpp
View file

@ -1,3 +1,8 @@
// Defines sigaction on msys:
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include "common.h"
#include "llama.h"
@ -6,6 +11,7 @@
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <string>

34
examples/perplexity/perplexity.cpp
View file

@ -2,6 +2,7 @@
#include "llama.h"
#include <cmath>
#include <ctime>
std::vector<float> softmax(const std::vector<float>& logits) {
std::vector<float> probs(logits.size());
@ -27,21 +28,29 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
int count = 0;
int seq_count = tokens.size() / params.n_ctx;
int n_vocab = llama_n_vocab(ctx);
double nll = 0.0;
fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count);
fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
for (int i = 0; i < seq_count; ++i) {
int start = i * params.n_ctx;
int end = start + params.n_ctx - 1; // TODO: this is not optimal, e.g. it makes the batch 511 instead of 512
// it is better to always be power of 2 for better performance
std::vector<llama_token> embd(tokens.begin() + start, tokens.begin() + end);
int end = start + params.n_ctx;
std::vector<float> logits;
int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch;
auto start_t = std::chrono::high_resolution_clock::now();
if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads, params.n_ethreads)) {
for (int j = 0; j < num_batches; ++j) {
int batch_start = start + j * params.n_batch;
int batch_size = std::min(end - batch_start, params.n_batch);
if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads, params.n_ethreads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return;
}
auto batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
}
auto end_t = std::chrono::high_resolution_clock::now();
if (i == 0) {
const float seconds = std::chrono::duration<float>(end_t - start_t).count();
@ -59,15 +68,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
// Example, we have a context window of 512, we will compute perplexity for each of the
// last 256 tokens. Then, we split the input up into context window size chunks to
// process the entire prompt.
auto logits = llama_get_logits(ctx);
for (int j = params.n_ctx / 2; j < params.n_ctx - 1; ++j) {
for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
// Calculate probability of next token, given the previous ones.
int n_vocab = llama_n_vocab(ctx);
std::vector<float> tok_logits(
logits + j * n_vocab,
logits + (j + 1) * n_vocab);
const float prob = softmax(tok_logits)[tokens[start + j + 1]];
logits.begin() + j * n_vocab,
logits.begin() + (j + 1) * n_vocab);
float prob = softmax(tok_logits)[tokens[start + j + 1]];
nll += -std::log(prob);
++count;
}
@ -82,11 +88,13 @@ int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
params.n_batch = 512;
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
params.perplexity = true;
params.n_batch = std::min(params.n_batch, params.n_ctx);
if (params.n_ctx > 2048) {
fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"

17
examples/quantize-stats/quantize-stats.cpp
View file

@ -1,6 +1,7 @@
#include "ggml.h"
#define LLAMA_API_INTERNAL
#include "llama.h"
#include "llama_internal.h"
#include <algorithm>
#include <cassert>
@ -15,9 +16,6 @@
#include <unordered_map>
#include <vector>
static const char * type_strs[] = { "q4_0", "q4_1", "i8", "i16", "i32", "f16", "f32" };
static_assert(sizeof(type_strs) == GGML_TYPE_COUNT * sizeof(char *), "Incomplete type list");
struct quantize_stats_params {
std::string model = "models/7B/ggml-model-f16.bin";
bool verbose = false;
@ -223,7 +221,7 @@ int main(int argc, char ** argv) {
break;
}
int j;
for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], type_strs[j]) != 0; j++) {
for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) i)) != 0; j++) {
// find match
}
if (j < GGML_TYPE_COUNT) {
@ -278,7 +276,7 @@ int main(int argc, char ** argv) {
continue;
}
if (params.verbose) {
printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), type_strs[kv_tensor.second->type], ggml_nelements(kv_tensor.second));
printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(kv_tensor.second->type), ggml_nelements(kv_tensor.second));
}
if (kv_tensor.second->type == GGML_TYPE_F16) {
is_f16 = true;
@ -303,13 +301,14 @@ int main(int argc, char ** argv) {
// loop throught quantization types
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
const ggml_type type = (ggml_type) i;
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
if (qfns.quantize_row_q && qfns.dequantize_row_q) {
if (params.verbose) {
printf("testing %s ...\n", type_strs[i]);
printf("testing %s ...\n", ggml_type_name(type));
}
error_stats global_stats {};
@ -321,7 +320,7 @@ int main(int argc, char ** argv) {
if (params.verbose) {
printf(" %s ...\n", kv_tensor.first.c_str());
}
std::string layer_name { type_strs[i] };
std::string layer_name { ggml_type_name(type) };
layer_name += "::" + kv_tensor.first;
test_roundtrip_on_layer(
layer_name,
@ -336,7 +335,7 @@ int main(int argc, char ** argv) {
);
}
print_error_stats(type_strs[i], global_stats, params.print_histogram);
print_error_stats(ggml_type_name(type), global_stats, params.print_histogram);
}
}

10
examples/quantize/quantize.cpp
View file

@ -5,15 +5,15 @@
#include <string>
// usage:
// ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
// ./quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
ggml_time_init();
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
fprintf(stderr, " type = 2 - q4_0\n");
fprintf(stderr, " type = 3 - q4_1\n");
fprintf(stderr, " type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0);
fprintf(stderr, " type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1);
return 1;
}
@ -27,7 +27,7 @@ int main(int argc, char ** argv) {
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const int itype = atoi(argv[3]);
const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]);
const int64_t t_main_start_us = ggml_time_us();
@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
{
const int64_t t_start_us = ggml_time_us();
if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype)) {
if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype)) {
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
return 1;
}

7
flake.nix
View file

@ -10,7 +10,6 @@
inherit system;
};
llama-python = pkgs.python310.withPackages (ps: with ps; [
torch
numpy
sentencepiece
]);
@ -28,10 +27,8 @@
];
installPhase = ''
mkdir -p $out/bin
mv bin/main $out/bin/llama
mv bin/quantize $out/bin/quantize
mv bin/embedding $out/bin/embedding
mv bin/perplexity $out/bin/perplexity
mv bin/* $out/bin/
mv $out/bin/main $out/bin/llama
echo "#!${llama-python}/bin/python" > $out/bin/convert-pth-to-ggml
cat ${./convert-pth-to-ggml.py} >> $out/bin/convert-pth-to-ggml

1321
ggml.c
View file

File diff suppressed because it is too large Load diff

32
ggml.h
View file

@ -182,6 +182,7 @@ extern "C" {
#define GGML_MAX_PARAMS 16
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_OPT 4
#define GGML_DEFAULT_N_THREADS 4
#ifdef __ARM_NEON
// we use the built-in 16-bit float type
@ -198,13 +199,15 @@ struct ggml_object;
struct ggml_context;
enum ggml_type {
GGML_TYPE_Q4_0,
GGML_TYPE_Q4_1,
// explicitly numbered values are used in llama.cpp files
GGML_TYPE_F32 = 0,
GGML_TYPE_F16 = 1,
GGML_TYPE_Q4_0 = 2,
GGML_TYPE_Q4_1 = 3,
GGML_TYPE_Q8_0 = 4,
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
GGML_TYPE_F16,
GGML_TYPE_F32,
GGML_TYPE_COUNT,
};
@ -251,6 +254,9 @@ enum ggml_op {
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
GGML_OP_MAP_UNARY,
GGML_OP_MAP_BINARY,
GGML_OP_COUNT,
};
@ -349,6 +355,8 @@ int ggml_blck_size (enum ggml_type type);
size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
const char * ggml_type_name(enum ggml_type type);
size_t ggml_element_size(const struct ggml_tensor * tensor);
struct ggml_context * ggml_init(struct ggml_init_params params);
@ -650,6 +658,21 @@ struct ggml_tensor * ggml_flash_ff(
struct ggml_tensor * c0,
struct ggml_tensor * c1);
// Mapping operations
typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_unary_op_f32_t fun);
struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_binary_op_f32_t fun);
//
// automatic differentiation
//
@ -814,6 +837,7 @@ typedef struct {
dequantize_row_q_t dequantize_row_q;
quantize_row_q_t quantize_row_q;
quantize_row_q_t quantize_row_q_reference;
quantize_row_q_t quantize_row_q_dot;
vec_dot_q_t vec_dot_q;
} quantize_fns_t;

100
llama.cpp
View file

@ -1,10 +1,15 @@
// Defines fileno on msys:
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include "llama_util.h"
#include "llama.h"
#include "llama_internal.h"
#include "ggml.h"
#include <array>
#include <ctime>
#include <cinttypes>
#include <fstream>
#include <random>
@ -77,7 +82,7 @@ struct llama_hparams {
uint32_t n_head = 32;
uint32_t n_layer = 32;
uint32_t n_rot = 64;
uint32_t f16 = 1;
enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
bool operator!=(const llama_hparams & other) const {
return memcmp(this, &other, sizeof(llama_hparams));
@ -257,22 +262,12 @@ static size_t checked_div(size_t a, size_t b) {
}
static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
std::string ret = "[" + std::to_string(ne.at(0));
char buf[256];
snprintf(buf, sizeof(buf), "%5u", ne.at(0));
for (size_t i = 1; i < ne.size(); i++) {
ret += " x " + std::to_string(ne.at(i));
}
ret += "]";
return ret;
}
static const char * llama_format_type(enum ggml_type type) {
switch (type) {
case GGML_TYPE_F32: return "f32";
case GGML_TYPE_F16: return "f16";
case GGML_TYPE_Q4_0: return "q4_0";
case GGML_TYPE_Q4_1: return "q4_1";
default: LLAMA_ASSERT(false);
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), " x %5u", ne.at(i));
}
return buf;
}
static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml_type type) {
@ -427,7 +422,7 @@ struct llama_file_loader {
hparams.n_head = file.read_u32();
hparams.n_layer = file.read_u32();
hparams.n_rot = file.read_u32();
hparams.f16 = file.read_u32();
hparams.ftype = (enum llama_ftype) file.read_u32();
}
void read_vocab() {
vocab.id_to_token.resize(hparams.n_vocab);
@ -453,20 +448,21 @@ struct llama_file_loader {
llama_load_tensor_shard shard;
uint32_t n_dims = file.read_u32();
uint32_t name_len = file.read_u32();
uint32_t ftype = file.read_u32();
shard.type = (enum ggml_type) file.read_u32();
shard.ne.resize(n_dims);
file.read_raw(shard.ne.data(), sizeof(shard.ne[0]) * n_dims);
std::string name = file.read_string(name_len);
if (n_dims < 1 || n_dims > 2) {
throw format("llama.cpp: tensor '%s' should not be %u-dimensional", name.c_str(), n_dims);
}
switch (ftype) {
case 0: shard.type = GGML_TYPE_F32; break;
case 1: shard.type = GGML_TYPE_F16; break;
case 2: shard.type = GGML_TYPE_Q4_0; break;
case 3: shard.type = GGML_TYPE_Q4_1; break;
switch (shard.type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
break;
default: {
throw format("unrecognized ftype %u\n", ftype);
throw format("unrecognized tensor type %u\n", shard.type);
}
}
@ -497,18 +493,18 @@ struct llama_file_loader {
struct llama_file_saver {
llama_file file;
llama_file_loader * any_file_loader;
llama_file_saver(const char * fname, llama_file_loader * any_file_loader, uint32_t new_f16)
llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype)
: file(fname, "wb"), any_file_loader(any_file_loader) {
fprintf(stderr, "llama.cpp: saving model to %s\n", fname);
write_magic();
write_hparams(new_f16);
write_hparams(new_ftype);
write_vocab();
}
void write_magic() {
file.write_u32('ggjt'); // magic
file.write_u32(1); // version
}
void write_hparams(uint32_t new_f16) {
void write_hparams(enum llama_ftype new_ftype) {
const llama_hparams & hparams = any_file_loader->hparams;
file.write_u32(hparams.n_vocab);
file.write_u32(hparams.n_embd);
@ -516,7 +512,7 @@ struct llama_file_saver {
file.write_u32(hparams.n_head);
file.write_u32(hparams.n_layer);
file.write_u32(hparams.n_rot);
file.write_u32(new_f16);
file.write_u32(new_ftype);
}
void write_vocab() {
if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) {
@ -531,17 +527,17 @@ struct llama_file_saver {
}
}
void write_tensor(llama_load_tensor & tensor, enum ggml_type new_type, const void * new_data, size_t new_size) {
uint32_t ftype;
switch (new_type) {
case GGML_TYPE_F32: ftype = 0; break;
case GGML_TYPE_F16: ftype = 1; break;
case GGML_TYPE_Q4_0: ftype = 2; break;
case GGML_TYPE_Q4_1: ftype = 3; break;
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
break;
default: LLAMA_ASSERT(false);
}
file.write_u32((uint32_t) tensor.ne.size());
file.write_u32((uint32_t) tensor.name.size());
file.write_u32(ftype);
file.write_u32(new_type);
file.write_raw(tensor.ne.data(), sizeof(tensor.ne[0]) * tensor.ne.size());
file.write_raw(tensor.name.data(), tensor.name.size());
file.seek(-file.tell() & 31, SEEK_CUR);
@ -815,6 +811,18 @@ static const char *llama_file_version_name(llama_file_version version) {
}
}
static const char *llama_ftype_name(enum llama_ftype ftype) {
switch (ftype) {
case LLAMA_FTYPE_ALL_F32: return "all F32";
case LLAMA_FTYPE_MOSTLY_F16: return "mostly F16";
case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0";
case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
return "mostly Q4_1, some F16";
default: return "unknown, may not work";
}
}
static const char *llama_model_type_name(e_model type) {
switch (type) {
case MODEL_7B: return "7B";
@ -867,7 +875,7 @@ static void llama_model_load_internal(
fprintf(stderr, "%s: n_head = %u\n", __func__, hparams.n_head);
fprintf(stderr, "%s: n_layer = %u\n", __func__, hparams.n_layer);
fprintf(stderr, "%s: n_rot = %u\n", __func__, hparams.n_rot);
fprintf(stderr, "%s: f16 = %u\n", __func__, hparams.f16);
fprintf(stderr, "%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
fprintf(stderr, "%s: n_ff = %u\n", __func__, n_ff);
fprintf(stderr, "%s: n_parts = %zu\n", __func__, ml->file_loaders.size());
fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type));
@ -1541,17 +1549,17 @@ static llama_vocab::id llama_sample_top_p_top_k(
// quantization
//
static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype) {
static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype) {
ggml_type quantized_type;
switch (itype) {
case 2: quantized_type = GGML_TYPE_Q4_0; break;
case 3: quantized_type = GGML_TYPE_Q4_1; break;
default: throw format("invalid quantization type %d\n", itype);
switch (ftype) {
case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break;
default: throw format("invalid output file type %d\n", ftype);
};
std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp.c_str(), /*use_mmap*/ false,
/*vocab_only*/ false));
llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), (uint32_t) itype);
llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype);
size_t total_size_org = 0;
size_t total_size_new = 0;
@ -1564,10 +1572,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
tensor.data = read_data.addr;
model_loader->load_data_for(tensor);
printf("[%zu/%zu] %36s - %s, type = %6s, ",
printf("[%4zu/%4zu] %36s - %16s, type = %6s, ",
++idx, model_loader->tensors_map.tensors.size(),
tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
llama_format_type(tensor.type));
ggml_type_name(tensor.type));
// This used to be a regex, but <regex> has an extreme cost to compile times.
bool quantize = tensor.name.rfind("weight") == tensor.name.size() - 6; // ends with 'weight'?
@ -1600,7 +1608,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data[i] = ggml_fp16_to_fp32(f16_data[i]);
}
} else {
throw format("type %s unsupported for integer quantization", llama_format_type(tensor.type));
throw format("type %s unsupported for integer quantization", ggml_type_name(tensor.type));
}
printf("quantizing .. ");
@ -1742,9 +1750,9 @@ void llama_free(struct llama_context * ctx) {
int llama_model_quantize(
const char * fname_inp,
const char * fname_out,
int itype) {
enum llama_ftype ftype) {
try {
llama_model_quantize_internal(fname_inp, fname_out, itype);
llama_model_quantize_internal(fname_inp, fname_out, ftype);
return 0;
} catch (const std::string & err) {
fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str());

22
llama.h
View file

@ -65,6 +65,15 @@ extern "C" {
void * progress_callback_user_data;
};
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0,
LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
};
LLAMA_API struct llama_context_params llama_context_default_params();
LLAMA_API bool llama_mmap_supported();
@ -85,7 +94,7 @@ extern "C" {
LLAMA_API int llama_model_quantize(
const char * fname_inp,
const char * fname_out,
int itype);
enum llama_ftype ftype);
// Returns the KV cache that will contain the context for the
// ongoing prediction with the model.
@ -171,4 +180,15 @@ extern "C" {
}
#endif
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL
#include <vector>
#include <string>
struct ggml_tensor;
std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif
#endif // LLAMA_H

12
llama_internal.h
View file

@ -1,12 +0,0 @@
// Internal header to be included by llama.cpp and tests/benchmarks only.
#ifndef LLAMA_INTERNAL_H
#define LLAMA_INTERNAL_H
#include <vector>
#include <string>
struct ggml_tensor;
std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif // LLAMA_INTERNAL_H

17
llama_util.h
View file

@ -26,7 +26,9 @@
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <io.h>
#include <stdio.h> // for _fseeki64
@ -41,8 +43,12 @@
} while (0)
#ifdef __GNUC__
#ifdef __MINGW32__
__attribute__((format(gnu_printf, 1, 2)))
#else
__attribute__((format(printf, 1, 2)))
#endif
#endif
static std::string format(const char * fmt, ...) {
va_list ap, ap2;
va_start(ap, fmt);
@ -55,7 +61,7 @@ static std::string format(const char * fmt, ...) {
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
};
}
struct llama_file {
// use FILE * so we don't have to re-open the file to mmap
@ -170,7 +176,6 @@ struct llama_mmap {
flags |= MAP_POPULATE;
#endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
close(fd);
if (addr == MAP_FAILED) {
throw format("mmap failed: %s", strerror(errno));
}
@ -209,6 +214,7 @@ struct llama_mmap {
throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
}
#if _WIN32_WINNT >= _WIN32_WINNT_WIN8
// Advise the kernel to preload the mapped memory
WIN32_MEMORY_RANGE_ENTRY range;
range.VirtualAddress = addr;
@ -217,6 +223,9 @@ struct llama_mmap {
fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
#else
#pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
#endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
}
~llama_mmap() {
@ -338,8 +347,8 @@ struct llama_mlock {
// Hopefully a megabyte is enough overhead:
size_t increment = size + 1048576;
// The minimum must be <= the maximum, so we need to increase both:
min_ws_size += size;
max_ws_size += size;
min_ws_size += increment;
max_ws_size += increment;
if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());

311
migrate-ggml-2023-03-30-pr613.py
View file

@ -1,311 +0,0 @@
# Migrate ggml file(s) with ggmf magic to ggml file with ggjt magic
#
# We caused a breaking change to the file format on 2023-03-30 in:
# https://github.com/ggerganov/llama.cpp/pull/613
#
# (1) If you still have the Meta LLaMA .pth files, then close this
# file now; you can just run `convert-pth-to-ggml.py` again to
# migrate to the new format. The tool is easier to use too. It
# isn't necessary anymore to manage split output files because
# the new format always combines things into a single file.
#
# (2) If you deleted the Meta LLaMA .pth files due to save on disk
# space, then this tool is intended to help you. Please check
# out the instructions below.
#
# USAGE
#
# python migrate-ggml-2023-03-30-pr613.py INPUT OUTPUT
#
# PREREQUISITES
#
# pip install numpy
# cd llama.cpp
# make -j4
#
# EXAMPLE (7B MODEL)
#
# # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
# python migrate-ggml-2023-03-30-pr613.py models/7B/ggml-model-f16.bin models/7B/ggml-model-f16-ggjt.bin
#
# # check that it works
# ./main -m models/7B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
#
# # you can delete the old files
# rm -f models/7B/ggml-model-f16.bin
# mv models/7B/ggml-model-f16-ggjt.bin models/7B/ggml-model-f16.bin
#
# EXAMPLE (13B MODEL)
#
# # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
# python migrate-ggml-2023-03-30-pr613.py models/13B/ggml-model-f16.bin models/13B/ggml-model-f16-ggjt.bin
#
# # check that it works
# ./main -m models/13B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
#
# # you can delete the old files
# rm -f models/13B/ggml-model-f16.bin*
# mv models/13B/ggml-model-f16-ggjt.bin models/13B/ggml-model-f16.bin
#
import argparse
import os
import sys
import json
import struct
import numpy as np
QK = 32
GGML_TYPE_Q4_0 = 0
GGML_TYPE_Q4_1 = 1
GGML_TYPE_I8 = 2
GGML_TYPE_I16 = 3
GGML_TYPE_I32 = 4
GGML_TYPE_F16 = 5
GGML_TYPE_F32 = 6
WTYPE_NAMES = {
0: "F32",
1: "F16",
2: "Q4_0",
3: "Q4_1",
}
WTYPES = {
0: GGML_TYPE_F32,
1: GGML_TYPE_F16,
2: GGML_TYPE_Q4_0,
3: GGML_TYPE_Q4_1,
}
GGML_BLCK_SIZE = {
GGML_TYPE_Q4_0: QK,
GGML_TYPE_Q4_1: QK,
GGML_TYPE_I8: 1,
GGML_TYPE_I16: 1,
GGML_TYPE_I32: 1,
GGML_TYPE_F16: 1,
GGML_TYPE_F32: 1,
}
GGML_TYPE_SIZE = {
GGML_TYPE_Q4_0: 4 + QK//2,
GGML_TYPE_Q4_1: 4*2 + QK//2,
GGML_TYPE_I8: 1,
GGML_TYPE_I16: 2,
GGML_TYPE_I32: 4,
GGML_TYPE_F16: 2,
GGML_TYPE_F32: 4,
}
HPARAMS = [
'magic', # int32
'version', # int32
'n_vocab', # int32
'n_embd', # int32
'n_mult', # int32
'n_head', # int32
'n_layer', # int32
'n_rot', # int32
'f16', # int32
]
def read_hparams(fin):
struct_fmt = "i" * len(HPARAMS)
struct_size = struct.calcsize(struct_fmt)
buf = fin.read(struct_size)
ints = struct.unpack(struct_fmt, buf)
hparams = dict(zip(HPARAMS, ints))
return hparams
def write_hparams(fout, hparams):
struct_fmt = "i" * len(HPARAMS)
struct_size = struct.calcsize(struct_fmt)
ints = [hparams[h] for h in HPARAMS]
fout.write(struct.pack(struct_fmt, *ints))
def read_tokens(fin, hparams):
tokens = []
for i in range(hparams['n_vocab']):
len_b = fin.read(4)
(length,) = struct.unpack("i", len_b)
word = fin.read(length)
score_b = fin.read(4)
(score,) = struct.unpack("f", score_b)
tokens.append((word, score))
return tokens
def write_tokens(fout, tokens):
for word, score in tokens:
fout.write(struct.pack("i", len(word)))
fout.write(word)
fout.write(struct.pack("f", score))
def ggml_nelements(shape):
r = 1
for i in shape:
r *= i
return r
def ggml_nbytes(shape, ftype):
x = ggml_nelements(shape)
t = WTYPES[ftype]
x *= GGML_TYPE_SIZE[t]
x //= GGML_BLCK_SIZE[t]
return x
def copy_tensors(fin, fout, part_id, n_parts):
while True:
b = fin.read(4)
if not b: break
(n_dims,) = struct.unpack("i", b)
b = fin.read(4)
(length,) = struct.unpack("i", b)
b = fin.read(4)
(ftype,) = struct.unpack("i", b)
assert n_dims in (1, 2)
partshape = list(range(n_dims))
for i in range(n_dims):
b = fin.read(4)
partshape[i] = struct.unpack("i", b)[0]
partshape = list(reversed(partshape))
name = fin.read(length)
data = fin.read(ggml_nbytes(partshape, ftype))
blck_size = GGML_BLCK_SIZE[WTYPES[ftype]]
type_size = GGML_TYPE_SIZE[WTYPES[ftype]]
print(f"Processing tensor {name} with shape: {partshape} and type: {WTYPE_NAMES[ftype]}")
# determine dimension along which multipart tensor is sharded
#
# split_dim 0 regex:
# - output.*
# - layers.*.attention.wq.weight
# - layers.*.attention.wk.weight
# - layers.*.attention.wv.weight
# - layers.*.feed_forward.w1.weight
# - layers.*.feed_forward.w3.weight
#
# split_dim 1 regex:
# - tok_embeddings.*
# - layers.*.attention.wo.weight
# - layers.*.feed_forward.w2.weight
#
if n_dims > 1:
split_dim = 1
if b"tok_embeddings" in name:
split_dim = 1
elif b"layers" in name:
if b"attention.wo.weight" in name:
split_dim = 1
elif b"feed_forward.w2.weight" in name:
split_dim = 1
else:
split_dim = 0
elif b"output" in name:
split_dim = 0
# output tensor header
fullshape = list(partshape)
if n_dims > 1:
fullshape[split_dim] *= n_parts
fout.write(struct.pack("iii", n_dims, len(name), ftype))
for dim in reversed(fullshape):
fout.write(struct.pack("i", dim))
fout.write(name)
# ensure tensor data is aligned
tensor_data_offset = fout.tell()
while tensor_data_offset % QK != 0:
fout.write(struct.pack("B", 0))
tensor_data_offset += 1
# output unified mappable tensor data
if n_dims == 1 or n_parts == 1:
# copy tensor which we thankfully received in one piece
if part_id == 0:
fout.write(data)
elif split_dim == 0:
# reassemble multifile tensor containing some of the rows
rows_per_chunk = partshape[0]
current_row = part_id * rows_per_chunk
bytes_per_row = fullshape[1] // blck_size * type_size
offset = current_row * bytes_per_row
fout.seek(tensor_data_offset + offset)
fout.write(data)
elif split_dim == 1:
# reassemble multifile tensor containing some of the cols
cols_per_chunk = partshape[1]
current_col = part_id * cols_per_chunk
bpr = partshape[1] // blck_size * type_size
bytes_per_row = fullshape[1] // blck_size * type_size
offset_current_col = current_col // blck_size * type_size
for row in range(partshape[0]):
offset_row = row * bytes_per_row
offset = offset_row + offset_current_col
fout.seek(tensor_data_offset + offset)
fout.write(data[row * bpr:row * bpr + bpr])
# advance file position to next tensor
fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype))
def parse_args():
parser = argparse.ArgumentParser(description='Migrate from GGML to new GGJT file format')
parser.add_argument('fin_path', help='your old ggml file (leave out the .1 .2 etc.)')
parser.add_argument('fout_path', help='your new ggjt file name')
return parser.parse_args()
def main():
args = parse_args()
assert args.fin_path
assert args.fout_path
assert args.fin_path != args.fout_path
with open(args.fin_path, "rb") as fin:
hparams = read_hparams(fin)
tokens = read_tokens(fin, hparams)
if hparams['magic'] == 0x67676a74: # ggjt
print(f"{args.fin_path}: input ggml has already been converted to 'ggjt' magic\n")
sys.exit(1)
if hparams['magic'] != 0x67676d66: # ggmf
print(f"{args.fin_path}: input ggml file doesn't have expected 'ggmf' magic: {hparams['magic']:#x}\n")
sys.exit(1)
hparams['magic'] = 0x67676a74 # ggjt
# count number of multipart files by convention
n_parts = 1
while True:
if os.path.exists(f"{args.fin_path}.{n_parts}"):
n_parts += 1
else:
break
# we output a single file for ggml
with open(args.fout_path, "wb") as fout:
write_hparams(fout, hparams)
write_tokens(fout, tokens)
offset_of_tensors = fout.tell()
# the tensors we load could be split across multiple files
for part_id in range(n_parts):
fout.seek(offset_of_tensors)
print(f"Processing part {part_id+1} of {n_parts}\n")
fin_path = args.fin_path
if part_id > 0:
fin_path += f".{part_id}"
with open(fin_path, "rb") as fin:
read_tokens(fin, read_hparams(fin))
copy_tensors(fin, fout, part_id, n_parts)
print(f"Done. Output file: {args.fout_path}\n")
if __name__ == "__main__":
main()

2
requirements.txt Normal file
View file

@ -0,0 +1,2 @@
numpy==1.24
sentencepiece==0.1.98