ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: llm-convert support chatglm family (#8643)

Browse source 
* convert chatglm
This commit is contained in:
Zhao Changmin 2023-08-01 11:16:18 +08:00 committed by GitHub
parent a607972c0b
commit 3e10260c6d
6 changed files with 471 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
python/llm/setup.py
View file

@ -49,7 +49,8 @@ llm_home = os.path.join(os.path.dirname(os.path.abspath(__file__)), "src")
github_artifact_dir = os.path.join(llm_home, '../llm-binary')
libs_dir = os.path.join(llm_home, "bigdl", "llm", "libs")
CONVERT_DEP = ['numpy >= 1.22', 'torch',
'transformers', 'sentencepiece', 'accelerate']
'transformers', 'sentencepiece',
'accelerate', 'tabulate']
windows_binarys = [
"llama.dll",
"gptneox.dll",

4
python/llm/src/bigdl/llm/convert_model.py
View file

@ -38,7 +38,7 @@ def llm_convert(model,
This function is able to:
1. Convert Hugging Face llama-like / gpt-neox-like / bloom-like / starcoder-like
PyTorch model to lower precision in BigDL-LLM optimized GGML format.
/ chatglm-like PyTorch model to lower precision in BigDL-LLM optimized GGML format.
2. Convert Hugging Face GPTQ format llama-like model to BigDL-LLM optimized
GGML format.
@ -56,7 +56,7 @@ def llm_convert(model,
:param outfile: Save path of output quantized model. You must pass a **directory** to
save all related output.
:param model_family: Which model family your input model belongs to.
Now ``llama``/``bloom``/``gptneox``/``starcoder`` has been supported.
Now ``llama``/``bloom``/``gptneox``/``starcoder``/``chatglm`` has been supported.
If ``model_format='gptq'``, only ``llama`` is supported.
:param dtype: Which quantized precision will be converted.
If ``model_format='pth'``, `int4` and `int8` are supported,

4
python/llm/src/bigdl/llm/ggml/convert.py
View file

@ -76,6 +76,10 @@ def _convert_starcoder(model_path, outfile_dir, outtype):
_convert_starcoder_hf_to_ggml(model_path, outfile_dir, outtype)
def _convert_chatglm(model_path, outfile_dir, outtype):
_convert_chatglm_hf_to_ggml(model_path, outfile_dir, outtype)
def _convert_to_ggml(model_path: str, outfile_dir: str,
model_family: str = 'llama', outtype: str="fp16"):
"""

13
python/llm/src/bigdl/llm/ggml/convert_model.py
View file

@ -16,7 +16,7 @@
import os
import time
from pathlib import Path
from bigdl.llm.ggml.convert import _convert_to_ggml
from bigdl.llm.ggml.convert import _convert_to_ggml, _convert_chatglm
from bigdl.llm.ggml.quantize import quantize
from bigdl.llm.utils.common import invalidInputError
import argparse
@ -54,9 +54,9 @@ def convert_model(input_path: str,
# make sure directory exists
os.makedirs(output_path, exist_ok=True)
# check input value
invalidInputError(model_family in ['llama', 'bloom', 'gptneox', 'starcoder'],
invalidInputError(model_family in ['llama', 'bloom', 'gptneox', 'starcoder', 'chatglm'],
"Now we only support quantization of model \
family('llama', 'bloom', 'gptneox', 'starcoder')",
family('llama', 'bloom', 'gptneox', 'starcoder', 'chatglm')",
"{} is not in the list.".format(model_family))
invalidInputError(os.path.isdir(output_path),
"The output_path {} was not a directory".format(output_path))
@ -78,6 +78,13 @@ def convert_model(input_path: str,
family('llama', 'gptneox', 'starcoder')",
"{} is not in the list.".format(model_family))
# chatglm merges convertion and quantization into one operation.
if model_family == 'chatglm':
_convert_chatglm(model_path=input_path,
outfile_dir=output_path,
outtype=dtype)
return
if tmp_path is not None:
model_name = Path(input_path).stem
tmp_ggml_file_path = os.path.join(tmp_path, f'{model_name}_{int(time.time())}')

439
python/llm/src/bigdl/llm/utils/convert_chatglm.py Normal file
View file

@ -0,0 +1,439 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# ===========================================================================
#
# This file is adapted from
# https://github.com/li-plus/chatglm.cpp/blob/main/convert.py
#
# MIT License
#
# Copyright (c) 2023 Georgi Gerganov
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""
Convert Hugging Face ChatGLM/ChatGLM2 models to GGML format
"""
import argparse
import platform
import struct
import sys
from enum import Enum
from pathlib import Path
import torch
from tabulate import tabulate
from tqdm import tqdm
from transformers import AutoModel, AutoTokenizer
from bigdl.llm.utils.common import invalidInputError
GGML_QK8_0 = 64
GGML_QK4_0 = 64
GGML_QK4_1 = 64
GGML_QK5_0 = 32
GGML_QK5_1 = 32
GGML_MEM_ALIGN = 16
if platform.system() == "Darwin":
# cpm_kernels doesn't support macOS but transformers will check missing packages, so mock it
sys.modules["cpm_kernels"] = object()
class GGMLType(Enum):
F32 = 0
F16 = 1
Q4_0 = 2
Q4_1 = 3
Q5_0 = 6
Q5_1 = 7
Q8_0 = 8
class ModelType(Enum):
CHATGLM = 1
CHATGLM2 = 2
def quantize_q8_0(tensor: torch.Tensor) -> torch.CharTensor:
# equivalent to ggml_quantize_q8_0 in ggml.c
invalidInputError(tensor.shape[1] % GGML_QK8_0 == 0,
"tensor.shape[1] should be divided by GGML_QK8_0(64), "
f"but get {tensor.shape[1]}")
tensor = tensor.view(-1, GGML_QK8_0)
scale = tensor.abs().max(dim=-1, keepdim=True).values / ((1 << 7) - 1)
tensor = (tensor / scale).round().clamp(min=-128, max=127).char()
# add scale into each block
tensor = torch.cat((scale.half().view(torch.int8), tensor), dim=-1)
return tensor
def quantize_q4_0(tensor: torch.Tensor) -> torch.CharTensor:
# equivalent to ggml_quantize_q4_0 in ggml.c
invalidInputError(tensor.shape[1] % GGML_QK4_1 == 0,
"tensor.shape[1] should be divided by GGML_QK4_1(64), "
f"but get {tensor.shape[1]}")
tensor = tensor.view(-1, GGML_QK4_0)
abs_max_indices = tensor.abs().max(dim=-1, keepdim=True).indices
max_values = torch.take_along_dim(tensor, abs_max_indices, dim=-1)
scale = max_values / -8
tensor = (tensor / scale + 8).round().clamp(min=0, max=15).char()
# compress two int4 weights into an int8
tensor = tensor[:, :GGML_QK4_0//2] | (tensor[:, GGML_QK4_0//2:] << 4)
# add scale into each block
tensor = torch.cat((scale.half().view(torch.int8), tensor), dim=-1)
return tensor
def quantize_q4_1(tensor: torch.Tensor) -> torch.CharTensor:
# equivalent to ggml_quantize_q4_1 in ggml.c
invalidInputError(tensor.shape[1] % GGML_QK4_1 == 0,
"tensor.shape[1] should be divided by GGML_QK4_1(64), "
f"but get {tensor.shape[1]}")
tensor = tensor.view(-1, GGML_QK4_1)
min_vals = tensor.min(dim=-1, keepdim=True).values
max_vals = tensor.max(dim=-1, keepdim=True).values
scale = (max_vals - min_vals) / ((1 << 4) - 1)
tensor = ((tensor - min_vals) / scale).round().clamp(min=0, max=15).char()
# compress two int4 weights into an int8
tensor = tensor[:, :GGML_QK4_1//2] | (tensor[:, GGML_QK4_1//2:] << 4)
# add scale & min into each block
tensor = torch.cat((scale.half().view(torch.int8),
min_vals.half().view(torch.int8), tensor), dim=-1)
return tensor
def quantize_q5_0(tensor: torch.Tensor) -> torch.CharTensor:
# equivalent to ggml_quantize_q5_0 in ggml.c
invalidInputError(tensor.shape[1] % GGML_QK5_0 == 0,
"tensor.shape[1] should be divided by GGML_QK5_1(32), "
f"but get {tensor.shape[1]}")
tensor = tensor.view(-1, GGML_QK5_0)
abs_max_indices = tensor.abs().max(dim=-1, keepdim=True).indices
max_values = torch.take_along_dim(tensor, abs_max_indices, dim=-1)
scale = max_values / -16
tensor = (tensor / scale + 16).round().clamp(min=0, max=31).char()
qs = (tensor[:, :16] & 0x0F) | (tensor[:, 16:] << 4)
qh = torch.zeros(tensor.shape[:-1], dtype=torch.int32)
for i in range(32):
qh |= ((tensor[:, i] & 0x10) >> 4).int() << i
# add scale into each block
tensor = torch.cat((scale.half().view(torch.int8), qh[..., None].view(torch.int8), qs), dim=-1)
return tensor
def quantize_q5_1(tensor: torch.Tensor) -> torch.CharTensor:
# equivalent to ggml_quantize_q5_1 in ggml.c
invalidInputError(tensor.shape[1] % GGML_QK5_1 == 0,
"tensor.shape[1] should be divided by GGML_QK5_1(32), "
f"but get {tensor.shape[1]}")
tensor = tensor.view(-1, GGML_QK5_1)
min_vals = tensor.min(dim=-1, keepdim=True).values
max_vals = tensor.max(dim=-1, keepdim=True).values
scale = (max_vals - min_vals) / ((1 << 5) - 1)
tensor = ((tensor - min_vals) / scale).round().clamp(min=0, max=31).char()
qs = (tensor[:, :16] & 0x0F) | (tensor[:, 16:] << 4)
qh = torch.zeros(tensor.shape[:-1], dtype=torch.int32)
for i in range(32):
qh |= ((tensor[:, i] & 0x10) >> 4).int() << i
# add scale & min into each block
tensor = torch.cat(
(scale.half().view(torch.int8), min_vals.half().view(torch.int8),
qh[..., None].view(torch.int8), qs), dim=-1
)
return tensor
def dump_tensor(f, name: str, tensor: torch.Tensor, ggml_type: GGMLType):
invalidInputError(tensor.dtype == torch.float32,
"tensor.dtype should be torch.float32, "
f"but get {tensor.dtype}")
# tensor name
f.write(struct.pack("i", len(name.encode())))
f.write(name.encode())
# tensor shape & dtype
f.write(struct.pack("i" * (2 + tensor.ndim), tensor.ndim, *tensor.shape, ggml_type.value))
# tensor data
if ggml_type == GGMLType.F32:
tensor = tensor.float()
elif ggml_type == GGMLType.F16:
tensor = tensor.half()
elif ggml_type == GGMLType.Q8_0:
tensor = quantize_q8_0(tensor)
elif ggml_type == GGMLType.Q4_0:
tensor = quantize_q4_0(tensor)
elif ggml_type == GGMLType.Q4_1:
tensor = quantize_q4_1(tensor)
elif ggml_type == GGMLType.Q5_0:
tensor = quantize_q5_0(tensor)
elif ggml_type == GGMLType.Q5_1:
tensor = quantize_q5_1(tensor)
else:
invalidInputError(False, f"Cannot dump tensor of dtype {tensor.dtype}")
# align address
aligned_pos = (f.tell() + (GGML_MEM_ALIGN - 1)) // GGML_MEM_ALIGN * GGML_MEM_ALIGN
f.seek(aligned_pos)
tensor.numpy().tofile(f)
def dump_state_dict(f, weight_names, state_dict, quantization_bit, ggml_type):
tensor_info = []
for name in tqdm(weight_names, desc="Dumping model state"):
tensor = state_dict[name]
if tensor.ndim == 2:
# 2d weight: should quantize it if needed
# step 1: de-quantize it back to float32
if tensor.dtype == torch.int8:
invalidInputError(quantization_bit in [4, 8],
"quantization_bit should be 4 or 8.")
scale = state_dict[f"{name}_scale"].float() # channel-wise scale
if quantization_bit == 4:
# convert int4 weight to int8
low_bits = ((tensor << 4) & 0xF0) >> 4
high_bits = (tensor & 0xF0) >> 4
tensor = torch.stack((high_bits, low_bits), dim=-1).view(tensor.shape[0], -1)
tensor = tensor * scale[:, None]
else:
tensor = tensor.float()
# step 2: quantize it into ggml format
tensor_ggml_type = ggml_type
else:
# 1d weight: convert it to float32
invalidInputError(tensor.ndim == 1,
"tensor.ndim should be 1")
tensor = tensor.float()
tensor_ggml_type = GGMLType.F32
dump_tensor(f, name, tensor, tensor_ggml_type)
tensor_info.append((name, tensor.shape, tensor_ggml_type.name))
print(tabulate(tensor_info, headers=["name", "shape", "dtype"], tablefmt="psql"))
class BaseConverter:
@classmethod
def convert(cls, model, tokenizer, ggml_type, save_path):
# convert all weights to fp16
with open(save_path, "wb") as f:
f.write(b"ggml") # magic
f.write(struct.pack("ii", cls.MODEL_TYPE.value, 1)) # model type & version
cls.dump_config(f, model.config, ggml_type)
cls.dump_tokenizer(f, tokenizer)
cls.dump_model(f, model, ggml_type)
print(f"{cls.MODEL_TYPE.name} GGML model saved to {save_path}")
class ChatGLMConverter(BaseConverter):
MODEL_TYPE = ModelType.CHATGLM
@staticmethod
def dump_config(f, config, ggml_type):
invalidInputError(config.position_encoding_2d,
"unimplemented: position_encoding_2d should be True")
invalidInputError((
config.inner_hidden_size == 4 * config.hidden_size
), "unimplemented: inner_hidden_size should be 4 times hidden_size")
config_values = [
ggml_type.value,
config.vocab_size,
config.hidden_size,
config.num_attention_heads,
config.num_layers,
config.inner_hidden_size,
config.max_sequence_length,
config.bos_token_id if config.bos_token_id is not None else -1,
config.eos_token_id if config.eos_token_id is not None else -1,
config.pad_token_id if config.pad_token_id is not None else -1,
config.sep_token_id if config.sep_token_id is not None else -1,
]
f.write(struct.pack("i" * len(config_values), *config_values))
@staticmethod
def dump_tokenizer(f, tokenizer):
serialized_model_proto = tokenizer.sp_tokenizer.text_tokenizer.sp.serialized_model_proto()
f.write(struct.pack("i", len(serialized_model_proto)))
f.write(serialized_model_proto)
@staticmethod
def dump_model(f, model, ggml_type):
invalidInputError(torch.allclose(
model.state_dict()["transformer.word_embeddings.weight"],
model.state_dict()["lm_head.weight"]
), "unimplemented: lm_head weight must be tied to input embedding")
weight_names = ["transformer.word_embeddings.weight"]
for i in range(model.config.num_layers):
weight_names += [
f"transformer.layers.{i}.input_layernorm.weight",
f"transformer.layers.{i}.input_layernorm.bias",
f"transformer.layers.{i}.attention.query_key_value.weight",
f"transformer.layers.{i}.attention.query_key_value.bias",
f"transformer.layers.{i}.attention.dense.weight",
f"transformer.layers.{i}.attention.dense.bias",
f"transformer.layers.{i}.post_attention_layernorm.weight",
f"transformer.layers.{i}.post_attention_layernorm.bias",
f"transformer.layers.{i}.mlp.dense_h_to_4h.weight",
f"transformer.layers.{i}.mlp.dense_h_to_4h.bias",
f"transformer.layers.{i}.mlp.dense_4h_to_h.weight",
f"transformer.layers.{i}.mlp.dense_4h_to_h.bias",
]
weight_names += [
"transformer.final_layernorm.weight",
"transformer.final_layernorm.bias",
]
dump_state_dict(f, weight_names, model.state_dict(),
model.config.quantization_bit, ggml_type)
class ChatGLM2Converter(BaseConverter):
MODEL_TYPE = ModelType.CHATGLM2
@staticmethod
def dump_config(f, config, ggml_type):
invalidInputError(config.add_bias_linear is False,
"unimplemented: add_bias_linear must be false")
invalidInputError(config.add_qkv_bias is True, "unimplemented: add_qkv_bias must be true")
invalidInputError((
config.apply_residual_connection_post_layernorm is False
), "unimplemented: apply_residual_connection_post_layernorm must be false")
invalidInputError((
config.kv_channels * config.num_attention_heads == config.hidden_size
), "unimplemented: invalid kv_channels")
invalidInputError(config.multi_query_attention is True,
"unimplemented: multi_query_attention must be true")
invalidInputError(config.original_rope is True,
"unimplemented: original_rope must be true")
invalidInputError(config.post_layer_norm is True,
"unimplemented: post_layer_norm must be true")
invalidInputError(config.rmsnorm is True, "unimplemented: rmsnorm must be true")
config_values = [
ggml_type.value,
config.padded_vocab_size,
config.hidden_size,
config.num_attention_heads,
config.num_layers,
config.ffn_hidden_size,
config.seq_length,
config.bos_token_id if config.bos_token_id is not None else -1,
config.eos_token_id if config.eos_token_id is not None else -1,
config.pad_token_id if config.pad_token_id is not None else -1,
config.sep_token_id if config.sep_token_id is not None else -1,
config.multi_query_group_num,
]
f.write(struct.pack("i" * len(config_values), *config_values))
@staticmethod
def dump_tokenizer(f, tokenizer):
serialized_model_proto = tokenizer.tokenizer.sp_model.serialized_model_proto()
f.write(struct.pack("i", len(serialized_model_proto)))
f.write(serialized_model_proto)
@staticmethod
def dump_model(f, model, ggml_type):
weight_names = ["transformer.embedding.word_embeddings.weight"]
for i in range(model.config.num_layers):
weight_names += [
f"transformer.encoder.layers.{i}.input_layernorm.weight",
f"transformer.encoder.layers.{i}.self_attention.query_key_value.weight",
f"transformer.encoder.layers.{i}.self_attention.query_key_value.bias",
f"transformer.encoder.layers.{i}.self_attention.dense.weight",
f"transformer.encoder.layers.{i}.post_attention_layernorm.weight",
f"transformer.encoder.layers.{i}.mlp.dense_h_to_4h.weight",
f"transformer.encoder.layers.{i}.mlp.dense_4h_to_h.weight",
]
weight_names += [
"transformer.encoder.final_layernorm.weight",
"transformer.output_layer.weight",
]
dump_state_dict(f, weight_names, model.state_dict(),
model.config.quantization_bit, ggml_type)
def _convert_chatglm_hf_to_ggml_(model_path, outfile_dir, outtype):
ggml_type = GGMLType[outtype.upper()]
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
if hasattr(model.config, "multi_query_attention"):
ChatGLM2Converter.convert(model, tokenizer, ggml_type, outfile_dir)
else:
ChatGLMConverter.convert(model, tokenizer, ggml_type, outfile_dir)
def main():
parser = argparse.ArgumentParser("chatglm-convert")
parser.add_argument(
"-i",
"--model_name_or_path",
default="THUDM/chatglm-6b",
type=str,
help="Model name or path used in AutoModel.from_pretrained",
)
parser.add_argument(
"-o", "--save_path", default="chatglm-ggml.bin",
type=Path, help="Path to save the generated GGML model"
)
parser.add_argument(
"-t",
"--type",
default="q4_0",
type=str,
choices=["f32", "f16", "q8_0", "q4_0", "q4_1", "q5_0", "q5_1"],
help="GGML model quantization type",
)
args = parser.parse_args()
ggml_type = GGMLType[args.type.upper()]
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, trust_remote_code=True)
model = AutoModel.from_pretrained(args.model_name_or_path, trust_remote_code=True)
if hasattr(model.config, "multi_query_attention"):
ChatGLM2Converter.convert(model, tokenizer, ggml_type, args.save_path)
else:
ChatGLMConverter.convert(model, tokenizer, ggml_type, args.save_path)
if __name__ == "__main__":
main()

15
python/llm/src/bigdl/llm/utils/convert_util.py
View file

@ -88,7 +88,8 @@ __all__ = ['Params',
'default_outfile',
'_convert_gptneox_hf_to_ggml',
'_convert_bloom_hf_to_ggml',
'_convert_starcoder_hf_to_ggml']
'_convert_starcoder_hf_to_ggml',
'_convert_chatglm_hf_to_ggml']
@dataclass(frozen=True)
@ -1586,3 +1587,15 @@ def _convert_starcoder_hf_to_ggml(model_path, outfile_dir, outtype):
data.tofile(fout)
fout.close()
def _convert_chatglm_hf_to_ggml(model_path, outfile_dir, outtype):
filestem = Path(model_path).stem
outfile = os.path.join(outfile_dir, f"ggml-{filestem}-{outtype}.bin")
invalidInputError(outtype in ["q4_0", "q4_1"],
"For now we only support quantization type 'q4_0' and 'q4_1' "
"in chatglm family.")
from bigdl.llm.utils.convert_chatglm import _convert_chatglm_hf_to_ggml_
_convert_chatglm_hf_to_ggml_(model_path,
outfile,
outtype)