llama.cpp/convert-llama-h5-to-gguf.py

318 lines
10 KiB
Python
Raw Normal View History

# HF llama --> gguf conversion, GQA/70b not supported
2023-07-29 09:20:05 +00:00
import gguf
import gguf_tensor_map as tmap
import os
2023-07-29 09:20:05 +00:00
import sys
import struct
import json
import numpy as np
from typing import Any, List
2023-07-29 09:20:05 +00:00
from pathlib import Path
import torch
2023-07-29 09:20:05 +00:00
from sentencepiece import SentencePieceProcessor
#NDArray = np.ndarray[Any, Any]
# compatible with python < 3.9
NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
2023-07-29 10:31:07 +00:00
2023-07-29 09:20:05 +00:00
def permute(weights: NDArray, n_head: int) -> NDArray:
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape))
2023-08-13 11:38:53 +00:00
def count_model_parts(dir_model: str) -> int:
num_parts = 0
for filename in os.listdir(dir_model):
if filename.startswith("pytorch_model-"):
num_parts += 1
if num_parts > 0:
print("gguf: found " + str(num_parts) + " model parts")
return num_parts
2023-07-29 10:31:07 +00:00
2023-08-13 11:38:53 +00:00
2023-07-29 09:20:05 +00:00
if len(sys.argv) < 3:
print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
print(" ftype == 0 -> float32")
print(" ftype == 1 -> float16")
sys.exit(1)
# output in the same directory as the model
dir_model = sys.argv[1]
last_dir = os.path.basename(os.path.normpath(dir_model))
2023-07-29 09:20:05 +00:00
# possible tensor data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if len(sys.argv) > 2:
ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1:
print("Invalid ftype: " + str(ftype))
sys.exit(1)
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
print("gguf: loading model "+last_dir)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
2023-07-29 09:20:05 +00:00
2023-07-31 01:02:00 +00:00
if hparams["architectures"][0] != "LlamaForCausalLM":
print("Model architecture not supported: " + hparams["architectures"][0])
2023-07-31 01:02:00 +00:00
sys.exit()
2023-07-29 09:20:05 +00:00
# get number of model parts
num_parts = count_model_parts(dir_model)
2023-07-29 09:20:05 +00:00
gguf_writer = gguf.GGUFWriter.open(fname_out)
print("gguf: get model metadata")
2023-07-29 09:20:05 +00:00
llm_arch = "llama"
hf_repo = hparams["_name_or_path"]
head_count = hparams["num_attention_heads"]
head_count_kv = hparams["num_key_value_heads"]
block_count = hparams["num_hidden_layers"]
2023-07-29 09:20:05 +00:00
gguf_writer.add_name(last_dir)
gguf_writer.add_architecture(llm_arch)
guff_writer.add_source_hf_repo(hf_repo)
gguf_writer.add_context_length(llm_arch, hparams["max_position_embeddings"])
gguf_writer.add_embedding_length(llm_arch, hparams["hidden_size"])
gguf_writer.add_block_count(llm_arch, block_count)
gguf_writer.add_feed_forward_length(llm_arch, hparams["intermediate_size"])
gguf_writer.add_rope_dimension_count(llm_arch, hparams["hidden_size"] // hparams["num_attention_heads"])
gguf_writer.add_head_count(llm_arch, head_count)
gguf_writer.add_head_count_kv(llm_arch, head_count_kv)
gguf_writer.add_layer_norm_rms_eps(llm_arch, hparams["rms_norm_eps"])
2023-07-29 09:20:05 +00:00
# TOKENIZATION
print("gguf: get tokenizer metadata")
2023-07-29 19:38:01 +00:00
2023-07-29 09:20:05 +00:00
tokens: List[str] = []
scores: List[float] = []
2023-07-29 10:31:07 +00:00
if Path(dir_model + "/tokenizer.model").is_file():
2023-07-29 19:38:01 +00:00
# vocab type sentencepiece
print("gguf: get sentencepiece tokenizer vocab and scores")
2023-07-29 10:31:07 +00:00
tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
2023-07-29 09:20:05 +00:00
for i in range(tokenizer.vocab_size()):
text: bytes
if tokenizer.is_unknown(i):
text = " \u2047 ".encode("utf-8")
elif tokenizer.is_control(i):
text = b""
if tokenizer.is_byte(i):
piece = tokenizer.id_to_piece(i)
if len(piece) != 6:
raise Exception(f"Invalid token: {piece}")
byte_value = int(piece[3:-1], 16)
text = struct.pack("B", byte_value)
else:
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
score: float = tokenizer.get_score(i)
2023-07-29 14:47:00 +00:00
tokens.append(text)
2023-07-29 10:31:07 +00:00
scores.append(score)
2023-07-29 09:20:05 +00:00
gguf_writer.add_tokenizer_model("llama")
gguf_writer.add_token_list(tokens)
gguf_writer.add_token_scores(scores)
if Path(dir_model + "/tokenizer.json").is_file():
with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
tokenizer = json.load(f)
if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
print("gguf: get special token ids")
with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
tokenizer_config = json.load(f)
# find special token ids
if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
for key in tokenizer["added_tokens"]:
if key["content"] == tokenizer_config["bos_token"]["content"]:
gguf_writer.add_bos_token_id(key["id"])
if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
for key in tokenizer["added_tokens"]:
if key["content"] == tokenizer_config["eos_token"]["content"]:
gguf_writer.add_eos_token_id(key["id"])
if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
for key in tokenizer["added_tokens"]:
if key["content"] == tokenizer_config["unk_token"]["content"]:
gguf_writer.add_unk_token_id(key["id"])
if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
for key in tokenizer["added_tokens"]:
if key["content"] == tokenizer_config["sep_token"]["content"]:
gguf_writer.add_sep_token_id(key["id"])
if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
for key in tokenizer["added_tokens"]:
if key["content"] == tokenizer_config["pad_token"]["content"]:
gguf_writer.add_pad_token_id(key["id"])
2023-07-29 09:20:05 +00:00
# TENSORS
tensor_map = tmap.get_tensor_map(block_count)
2023-07-29 09:20:05 +00:00
# tensor info
print("gguf: get tensor metadata")
2023-07-29 09:20:05 +00:00
if num_parts == 0:
part_names = ("pytorch_model.bin",)
else:
part_names = (
f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
)
2023-07-29 09:20:05 +00:00
for part_name in part_names:
2023-08-13 11:38:53 +00:00
print("gguf: loading model part '" + part_name + "'")
model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
2023-07-29 09:20:05 +00:00
for name in model_part.keys():
data = model_part[name]
2023-07-29 09:20:05 +00:00
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
2023-07-29 09:20:05 +00:00
# convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32)
2023-07-29 14:47:00 +00:00
data = data.squeeze().numpy()
# permute these
if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
2023-08-13 11:38:53 +00:00
data = permute(data, head_count)
# map tensor names
if name.endswith(".weight") and name[:-7] in tensor_map:
name = tensor_map[name[:-7]] + ".weight"
elif name.endswith(".bias") and name[:-5] in tensor_map:
name = tensor_map[name[:-5]] + ".bias"
else:
2023-08-13 11:38:53 +00:00
print("Can not map tensor '" + name + "'")
sys.exit()
n_dims = len(data.shape)
2023-08-13 11:38:53 +00:00
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if ftype == 0 and data.dtype == np.float16:
data_dtype = np.float32
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data_dtype = np.float32
# if f16 desired, convert any float32 2-dim weight tensors to float16
if ftype == 1 and data.dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data_dtype = np.float16
data_nbytes = data.size * 2 if data_dtype == np.float16 else data.size * 4
gguf_writer.add_tensor_info(name, data.shape, data_dtype, data_nbytes)
2023-07-29 09:20:05 +00:00
print("gguf: write header")
gguf_writer.write_header_to_file()
print("gguf: write metadata")
gguf_writer.write_kv_data_to_file()
print("gguf: write tensor metadata")
gguf_writer.write_ti_data_to_file()
2023-07-29 09:20:05 +00:00
# tensor data
print("gguf: convert and write tensor data")
2023-07-29 09:20:05 +00:00
if num_parts == 0:
part_names = ("pytorch_model.bin",)
else:
part_names = (
f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
)
2023-07-29 09:20:05 +00:00
for part_name in part_names:
2023-08-13 11:38:53 +00:00
print("gguf: loading model part '" + part_name + "'")
model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
2023-07-29 09:20:05 +00:00
for name in model_part.keys():
data = model_part[name]
2023-07-29 09:20:05 +00:00
2023-08-13 11:38:53 +00:00
<< << << < HEAD
n_dims = len(data.shape)
data_dtype = data.dtype
2023-08-13 11:38:53 +00:00
== == == =
old_dtype = data.dtype
2023-08-13 11:38:53 +00:00
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
>>>>>> > 17800cd80fec468411481dc34a51d42a936442f1
2023-08-13 11:38:53 +00:00
# convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32)
2023-08-13 11:38:53 +00:00
data = data.squeeze().numpy()
2023-08-13 11:38:53 +00:00
# permute these
if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
data = permute(data, head_count)
2023-08-13 11:38:53 +00:00
# map tensor names
if name.endswith(".weight") and name[:-7] in tensor_map:
name = tensor_map[name[:-7]] + ".weight"
elif name.endswith(".bias") and name[:-5] in tensor_map:
name = tensor_map[name[:-5]] + ".bias"
else:
print("Can not map tensor '" + name + "'" )
sys.exit()
2023-08-13 11:38:53 +00:00
n_dims = len(data.shape)
data_dtype = data.dtype
2023-08-13 11:38:53 +00:00
# if f32 desired, convert any float16 to float32
if ftype == 0 and data.dtype == np.float16:
data = data.astype(np.float32)
2023-08-13 11:38:53 +00:00
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
2023-08-13 11:38:53 +00:00
# if f16 desired, convert any float32 2-dim weight tensors to float16
if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
2023-08-13 11:38:53 +00:00
print(name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype))
2023-07-29 09:20:05 +00:00
2023-08-13 11:38:53 +00:00
gguf_writer.write_tensor_to_file(data)
2023-07-29 09:20:05 +00:00
gguf_writer.close()
print("gguf: model successfully exported to '" + fname_out + "'")
2023-07-29 09:20:05 +00:00
print("")