mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-12 09:59:41 +00:00
arranged files, updated kobold lite, modified makefile for extra link args on linux, started RWKV implementation
This commit is contained in:
Concedo
parent
9581171a9f
commit
763ad172c0
21 changed files with 13597 additions and 46 deletions
183
otherarch/tools/convert_hf_gpt2.py
Normal file
183
otherarch/tools/convert_hf_gpt2.py
Normal file
|
|
@ -0,0 +1,183 @@
|
|||
# Convert Cerebras models to ggml format
|
||||
#
|
||||
# ref: https://www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/
|
||||
#
|
||||
|
||||
import sys
|
||||
import struct
|
||||
import json
|
||||
import torch
|
||||
import numpy as np
|
||||
import re
|
||||
|
||||
from transformers import GPTJForCausalLM, AutoModelForCausalLM
|
||||
|
||||
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
if len(sys.argv) < 2:
|
||||
print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
|
||||
sys.exit(1)
|
||||
|
||||
# output in the same directory as the model
|
||||
dir_model = sys.argv[1]
|
||||
fname_out = sys.argv[1] + "/ggml-model-f16.bin"
|
||||
|
||||
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
|
||||
encoder = json.load(f)
|
||||
|
||||
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
# use 16-bit or 32-bit floats
|
||||
use_f16 = True
|
||||
if len(sys.argv) > 2:
|
||||
use_f16 = False
|
||||
fname_out = sys.argv[1] + "/ggml-model-f32.bin"
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
|
||||
#print (model)
|
||||
|
||||
list_vars = model.state_dict()
|
||||
#print (list_vars)
|
||||
|
||||
print(hparams)
|
||||
|
||||
fout = open(fname_out, "wb")
|
||||
|
||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||
fout.write(struct.pack("i", hparams["vocab_size"]))
|
||||
fout.write(struct.pack("i", hparams["n_positions"]))
|
||||
fout.write(struct.pack("i", hparams["n_embd"]))
|
||||
fout.write(struct.pack("i", hparams["n_head"]))
|
||||
fout.write(struct.pack("i", hparams["n_layer"]))
|
||||
fout.write(struct.pack("i", use_f16))
|
||||
|
||||
byte_encoder = bytes_to_unicode()
|
||||
byte_decoder = {v:k for k, v in byte_encoder.items()}
|
||||
|
||||
fout.write(struct.pack("i", len(encoder)))
|
||||
|
||||
for key in encoder:
|
||||
text = bytearray([byte_decoder[c] for c in key])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
for name in list_vars.keys():
|
||||
data = list_vars[name].squeeze().numpy()
|
||||
print("Processing variable: " + name + " with shape: ", data.shape)
|
||||
|
||||
# rename headers to keep compatibility
|
||||
if name == "transformer.ln_f.weight":
|
||||
name = "model/ln_f/g"
|
||||
elif name == "transformer.ln_f.bias":
|
||||
name = "model/ln_f/b"
|
||||
elif name == "transformer.wte.weight":
|
||||
name = "model/wte"
|
||||
elif name == "transformer.wpe.weight":
|
||||
name = "model/wpe"
|
||||
elif name == "lm_head.weight":
|
||||
name = "model/lm_head"
|
||||
elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/g"
|
||||
elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/b"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/w"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/b"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/w"
|
||||
elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/b"
|
||||
elif re.match(r"transformer.h.\d+.ln_2.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/g"
|
||||
elif re.match(r"transformer.h.\d+.ln_2.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/b"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/w"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/b"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/w"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/b"
|
||||
else:
|
||||
print("Unrecognized variable name. %s", name)
|
||||
|
||||
# we don't need these
|
||||
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
|
||||
print(" Skipping variable: " + name)
|
||||
continue
|
||||
|
||||
n_dims = len(data.shape);
|
||||
|
||||
# ftype == 0 -> float32, ftype == 1 -> float16
|
||||
ftype = 0;
|
||||
if use_f16:
|
||||
if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
|
||||
print(" Converting to float16")
|
||||
data = data.astype(np.float16)
|
||||
ftype = 1
|
||||
else:
|
||||
print(" Converting to float32")
|
||||
data = data.astype(np.float32)
|
||||
ftype = 0
|
||||
|
||||
# for efficiency - transpose the projection matrices
|
||||
# "model/h.*/attn/c_attn/w"
|
||||
# "model/h.*/attn/c_proj/w"
|
||||
# "model/h.*/mlp/c_fc/w"
|
||||
# "model/h.*/mlp/c_proj/w"
|
||||
if name[-14:] == "/attn/c_attn/w" or \
|
||||
name[-14:] == "/attn/c_proj/w" or \
|
||||
name[-11:] == "/mlp/c_fc/w" or \
|
||||
name[-13:] == "/mlp/c_proj/w":
|
||||
print(" Transposing")
|
||||
data = data.transpose()
|
||||
|
||||
# header
|
||||
str = name.encode('utf-8')
|
||||
fout.write(struct.pack("iii", n_dims, len(str), ftype))
|
||||
for i in range(n_dims):
|
||||
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
|
||||
fout.write(str);
|
||||
|
||||
# data
|
||||
data.tofile(fout)
|
||||
|
||||
fout.close()
|
||||
|
||||
print("Done. Output file: " + fname_out)
|
||||
print("")
|
||||
Loading…
Add table
Add a link
Reference in a new issue