lora finetune (need to be refactored)
This commit is contained in:
josc146
0
finetune/lora/src/__init__.py
vendored
Normal file
0
finetune/lora/src/__init__.py
vendored
Normal file
269
finetune/lora/src/binidx.py
vendored
Normal file
269
finetune/lora/src/binidx.py
vendored
Normal file
@@ -0,0 +1,269 @@
|
||||
from lib2to3.pgen2 import token
|
||||
import os
|
||||
import torch
|
||||
import numpy as np
|
||||
import shutil
|
||||
import struct
|
||||
from functools import lru_cache
|
||||
from itertools import accumulate
|
||||
|
||||
def print_rank_0(*message):
|
||||
pass
|
||||
# """If distributed is initialized print only on rank 0."""
|
||||
# if torch.distributed.is_initialized():
|
||||
# if torch.distributed.get_rank() == 0:
|
||||
# print(*message, flush=True)
|
||||
# else:
|
||||
# print(*message, flush=True)
|
||||
|
||||
def _warmup_mmap_file(path):
|
||||
pass
|
||||
# with open(path, "rb") as stream:
|
||||
# while stream.read(100 * 1024 * 1024):
|
||||
# pass
|
||||
|
||||
dtypes = {
|
||||
1: np.uint8,
|
||||
2: np.int8,
|
||||
3: np.int16,
|
||||
4: np.int32,
|
||||
5: np.int64,
|
||||
6: float,
|
||||
7: np.double,
|
||||
8: np.uint16,
|
||||
}
|
||||
|
||||
def code(dtype):
|
||||
for k in dtypes.keys():
|
||||
if dtypes[k] == dtype:
|
||||
return k
|
||||
raise ValueError(dtype)
|
||||
|
||||
def index_file_path(prefix_path):
|
||||
return prefix_path + ".idx"
|
||||
|
||||
def data_file_path(prefix_path):
|
||||
return prefix_path + ".bin"
|
||||
|
||||
class MMapIndexedDataset(torch.utils.data.Dataset):
|
||||
class Index(object):
|
||||
_HDR_MAGIC = b"MMIDIDX\x00\x00"
|
||||
|
||||
@classmethod
|
||||
def writer(cls, path, dtype):
|
||||
class _Writer(object):
|
||||
def __enter__(self):
|
||||
self._file = open(path, "wb")
|
||||
|
||||
# Write Magic string so we can check the file format then opening it again.
|
||||
self._file.write(cls._HDR_MAGIC)
|
||||
# Write version number
|
||||
# Little endian unsigned 64 Bit integer
|
||||
self._file.write(struct.pack("<Q", 1))
|
||||
# Little endian unsigned 8 Bit integer
|
||||
self._file.write(struct.pack("<B", code(dtype)))
|
||||
|
||||
return self
|
||||
|
||||
@staticmethod
|
||||
def _get_pointers(sizes):
|
||||
dtype_size = dtype().itemsize
|
||||
address = 0
|
||||
pointers = []
|
||||
|
||||
for size in sizes:
|
||||
pointers.append(address)
|
||||
address += size * dtype_size
|
||||
|
||||
return pointers
|
||||
|
||||
def write(self, sizes, doc_idx):
|
||||
pointers = self._get_pointers(sizes)
|
||||
|
||||
# Little endian unsigned 64 Bit integer
|
||||
self._file.write(struct.pack("<Q", len(sizes)))
|
||||
# Little endian unsigned 64 Bit integer
|
||||
self._file.write(struct.pack("<Q", len(doc_idx)))
|
||||
|
||||
sizes = np.array(sizes, dtype=np.int32)
|
||||
self._file.write(sizes.tobytes(order="C"))
|
||||
del sizes
|
||||
|
||||
pointers = np.array(pointers, dtype=np.int64)
|
||||
self._file.write(pointers.tobytes(order="C"))
|
||||
del pointers
|
||||
|
||||
doc_idx = np.array(doc_idx, dtype=np.int64)
|
||||
self._file.write(doc_idx.tobytes(order="C"))
|
||||
|
||||
def __exit__(self, exc_type, exc_val, exc_tb):
|
||||
self._file.close()
|
||||
|
||||
return _Writer()
|
||||
|
||||
def __init__(self, path, skip_warmup=False):
|
||||
with open(path, "rb") as stream:
|
||||
magic_test = stream.read(9)
|
||||
assert self._HDR_MAGIC == magic_test, (
|
||||
"Index file doesn't match expected format. "
|
||||
"Make sure that --dataset-impl is configured properly."
|
||||
)
|
||||
# Little endian unsigned 64 Bit integer
|
||||
version = struct.unpack("<Q", stream.read(8))
|
||||
assert (1,) == version
|
||||
|
||||
# Little endian unsigned 8 Bit integer
|
||||
(dtype_code,) = struct.unpack("<B", stream.read(1))
|
||||
self._dtype = dtypes[dtype_code]
|
||||
self._dtype_size = self._dtype().itemsize
|
||||
|
||||
self._len = struct.unpack("<Q", stream.read(8))[0]
|
||||
self._doc_count = struct.unpack("<Q", stream.read(8))[0]
|
||||
offset = stream.tell()
|
||||
|
||||
if not skip_warmup:
|
||||
print_rank_0(" warming up index mmap file...")
|
||||
_warmup_mmap_file(path)
|
||||
|
||||
self._bin_buffer_mmap = np.memmap(path, mode="r", order="C")
|
||||
self._bin_buffer = memoryview(self._bin_buffer_mmap)
|
||||
print_rank_0(" reading sizes...")
|
||||
self._sizes = np.frombuffer(
|
||||
self._bin_buffer, dtype=np.int32, count=self._len, offset=offset
|
||||
)
|
||||
print_rank_0(" reading pointers...")
|
||||
self._pointers = np.frombuffer(
|
||||
self._bin_buffer,
|
||||
dtype=np.int64,
|
||||
count=self._len,
|
||||
offset=offset + self._sizes.nbytes,
|
||||
)
|
||||
print_rank_0(" reading document index...")
|
||||
self._doc_idx = np.frombuffer(
|
||||
self._bin_buffer,
|
||||
dtype=np.int64,
|
||||
count=self._doc_count,
|
||||
offset=offset + self._sizes.nbytes + self._pointers.nbytes,
|
||||
)
|
||||
|
||||
def __del__(self):
|
||||
self._bin_buffer_mmap._mmap.close()
|
||||
del self._bin_buffer_mmap
|
||||
|
||||
@property
|
||||
def dtype(self):
|
||||
return self._dtype
|
||||
|
||||
@property
|
||||
def sizes(self):
|
||||
return self._sizes
|
||||
|
||||
@property
|
||||
def doc_idx(self):
|
||||
return self._doc_idx
|
||||
|
||||
@lru_cache(maxsize=8)
|
||||
def __getitem__(self, i):
|
||||
return self._pointers[i], self._sizes[i]
|
||||
|
||||
def __len__(self):
|
||||
return self._len
|
||||
|
||||
def __init__(self, path, skip_warmup=False):
|
||||
super().__init__()
|
||||
|
||||
self._path = None
|
||||
self._index = None
|
||||
self._bin_buffer = None
|
||||
|
||||
self._do_init(path, skip_warmup)
|
||||
|
||||
def __getstate__(self):
|
||||
return self._path
|
||||
|
||||
def __setstate__(self, state):
|
||||
self._do_init(state)
|
||||
|
||||
def _do_init(self, path, skip_warmup):
|
||||
self._path = path
|
||||
self._index = self.Index(index_file_path(self._path), skip_warmup)
|
||||
|
||||
if not skip_warmup:
|
||||
print_rank_0(" warming up data mmap file...")
|
||||
_warmup_mmap_file(data_file_path(self._path))
|
||||
print_rank_0(" creating numpy buffer of mmap...")
|
||||
self._bin_buffer_mmap = np.memmap(
|
||||
data_file_path(self._path), mode="r", order="C"
|
||||
)
|
||||
print_rank_0(" creating memory view of numpy buffer...")
|
||||
self._bin_buffer = memoryview(self._bin_buffer_mmap)
|
||||
|
||||
def __del__(self):
|
||||
self._bin_buffer_mmap._mmap.close()
|
||||
del self._bin_buffer_mmap
|
||||
del self._index
|
||||
|
||||
def __len__(self):
|
||||
return len(self._index)
|
||||
|
||||
# @lru_cache(maxsize=8)
|
||||
def __getitem__(self, idx):
|
||||
if isinstance(idx, int):
|
||||
ptr, size = self._index[idx]
|
||||
np_array = np.frombuffer(
|
||||
self._bin_buffer, dtype=self._index.dtype, count=size, offset=ptr
|
||||
)
|
||||
return np_array
|
||||
elif isinstance(idx, slice):
|
||||
start, stop, step = idx.indices(len(self))
|
||||
if step != 1:
|
||||
raise ValueError(
|
||||
"Slices into indexed_dataset must be contiguous")
|
||||
ptr = self._index._pointers[start]
|
||||
sizes = self._index._sizes[idx]
|
||||
offsets = list(accumulate(sizes))
|
||||
total_size = sum(sizes)
|
||||
np_array = np.frombuffer(
|
||||
self._bin_buffer, dtype=self._index.dtype, count=total_size, offset=ptr
|
||||
)
|
||||
sents = np.split(np_array, offsets[:-1])
|
||||
return sents
|
||||
|
||||
def get(self, idx, offset=0, length=None):
|
||||
"""Retrieves a single item from the dataset with the option to only
|
||||
return a portion of the item.
|
||||
|
||||
get(idx) is the same as [idx] but get() does not support slicing.
|
||||
"""
|
||||
ptr, size = self._index[idx]
|
||||
if length is None:
|
||||
length = size - offset
|
||||
ptr += offset * np.dtype(self._index.dtype).itemsize
|
||||
np_array = np.frombuffer(
|
||||
self._bin_buffer, dtype=self._index.dtype, count=length, offset=ptr
|
||||
)
|
||||
return np_array
|
||||
|
||||
@property
|
||||
def sizes(self):
|
||||
return self._index.sizes
|
||||
|
||||
@property
|
||||
def doc_idx(self):
|
||||
return self._index.doc_idx
|
||||
|
||||
def get_doc_idx(self):
|
||||
return self._index._doc_idx
|
||||
|
||||
def set_doc_idx(self, doc_idx_):
|
||||
self._index._doc_idx = doc_idx_
|
||||
|
||||
@property
|
||||
def supports_prefetch(self):
|
||||
return False
|
||||
|
||||
@staticmethod
|
||||
def exists(path):
|
||||
return os.path.exists(index_file_path(path)) and os.path.exists(
|
||||
data_file_path(path)
|
||||
)
|
||||
224
finetune/lora/src/dataset.py
vendored
Normal file
224
finetune/lora/src/dataset.py
vendored
Normal file
@@ -0,0 +1,224 @@
|
||||
########################################################################################################
|
||||
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
|
||||
########################################################################################################
|
||||
|
||||
import json, math, random, os, sys
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import Dataset
|
||||
from pytorch_lightning.utilities import rank_zero_info
|
||||
from .binidx import MMapIndexedDataset
|
||||
from .utils import MaybeIsPrime
|
||||
|
||||
|
||||
class MyDataset(Dataset):
|
||||
def __init__(self, args):
|
||||
self.args = args
|
||||
|
||||
if args.data_type == "binidx":
|
||||
self.vocab_size = args.vocab_size
|
||||
rank_zero_info(f"Current vocab size = {self.vocab_size} (make sure it's correct)")
|
||||
|
||||
if args.data_file.endswith('/'):
|
||||
d_all = []
|
||||
for p in os.listdir(args.data_file):
|
||||
if p.endswith(".idx"):
|
||||
d_all += [p[:-4]]
|
||||
d_all.sort()
|
||||
rank_zero_info(d_all)
|
||||
exit(0)
|
||||
else:
|
||||
self.data = MMapIndexedDataset(args.data_file)
|
||||
self.data_size = len(self.data._bin_buffer) // self.data._index._dtype_size
|
||||
rank_zero_info(f"Data has {self.data_size} tokens.")
|
||||
|
||||
if args.my_qa_mask > 0:
|
||||
self.data_pile = MMapIndexedDataset('/fsx/BlinkDL/pile/pile_20B_tokenizer_text_document')
|
||||
self.data_pile_size = len(self.data_pile._bin_buffer) // self.data._index._dtype_size
|
||||
|
||||
if args.my_pile_stage > 0:
|
||||
# assert self.data_size == 332115325534 and self.vocab_size == 50277
|
||||
self.samples_per_epoch = args.epoch_steps * args.real_bsz
|
||||
assert self.samples_per_epoch == 40320
|
||||
rank_zero_info(f"########## Pile 20b-tokenized stage {args.my_pile_stage} ##########")
|
||||
dataset_slot = self.data_size // args.ctx_len
|
||||
if args.my_pile_stage != 4:
|
||||
assert MaybeIsPrime(args.magic_prime)
|
||||
assert args.magic_prime % 3 == 2
|
||||
assert args.magic_prime / dataset_slot > 0.99 and args.magic_prime / dataset_slot <= 1
|
||||
elif args.data_type == "numpy":
|
||||
self.data = np.load(args.data_file).astype("int")
|
||||
self.vocab_size = args.vocab_size
|
||||
rank_zero_info("Current vocab size =", self.vocab_size, "(make sure it's correct)")
|
||||
self.data_size = len(self.data)
|
||||
rank_zero_info(f"Data has {self.data_size} tokens.")
|
||||
elif args.data_type == "uint16":
|
||||
self.data = np.fromfile(args.data_file, dtype=np.uint16).astype("int32").reshape(-1, args.my_sample_len)
|
||||
self.vocab_size = args.vocab_size
|
||||
rank_zero_info("Current vocab size =", self.vocab_size, "(make sure it's correct)")
|
||||
self.data_size = self.data.shape[0]
|
||||
rank_zero_info(f"Data has {self.data_size} samples.")
|
||||
elif args.data_type == "wds_img":
|
||||
self.vocab_size = -1
|
||||
self.data_size = -1
|
||||
self.data = None
|
||||
self.error_count = 0
|
||||
else:
|
||||
if args.data_type == "dummy":
|
||||
rank_zero_info("Building dummy data...")
|
||||
self.data = ""
|
||||
for i in range(100000):
|
||||
aa = (i) % 10000
|
||||
bb = (i * i) % 10000
|
||||
cc = aa + bb
|
||||
self.data += f".{aa}+{bb}={cc}."
|
||||
else:
|
||||
self.data = open(args.data_file, "r", encoding=args.data_type).read()
|
||||
rank_zero_info("Building token list...")
|
||||
unique = sorted(list(set(self.data)))
|
||||
self.vocab_size = len(unique)
|
||||
# rank_zero_info()
|
||||
# for u in unique:
|
||||
# print(u, end=' ')
|
||||
# rank_zero_info('\n\n')
|
||||
xx = 0
|
||||
xxObj = {}
|
||||
for u in unique:
|
||||
xxObj[xx] = u
|
||||
xx += 1
|
||||
with open(f"{args.proj_dir}/vocab.json", "w", encoding="utf-16le") as vocab_file:
|
||||
vocab_file.write(json.dumps(xxObj, ensure_ascii=False))
|
||||
self.data_size = len(self.data)
|
||||
rank_zero_info(f"Data has {self.data_size} tokens, {self.vocab_size} vocab size.")
|
||||
self.stoi = {ch: i for i, ch in enumerate(unique)}
|
||||
self.itos = {i: ch for i, ch in enumerate(unique)}
|
||||
|
||||
def __len__(self):
|
||||
return self.args.epoch_steps * self.args.micro_bsz
|
||||
|
||||
def __getitem__(self, idx):
|
||||
args = self.args
|
||||
rank = self.global_rank
|
||||
epoch = self.real_epoch
|
||||
world_size = self.world_size
|
||||
# print(f"epoch {epoch} idx {idx} rank {rank}/{world_size}")
|
||||
|
||||
if args.data_type == "wds_img":
|
||||
def init_wds(self, bias=0):
|
||||
def identity(x):
|
||||
return x
|
||||
import webdataset as wds
|
||||
import torchvision.transforms as transforms
|
||||
# img_transform = transforms.Compose(
|
||||
# [transforms.CenterCrop(256)]
|
||||
# )
|
||||
img_transform = transforms.Compose([
|
||||
transforms.CenterCrop(512),
|
||||
transforms.Resize((args.my_img_size))
|
||||
])
|
||||
self.data_raw = wds.WebDataset(args.data_file, resampled=True).shuffle(10000, initial=1000, rng=random.Random(epoch*100000+rank+bias*1e9)).decode("torchrgb").to_tuple("jpg", "json", "txt").map_tuple(img_transform, identity, identity)
|
||||
for pp in self.data_raw.pipeline:
|
||||
if 'Resampled' in str(pp):
|
||||
pp.deterministic = True
|
||||
def worker_seed():
|
||||
return rank*100000+epoch+bias*1e9
|
||||
pp.worker_seed = worker_seed
|
||||
self.data = iter(self.data_raw)
|
||||
# print(f"WebDataset loaded for rank {rank} epoch {epoch}")
|
||||
if self.data == None:
|
||||
init_wds(self)
|
||||
trial = 0
|
||||
while trial < 10:
|
||||
try:
|
||||
dd = next(self.data) # jpg, json, txt
|
||||
break
|
||||
except:
|
||||
print(f'[dataloader error - epoch {epoch} rank {rank} - trying a new shuffle]')
|
||||
self.error_count += 1
|
||||
init_wds(self, self.error_count)
|
||||
trial += 1
|
||||
pass
|
||||
# print(f"epoch {epoch} idx {idx} rank {rank}/{world_size} {dd[2]}")
|
||||
# with open(f"sample_{rank}.txt", "a", encoding="utf-8") as tmp:
|
||||
# tmp.write(f"epoch {epoch} idx {idx} rank {rank}/{world_size} {int(dd[1]['key'])}\n")
|
||||
return dd[0], dd[2]
|
||||
else:
|
||||
if args.data_type == "uint16":
|
||||
i = np.random.randint(0, self.data_size-1)
|
||||
dix = self.data[i]
|
||||
x = torch.tensor(dix[:-1], dtype=torch.long)
|
||||
y = torch.tensor(dix[1:], dtype=torch.long)
|
||||
else:
|
||||
ctx_len = args.ctx_len
|
||||
req_len = ctx_len + 1
|
||||
magic_prime = args.magic_prime
|
||||
data = self.data
|
||||
|
||||
if args.my_pile_stage > 0 and args.my_pile_stage != 4:
|
||||
ii = 1 + epoch * self.samples_per_epoch + (idx * world_size) + rank
|
||||
|
||||
if args.my_qa_mask > 0:
|
||||
ii_orig = ii
|
||||
if ii % 2 == 0:
|
||||
ii = (ii // 2) * args.magic_prime
|
||||
if args.ctx_len == 1024:
|
||||
magic_prime = 324331313
|
||||
elif args.ctx_len == 2048:
|
||||
magic_prime = 162165671
|
||||
elif args.ctx_len == 4096:
|
||||
magic_prime = 81082817
|
||||
data = self.data_pile
|
||||
else:
|
||||
ii = ii // 2
|
||||
|
||||
factor = (math.sqrt(5) - 1) / 2
|
||||
factor = int(magic_prime * factor)
|
||||
i = ((factor * ii * ii * ii) % magic_prime) * ctx_len
|
||||
if (args.my_qa_mask == 0) or (data == self.data_pile):
|
||||
i = i + args.my_pile_shift
|
||||
# print(f"epoch {epoch} idx {idx} rank {rank}/{world_size} ii {ii} pos {round(i / self.data_size, 3)}")
|
||||
else:
|
||||
# cheat: pick a random spot in dataset
|
||||
i = np.random.randint(0, self.data_size - req_len)
|
||||
|
||||
if args.data_type == "binidx":
|
||||
dix = data.get(idx=0, offset=i, length=req_len).astype(int)
|
||||
elif args.data_type == "numpy":
|
||||
dix = data[i : i + req_len]
|
||||
else:
|
||||
dix = [self.stoi[s] for s in data[i : i + req_len]]
|
||||
|
||||
if args.my_qa_mask == 1:
|
||||
if data == self.data_pile:
|
||||
z = [1] * ctx_len
|
||||
else:
|
||||
z = [0] * ctx_len
|
||||
z_sum = 0
|
||||
isGood = False
|
||||
for i in range(3, ctx_len):
|
||||
if dix[i] == 27 and dix[i-1] == 34 and dix[i-2] == 187 and dix[i-3] == 187:
|
||||
isGood = True
|
||||
if dix[i] == 0:
|
||||
isGood = False
|
||||
if isGood:
|
||||
z[i] = 1
|
||||
z_sum += 1
|
||||
if z_sum == 0:
|
||||
z = [1] * ctx_len
|
||||
i = np.random.randint(0, self.data_pile_size - req_len)
|
||||
dix = self.data_pile.get(idx=0, offset=i, length=req_len).astype(int)
|
||||
z = torch.tensor(z, dtype=torch.bfloat16)
|
||||
|
||||
x = torch.tensor(dix[:-1], dtype=torch.long)
|
||||
y = torch.tensor(dix[1:], dtype=torch.long)
|
||||
|
||||
# if ii_orig < 50:
|
||||
# # if rank == 1:
|
||||
# print('rank', rank, 'i', ii_orig, ii, i, 'x', x[:5], '...', x[-5:])
|
||||
# else:
|
||||
# exit(0)
|
||||
|
||||
if args.my_qa_mask == 1:
|
||||
return x, y, z
|
||||
|
||||
return x, y
|
||||
678
finetune/lora/src/model.py
vendored
Normal file
678
finetune/lora/src/model.py
vendored
Normal file
@@ -0,0 +1,678 @@
|
||||
########################################################################################################
|
||||
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
|
||||
########################################################################################################
|
||||
|
||||
import functools
|
||||
import os, math, gc, importlib
|
||||
import torch
|
||||
# torch._C._jit_set_profiling_executor(True)
|
||||
# torch._C._jit_set_profiling_mode(True)
|
||||
import torch.nn as nn
|
||||
from torch.utils.checkpoint import checkpoint as torch_checkpoint
|
||||
from torch.nn import functional as F
|
||||
import pytorch_lightning as pl
|
||||
from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
|
||||
from pytorch_lightning.strategies import DeepSpeedStrategy
|
||||
if importlib.util.find_spec('deepspeed'):
|
||||
import deepspeed
|
||||
from deepspeed.ops.adam import DeepSpeedCPUAdam, FusedAdam
|
||||
|
||||
# from deepspeed.runtime.fp16.onebit.zoadam import ZeroOneAdam
|
||||
|
||||
LORA_CONFIG = {
|
||||
"r": 0,
|
||||
"alpha": 0,
|
||||
"dropout": 0,
|
||||
"parts": {"att", "ln", "time"},
|
||||
}
|
||||
|
||||
|
||||
try:
|
||||
print('RWKV_MY_TESTING', os.environ["RWKV_MY_TESTING"])
|
||||
except:
|
||||
os.environ["RWKV_MY_TESTING"] = ''
|
||||
|
||||
def __nop(ob):
|
||||
return ob
|
||||
|
||||
|
||||
MyModule = nn.Module
|
||||
MyFunction = __nop
|
||||
if os.environ["RWKV_JIT_ON"] == "1":
|
||||
MyModule = torch.jit.ScriptModule
|
||||
MyFunction = torch.jit.script_method
|
||||
|
||||
|
||||
########################################################################################################
|
||||
# CUDA Kernel
|
||||
########################################################################################################
|
||||
|
||||
T_MAX = int(os.environ["RWKV_T_MAX"]) # TAKES LOTS OF VRAM!
|
||||
# it's possible to go beyond CUDA limitations if you slice the ctx and pass the hidden state in each slice
|
||||
|
||||
from torch.utils.cpp_extension import load
|
||||
|
||||
if os.environ["RWKV_FLOAT_MODE"] == "bf16":
|
||||
wkv_cuda = load(name=f"wkv_{T_MAX}_bf16", sources=["finetune/lora/cuda/wkv_op_bf16.cpp", "finetune/lora/cuda/wkv_cuda_bf16.cu"], verbose=True, extra_cuda_cflags=["-t 4", "-std=c++17", "-res-usage", "--maxrregcount 60", "--use_fast_math", "-O3", "-Xptxas -O3", "--extra-device-vectorization", f"-DTmax={T_MAX}"])
|
||||
class WKV(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(ctx, B, T, C, w, u, k, v):
|
||||
ctx.B = B
|
||||
ctx.T = T
|
||||
ctx.C = C
|
||||
assert T <= T_MAX
|
||||
assert B * C % min(C, 32) == 0
|
||||
w = -torch.exp(w.float().contiguous())
|
||||
u = u.contiguous()
|
||||
k = k.contiguous()
|
||||
v = v.contiguous()
|
||||
y = torch.empty((B, T, C), device=w.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
|
||||
wkv_cuda.forward(B, T, C, w, u, k, v, y)
|
||||
ctx.save_for_backward(w, u, k, v, y)
|
||||
return y
|
||||
@staticmethod
|
||||
def backward(ctx, gy):
|
||||
B = ctx.B
|
||||
T = ctx.T
|
||||
C = ctx.C
|
||||
assert T <= T_MAX
|
||||
assert B * C % min(C, 32) == 0
|
||||
w, u, k, v, y = ctx.saved_tensors
|
||||
gw = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
|
||||
gu = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
|
||||
gk = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
|
||||
gv = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
|
||||
wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.contiguous(), gw, gu, gk, gv)
|
||||
gw = torch.sum(gw, dim=0)
|
||||
gu = torch.sum(gu, dim=0)
|
||||
return (None, None, None, gw, gu, gk, gv)
|
||||
else:
|
||||
wkv_cuda = load(name=f"wkv_{T_MAX}", sources=["finetune/lora/cuda/wkv_op.cpp", "finetune/lora/cuda/wkv_cuda.cu"], verbose=True, extra_cuda_cflags=["-res-usage", "--maxrregcount 60", "--use_fast_math", "-O3", "-Xptxas -O3", "--extra-device-vectorization", f"-DTmax={T_MAX}"])
|
||||
class WKV(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(ctx, B, T, C, w, u, k, v):
|
||||
ctx.B = B
|
||||
ctx.T = T
|
||||
ctx.C = C
|
||||
assert T <= T_MAX
|
||||
assert B * C % min(C, 32) == 0
|
||||
if "32" in os.environ["RWKV_FLOAT_MODE"]:
|
||||
w = -torch.exp(w.contiguous())
|
||||
u = u.contiguous()
|
||||
k = k.contiguous()
|
||||
v = v.contiguous()
|
||||
else:
|
||||
w = -torch.exp(w.float().contiguous())
|
||||
u = u.float().contiguous()
|
||||
k = k.float().contiguous()
|
||||
v = v.float().contiguous()
|
||||
y = torch.empty((B, T, C), device=w.device, memory_format=torch.contiguous_format)
|
||||
wkv_cuda.forward(B, T, C, w, u, k, v, y)
|
||||
ctx.save_for_backward(w, u, k, v, y)
|
||||
if "32" in os.environ["RWKV_FLOAT_MODE"]:
|
||||
return y
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
|
||||
return y.half()
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
|
||||
return y.bfloat16()
|
||||
@staticmethod
|
||||
def backward(ctx, gy):
|
||||
B = ctx.B
|
||||
T = ctx.T
|
||||
C = ctx.C
|
||||
assert T <= T_MAX
|
||||
assert B * C % min(C, 32) == 0
|
||||
w, u, k, v, y = ctx.saved_tensors
|
||||
gw = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format)
|
||||
gu = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format)
|
||||
gk = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format)
|
||||
gv = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format)
|
||||
if "32" in os.environ["RWKV_FLOAT_MODE"]:
|
||||
wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.contiguous(), gw, gu, gk, gv)
|
||||
else:
|
||||
wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.float().contiguous(), gw, gu, gk, gv)
|
||||
gw = torch.sum(gw, dim=0)
|
||||
gu = torch.sum(gu, dim=0)
|
||||
if "32" in os.environ["RWKV_FLOAT_MODE"]:
|
||||
return (None, None, None, gw, gu, gk, gv)
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
|
||||
return (None, None, None, gw.half(), gu.half(), gk.half(), gv.half())
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
|
||||
return (None, None, None, gw.bfloat16(), gu.bfloat16(), gk.bfloat16(), gv.bfloat16())
|
||||
|
||||
|
||||
def RUN_CUDA(B, T, C, w, u, k, v):
|
||||
return WKV.apply(B, T, C, w, u, k, v)
|
||||
|
||||
|
||||
########################################################################################################
|
||||
# LoRA
|
||||
########################################################################################################
|
||||
|
||||
|
||||
class LoraLinear(nn.Module):
|
||||
|
||||
def __init__(self, in_features: int, out_features: int, bias: bool):
|
||||
super().__init__()
|
||||
|
||||
self.weight = nn.Parameter(torch.empty((out_features, in_features)))
|
||||
assert bias == False, "Biased LoraLinear not supported"
|
||||
|
||||
r, alpha, dropout = LORA_CONFIG["r"], LORA_CONFIG[
|
||||
"alpha"], LORA_CONFIG["dropout"]
|
||||
self.lora_A = nn.Parameter(torch.empty(r, in_features))
|
||||
self.lora_B = nn.Parameter(torch.empty(out_features, r))
|
||||
self.lora_dropout = nn.Dropout(dropout)
|
||||
self.scaling = alpha / r
|
||||
|
||||
nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
|
||||
nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
|
||||
nn.init.zeros_(self.lora_B)
|
||||
|
||||
def forward(self, x):
|
||||
return (
|
||||
F.linear(x, self.weight) + self.scaling *
|
||||
F.linear(F.linear(self.lora_dropout(x), self.lora_A), self.lora_B))
|
||||
|
||||
|
||||
@functools.wraps(LoraLinear)
|
||||
def make_linear_att(*args, **kwargs):
|
||||
if "att" in LORA_CONFIG["parts"] and LORA_CONFIG["r"] > 0:
|
||||
return LoraLinear(*args, **kwargs)
|
||||
else:
|
||||
return nn.Linear(*args, **kwargs)
|
||||
|
||||
|
||||
@functools.wraps(LoraLinear)
|
||||
def make_linear_ffn(*args, **kwargs):
|
||||
if "ffn" in LORA_CONFIG["parts"] and LORA_CONFIG["r"] > 0:
|
||||
return LoraLinear(*args, **kwargs)
|
||||
else:
|
||||
return nn.Linear(*args, **kwargs)
|
||||
|
||||
|
||||
########################################################################################################
|
||||
# RWKV: RWKV Time-mix + RWKV Channel-mix
|
||||
########################################################################################################
|
||||
|
||||
|
||||
class RWKV_TimeMix(MyModule):
|
||||
def __init__(self, args, layer_id):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
self.layer_id = layer_id
|
||||
self.ctx_len = args.ctx_len
|
||||
self.n_embd = args.n_embd
|
||||
|
||||
with torch.no_grad(): # fancy init
|
||||
ratio_0_to_1 = layer_id / (args.n_layer - 1) # 0 to 1
|
||||
ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer) # 1 to ~0
|
||||
ddd = torch.ones(1, 1, args.n_embd)
|
||||
for i in range(args.n_embd):
|
||||
ddd[0, 0, i] = i / args.n_embd
|
||||
|
||||
# fancy time_decay
|
||||
decay_speed = torch.ones(args.dim_att)
|
||||
for h in range(args.dim_att):
|
||||
decay_speed[h] = -5 + 8 * (h / (args.dim_att - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
|
||||
self.time_decay = nn.Parameter(decay_speed)
|
||||
# print(layer_id, self.time_decay.flatten()[:3].cpu().numpy(), '...', self.time_decay.flatten()[-3:].cpu().numpy())
|
||||
|
||||
# fancy time_first
|
||||
zigzag = torch.tensor([(i + 1) % 3 - 1 for i in range(args.dim_att)]) * 0.5
|
||||
self.time_first = nn.Parameter(torch.ones(args.dim_att) * math.log(0.3) + zigzag)
|
||||
|
||||
# fancy time_mix
|
||||
self.time_mix_k = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
|
||||
self.time_mix_v = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
|
||||
self.time_mix_r = nn.Parameter(torch.pow(ddd, 0.5 * ratio_1_to_almost0))
|
||||
|
||||
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
||||
|
||||
self.key = make_linear_att(args.n_embd, args.dim_att, bias=False)
|
||||
self.value = make_linear_att(args.n_embd, args.dim_att, bias=False)
|
||||
self.receptance = make_linear_att(args.n_embd, args.dim_att, bias=False)
|
||||
|
||||
self.output = nn.Linear(args.dim_att, args.n_embd, bias=False)
|
||||
|
||||
if 'a' in os.environ["RWKV_MY_TESTING"]:
|
||||
self.register_buffer("att_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
|
||||
d_qkv = args.n_embd // 16
|
||||
self.qq = nn.Linear(args.n_embd, d_qkv, bias=False)
|
||||
self.kk = nn.Linear(args.n_embd, d_qkv, bias=False)
|
||||
self.vv = nn.Linear(args.n_embd, d_qkv, bias=False)
|
||||
self.oo = nn.Linear(d_qkv, args.n_embd, bias=False)
|
||||
with torch.no_grad():
|
||||
self.time_mix_qq = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
|
||||
self.time_mix_kk = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
|
||||
self.time_mix_vv = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
|
||||
|
||||
if 'a' not in os.environ["RWKV_MY_TESTING"]:
|
||||
@MyFunction
|
||||
def jit_func(self, x):
|
||||
xx = self.time_shift(x) # Mix x with the previous timestep to produce xk, xv, xr
|
||||
xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
|
||||
xv = x * self.time_mix_v + xx * (1 - self.time_mix_v)
|
||||
xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
|
||||
k = self.key(xk)
|
||||
v = self.value(xv)
|
||||
r = self.receptance(xr)
|
||||
sr = torch.sigmoid(r)
|
||||
return sr, k, v
|
||||
|
||||
def forward(self, x):
|
||||
B, T, C = x.size() # x = (Batch,Time,Channel)
|
||||
sr, k, v = self.jit_func(x)
|
||||
rwkv = sr * RUN_CUDA(B, T, self.args.dim_att, self.time_decay, self.time_first, k, v)
|
||||
return self.output(rwkv)
|
||||
|
||||
if 'a' in os.environ["RWKV_MY_TESTING"]:
|
||||
@MyFunction
|
||||
def QKV(self, q, k, v):
|
||||
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
|
||||
att = att.masked_fill(self.att_mask == 0, float('-inf'))
|
||||
att = F.softmax(att, dim = -1)
|
||||
x = att @ v
|
||||
return x
|
||||
|
||||
@MyFunction
|
||||
def jit_funcQKV(self, x):
|
||||
xx = self.time_shift(x) # Mix x with the previous timestep to produce xk, xv, xr
|
||||
xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
|
||||
xv = x * self.time_mix_v + xx * (1 - self.time_mix_v)
|
||||
xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
|
||||
xqq = x * self.time_mix_qq + xx * (1 - self.time_mix_qq)
|
||||
xkk = x * self.time_mix_kk + xx * (1 - self.time_mix_kk)
|
||||
xvv = x * self.time_mix_vv + xx * (1 - self.time_mix_vv)
|
||||
k = self.key(xk)
|
||||
v = self.value(xv)
|
||||
r = self.receptance(xr)
|
||||
sr = torch.sigmoid(r)
|
||||
qq = self.qq(xqq)
|
||||
kk = self.kk(xkk)
|
||||
vv = self.vv(xvv)
|
||||
return sr, k, v, qq, kk, vv
|
||||
|
||||
def forward(self, x):
|
||||
B, T, C = x.size() # x = (Batch,Time,Channel)
|
||||
sr, k, v, qq, kk, vv = self.jit_funcQKV(x)
|
||||
rwkv = sr * RUN_CUDA(B, T, self.args.dim_att, self.time_decay, self.time_first, k, v)
|
||||
rwkv = self.output(rwkv) + self.oo(self.QKV(qq, kk, vv))
|
||||
return rwkv
|
||||
|
||||
########################################################################################################
|
||||
|
||||
class RWKV_ChannelMix(MyModule):
|
||||
def __init__(self, args, layer_id):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
self.layer_id = layer_id
|
||||
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
||||
|
||||
with torch.no_grad(): # fancy init of time_mix
|
||||
ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer) # 1 to ~0
|
||||
ddd = torch.ones(1, 1, args.n_embd)
|
||||
for i in range(args.n_embd):
|
||||
ddd[0, 0, i] = i / args.n_embd
|
||||
self.time_mix_k = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
|
||||
self.time_mix_r = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
|
||||
|
||||
self.key = make_linear_ffn(args.n_embd, args.dim_ffn, bias=False)
|
||||
self.receptance = make_linear_ffn(args.n_embd, args.n_embd, bias=False)
|
||||
self.value = make_linear_ffn(args.dim_ffn, args.n_embd, bias=False)
|
||||
|
||||
@MyFunction
|
||||
def forward(self, x):
|
||||
xx = self.time_shift(x)
|
||||
xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
|
||||
xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
|
||||
k = self.key(xk)
|
||||
k = torch.square(torch.relu(k))
|
||||
kv = self.value(k)
|
||||
return torch.sigmoid(self.receptance(xr)) * kv
|
||||
|
||||
class MishGLU(MyModule):
|
||||
def __init__(self, args, layer_id):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
self.layer_id = layer_id
|
||||
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
||||
|
||||
with torch.no_grad():
|
||||
ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer)
|
||||
|
||||
x = torch.ones(1, 1, args.n_embd)
|
||||
for i in range(args.n_embd):
|
||||
x[0, 0, i] = i / args.n_embd
|
||||
|
||||
self.time_mix_k = nn.Parameter(torch.pow(x, ratio_1_to_almost0))
|
||||
self.time_mix_r = nn.Parameter(torch.pow(x, ratio_1_to_almost0))
|
||||
self.aa = nn.Linear(args.n_embd, args.dim_ffn, bias=False)
|
||||
self.bb = nn.Linear(args.n_embd, args.dim_ffn, bias=False)
|
||||
self.value = nn.Linear(args.dim_ffn, args.n_embd, bias=False)
|
||||
|
||||
@MyFunction
|
||||
def forward(self, x):
|
||||
xx = self.time_shift(x)
|
||||
xa = x * self.time_mix_k + xx * (1 - self.time_mix_k)
|
||||
xb = x * self.time_mix_r + xx * (1 - self.time_mix_r)
|
||||
a = self.aa(xa)
|
||||
b = self.bb(xb)
|
||||
return self.value(a * F.mish(b))
|
||||
|
||||
########################################################################################################
|
||||
# The RWKV Model with our blocks
|
||||
########################################################################################################
|
||||
|
||||
|
||||
class Block(nn.Module):
|
||||
def __init__(self, args, layer_id):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
self.layer_id = layer_id
|
||||
|
||||
self.ln1 = nn.LayerNorm(args.n_embd)
|
||||
self.ln2 = nn.LayerNorm(args.n_embd)
|
||||
|
||||
if self.layer_id == 0:
|
||||
self.ln0 = nn.LayerNorm(args.n_embd)
|
||||
if args.my_pos_emb > 0:
|
||||
self.pos_emb_x = nn.Parameter(torch.zeros((1,args.my_pos_emb,args.n_embd)))
|
||||
self.pos_emb_y = nn.Parameter(torch.zeros((args.my_pos_emb,1,args.n_embd)))
|
||||
|
||||
if self.layer_id == 0 and self.args.pre_ffn > 0:
|
||||
self.ffnPre = RWKV_ChannelMix(args, 0)
|
||||
else:
|
||||
self.att = RWKV_TimeMix(args, layer_id)
|
||||
|
||||
if 'g' in os.environ["RWKV_MY_TESTING"]:
|
||||
self.ffn = MishGLU(args, layer_id)
|
||||
else:
|
||||
self.ffn = RWKV_ChannelMix(args, layer_id)
|
||||
|
||||
if args.tiny_att_dim > 0 and self.layer_id == args.tiny_att_layer:
|
||||
self.tiny_ln = nn.LayerNorm(args.n_embd)
|
||||
self.tiny_q = nn.Linear(args.n_embd, args.tiny_att_dim, bias=False)
|
||||
self.tiny_k = nn.Linear(args.n_embd, args.tiny_att_dim, bias=False)
|
||||
self.tiny_v = nn.Linear(args.n_embd, args.n_embd, bias=False)
|
||||
self.register_buffer("tiny_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
|
||||
|
||||
def forward(self, x, x_emb=None):
|
||||
args = self.args
|
||||
B, T, C = x.size()
|
||||
if self.layer_id == 0:
|
||||
x = self.ln0(x)
|
||||
if args.my_pos_emb > 0:
|
||||
pos_emb = (self.pos_emb_x + self.pos_emb_y).reshape(T+1, -1)[:-1,:]
|
||||
x = x + pos_emb
|
||||
|
||||
if self.layer_id == 0 and args.pre_ffn > 0:
|
||||
x = x + self.ffnPre(self.ln1(x))
|
||||
else:
|
||||
x = x + self.att(self.ln1(x))
|
||||
x = x + self.ffn(self.ln2(x))
|
||||
|
||||
if args.tiny_att_dim > 0 and self.layer_id == args.tiny_att_layer:
|
||||
xx = self.tiny_ln(x)
|
||||
q = self.tiny_q(xx)[:, :T, :]
|
||||
k = self.tiny_k(xx)[:, :T, :]
|
||||
c = (q @ k.transpose(-2, -1)) * (args.tiny_att_dim ** (-0.5))
|
||||
c = c.masked_fill(self.tiny_mask[:T, :T] == 0, 0)
|
||||
x = x + c @ self.tiny_v(x_emb)
|
||||
return x
|
||||
|
||||
|
||||
class L2Wrap(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(ctx, loss, y):
|
||||
ctx.save_for_backward(y)
|
||||
return loss
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, grad_output):
|
||||
y = ctx.saved_tensors[0]
|
||||
# to encourage the logits to be close to 0
|
||||
factor = 1e-4 / (y.shape[0] * y.shape[1])
|
||||
maxx, ids = torch.max(y, -1, keepdim=True)
|
||||
gy = torch.zeros_like(y)
|
||||
gy.scatter_(-1, ids, maxx * factor)
|
||||
return (grad_output, gy)
|
||||
|
||||
|
||||
class RWKV(pl.LightningModule):
|
||||
def __init__(self, args):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
if not hasattr(args, 'dim_att'):
|
||||
args.dim_att = args.n_embd
|
||||
if not hasattr(args, 'dim_ffn'):
|
||||
args.dim_ffn = args.n_embd * 4
|
||||
if not hasattr(args, 'tiny_att_layer'):
|
||||
args.tiny_att_layer = -1
|
||||
if not hasattr(args, 'tiny_att_dim'):
|
||||
args.tiny_att_dim = -1
|
||||
|
||||
self.emb = nn.Embedding(args.vocab_size, args.n_embd)
|
||||
|
||||
self.blocks = nn.ModuleList([Block(args, i) for i in range(args.n_layer)])
|
||||
|
||||
self.ln_out = nn.LayerNorm(args.n_embd)
|
||||
self.head = nn.Linear(args.n_embd, args.vocab_size, bias=False)
|
||||
|
||||
if args.head_qk > 0:
|
||||
self.head_q = nn.Linear(args.n_embd, args.head_qk, bias=False)
|
||||
self.head_k = nn.Linear(args.n_embd, args.head_qk, bias=False)
|
||||
self.register_buffer("copy_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
|
||||
|
||||
def configure_optimizers(self):
|
||||
args = self.args
|
||||
if args.layerwise_lr > 0:
|
||||
lr_1x = set()
|
||||
lr_2x = set()
|
||||
lr_3x = set()
|
||||
for n, p in self.named_parameters():
|
||||
if "time_mix" in n:
|
||||
if args.my_pile_stage == 2:
|
||||
lr_2x.add(n)
|
||||
else:
|
||||
lr_1x.add(n)
|
||||
elif "time_decay" in n:
|
||||
if args.my_pile_stage == 2:
|
||||
lr_3x.add(n)
|
||||
else:
|
||||
lr_2x.add(n)
|
||||
elif "time_first" in n:
|
||||
lr_3x.add(n)
|
||||
else:
|
||||
lr_1x.add(n)
|
||||
lr_1x = sorted(list(lr_1x))
|
||||
lr_2x = sorted(list(lr_2x))
|
||||
lr_3x = sorted(list(lr_3x))
|
||||
# print('1x', lr_1x)
|
||||
# print('2x', lr_2x)
|
||||
# print('3x', lr_3x)
|
||||
param_dict = {n: p for n, p in self.named_parameters()}
|
||||
if args.my_pile_stage == 2:
|
||||
optim_groups = [
|
||||
{"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
|
||||
{"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 5.0},# test: 2e-3 / args.lr_init},
|
||||
{"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 5.0},# test: 3e-3 / args.lr_init},
|
||||
]
|
||||
else:
|
||||
optim_groups = [
|
||||
{"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
|
||||
{"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 2.0},
|
||||
{"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 3.0},
|
||||
]
|
||||
else:
|
||||
optim_groups = [
|
||||
{"params": [p for n, p in self.named_parameters()], "weight_decay": 0.0},
|
||||
]
|
||||
|
||||
for g in optim_groups:
|
||||
g["params"] = [p for p in g["params"] if p.requires_grad]
|
||||
optim_groups = [g for g in optim_groups if len(g["params"]) > 0]
|
||||
|
||||
if self.deepspeed_offload:
|
||||
return DeepSpeedCPUAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adamw_mode=False, weight_decay=0, amsgrad=False)
|
||||
return FusedAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adam_w_mode=False, weight_decay=0, amsgrad=False)
|
||||
# return ZeroOneAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, weight_decay=0, amsgrad=False, cuda_aware=False)
|
||||
|
||||
@property
|
||||
def deepspeed_offload(self) -> bool:
|
||||
strategy = self.trainer.strategy
|
||||
if isinstance(strategy, DeepSpeedStrategy):
|
||||
cfg = strategy.config["zero_optimization"]
|
||||
return cfg.get("offload_optimizer") or cfg.get("offload_param")
|
||||
return False
|
||||
|
||||
def forward(self, idx):
|
||||
args = self.args
|
||||
B, T = idx.size()
|
||||
assert T <= args.ctx_len, "Cannot forward, model ctx_len is exhausted."
|
||||
|
||||
x = self.emb(idx)
|
||||
x_emb = x
|
||||
|
||||
if args.tiny_att_dim > 0:
|
||||
for block in self.blocks:
|
||||
if args.grad_cp == 1:
|
||||
if args.lora:
|
||||
x = torch_checkpoint(block, x, x_emb, use_reentrant=False)
|
||||
else:
|
||||
x = deepspeed.checkpointing.checkpoint(block, x, x_emb)
|
||||
else:
|
||||
x = block(x, x_emb)
|
||||
else:
|
||||
for block in self.blocks:
|
||||
if args.grad_cp == 1:
|
||||
if args.lora:
|
||||
x = torch_checkpoint(block, x, x_emb, use_reentrant=False)
|
||||
else:
|
||||
x = deepspeed.checkpointing.checkpoint(block, x)
|
||||
else:
|
||||
x = block(x)
|
||||
|
||||
x = self.ln_out(x)
|
||||
|
||||
if args.head_qk > 0:
|
||||
q = self.head_q(x)[:, :T, :]
|
||||
k = self.head_k(x)[:, :T, :]
|
||||
c = (q @ k.transpose(-2, -1)) * (1.0 / args.head_qk)
|
||||
c = c.masked_fill(self.copy_mask[:T, :T] == 0, 0)
|
||||
|
||||
if "32" in os.environ["RWKV_FLOAT_MODE"]:
|
||||
c = c @ F.one_hot(idx, num_classes=args.vocab_size)
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
|
||||
c = c @ F.one_hot(idx, num_classes=args.vocab_size).half()
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
|
||||
c = c @ F.one_hot(idx, num_classes=args.vocab_size).bfloat16()
|
||||
|
||||
x = self.head(x) + c
|
||||
else:
|
||||
x = self.head(x)
|
||||
|
||||
return x
|
||||
|
||||
def training_step(self, batch, batch_idx):
|
||||
args = self.args
|
||||
if args.my_qa_mask != 1:
|
||||
idx, targets = batch
|
||||
logits = self(idx)
|
||||
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
|
||||
else:
|
||||
idx, targets, mask = batch
|
||||
mask = mask.view(-1)
|
||||
sum_mask = torch.sum(mask).item()
|
||||
# if sum_mask == 0:
|
||||
# return torch.tensor([0.0], requires_grad=True)
|
||||
|
||||
logits = self(idx)
|
||||
if sum_mask == mask.shape[0]:
|
||||
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
|
||||
# print('rank', self.global_rank, 'loss', loss.item())
|
||||
else:
|
||||
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), reduction='none')
|
||||
# loss_raw = loss
|
||||
loss = torch.sum(loss * mask) / sum_mask
|
||||
|
||||
# torch.set_printoptions(threshold=10000)
|
||||
# if True: #self.global_rank == 1:
|
||||
# tmp = ''
|
||||
# sss = 0
|
||||
# ccc = 0
|
||||
# for i in range(mask.shape[0]):
|
||||
# if mask[i] > 0:
|
||||
# tmp += str(idx.view(-1)[i].item()) + ','
|
||||
# sss += loss_raw.view(-1)[i].float().item()
|
||||
# ccc += 1
|
||||
# print('rank', self.global_rank, 'loss', loss.item(), 'lavg', sss / ccc)#, 'tmp', tmp, 'input', idx)
|
||||
|
||||
return L2Wrap.apply(loss, logits)
|
||||
|
||||
def training_step_end(self, batch_parts):
|
||||
all = self.all_gather(batch_parts)
|
||||
if self.trainer.is_global_zero:
|
||||
self.trainer.my_loss_all = all
|
||||
|
||||
def generate_init_weight(self):
|
||||
print(
|
||||
f"""
|
||||
############################################################################
|
||||
#
|
||||
# Init model weight (slow for large models)...
|
||||
#
|
||||
############################################################################
|
||||
"""
|
||||
)
|
||||
m = {}
|
||||
for n in self.state_dict():
|
||||
p = self.state_dict()[n]
|
||||
shape = p.shape
|
||||
|
||||
gain = 1.0
|
||||
scale = 1.0
|
||||
if "ln_" in n or ".ln" in n or "time_" in n or "_mask" in n or "pos_emb" in n or '.mask.' in n:
|
||||
m[n] = p
|
||||
else:
|
||||
if n == "emb.weight":
|
||||
scale = -1 * self.args.lr_init
|
||||
else:
|
||||
if shape[0] > shape[1]:
|
||||
gain = math.sqrt(shape[0] / shape[1])
|
||||
for kk in [".att.key.", ".att.receptance.", ".att.output.", ".att.key.", ".ffn.value.", ".ffn.receptance.", ".ffnPre.value.", ".ffnPre.receptance.", "head_q.", '.oo.', '.rr.']:
|
||||
if kk in n:
|
||||
scale = 0
|
||||
if n == "head.weight":
|
||||
scale = 0.5
|
||||
if "head_k." in n:
|
||||
scale = 0.1
|
||||
if "head_q." in n:
|
||||
scale = 0
|
||||
|
||||
print(f"{str(shape[0]).ljust(5)} {str(shape[1]).ljust(5)} {str(scale).ljust(4)} {n}")
|
||||
|
||||
if self.args.accelerator.upper() == "GPU":
|
||||
m[n] = torch.empty((shape[0], shape[1]), device="cuda")
|
||||
else:
|
||||
m[n] = torch.empty((shape[0], shape[1]))
|
||||
|
||||
if scale == 0:
|
||||
nn.init.zeros_(m[n])
|
||||
elif scale < 0:
|
||||
nn.init.uniform_(m[n], a=scale, b=-scale)
|
||||
else:
|
||||
nn.init.orthogonal_(m[n], gain=gain * scale)
|
||||
|
||||
m[n] = m[n].cpu()
|
||||
if os.environ["RWKV_FLOAT_MODE"] == "fp16":
|
||||
m[n] = m[n].half()
|
||||
elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
|
||||
m[n] = m[n].bfloat16()
|
||||
|
||||
# if n == "emb.weight":
|
||||
# print(m[n])
|
||||
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
return m
|
||||
203
finetune/lora/src/trainer.py
vendored
Normal file
203
finetune/lora/src/trainer.py
vendored
Normal file
@@ -0,0 +1,203 @@
|
||||
import os, math, time, datetime, subprocess
|
||||
import torch
|
||||
from torch.utils.data import DataLoader
|
||||
import pytorch_lightning as pl
|
||||
from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
|
||||
from .model import LORA_CONFIG
|
||||
|
||||
def my_save(dd, ff):
|
||||
if '14b-run1' not in ff:
|
||||
torch.save(dd, ff)
|
||||
else:
|
||||
fn = ff.split('/')[-1]
|
||||
fff = '/dev/shm/' + fn
|
||||
torch.save(dd, fff)
|
||||
subprocess.Popen(f" aws s3 mv {fff} s3://rwkv-14b-4k/{fn} --quiet", shell=True)
|
||||
|
||||
class train_callback(pl.Callback):
|
||||
def __init__(self, args):
|
||||
super().__init__()
|
||||
self.args = args
|
||||
|
||||
def on_train_batch_start(self, trainer, pl_module, batch, batch_idx):
|
||||
args = self.args
|
||||
# if args.cuda_cleanup > 0:
|
||||
# torch.cuda.empty_cache()
|
||||
real_step = trainer.global_step + args.epoch_begin * args.epoch_steps
|
||||
|
||||
# LR schedule
|
||||
w_step = args.warmup_steps
|
||||
if args.lr_final == args.lr_init or args.epoch_count == 0:
|
||||
lr = args.lr_init
|
||||
else:
|
||||
decay_step = real_step - args.my_pile_edecay * args.epoch_steps
|
||||
decay_total = (args.epoch_count - args.my_pile_edecay) * args.epoch_steps
|
||||
progress = (decay_step - w_step + 1) / (decay_total - w_step)
|
||||
progress = min(1, max(0, progress))
|
||||
|
||||
if args.lr_final == 0 or args.lr_init == 0: # linear decay
|
||||
lr = args.lr_init + (args.lr_final - args.lr_init) * progress
|
||||
else: # exp decay
|
||||
lr = args.lr_init * math.exp(math.log(args.lr_final / args.lr_init) * pow(progress, 1))
|
||||
|
||||
if trainer.global_step < w_step:
|
||||
lr = lr * (0.2 + 0.8 * trainer.global_step / w_step)
|
||||
# if trainer.is_global_zero:
|
||||
# print(trainer.global_step, decay_step, decay_total, w_step, progress, lr)
|
||||
|
||||
for param_group in trainer.optimizers[0].param_groups:
|
||||
if args.layerwise_lr > 0:
|
||||
param_group["lr"] = lr * param_group["my_lr_scale"]
|
||||
# print(param_group["lr"], param_group["my_lr_scale"])
|
||||
else:
|
||||
param_group["lr"] = lr
|
||||
|
||||
trainer.my_lr = lr
|
||||
# rank_zero_info(f"{real_step} {lr}")
|
||||
|
||||
if trainer.global_step == 0:
|
||||
if trainer.is_global_zero: # logging
|
||||
trainer.my_loss_sum = 0
|
||||
trainer.my_loss_count = 0
|
||||
trainer.my_log = open(args.proj_dir + "/train_log.txt", "a")
|
||||
trainer.my_log.write(f"NEW RUN {args.my_timestamp}\n{vars(self.args)}\n")
|
||||
try:
|
||||
print(f"\n{trainer.strategy.config}\n")
|
||||
trainer.my_log.write(f"{trainer.strategy.config}\n")
|
||||
except:
|
||||
pass
|
||||
trainer.my_log.flush()
|
||||
if len(args.wandb) > 0:
|
||||
print("Login to wandb...")
|
||||
import wandb
|
||||
wandb.init(
|
||||
project=args.wandb,
|
||||
name=args.run_name + " " + args.my_timestamp,
|
||||
config=args,
|
||||
save_code=False,
|
||||
)
|
||||
trainer.my_wandb = wandb
|
||||
|
||||
def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
|
||||
args = self.args
|
||||
if trainer.is_global_zero: # logging
|
||||
t_now = time.time_ns()
|
||||
token_per_step = args.ctx_len * args.real_bsz
|
||||
real_step = trainer.global_step + args.epoch_begin * args.epoch_steps
|
||||
kt_s = 0
|
||||
try:
|
||||
t_cost = (t_now - trainer.my_time_ns) / 1e9
|
||||
kt_s = token_per_step / t_cost / 1000
|
||||
self.log("REAL it/s", 1.0 / t_cost, prog_bar=True, on_step=True)
|
||||
self.log("Kt/s", kt_s, prog_bar=True, on_step=True)
|
||||
except:
|
||||
pass
|
||||
trainer.my_time_ns = t_now
|
||||
trainer.my_loss = trainer.my_loss_all.float().mean().item()
|
||||
trainer.my_loss_sum += trainer.my_loss
|
||||
trainer.my_loss_count += 1
|
||||
trainer.my_epoch_loss = trainer.my_loss_sum / trainer.my_loss_count
|
||||
self.log("lr", trainer.my_lr, prog_bar=True, on_step=True)
|
||||
self.log("loss", trainer.my_epoch_loss, prog_bar=True, on_step=True)
|
||||
# self.log("s", real_step, prog_bar=True, on_step=True)
|
||||
|
||||
if len(args.wandb) > 0:
|
||||
lll = {"loss": trainer.my_loss, "lr": trainer.my_lr, "Gtokens": real_step * token_per_step / 1e9}
|
||||
if kt_s > 0:
|
||||
lll["kt/s"] = kt_s
|
||||
trainer.my_wandb.log(lll, step=int(real_step))
|
||||
if args.magic_prime > 0:
|
||||
expand_factor = 2 if args.my_qa_mask > 0 else 1
|
||||
if int(real_step) == int(args.magic_prime * expand_factor // args.real_bsz) - 1:
|
||||
to_save_dict = pl_module.state_dict()
|
||||
my_save(
|
||||
to_save_dict,
|
||||
f"{args.proj_dir}/rwkv-final.pth",
|
||||
)
|
||||
|
||||
|
||||
def on_train_epoch_start(self, trainer, pl_module):
|
||||
args = self.args
|
||||
dataset = trainer.train_dataloader.dataset.datasets
|
||||
assert "MyDataset" in str(dataset)
|
||||
dataset.global_rank = trainer.global_rank
|
||||
dataset.real_epoch = int(args.epoch_begin + trainer.current_epoch)
|
||||
dataset.world_size = trainer.world_size
|
||||
# print(f'########## world_size {dataset.world_size} global_rank {dataset.global_rank} real_epoch {dataset.real_epoch} ##########')
|
||||
|
||||
def on_train_epoch_end(self, trainer, pl_module):
|
||||
args = self.args
|
||||
if trainer.is_global_zero: # logging & save state_dict
|
||||
if (args.epoch_save > 0 and trainer.current_epoch % args.epoch_save == 0) or trainer.current_epoch == args.epoch_count - 1:
|
||||
if args.data_type == 'wds_img':
|
||||
raw_dict = pl_module.state_dict()
|
||||
to_save_dict = {}
|
||||
for k in raw_dict:
|
||||
if k.startswith('encoder.') or k.startswith('decoder.'):
|
||||
to_save_dict[k] = raw_dict[k]
|
||||
else:
|
||||
to_save_dict = pl_module.state_dict()
|
||||
|
||||
if args.lora:
|
||||
enable_time_finetune = 'time' in LORA_CONFIG["parts"]
|
||||
enable_ln_finetune = 'ln' in LORA_CONFIG["parts"]
|
||||
lora_dict = {}
|
||||
for name, state in to_save_dict.items():
|
||||
if ('.lora_' in name
|
||||
or (enable_time_finetune and '.time_' in name)
|
||||
or (enable_ln_finetune and '.ln' in name)):
|
||||
lora_dict[name] = state
|
||||
to_save_dict = lora_dict
|
||||
|
||||
try:
|
||||
my_save(
|
||||
to_save_dict,
|
||||
f"{args.proj_dir}/rwkv-{args.epoch_begin + trainer.current_epoch}.pth",
|
||||
)
|
||||
except Exception as e:
|
||||
print('Error\n\n', e, '\n\n')
|
||||
trainer.my_log.write(f"{args.epoch_begin + trainer.current_epoch} {trainer.my_epoch_loss:.6f} {math.exp(trainer.my_epoch_loss):.4f} {trainer.my_lr:.8f} {datetime.datetime.now()} {trainer.current_epoch}\n")
|
||||
trainer.my_log.flush()
|
||||
|
||||
trainer.my_loss_sum = 0
|
||||
trainer.my_loss_count = 0
|
||||
|
||||
|
||||
@rank_zero_only
|
||||
def generate_init_weight(model, init_weight_name):
|
||||
mm = model.generate_init_weight()
|
||||
|
||||
if model.args.my_pile_stage == 1:
|
||||
if len(model.args.load_model) > 0:
|
||||
print(f"Combine weights from {model.args.load_model}...")
|
||||
load_dict = torch.load(model.args.load_model, map_location="cpu")
|
||||
for k in load_dict:
|
||||
assert k in mm
|
||||
src = load_dict[k]
|
||||
try:
|
||||
mm[k] = src.reshape(mm[k].shape)
|
||||
except:
|
||||
tmp = mm[k].squeeze().clone()
|
||||
print(k, src.shape, '-->', mm[k].shape)
|
||||
ss = src.shape[0]
|
||||
dd = tmp.shape[0]
|
||||
for i in range(dd):
|
||||
pos = i / dd * ss
|
||||
if pos >= ss - 1:
|
||||
tmp[i] = src[ss-1]
|
||||
else:
|
||||
p0 = int(math.floor(pos))
|
||||
ii = pos - p0
|
||||
tmp[i] = src[p0] * (1-ii) + src[p0+1] * (ii)
|
||||
mm[k] = tmp.reshape(mm[k].shape)
|
||||
sss = src.squeeze().float().cpu().numpy()
|
||||
print(sss[:10], '...', sss[-10:])
|
||||
mmm = mm[k].squeeze().float().cpu().numpy()
|
||||
print(mmm[:10], '...', mmm[-10:])
|
||||
|
||||
print(f"Save to {init_weight_name}...")
|
||||
torch.save(mm, init_weight_name)
|
||||
|
||||
if model.args.my_pile_stage == 1:
|
||||
print("Done. Now go for stage 2.")
|
||||
exit(0)
|
||||
130
finetune/lora/src/utils.py
vendored
Normal file
130
finetune/lora/src/utils.py
vendored
Normal file
@@ -0,0 +1,130 @@
|
||||
import json, time, random, os
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.nn import functional as F
|
||||
|
||||
time_slot = {}
|
||||
time_ref = time.time_ns()
|
||||
|
||||
def record_time(name):
|
||||
if name not in time_slot:
|
||||
time_slot[name] = 1e20
|
||||
tt = (time.time_ns() - time_ref) / 1e9
|
||||
if tt < time_slot[name]:
|
||||
time_slot[name] = tt
|
||||
|
||||
class TOKENIZER():
|
||||
def __init__(self, WORD_NAME, UNKNOWN_CHAR='\ue083'):
|
||||
if 'list' in str(type(WORD_NAME)):
|
||||
self.charMode = False
|
||||
if WORD_NAME[0] == WORD_NAME[1]:
|
||||
from transformers import PreTrainedTokenizerFast
|
||||
self.tokenizer = PreTrainedTokenizerFast(tokenizer_file=WORD_NAME[0])
|
||||
else:
|
||||
from transformers import GPT2TokenizerFast
|
||||
self.tokenizer = GPT2TokenizerFast(WORD_NAME[0], WORD_NAME[1])
|
||||
self.vocab_size = len(self.tokenizer)
|
||||
else:
|
||||
self.charMode = True
|
||||
with open(WORD_NAME + '.json', "r", encoding="utf-16") as result_file:
|
||||
self.word_table = json.load(result_file)
|
||||
|
||||
self.vocab_size = len(self.word_table)
|
||||
|
||||
self.stoi = {v: int(k) for k, v in self.word_table.items()}
|
||||
self.itos = {int(k): v for k, v in self.word_table.items()}
|
||||
|
||||
self.UNKNOWN_CHAR = self.stoi[UNKNOWN_CHAR]
|
||||
|
||||
def refine_context(self, context):
|
||||
context = context.strip().split('\n')
|
||||
for c in range(len(context)):
|
||||
context[c] = context[c].strip().strip('\u3000').strip('\r')
|
||||
context = list(filter(lambda c: c != '', context))
|
||||
context = '\n' + ('\n'.join(context)).strip()
|
||||
if context == '':
|
||||
context = '\n'
|
||||
return context
|
||||
|
||||
def sample_logits(self, out, x, ctx_len, temperature=1.0, top_p_usual=None, top_p_newline=None):
|
||||
# out[self.UNKNOWN_CHAR] = -float('Inf')
|
||||
lastChar = int(x[-1])
|
||||
|
||||
probs = F.softmax(out, dim=-1)
|
||||
|
||||
if self.charMode:
|
||||
if self.itos[lastChar] == '\n':
|
||||
top_p = top_p_newline
|
||||
else:
|
||||
top_p = top_p_usual
|
||||
else:
|
||||
top_p = top_p_usual
|
||||
|
||||
if os.environ["RWKV_RUN_DEVICE"] == "cpu":
|
||||
probs = probs.numpy()
|
||||
sorted_probs = np.sort(probs)[::-1]
|
||||
cumulative_probs = np.cumsum(sorted_probs)
|
||||
cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
|
||||
probs[probs < cutoff] = 0
|
||||
if temperature != 1.0:
|
||||
probs = probs.pow(1.0 / temperature)
|
||||
probs = probs / np.sum(probs)
|
||||
out = np.random.choice(a=len(probs), p=probs)
|
||||
return out
|
||||
else:
|
||||
sorted_probs = torch.sort(probs, descending=True)[0]
|
||||
cumulative_probs = torch.cumsum(sorted_probs, dim=-1).cpu().numpy()
|
||||
cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
|
||||
probs[probs < cutoff] = 0
|
||||
if temperature != 1.0:
|
||||
probs = probs.pow(1.0 / temperature)
|
||||
out = torch.multinomial(probs, num_samples=1)[0]
|
||||
return out
|
||||
|
||||
def MaybeIsPrime(number):
|
||||
if FermatPrimalityTest(number) and MillerRabinPrimalityTest(number):
|
||||
return True
|
||||
else:
|
||||
return False
|
||||
|
||||
|
||||
def FermatPrimalityTest(number):
|
||||
if number > 1:
|
||||
for time in range(3):
|
||||
randomNumber = random.randint(2, number) - 1
|
||||
if pow(randomNumber, number - 1, number) != 1:
|
||||
return False
|
||||
return True
|
||||
else:
|
||||
return False
|
||||
|
||||
|
||||
def MillerRabinPrimalityTest(number):
|
||||
if number == 2:
|
||||
return True
|
||||
elif number == 1 or number % 2 == 0:
|
||||
return False
|
||||
oddPartOfNumber = number - 1
|
||||
timesTwoDividNumber = 0
|
||||
while oddPartOfNumber % 2 == 0:
|
||||
oddPartOfNumber = oddPartOfNumber // 2
|
||||
timesTwoDividNumber = timesTwoDividNumber + 1
|
||||
|
||||
for time in range(3):
|
||||
while True:
|
||||
randomNumber = random.randint(2, number) - 1
|
||||
if randomNumber != 0 and randomNumber != 1:
|
||||
break
|
||||
|
||||
randomNumberWithPower = pow(randomNumber, oddPartOfNumber, number)
|
||||
|
||||
if (randomNumberWithPower != 1) and (randomNumberWithPower != number - 1):
|
||||
iterationNumber = 1
|
||||
|
||||
while (iterationNumber <= timesTwoDividNumber - 1) and (randomNumberWithPower != number - 1):
|
||||
randomNumberWithPower = pow(randomNumberWithPower, 2, number)
|
||||
iterationNumber = iterationNumber + 1
|
||||
if randomNumberWithPower != (number - 1):
|
||||
return False
|
||||
|
||||
return True
|
||||
Reference in New Issue
Block a user