Skip to content

C
chatCSDN

CSDN 技术社区 / ai / chatCSDN

通知 107
Star 8
Fork 2
C
chatCSDN
提交 f25b7ce4 编写于 作者: U u010280923
浏览文件
操作

debug

上级 a06be6de
隐藏空白更改
内联 并排
Showing with 68 addition and 263 deletion
forward_demo.py
浏览文件 @ f25b7ce4
#!/usr/bin/env python import os, sys, torch
# -*- coding: utf-8 -*- import numpy as np
# @Time : 2023/3/1 11:54 np.set_printoptions(precision=4, suppress=True, linewidth=200)
# @Author : clong
# @File : train_sft.py
# current_path = os.path.dirname(os.path.abspath(__file__))
# sys.path.append(f'{current_path}/rwkv_pip_package/src')
# Tune these below (test True/False for all of them) to find the fastest setting:
# torch._C._jit_set_profiling_executor(True)
# torch._C._jit_set_profiling_mode(True)
# torch._C._jit_override_can_fuse_on_cpu(True)
# torch._C._jit_override_can_fuse_on_gpu(True)
# torch._C._jit_set_texpr_fuser_enabled(False)
# torch._C._jit_set_nvfuser_enabled(False)
######################################################################################################## ########################################################################################################
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
########################################################################################################
if __name__ == "__main__":
from argparse import ArgumentParser
from pytorch_lightning import Trainer
from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
rank_zero_info("########## work in progress ##########")
########################################################################################################
#
# example: train a simple L12-D768 RWKV on dummy data
#
# python train.py --load_model "" --wandb "" --proj_dir "out" \
# --data_file "" --data_type "dummy" --vocab_size 0 \
# --ctx_len 128 --epoch_steps 1000 --epoch_count 20 --epoch_begin 0 --epoch_save 10 \
# --micro_bsz 16 --n_layer 12 --n_embd 768 --pre_ffn 0 --head_qk 0 \
# --lr_init 6e-4 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.99 --adam_eps 1e-8 \
# --accelerator gpu --devices 1 --precision bf16 --strategy ddp_find_unused_parameters_false --grad_cp 0
# example: train a simple L6-D512 RWKV from scratch on enwik8
#
# python train.py --load_model "" --wandb "" --proj_dir "out" \
# --data_file "../data/enwik8" --data_type "utf-8" --vocab_size 0 \
# --ctx_len 512 --epoch_steps 5000 --epoch_count 500 --epoch_begin 0 --epoch_save 5 \
# --micro_bsz 12 --n_layer 6 --n_embd 512 --pre_ffn 0 --head_qk 0 \
# --lr_init 8e-4 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.99 --adam_eps 1e-8 \
# --accelerator gpu --devices 1 --precision bf16 --strategy ddp_find_unused_parameters_false --grad_cp 0
# example: fine-tune RWKV 1.5B using 8xA100 40G = 1.76it/s = 115k token/s, VRAM 37477M
#
# python train.py --load_model "/fsx/BlinkDL/CODE/FP16/out_1b2/all-8040.pth" --wandb "" --proj_dir "out" \
# --data_file "../data/train.npy" --data_type "numpy" --vocab_size 50277 \
# --ctx_len 1024 --epoch_steps 1000 --epoch_count 1000 --epoch_begin 0 --epoch_save 5 \
# --micro_bsz 8 --n_layer 24 --n_embd 2048 --pre_ffn 0 --head_qk 0 \
# --lr_init 1e-5 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.999 --adam_eps 1e-8 \
# --accelerator gpu --devices 8 --precision bf16 --strategy deepspeed_stage_2 --grad_cp 0
# example: fine-tune RWKV 1.5B using 1 GPU fp16 (VRAM 16G) NOTE: fp16 might overflow
#
# python train.py --load_model "/fsx/BlinkDL/CODE/FP16/out_1b2/all-8040.pth" --wandb "" --proj_dir "out" \
# --data_file "../data/train.npy" --data_type "numpy" --vocab_size 50277 \
# --ctx_len 1024 --epoch_steps 200 --epoch_count 1000 --epoch_begin 0 --epoch_save 1 \
# --micro_bsz 11 --n_layer 24 --n_embd 2048 --pre_ffn 0 --head_qk 0 \
# --lr_init 1e-5 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.999 --adam_eps 1e-8 \
# --accelerator gpu --devices 1 --precision fp16 --strategy deepspeed_stage_2_offload --grad_cp 1
parser = ArgumentParser()
parser.add_argument("--load_model", default="", type=str) # full path, with .pth
parser.add_argument("--wandb", default="", type=str) # wandb project name. if "" then don't use wandb
parser.add_argument("--proj_dir", default="out", type=str)
parser.add_argument("--random_seed", default="-1", type=int)
parser.add_argument("--data_file", default="", type=str)
parser.add_argument("--data_type", default="utf-8", type=str)
parser.add_argument("--vocab_size", default=0, type=int) # vocab_size = 0 means auto (for char-level LM and .txt data)
parser.add_argument("--ctx_len", default=1024, type=int)
parser.add_argument("--epoch_steps", default=1000, type=int) # a mini "epoch" has [epoch_steps] steps
parser.add_argument("--epoch_count", default=500, type=int) # train for this many "epochs". will continue afterwards with lr = lr_final
parser.add_argument("--epoch_begin", default=0, type=int) # if you load a model trained for x "epochs", set epoch_begin = x
parser.add_argument("--epoch_save", default=5, type=int) # save the model every [epoch_save] "epochs"
parser.add_argument("--micro_bsz", default=12, type=int) # micro batch size (batch size per GPU)
parser.add_argument("--n_layer", default=6, type=int)
parser.add_argument("--n_embd", default=512, type=int)
parser.add_argument("--dim_att", default=0, type=int)
parser.add_argument("--dim_ffn", default=0, type=int)
parser.add_argument("--pre_ffn", default=0, type=int) # replace first att layer by ffn (sometimes better)
parser.add_argument("--head_qk", default=0, type=int) # my headQK trick
parser.add_argument("--tiny_att_dim", default=0, type=int) # tiny attention dim
parser.add_argument("--tiny_att_layer", default=-999, type=int) # tiny attention @ which layer
parser.add_argument("--lr_init", default=6e-4, type=float) # 6e-4 for L12-D768, 4e-4 for L24-D1024, 3e-4 for L24-D2048
parser.add_argument("--lr_final", default=1e-5, type=float)
parser.add_argument("--warmup_steps", default=0, type=int) # try 50 if you load a model
parser.add_argument("--beta1", default=0.9, type=float)
parser.add_argument("--beta2", default=0.99, type=float) # use 0.999 when your model is close to convergence
parser.add_argument("--adam_eps", default=1e-8, type=float)
parser.add_argument("--grad_cp", default=0, type=int) # gradient checkpt: saves VRAM, but slower
parser.add_argument("--my_pile_stage", default=0, type=int) # my special pile mode
parser.add_argument("--my_pile_shift", default=-1, type=int) # my special pile mode - text shift
parser.add_argument("--my_pile_edecay", default=0, type=int)
parser.add_argument("--layerwise_lr", default=1, type=int) # layerwise lr for faster convergence (but slower it/s)
parser.add_argument("--ds_bucket_mb", default=200, type=int) # deepspeed bucket size in MB. 200 seems enough
# parser.add_argument("--cuda_cleanup", default=0, type=int) # extra cuda cleanup (sometimes helpful)
parser.add_argument("--my_img_version", default=0, type=str)
parser.add_argument("--my_img_size", default=0, type=int)
parser.add_argument("--my_img_bit", default=0, type=int)
parser.add_argument("--my_img_clip", default='x', type=str)
parser.add_argument("--my_img_clip_scale", default=1, type=float)
parser.add_argument("--my_img_l1_scale", default=0, type=float)
parser.add_argument("--my_img_encoder", default='x', type=str)
# parser.add_argument("--my_img_noise_scale", default=0, type=float)
parser.add_argument("--my_sample_len", default=0, type=int)
parser.add_argument("--my_ffn_shift", default=1, type=int)
parser.add_argument("--my_att_shift", default=1, type=int)
parser.add_argument("--my_pos_emb", default=0, type=int)
parser.add_argument("--load_partial", default=0, type=int)
parser.add_argument("--magic_prime", default=0, type=int)
parser.add_argument("--my_qa_mask", default=0, type=int)
parser.add_argument("--my_testing", default='', type=str)
parser = Trainer.add_argparse_args(parser)
args = parser.parse_args()
########################################################################################################
import os, warnings, math, datetime, sys, time
import numpy as np
import torch
from torch.utils.data import DataLoader
import deepspeed
import pytorch_lightning as pl
from pytorch_lightning import seed_everything
if args.random_seed >= 0:
print(f"########## WARNING: GLOBAL SEED {args.random_seed} THIS WILL AFFECT MULTIGPU SAMPLING ##########\n" * 3)
seed_everything(args.random_seed)
np.set_printoptions(precision=4, suppress=True, linewidth=200)
warnings.filterwarnings("ignore", ".*Consider increasing the value of the `num_workers` argument*")
warnings.filterwarnings("ignore", ".*The progress bar already tracks a metric with the*")
# os.environ["WDS_SHOW_SEED"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "false"
args.my_timestamp = datetime.datetime.today().strftime("%Y-%m-%d-%H-%M-%S")
args.enable_checkpointing = False
args.replace_sampler_ddp = False
args.logger = False
args.gradient_clip_val = 1.0
args.num_sanity_val_steps = 0
args.check_val_every_n_epoch = int(1e20)
args.log_every_n_steps = int(1e20)
args.max_epochs = -1 # continue forever
args.betas = (args.beta1, args.beta2)
args.real_bsz = int(args.num_nodes) * int(args.devices) * args.micro_bsz
os.environ["RWKV_T_MAX"] = str(args.ctx_len)
os.environ["RWKV_MY_TESTING"] = args.my_testing
if args.dim_att <= 0:
args.dim_att = args.n_embd
if args.dim_ffn <= 0:
args.dim_ffn = args.n_embd * 4
args.run_name = f"{args.vocab_size} ctx{args.ctx_len} L{args.n_layer} D{args.n_embd}"
if not os.path.exists(args.proj_dir):
os.makedirs(args.proj_dir)
samples_per_epoch = args.epoch_steps * args.real_bsz
tokens_per_epoch = samples_per_epoch * args.ctx_len
rank_zero_info(
f"""
############################################################################
#
# RWKV-4 {args.precision.upper()} on {args.num_nodes}x{args.devices} {args.accelerator.upper()}, bsz {args.num_nodes}x{args.devices}x{args.micro_bsz}={args.real_bsz}, {args.strategy} {'with grad_cp' if args.grad_cp > 0 else ''}
# #
# Data = {args.data_file} ({args.data_type}), ProjDir = {args.proj_dir} # Use '/' in model path, instead of '\'. Use ctx4096 models if you need long ctx.
# #
# Epoch = {args.epoch_begin} to {args.epoch_begin + args.epoch_count - 1} (will continue afterwards), save every {args.epoch_save} epoch # fp16 = good for GPU (!!! DOES NOT support CPU !!!)
# fp32 = good for CPU
# bf16 = worse accuracy, supports CPU
# xxxi8 (example: fp16i8) = xxx with int8 quantization to save 50% VRAM/RAM, slower, slightly less accuracy
# #
# Each "epoch" = {args.epoch_steps} steps, {samples_per_epoch} samples, {tokens_per_epoch} tokens # Read https://pypi.org/project/rwkv/ for Strategy Guide
# #
# Model = {args.n_layer} n_layer, {args.n_embd} n_embd, {args.ctx_len} ctx_len ########################################################################################################
# # set these before import RWKV
# Adam = lr {args.lr_init} to {args.lr_final}, warmup {args.warmup_steps} steps, beta {args.betas}, eps {args.adam_eps} os.environ['RWKV_JIT_ON'] = '1'
# os.environ["RWKV_CUDA_ON"] = '0' # if '1' then compile CUDA kernel for seq mode (much faster)
# Found torch {torch.__version__}, recommend 1.12.1+cu116 or newer os.environ["RWKV_T_MAX"] = '1024'
# Found deepspeed {deepspeed.__version__}, recommend 0.7.0 (faster than newer versions)
# Found pytorch_lightning {pl.__version__}, recommend 1.7.4 or newer from src.model import RWKV # pip install rwkv
# # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-169m/RWKV-4-Pile-169M-20220807-8023', strategy='cuda fp16')
############################################################################ # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-169m/RWKV-4-Pile-169M-20220807-8023', strategy='cuda fp16i8')
""" model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-169m/RWKV-4-Pile-169M-20220807-8023', strategy='cpu fp32')
) # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-169m/RWKV-4-Pile-169M-20220807-8023', strategy='cpu fp32 *3 -> cuda fp16 *6+')
rank_zero_info(str(vars(args)) + "\n") # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cpu fp32')
# model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cuda fp16')
assert args.data_type in ["utf-8", "utf-16le", "numpy", "binidx", "dummy", "wds_img", "uint16"] # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cuda fp16 *8 -> cpu fp32')
# model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cuda:0 fp16 -> cuda:1 fp16 -> cpu fp32 *1')
if args.lr_final == 0 or args.lr_init == 0: # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-1b5/RWKV-4-Pile-1B5-20220903-8040', strategy='cuda fp16 *6+')
rank_zero_info("\n\nNote: lr_final = 0 or lr_init = 0. Using linear LR schedule instead.\n\n") # model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-14b/RWKV-4-Pile-14B-20230213-8019', strategy='cuda fp16 *0+ -> cpu fp32 *1')
# model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-3b/RWKV-4-Pile-3B-20221110-ctx4096', strategy='cuda:0 fp16 *25 -> cuda:1 fp16')
assert args.precision in ["fp32", "tf32", "fp16", "bf16"]
os.environ["RWKV_FLOAT_MODE"] = args.precision out, state = model.forward([187, 510, 1563, 310, 247], None)
if args.precision == "fp32": print(out.detach().cpu().numpy()) # get logits
rank_zero_info("\n\nNote: you are using fp32 (very slow). Try bf16 / tf32 for faster training.\n\n") out, state = model.forward([187, 510], None)
if args.precision == "fp16": out, state = model.forward([1563], state) # RNN has state (use deepcopy to clone states)
rank_zero_info("\n\nNote: you are using fp16 (might overflow). Try bf16 / tf32 for stable training.\n\n") out, state = model.forward([310, 247], state)
print(out.detach().cpu().numpy()) # same result as above
os.environ["RWKV_JIT_ON"] = "1"
if "deepspeed_stage_3" in args.strategy: print('\n')
os.environ["RWKV_JIT_ON"] = "0"
from src.utils import PIPELINE, PIPELINE_ARGS
torch.backends.cudnn.benchmark = True pipeline = PIPELINE(model, "20B_tokenizer.json")
torch.backends.cudnn.enabled = True
if args.precision == "fp32": ctx = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
torch.backends.cudnn.allow_tf32 = False print(ctx, end='')
torch.backends.cuda.matmul.allow_tf32 = False
else: def my_print(s):
torch.backends.cudnn.allow_tf32 = True print(s, end='', flush=True)
torch.backends.cuda.matmul.allow_tf32 = True
# For alpha_frequency and alpha_presence, see "Frequency and presence penalties":
if "32" in args.precision: # https://platform.openai.com/docs/api-reference/parameter-details
args.precision = 32
elif args.precision == "fp16": args = PIPELINE_ARGS(temperature = 1.0, top_p = 0.7,
args.precision = 16 alpha_frequency = 0.25,
else: alpha_presence = 0.25,
args.precision = "bf16" token_ban = [0], # ban the generation of some tokens
token_stop = []) # stop generation whenever you see any token here
########################################################################################################
from src.trainer import train_callback, generate_init_weight
args.vocab_size = 50277
from src.model import RWKV
model = RWKV(args)
if len(args.load_model) == 0:
rank_zero_info(f"SFT must load model, please input ")
exit(1)
rank_zero_info(f"########## Loading {args.load_model}... ##########")
try:
load_dict = torch.load(args.load_model, map_location="cpu")
except:
rank_zero_info(f"Bad checkpoint {args.load_model}")
exit(1)
if args.load_partial == 1:
load_keys = load_dict.keys()
for k in model.state_dict():
if k not in load_keys:
load_dict[k] = model.state_dict()[k]
model.load_state_dict(load_dict)
trainer = Trainer.from_argparse_args(
args,
callbacks=[train_callback(args)],
)
if trainer.global_rank == 0:
for n in model.state_dict():
shape = model.state_dict()[n].shape
shape = [i for i in shape if i != 1]
if len(shape) > 1:
print(f"{str(shape[0]).ljust(5)} {str(shape[1]).ljust(5)} {n}")
else:
print(f"{str(shape[0]).ljust(5)} {n}")
if "deepspeed" in args.strategy:
trainer.strategy.config["zero_optimization"]["allgather_bucket_size"] = args.ds_bucket_mb * 1000 * 1000
trainer.strategy.config["zero_optimization"]["reduce_bucket_size"] = args.ds_bucket_mb * 1000 * 1000
model.deepspeed_offload()
seq = torch.randint(0, 50277, (1, 100))
model(seq)
import ipdb ########################################################################################################
ipdb.set_trace() # 1. set os.environ["RWKV_CUDA_ON"] = '1' if possible, for faster preprocess of a long ctx.
# 2. Reuse the state (use deepcopy to clone it) when you are running the same ctx multiple times.
pipeline.generate(ctx, token_count=200, args=args, callback=my_print)
print('\n')
\ No newline at end of file
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或