def run_predict_pipeline(args_opt):
device_id = int(os.getenv("DEVICE_ID"))
rank_id_str = os.getenv('RANK_ID', '0')
rank_id = int(rank_id_str[rank_id_str.rfind('-') + 1:])
print('rank_id:{}'.format(rank_id), "rank_id str:{}".format(rank_id_str))
device_id = int(os.getenv('DEVICE_ID'))
local_rank = rank_id
print('local_rank:{}, device id:{} start to run...'.format(
local_rank, device_id),
flush=True)
context.set_context(save_graphs=False,
mode=context.GRAPH_MODE,
device_target="Ascend",
device_id=device_id)
context.set_context(variable_memory_max_size="30GB")
if args_opt.distribute == "true":
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
print("device_id is {}, rank_id is {}, device_num is {}".format(
device_id, rank, device_num))
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL,
gradients_mean=False,
device_num=device_num,
full_batch=True,
loss_repeated_mean=True,
enable_parallel_optimizer=False,
pipeline_stages=args_opt.stage_num)
set_algo_parameters(elementwise_op_strategy_follow=True)
_set_multi_subgraphs()
else:
rank = 0
device_num = 1
model_parallel_num = args_opt.tensor_model_parallel_num
stage_device_num = int(device_num / args_opt.stage_num)
data_parallel_num = int(stage_device_num / model_parallel_num)
per_batch_size = args_opt.per_batch_size
batch_size = per_batch_size * data_parallel_num * args_opt.micro_size
config = PANGUALPHAConfig(data_parallel_num=data_parallel_num,
model_parallel_num=model_parallel_num,
batch_size=batch_size,
seq_length=args_opt.seq_length,
vocab_size=args_opt.vocab_size,
embedding_size=args_opt.embedding_size,
num_layers=args_opt.num_layers,
num_heads=args_opt.num_heads,
expand_ratio=4,
post_layernorm_residual=False,
dropout_rate=0.0,
compute_dtype=mstype.float16,
use_past=False,
self_layernorm=True,
forward_reduce_scatter=True,
stage_num=args_opt.stage_num,
micro_size=args_opt.micro_size,
word_emb_dp=False)
print("===config is: ", config, flush=True)
print("=====args_opt is: ", args_opt, flush=True)
per_stage_layers = config.num_layers // config.stage_num
per_stage_devices = device_num // config.stage_num
self_stage = rank_id // per_stage_devices
# all cards will save ckpt
train_stage_num = 16
train_device_num = 1024
train_mp = 16
ckpt_name = args_opt.load_ckpt_name
train_per_stage_num = train_device_num // train_stage_num
if config.mp != train_mp:
raise ValueError(
"the model parallel num is not equal to training model parallel num"
)
concat_stage_num = train_stage_num // config.stage_num
pangu_alpha = PANGUALPHAPipeline(config)
eval_net = EvalNet(pangu_alpha)
eval_net.set_train(False)
model_predict = Model(eval_net)
inputs_np = Tensor(np.ones(shape=(1, config.seq_length)), mstype.int32)
model_predict.infer_predict_layout(inputs_np)
print("======start load_distributed checkpoint", flush=True)
for i in range(self_stage * concat_stage_num,
(self_stage + 1) * concat_stage_num):
stage_position = local_rank % (config.mp * config.dp)
ckpt_rank = i * train_per_stage_num + stage_position # 訓練時候的rank號
ckpt_dir = os.path.join(args_opt.load_ckpt_path,
f"rank_{(ckpt_rank)}") # 命名還是以訓練時候的rank號命名
local_ckpt_file = os.path.join(ckpt_dir, ckpt_name)
if not os.path.exists(local_ckpt_file):
raise ValueError("Ckpt file not exits,", local_ckpt_file)
params_dict = load_checkpoint(local_ckpt_file, filter_prefix="adam")
load_param_into_net(eval_net, params_dict)
print("================load param ok=================", flush=True)
# here predict with fake input
model_predict.predict(inputs_np)
def run_predict_no_pipeline(args_opt):
device_id = int(os.getenv("DEVICE_ID"))
rank_id_str = os.getenv('RANK_ID', '0')
rank_id = int(rank_id_str[rank_id_str.rfind('-') + 1:])
print('rank_id:{}'.format(rank_id), "rank_id str:{}".format(rank_id_str))
device_id = int(os.getenv('DEVICE_ID'))
local_rank = rank_id
print('local_rank:{}, device id:{} start to run...'.format(
local_rank, device_id),
flush=True)
context.set_context(save_graphs=False,
mode=context.GRAPH_MODE,
device_target="Ascend",
device_id=device_id)
context.set_context(variable_memory_max_size="30GB")
if args_opt.distribute == "true":
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
print("device_id is {}, rank_id is {}, device_num is {}".format(
device_id, rank, device_num))
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL,
gradients_mean=False,
device_num=device_num,
full_batch=True,
loss_repeated_mean=True,
enable_parallel_optimizer=False,
strategy_ckpt_load_file=args_opt.strategy_load_ckpt_path,
pipeline_stages=args_opt.stage_num)
set_algo_parameters(elementwise_op_strategy_follow=True)
_set_multi_subgraphs()
else:
rank = 0
device_num = 1
model_parallel_num = args_opt.tensor_model_parallel_num
data_parallel_num = int(device_num / model_parallel_num)
per_batch_size = args_opt.per_batch_size
batch_size = per_batch_size * data_parallel_num
config = PANGUALPHAConfig(data_parallel_num=data_parallel_num,
model_parallel_num=model_parallel_num,
batch_size=batch_size,
seq_length=args_opt.seq_length,
vocab_size=args_opt.vocab_size,
embedding_size=args_opt.embedding_size,
num_layers=args_opt.num_layers,
num_heads=args_opt.num_heads,
expand_ratio=4,
post_layernorm_residual=False,
dropout_rate=0.0,
compute_dtype=mstype.float16,
use_past=False,
self_layernorm=True,
forward_reduce_scatter=True,
stage_num=args_opt.stage_num,
micro_size=args_opt.micro_size,
eod_reset=False,
word_emb_dp=True,
load_ckpt_path=args_opt.load_ckpt_path)
print("===config is: ", config, flush=True)
print("=====args_opt is: ", args_opt, flush=True)
ckpt_name = args_opt.load_ckpt_name
pangu_alpha = PANGUALPHA(config)
eval_net = EvalNet(pangu_alpha)
eval_net.set_train(False)
model_predict = Model(eval_net)
inputs_np = Tensor(np.ones(shape=(config.batch_size, config.seq_length)),
mstype.int32)
predict_layout = model_predict.infer_predict_layout(inputs_np)
print("======start load_distributed checkpoint", flush=True)
# For 2.6B and 13B models, the number of ckpt files is 512.
ckpt_name = 'filerted'
ckpt_file_list = [
os.path.join(args_opt.load_ckpt_path, f"{ckpt_name}_{ckpt_rank}.ckpt")
for ckpt_rank in range(0, 512)
]
print(f"Loading from path {ckpt_file_list[0]}", flush=True)
load_distributed_checkpoint(eval_net, ckpt_file_list, predict_layout)
print("================load param ok=================", flush=True)
from tokenization_jieba import JIEBATokenizer
from generate import generate
tokenizer = JIEBATokenizer(
os.path.join(args_opt.tokenizer_path, 'vocab.vocab'),
os.path.join(args_opt.tokenizer_path, 'vocab.model'))
sample = "今天是一个好天气"
tokenized_token = tokenizer.tokenize(sample)
start_sentence = tokenizer.convert_tokens_to_ids(tokenized_token)
input_ids = np.array(start_sentence).reshape(1, -1)
output_ids = generate(model_predict, input_ids, config.seq_length, 9)
output_samples = tokenizer.convert_ids_to_tokens(output_ids.tolist())
print('Output is:', output_samples, flush=True)