github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-06 11:32:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

[misc] update pre-commit and run all files (#4752)

Browse Source 
* [misc] update pre-commit

* [misc] run pre-commit

* [misc] remove useless configuration files

* [misc] ignore cuda for clang-format
This commit is contained in:
Hongxin Liu
2023-09-19 14:20:26 +08:00
committed by GitHub
parent 3c6b831c26
commit 079bf3cb26
1268 changed files with 50037 additions and 38444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
examples/language/gpt/experiments/pipeline_parallel/model_zoo.py
View File

@@ -4,22 +4,25 @@ from transformers import GPT2Config, GPT2LMHeadModel
## Define the Model and Loss Based on Huggingface transformers GPT2LMHeadModel
class GPTLMModel(nn.Module):
def __init__(self,
hidden_size=768,
num_layers=12,
num_attention_heads=12,
max_seq_len=1024,
vocab_size=50257,
checkpoint=False):
def __init__(
self,
hidden_size=768,
num_layers=12,
num_attention_heads=12,
max_seq_len=1024,
vocab_size=50257,
checkpoint=False,
):
super().__init__()
self.checkpoint = checkpoint
self.config = GPT2Config(n_embd=hidden_size,
n_layer=num_layers,
n_head=num_attention_heads,
n_positions=max_seq_len,
n_ctx=max_seq_len,
vocab_size=vocab_size)
self.config = GPT2Config(
n_embd=hidden_size,
n_layer=num_layers,
n_head=num_attention_heads,
n_positions=max_seq_len,
n_ctx=max_seq_len,
vocab_size=vocab_size,
)
self.model = GPT2LMHeadModel(self.config)
if checkpoint:
self.model.gradient_checkpointing_enable()
@@ -70,4 +73,4 @@ def model_builder(model_size: str) -> callable:
raise TypeError(f"model_builder {model_size}")
__all__ = ['model_builder']
__all__ = ["model_builder"]

79
examples/language/gpt/experiments/pipeline_parallel/train_gpt_pp.py
View File

@@ -3,41 +3,34 @@ import time
from functools import partial
import torch
from model_zoo import model_builder
from torch import nn
from tqdm import tqdm
from colossalai.fx import ColoTracer
from colossalai.fx.passes.adding_split_node_pass import (
avgnode_split_pass,
gpipe_dp_split_pass,
split_with_split_nodes_pass,
)
from colossalai.fx.passes.adding_split_node_pass import gpipe_dp_split_pass, split_with_split_nodes_pass
from colossalai.fx.passes.meta_info_prop import MetaInfoProp
from colossalai.legacy.pipeline.middleware.adaptor import get_fx_topology
from colossalai.legacy.pipeline.rpc._pipeline_schedule import FillDrainPipelineEngine, OneFOneBPipelineEngine
from colossalai.legacy.pipeline.rpc._pipeline_schedule import FillDrainPipelineEngine
from colossalai.legacy.pipeline.rpc.utils import rpc_run
from colossalai.logging import disable_existing_loggers, get_dist_logger
from colossalai.nn.optimizer import HybridAdam
from model_zoo import model_builder
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--model_type', type=str, default="gpt2_medium")
parser.add_argument('--world_size', type=int, default=2)
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--dp_degree', type=int, default=1)
parser.add_argument('--tp_degree', type=int, default=1)
parser.add_argument('--num_microbatches', type=int, default=2)
parser.add_argument('--device', type=str, choices=['cpu', 'cuda'], default='cuda')
parser.add_argument('--master_addr', type=str, default='localhost')
parser.add_argument('--master_port', type=str, default='29011')
parser.add_argument('--num_worker_threads', type=int, default=128)
parser.add_argument("--model_type", type=str, default="gpt2_medium")
parser.add_argument("--world_size", type=int, default=2)
parser.add_argument("--batch_size", type=int, default=16)
parser.add_argument("--dp_degree", type=int, default=1)
parser.add_argument("--tp_degree", type=int, default=1)
parser.add_argument("--num_microbatches", type=int, default=2)
parser.add_argument("--device", type=str, choices=["cpu", "cuda"], default="cuda")
parser.add_argument("--master_addr", type=str, default="localhost")
parser.add_argument("--master_port", type=str, default="29011")
parser.add_argument("--num_worker_threads", type=int, default=128)
return parser.parse_args()
class GPTLMLoss(nn.Module):
def __init__(self):
super().__init__()
self.loss_fn = nn.CrossEntropyLoss()
@@ -63,16 +56,16 @@ def get_tflops(model_numel, batch_size, seq_len, step_time):
# Create annotated model which is noted where to be splitted.
def get_annotated_model(model, data_kwargs, num_stages, num_microbatches):
tracer = ColoTracer()
meta_args = {k: v.to('meta') for k, v in data_kwargs.items()}
meta_args = {k: v.to("meta") for k, v in data_kwargs.items()}
graph = tracer.trace(root=model, meta_args=meta_args)
gm = torch.fx.GraphModule(model, graph, model.__class__.__name__)
interp_meta_args = tuple([v.to('meta') for k, v in data_kwargs.items()])
interp_meta_args = tuple([v.to("meta") for k, v in data_kwargs.items()])
interp = MetaInfoProp(gm)
interp.run(*interp_meta_args)
#annotated_model = avgnode_split_pass(gm, num_stages)
annotated_model = gpipe_dp_split_pass(gm, num_stages, num_microbatches, mode='block', block_limit=0.01)
# annotated_model = avgnode_split_pass(gm, num_stages)
annotated_model = gpipe_dp_split_pass(gm, num_stages, num_microbatches, mode="block", block_limit=0.01)
return annotated_model
@@ -83,7 +76,7 @@ def create_partition_module(pp_rank: int, num_stages: int, model, data_kwargs, n
topo = get_fx_topology(top_module)
for submodule in split_submodules:
if isinstance(submodule, torch.fx.GraphModule):
setattr(submodule, '_topo', topo)
setattr(submodule, "_topo", topo)
return split_submodules[pp_rank + 1]
@@ -107,8 +100,10 @@ def run_master(args):
disable_existing_loggers()
logger = get_dist_logger()
logger.info(f"{args.model_type}, batch size {batch_size}, num stage {stage_num}, num microbatch {num_microbatches}",
ranks=[0])
logger.info(
f"{args.model_type}, batch size {batch_size}, num stage {stage_num}, num microbatch {num_microbatches}",
ranks=[0],
)
torch.manual_seed(123)
@@ -117,26 +112,28 @@ def run_master(args):
# warm up pipeline fx partition
input_ids, attn_mask = get_data(batch_size, SEQ_LEN, VOCAB_SIZE)
warmup_data_kwargs = {'input_ids': input_ids, 'attention_mask': attn_mask}
warmup_data_kwargs = {"input_ids": input_ids, "attention_mask": attn_mask}
# create model
logger.info(f'start model_builder')
logger.info(f"start model_builder")
model = model_builder(model_type)(checkpoint=False)
logger.info(f'end model_builder')
logger.info(f"end model_builder")
# set 1f1b pipeline engine
pp_engine = FillDrainPipelineEngine(partition_fn=partial(partition, model, warmup_data_kwargs, num_microbatches),
stage_num=stage_num,
num_microbatches=num_microbatches,
device=device,
chunk=1,
criterion=criterion,
metric=None,
checkpoint=False)
pp_engine = FillDrainPipelineEngine(
partition_fn=partial(partition, model, warmup_data_kwargs, num_microbatches),
stage_num=stage_num,
num_microbatches=num_microbatches,
device=device,
chunk=1,
criterion=criterion,
metric=None,
checkpoint=False,
)
partition_numels = pp_engine.remote_numels()
for rank, numel in partition_numels.items():
logger.info(f'{rank=} numel in the partition:{numel}')
logger.info(f"{rank=} numel in the partition:{numel}")
# build optim
pp_engine.initialize_optimizer(torch.optim.Adam, lr=1e-3)
@@ -145,7 +142,7 @@ def run_master(args):
for n in range(NUM_STEPS):
# we just use randomly generated data here
input_ids, attn_mask = get_data(batch_size, SEQ_LEN, VOCAB_SIZE)
batch = {'input_ids': input_ids, 'attention_mask': attn_mask}
batch = {"input_ids": input_ids, "attention_mask": attn_mask}
start = time.time()
outputs = pp_engine.forward_backward(batch=batch, labels=input_ids, forward_only=False)
@@ -175,6 +172,6 @@ def run_master(args):
logger.info(f"Avg TFLOPS per GPU is {sum(gpu_tflops) / world_size:.3f}")
if __name__ == '__main__':
if __name__ == "__main__":
args = parse_args()
rpc_run(args, run_master)