-
Notifications
You must be signed in to change notification settings - Fork 0
/
lightning_fsdp_inference.py
329 lines (295 loc) · 12.4 KB
/
lightning_fsdp_inference.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""The LightningModule - an nn.Module with many additional features."""
import os
from datetime import timedelta
import lightning.pytorch as pl
import torch
import transformers
from config import parse_args
from lightning.pytorch.loggers import CSVLogger
from lightning.pytorch.strategies import FSDPStrategy
from torch.distributed.fsdp import BackwardPrefetch, MixedPrecision
from train_utils import get_training_logger
# from transformers.models.llama.modeling_llama import LlamaDecoderLayer
from transformers.models.mixtral.modeling_mixtral import MixtralDecoderLayer, MixtralSdpaAttention, MixtralSparseMoeBlock
from transformers.models.mistral.modeling_mistral import MistralDecoderLayer
# import torch._dynamo
# torch._dynamo.config.suppress_errors = True
from sentencepiece import SentencePieceProcessor
from model import Transformer, TransformerBlock
def _load_model(checkpoint_path, device, precision):
from pathlib import Path
checkpoint_path = Path(checkpoint_path + "/model.pth")
with torch.device('meta'):
print(checkpoint_path.parent.name)
model = Transformer.from_name(checkpoint_path.parent.name)
checkpoint = torch.load(str(checkpoint_path), mmap=True, weights_only=True)
model.load_state_dict(checkpoint, assign=True)
model = model.to(device=device, dtype=precision)
return model # model.eval()
class LanguageModel(pl.LightningModule):
def __init__(
self, model_path, tokenizer, lr, warmup_ratio, weight_decay, enable_fp8
):
super().__init__()
self.model_path = model_path
self.tokenizer = tokenizer
self.model = None
self.lr = lr
self.warmup_ratio = warmup_ratio
self.weight_decay = weight_decay
self.mask_dict = {}
self.num_correct = 0
self.num_total = 0
self.enable_fp8 = enable_fp8
def configure_model(self):
# https://lightning.ai/docs/pytorch/stable/advanced/model_parallel/fsdp.html#speed-up-model-initialization
if self.model is not None:
return
from transformers import AutoModelForCausalLM
self.model = AutoModelForCausalLM.from_pretrained(
self.model_path, use_cache=False, torch_dtype=None, low_cpu_mem_usage=True,
)
# self.model = _load_model(self.model_path, device="cpu", precision=torch.float32)
if self.enable_fp8:
# Check the PyTorch version
torch_version = torch.__version__
# Convert the version string to a tuple of integers for comparison
version_tuple = tuple(map(int, torch_version.split('.')[:2]))
if version_tuple < (2, 4):
from train_utils import patch_torch
patch_torch()
from torchao.float8 import ( # precompute_float8_dynamic_scale_for_fsdp, # specific to fsdp2 + dynamic scaling, apply after each training loop iter
CastConfig,
Float8LinearConfig,
ScalingType,
convert_to_float8_training,
)
config = Float8LinearConfig(
enable_amax_init=True, # only needed for autocast + compile + FSDP + float8 delayed
enable_pre_and_post_forward=False, # only needed for autocast + compile + FSDP + float8 delayed
# enable_fsdp_float8_all_gather=True,
# enable_amax_init=False,
cast_config_input=CastConfig(scaling_type=ScalingType.DYNAMIC),
cast_config_weight=CastConfig(scaling_type=ScalingType.DELAYED),
cast_config_grad_output=CastConfig(scaling_type=ScalingType.DELAYED),
)
# convert all `torch.nn.Linear` modules to `Float8Linear`, specifying scaling
# type
def module_filter_fn(mod: torch.nn.Module, fqn: str):
# don't convert the output module
if "lm_head" in fqn:
return False
# don't convert linear modules with weight dimensions not divisible by 16
if isinstance(mod, torch.nn.Linear):
if "block_sparse_moe.gate" in fqn:
# if mod.in_features % 16 != 0 or mod.out_features % 16 != 0:
return False
return True
convert_to_float8_training(
self.model,
config=config,
module_filter_fn=module_filter_fn,
)
# self.model = torch.compile(self.model) # fullgraph=True
self.model.eval()
def forward(self, input_ids, attention_mask, labels=None, **kwargs):
return self.model(
input_ids=input_ids, attention_mask=attention_mask, labels=labels, **kwargs
)
def training_step(self, batch):
outputs = self.model(
input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
labels=batch["labels"],
)
loss = outputs.loss
self.log_dict(
{"train_loss": loss},
on_step=True,
on_epoch=False,
prog_bar=True,
logger=True,
rank_zero_only=True,
sync_dist=False,
)
return loss
def on_validation_epoch_end(self) -> None:
self.log(
"val_accuracy_epoch",
self.num_correct / self.num_total,
on_epoch=True,
prog_bar=True,
logger=True,
rank_zero_only=True,
sync_dist=True,
)
self.num_correct = 0
self.num_total = 0
def validation_step(self, batch):
outputs = self.model(
input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
labels=batch["labels"],
)
loss = outputs.loss
shift_logits = outputs.logits[..., :-1, :].contiguous()
shift_labels = batch["labels"][..., 1:].contiguous()
mask = shift_labels != -100
correct = (shift_logits.argmax(dim=-1) == shift_labels) & mask
self.num_correct += correct.sum().item()
self.num_total += mask.sum().item()
self.log_dict(
{"val_loss": loss},
on_step=True,
on_epoch=True,
prog_bar=True,
logger=True,
rank_zero_only=True,
sync_dist=True,
)
return loss
def predict_step(self, batch):
#if self.model.model.layers[0].input_layernorm.weight.numel() > 0:
# print("QQQ check model.model.layers[0].input_layernorm.weight max", self.model.model.layers[0].input_layernorm.weight)
#print("QQQ input_layernorm shape", self.model.model.layers[0].input_layernorm.weight.shape)
#print("QQQ input layernorm layer", self.model.model.layers[0].input_layernorm)
#if self.model.model.layers[0].self_attn.q_proj.weight.numel() > 0:
# print("QQQ check model.model.layers[0].self_attn_q_proj.weight max", self.model.model.layers[0].self_attn.q_proj.weight)
outputs = self.model(
input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
# labels=batch["labels"],
)
loss = 0.0 # outputs.loss
shift_logits = outputs.logits[..., -2:-1, :].contiguous()
shift_labels = batch["labels"][..., -1:].contiguous()
mask = shift_labels != -100
correct = (shift_logits.argmax(dim=-1) == shift_labels) & mask
self.num_correct += correct.sum().item()
self.num_total += mask.sum().item()
return loss
def on_predict_epoch_end(self) -> None:
print(f"Final pred_accuracy_epoch: {self.num_correct / self.num_total}")
def train():
pl.seed_everything(42)
args = parse_args()
# training_logger = get_training_logger(run_name=args.output_dir)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
layers = {MixtralDecoderLayer}
# layers = {TransformerBlock}
# layers = {MixtralSdpaAttention, MixtralSparseMoeBlock}
# layers = {LlamaDecoderLayer}
# layers = {MistralDecoderLayer}
import functools
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy
my_auto_wrap_policy = functools.partial(
size_based_auto_wrap_policy, min_num_params=20000
)
from torch.distributed.fsdp.api import CPUOffload, ShardingStrategy
fsdp_strategy = FSDPStrategy(
# cpu_offload=CPUOffload(offload_params=True),
auto_wrap_policy=layers,
# auto_wrap_policy=my_auto_wrap_policy,
sharding_strategy="FULL_SHARD",
# backward_prefetch=BackwardPrefetch.BACKWARD_PRE,
sync_module_states=True,
limit_all_gathers=True,
# state_dict_type="sharded",
# activation_checkpointing_policy=layers,
# we set mixed precision here instead of passing precision to PL trainer.
# precision="bf16-true" in PL trainer means pure half precision (including optimizer update etc.)
# while precision="bf16-mixed" results in unshard allgather performed in fp32:
# https:/Lightning-AI/pytorch-lightning/blobeieeccnhcruhfrrtcfhdevtlvvgnrhnkrjjhbbkdvegj
# /bf25167bbf64f50ba335aa759318946b21775cd2/src/lightning/fabric/plugins/precision/fsdp.py#L83
mixed_precision=MixedPrecision(
param_dtype=torch.bfloat16,
# reduce_dtype=torch.bfloat16,
# buffer_dtype=torch.float32
),
)
fsdp_strategy._timeout = timedelta(seconds=7200)
trainer = pl.Trainer(
accelerator="cuda",
strategy=fsdp_strategy,
devices=torch.cuda.device_count() if args.num_gpus is None else args.num_gpus,
enable_checkpointing=True,
default_root_dir=args.output_dir,
log_every_n_steps=1,
max_epochs=args.num_epoch,
logger=[
# training_logger,
CSVLogger(args.output_dir, flush_logs_every_n_steps=10),
],
callbacks=[],
val_check_interval=args.val_log_step,
# precision="bf16-true",
)
tokenizer = transformers.AutoTokenizer.from_pretrained(
args.model_path, padding_side="left", truncation_side="left"
)
tokenizer.pad_token = tokenizer.eos_token
if args.dataset == "cnn_dailymail":
from data_utils import CNNModule
data_module = CNNModule(
tokenizer=tokenizer,
data_path="/shared/public/data/cnn_dailymail/",
max_length=args.max_length,
batch_size=args.batch_size,
n_train=args.n_train,
n_val=args.n_val,
)
elif args.dataset == "mmlu":
from data_utils import MMLUModule
data_module = MMLUModule(
tokenizer=tokenizer,
data_path="/export/home/qsong/mmlu",
max_length=args.max_length,
batch_size=args.batch_size,
n_train=args.n_train,
n_val=args.n_val,
exp_len=args.max_length,
)
else:
raise ValueError("Unkown dataset.")
model = LanguageModel(
model_path=args.model_path,
tokenizer=tokenizer,
lr=args.lr,
warmup_ratio=args.warmup_ratio,
weight_decay=args.weight_decay,
enable_fp8=args.enable_fp8,
)
model = torch.compile(model)
# trainer.fit(model, datamodule=data_module)
# trainer.save_checkpoint(f"{args.output_dir}/model.ckpt")
# inference and record the time
# init_start_event = torch.cuda.Event(enable_timing=True)
# init_end_event = torch.cuda.Event(enable_timing=True)
# init_start_event.record()
with torch.inference_mode():
trainer.predict(model, datamodule=data_module)
# init_end_event.record()
print(model.model)
# torch.cuda.synchronize()
# if int(os.environ["RANK"]) == 0:
# torch.cuda.synchronize()
# print(
# f"CUDA event elapsed time: {init_start_event.elapsed_time(init_end_event) / 1000}sec"
# )
# print(f"{model}")
if __name__ == "__main__":
train()