Spaces:
Runtime error
Runtime error
File size: 12,322 Bytes
bb5cd12 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 |
import argparse
import torch.distributed as dist
from transformers import GPT2TokenizerFast
import deepspeed
from pathlib import Path
import wandb
import os
import yaml
import torch
from collections import defaultdict
from torchtyping import TensorType
import gdown
def is_main():
if dist.is_initialized():
return dist.get_rank() == 0
return True
def print_main(*msg):
if is_main():
print(*msg)
def reduce_losses(losses):
"""Reduce a tensor of losses across all GPUs."""
if dist.is_initialized():
losses = losses.detach().clone()
# We use `all_reduce` because it is better supported than `reduce`
dist.all_reduce(losses, dist.ReduceOp.SUM)
return losses / dist.get_world_size()
else:
return losses
def cycle(loader):
while True:
for data in loader:
yield data
def get_tokenizer(name="gpt2", sequence_length=2048):
"""
Gets tokenizer for LM
"""
if name == "gpt2":
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
tokenizer.pad_token_id = tokenizer.eos_token
tokenizer.padding_side = "right"
tokenizer.model_max_length = sequence_length
# setup lm settings
tokenizer.add_special_tokens(
{"cls_token": "<|image|>"}
) # add special image token to tokenizer
else:
raise ValueError(f"Tokenizer {name} not recognized")
return tokenizer
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--config", type=str, required=False, help="path to your training config"
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local rank passed from distributed launcher",
)
deepspeed.add_config_arguments(parser)
args = parser.parse_args()
args.deepspeed = True
return args
def wandb_log(*args, **kwargs):
if is_main():
wandb.log(*args, **kwargs)
def wandb_init(*args, **kwargs):
if is_main():
wandb.init(*args, **kwargs)
def save_model(model_engine, save_dir, global_step, config=None):
os.makedirs(save_dir, exist_ok=True)
if config is not None:
config = config.to_dict()
with open(str(Path(save_dir) / "config.yml"), "w") as f:
yaml.dump(config, f, default_flow_style=False)
sd = {"global_step": global_step, "config": config}
model_engine.save_checkpoint(save_dir, client_state=sd)
def load_model(
model_engine, load_dir, load_optimizer_states=True, load_lr_scheduler_states=True
):
"""
Loads a model from disk and returns the global step to resume from if loading was successful, otherwise returns 0
"""
try:
load_path, sd = model_engine.load_checkpoint(
load_dir,
load_optimizer_states=load_optimizer_states,
load_lr_scheduler_states=load_lr_scheduler_states,
)
except AssertionError as e:
load_path = None
print(e)
if load_path is None:
print("Model loading failed - starting from global step 0")
return 0
return sd["global_step"]
def get_params_for_weight_decay_optimization(module, config):
"""
Divide params into with-weight-decay and without-weight-decay groups.
Layernorms and biases will have no weight decay but the rest will.
"""
weight_decay_params = {"params": []}
no_weight_decay_params = {"params": [], "weight_decay": 0.0}
blacklist_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
for module_ in module.modules():
if isinstance(module_, blacklist_modules) or (
config.weight_decay == 0.0
): # also include all parameters here if no weight decay is being done
no_weight_decay_params["params"].extend(
[
p
for p in list(module_._parameters.values())
if (p is not None) and p.requires_grad
]
)
else:
for n, p in list(module_._parameters.items()):
if p is not None and p.requires_grad:
if n != "bias":
weight_decay_params["params"].append(p)
else:
no_weight_decay_params["params"].append(p)
param_dict = {
pn: p
for pn, p in module.named_parameters()
if p is not None and p.requires_grad
}
assert len(no_weight_decay_params["params"]) + len(
weight_decay_params["params"]
) == len(
param_dict.keys()
), "Number of params in both groups != total number of trainable params"
if config.weight_decay == 0.0:
# only return a single param group if no weight decay is being used anyway
return [no_weight_decay_params]
return [weight_decay_params, no_weight_decay_params]
def configure_param_groups(model, config):
"""
Configures the different parameter groups in the model for training.
If a separate learning rate for the image prefix is provided, we separate out the groups here.
Additionally, parameters to which weight decay shouldn't be applied (layernorms / biases) are separated.
"""
if config.image_enc_lr is not None:
# get the params for the image prefix / proj
image_enc_params = get_params_for_weight_decay_optimization(
model.image_prefix.enc, config
)
for pdict in image_enc_params:
pdict["lr"] = config.image_enc_lr
image_proj_params = get_params_for_weight_decay_optimization(
model.image_prefix.proj, config
)
# get the params for layernorm if it exists
if config.use_image_embed_layernorm:
image_ln_params = get_params_for_weight_decay_optimization(
model.image_prefix.ln, config
)
image_proj_params += image_ln_params
# get the params for the lm
lm_params = get_params_for_weight_decay_optimization(model.lm, config)
# get params for class head if it exists
class_params = []
if hasattr(model, "class_head") and model.class_head is not None:
class_params = get_params_for_weight_decay_optimization(
model.class_head, config
)
all_params = []
for p in image_enc_params + lm_params + image_proj_params + class_params:
if p["params"]:
all_params.append(p)
else:
all_params = get_params_for_weight_decay_optimization(model, config)
# merge param dicts with shared lr / wd values
d = defaultdict(dict)
for param_group in all_params:
lr = param_group.get("lr", None)
wd = param_group.get("weight_decay", None)
key = f"lr_{lr}_wd_{wd}"
if d[key].get("params") is None:
d[key]["params"] = []
d[key]["params"].extend(param_group["params"])
if lr is not None:
d[key]["lr"] = lr
if wd is not None:
d[key]["weight_decay"] = wd
all_params = list(d.values())
n_params = sum([len(d["params"]) for d in all_params])
param_dict = {
pn: p for pn, p in model.named_parameters() if p is not None and p.requires_grad
}
assert n_params == len(
param_dict
), f"Some parameters are missing from param groups ({n_params} | {len(param_dict)})"
# if we're using multiple param groups, set the min / max lr for each one[]
# appropriately in deepspeed's scheduler
config.deepspeed_config_params["scheduler"]["params"]["warmup_min_lr"] = [
config.min_lr for _ in all_params
]
config.deepspeed_config_params["scheduler"]["params"]["warmup_max_lr"] = [
d.get("lr", config.lr) for d in all_params
]
return all_params
def count_parameters(model):
"""
Counts the number of trainable parameters in a model
"""
return sum(p.numel() for p in model.parameters() if p.requires_grad)
def log_table(name, model_outputs, gt_answers_list, global_step):
results_table = wandb.Table(columns=["model output", "ground truth(s)"])
for o, gt in zip(model_outputs, gt_answers_list):
results_table.add_data(o, gt)
wandb_log({f"eval/{name}": results_table}, step=global_step)
def get_world_info():
local_rank = int(os.environ["LOCAL_RANK"])
rank = int(os.environ["RANK"])
world_size = int(os.environ["WORLD_SIZE"])
return local_rank, rank, world_size
def init_distributed(backend="nccl"):
if not torch.distributed.is_initialized():
deepspeed.init_distributed(
dist_backend=backend, verbose=True, auto_mpi_discovery=True
)
local_rank, rank, world_size = get_world_info()
torch.cuda.set_device(local_rank)
return local_rank, rank, world_size
def collate_fn_classification(batch_data, seq_len=2048):
# for nvlr2: list(zip*(batch_data)) = [l_images, r_images, captions, class_labels]
image_list = list(zip(*batch_data))[:-2]
captions, class_labels = list(zip(*batch_data))[-2:]
# images, captions, class_labels = list(zip(*batch_data))
images_list = [torch.cat(image) for image in image_list]
captions = torch.cat([i[:, :seq_len] for i in captions])
class_labels = torch.stack(class_labels)
return images_list, captions, class_labels
def infer_checkpoint_path_from_config(config):
checkpoint_folder = config.save
if checkpoint_folder is None:
raise ValueError(
"No checkpoint folder specified in config. Please provide a checkpoint."
)
# check for 'latest' tag in checkpoint folder
if (Path(checkpoint_folder) / "latest").exists():
latest_ckpt = (Path(checkpoint_folder) / "latest").read_text().strip()
else:
raise ValueError(
f"No checkpoint found in {checkpoint_folder}. Please provide a checkpoint."
)
checkpoint_path = str(
Path(checkpoint_folder) / latest_ckpt / "mp_rank_00_model_states.pt"
)
if not Path(checkpoint_path).exists():
raise ValueError(
f"No checkpoint found in {checkpoint_path}. Please provide a checkpoint."
)
return checkpoint_path
# [tensor_1, tensor_2], tensor_3, tensor_4 = to_cuda_half([tensor_1, tensor_2], tensor_3, tensor_4)
# probably not working yet
def to_cuda_half(*args):
cuda_half_args = []
for x in args:
if isinstance(x, list):
x_cuda_half = to_cuda_half(*x)
cuda_half_args.append(x_cuda_half)
elif isinstance(x, tuple):
x_cuda_half = to_cuda_half(*x)
cuda_half_args.append(x_cuda_half)
else:
if x.dtype in [torch.float32, torch.float16]:
cuda_half_args.append(x.cuda().half())
elif x.dtype == torch.long:
cuda_half_args.append(x.cuda())
if len(cuda_half_args) == 1:
return cuda_half_args[0]
else:
return cuda_half_args
def build_labels(
input_embeddings: TensorType["b", "s", "d"],
captions: TensorType["b", "s"],
eos_token,
device,
) -> TensorType["b", "s"]:
"""
Builds labels from input embeddings.
Masks out the labels with -100 in positions up to the seq length of the embeddings, so loss is only computed for captions,
and not for image tokens.
Additionally, masks out everything *after* the first eos token.
"""
shape = input_embeddings.shape[:2] # b, s
assert captions.shape[1] >= shape[1]
# make sure to add masked embedding tokens in the appropriate locations in the labels
embedding_tokens = torch.zeros(shape, dtype=torch.int64).to(device) - 100
labels = torch.cat(
(embedding_tokens, captions[:, : -shape[1]]), dim=1
) # we truncate the sequence length of the captions, as they are always padded to the full sequence length
# mask out repeating eos tokens
for label in labels:
for k, token in enumerate(label):
if token == eos_token:
label[k + 1 :] = -100
break
return labels
def is_url(string):
return string.startswith("http://") or string.startswith("https://")
def download_checkpoint(checkpoint_url, save_as):
gdown.download(url = checkpoint_url, output = save_as, quiet=False)
|