Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Minor usability improvements to tinyllama pretraining script #749

Merged
merged 8 commits into from
Nov 21, 2023
Merged

Minor usability improvements to tinyllama pretraining script #749

Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
85cff8e
tokens
awaelchli Nov 17, 2023
881d4d6
updates
awaelchli Nov 21, 2023
7cfdd5f
fix
awaelchli Nov 21, 2023
24e4cfb
fix
awaelchli Nov 21, 2023
10225ed
update
awaelchli Nov 21, 2023
0a92638
update
awaelchli Nov 21, 2023
72f4d87
Update pretrain/tinyllama.py
awaelchli Nov 21, 2023
b989a94
Update pretrain/tinyllama.py
awaelchli Nov 21, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 10 additions & 8 deletions pretrain/tinyllama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,8 +37,7 @@

global_batch_size = 512
learning_rate = 4e-4
micro_batch_size = 8
max_steps = 715256 * 2
max_tokens = int(3 * 1e12)
awaelchli marked this conversation as resolved.
Show resolved Hide resolved
awaelchli marked this conversation as resolved.
Show resolved Hide resolved
warmup_steps = 2000
log_step_interval = 2
eval_iters = 100
Expand Down Expand Up @@ -67,9 +66,9 @@

def setup(resume: Union[bool, Path] = False):
if use_wandb:
logger = WandbLogger(project="tinyllama", name="training", resume=(resume is not False))
logger = WandbLogger(project="tinyllama", name=name, resume=(resume is not False))
else:
logger = CSVLogger(root_dir="logs", name="tinyllama")
logger = CSVLogger(root_dir="logs", name=name)

if devices > 1:
strategy = FSDPStrategy(
Expand All @@ -78,6 +77,7 @@ def setup(resume: Union[bool, Path] = False):
state_dict_type="full",
limit_all_gathers=True,
cpu_offload=False,
sharding_strategy="HYBRID_SHARD",
)
else:
strategy = "auto"
Expand Down Expand Up @@ -155,6 +155,11 @@ def train(fabric, state, train_dataloader, val_dataloader, resume):
curr_iter = 0

for train_data in train_dataloader:

total_tokens = state["iter_num"] * micro_batch_size * model.config.block_size * fabric.world_size
if state["iter_num"] >= max_iters > 0 or total_tokens >= max_tokens > 0:
break

# resume data loader state by fast-forwarding through all seen batches
# drop this once streaming dataset supports proper resuming
if resume:
Expand All @@ -170,9 +175,6 @@ def train(fabric, state, train_dataloader, val_dataloader, resume):
f"Took {time.perf_counter() - total_t0:.1f} seconds to reach iteration {initial_iter}."
)

if state["iter_num"] >= max_iters:
break

# determine and set the learning rate for this iteration
lr = get_lr(state["iter_num"]) if decay_lr else learning_rate
for param_group in optimizer.param_groups:
Expand Down Expand Up @@ -215,7 +217,7 @@ def train(fabric, state, train_dataloader, val_dataloader, resume):
(t1 - total_t0) / (state["iter_num"] - initial_iter) * (max_iters - state["iter_num"])
),
"tokens": state["iter_num"] * micro_batch_size * model.config.block_size,
"total_tokens": state["iter_num"] * micro_batch_size * model.config.block_size * fabric.world_size,
"total_tokens": total_tokens,
}

fabric.print(
Expand Down
Loading