Skip to content

[TB] add throughput graphs #210

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

Merged
stas00 merged 6 commits into from
Dec 7, 2021
Merged

[TB] add throughput graphs #210

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
bcb8bd8
add throughput graphs
stas00 Nov 27, 2021
8b2b055
add TFLOPs graph
stas00 Nov 27, 2021
fd4f31e
Merge remote-tracking branch 'origin/main' into tb-throughput
stas00 Nov 27, 2021
2cfd490
add estimated
stas00 Nov 29, 2021
4fdeeba
add tokens per sec
stas00 Nov 29, 2021
d7cbde2
Merge remote-tracking branch 'origin/main' into tb-throughput
stas00 Dec 7, 2021
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
34 changes: 34 additions & 0 deletions megatron/training.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,7 @@ def pretrain(train_valid_test_dataset_provider,
# Model, optimizer, and learning rate.
timers('model-and-optimizer-setup').start()
model, optimizer, lr_scheduler = setup_model_and_optimizer(model_provider)
args.parameters_in_billions_no_embedding = get_parameters_in_billions(model, exclude_embeddings=True)
print_rank_0(f'estimated model parameters: {get_parameters_in_billions(model)}')
print_rank_0(f'estimated model parameters without embeddings: {get_parameters_in_billions(model, exclude_embeddings=True)}')
timers('model-and-optimizer-setup').stop()
Expand Down Expand Up @@ -661,6 +662,26 @@ def add_to_logging(name):
if iteration % args.log_interval == 0:
elapsed_time = timers('interval-time').elapsed()
elapsed_time_per_iteration = elapsed_time / total_iterations

seq_len = args.curriculum_seqlen if args.curriculum_learning else args.seq_length

# throughput
samples_per_sec = batch_size / (elapsed_time_per_iteration * 1e3)
samples_per_sec_per_replica = samples_per_sec / args.data_parallel_size
tokens_per_sec = samples_per_sec * seq_len
tokens_per_sec_per_replica = tokens_per_sec / args.data_parallel_size

# general TFLOPs formula
# model_size_in_B * 4 * 2 * seqlen * global_batch_size / (time_in_sec_per_interation * total_gpus * 1e3)
#
# The factor of 4 is when used with activation check-pointing,
# otherwise it will be 3, but for 200B model, activation check-pointing will always be on.
#
# here:
# model_size_in_B * 4 * 2 * seqlen * batch_size / (time_in_msec_per_interation * total_gpus)
checkpoint_activations_factor = 4 if args.checkpoint_activations else 3
tflops = args.parameters_in_billions_no_embedding * checkpoint_activations_factor * 2 * seq_len * batch_size / (elapsed_time_per_iteration * args.world_size)

# only the last rank process has a non-None _GLOBAL_TENSORBOARD_WRITER
if writer and is_last_rank():
if args.log_timers_to_tensorboard:
Expand All @@ -670,6 +691,17 @@ def add_to_logging(name):
elapsed_time_per_iteration, args.consumed_train_samples)
writer.add_scalar('iteration-time/iteration-time vs tokens',

@thomasw21 thomasw21 Nov 29, 2021 •
edited
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should remove this one then as it doesn't make sense IMO, nor does the iteration-time/iteration-time vs samples. I think this just prints out how much time it took to run a step at a specific token timestamp.

@VictorSanh VictorSanh Nov 29, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

does this metric differ from tokens/samples per second?

@stas00 stas00 Nov 29, 2021 •
edited
Loading

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, other than being the inverse, it's different while the batch size is changing.

samples_per_sec = batch_size / (elapsed_time_per_iteration * 1e3)

elapsed_time_per_iteration, args.consumed_train_tokens)
writer.add_scalar('iteration-time/samples per second',

@thomasw21 thomasw21 Nov 29, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps a better metric would be to print out tokens per second? Typically samples per second is complicated with CL (samples change in terms of size).

@thomasw21 thomasw21 Nov 29, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually isn't that just the inverse of what we'd expect in iteration-time/iteration-time vs samples?

@VictorSanh VictorSanh Nov 29, 2021 •
edited
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we can have both. tokens per second is likely going to be impacted by CL anyway (shorter sequences)

@stas00 stas00 Nov 29, 2021

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually isn't that just the inverse of what we'd expect in iteration-time/iteration-time vs samples?

It isn't while bs is ramping up:

samples_per_sec = batch_size / (elapsed_time_per_iteration * 1e3)

@stas00 stas00 Nov 29, 2021

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

added:

  • tokens per second
  • tokens per second per replica

can remove the 2 samples per sec entries if wanted.

@thomasw21 thomasw21 Nov 30, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah sorry it's not the inverse right now. But I would have expected iteration-time vs samples to return the time it took to process a single sample, which is the inverse of the number of samples you process per second.

IMO, iteration-time vs samples doesn't bring much information besides it increases linearly during batch size rampup and then stays relatively constant. We should already see those behaviours in batch-size/batch-size vs samples. Though sometimes we notice bad allocations where the iteration time doubles (https://huggingface.co/bigscience/tr7c-1B3-alibi-logs/tensorboard) but we should be able to see the same thing with the new logs X per second.

@VictorSanh VictorSanh Nov 30, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fwiw, a lot of graphs are redundant in the TBs. It doesn't cost much to log (time and disk space) so I am not worried about information that can be inferred from another logging

samples_per_sec, args.iteration)
writer.add_scalar('iteration-time/samples per second per replica',
samples_per_sec_per_replica, args.iteration)
writer.add_scalar('iteration-time/tokens per second',
tokens_per_sec, args.iteration)
writer.add_scalar('iteration-time/tokens per second per replica',
tokens_per_sec_per_replica, args.iteration)
writer.add_scalar('iteration-time/TFLOPs per gpu (estimated)',
tflops, args.iteration)

log_string = ' iteration {:8d}/{:8d} |'.format(
iteration, args.train_iters)
log_string += ' consumed samples: {:12d} |'.format(
Expand Down Expand Up @@ -701,6 +733,8 @@ def add_to_logging(name):
total_loss_dict[skipped_iters_key])
log_string += ' number of nan iterations: {:3d} |'.format(
total_loss_dict[nan_iters_key])
log_string += ' samples per second: {:.3f} |'.format(samples_per_sec)
log_string += ' TFLOPs: {:.2f} |'.format(tflops)
total_loss_dict[advanced_iters_key] = 0
total_loss_dict[skipped_iters_key] = 0
total_loss_dict[nan_iters_key] = 0
Expand Down