Skip to content

[DO NOT MERGE] correctly measure average step w/ block_until_ready #46

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

Draft
emergenz wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

[DO NOT MERGE] correctly measure average step w/ block_until_ready #46

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
55a787f
feat: correctly measure average step w/ block_until_ready
emergenz Jun 27, 2025
a7c5361
Merge branch 'main' into correct-step-time-logging
emergenz Jul 10, 2025
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
15 changes: 14 additions & 1 deletion train_dynamics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
from dataclasses import dataclass, field
import os
import time

import einops
from flax.training.train_state import TrainState
Expand Down Expand Up @@ -70,6 +71,7 @@ class Args:
log_checkpoint_interval: int = 25000
log_checkpoint_keep_period: int = 20000
log_gradients: bool = False
time_measurement_interval: int = 50


args = tyro.cli(Args)
Expand Down Expand Up @@ -242,6 +244,7 @@ def train_step(state, inputs):
prefetch_buffer_size=1,
seed=args.seed,
)
step = 0
initial_state = grain_dataloader._create_initial_state()
grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)

Expand All @@ -268,6 +271,7 @@ def train_step(state, inputs):

# --- TRAIN LOOP ---
dataloader = (jax.make_array_from_process_local_data(videos_sharding, elem) for elem in grain_iterator) # type: ignore
start_time = time.time()
while step < args.num_steps:
for videos in dataloader:
# --- Train step ---
Expand All @@ -280,7 +284,16 @@ def train_step(state, inputs):
mask_rng=_rng_mask,
)
train_state, loss, recon, metrics = train_step(train_state, inputs)
print(f"Step {step}, loss: {loss}")
if step % args.time_measurement_interval == 0:
jax.block_until_ready(train_state)
elapsed_time = (time.time() - start_time)
avg_step_time = elapsed_time / args.time_measurement_interval
print(f"Step {step}, loss: {loss}, avg step time: {avg_step_time:.2f}s")
if args.log and jax.process_index() == 0:
wandb.log({"avg_step_time_s": avg_step_time, "step": step})
start_time = time.time()
else:
print(f"Step {step}, loss: {loss}")
step += 1

# --- Logging ---
Expand Down
14 changes: 13 additions & 1 deletion train_lam.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
from dataclasses import dataclass, field
import os
import time

import einops
from flax.training import orbax_utils
Expand Down Expand Up @@ -59,6 +60,7 @@ class Args:
log_image_interval: int = 250
ckpt_dir: str = ""
log_checkpoint_interval: int = 10000
time_measurement_interval: int = 50
log_checkpoint_keep_period: int = 20000


Expand Down Expand Up @@ -257,6 +259,7 @@ def train_step(state, inputs, action_last_active):

# --- TRAIN LOOP ---
dataloader = (jax.make_array_from_process_local_data(videos_sharding, elem) for elem in grain_iterator) # type: ignore
start_time = time.time()
print(f"Starting training from step {step}...")
while step < args.num_steps:
for videos in dataloader:
Expand All @@ -267,7 +270,16 @@ def train_step(state, inputs, action_last_active):
train_state, loss, recon, action_last_active, metrics = train_step(
train_state, inputs, action_last_active
)
print(f"Step {step}, loss: {loss}")
if step % args.time_measurement_interval == 0:
jax.block_until_ready(train_state)
elapsed_time = (time.time() - start_time)
avg_step_time = elapsed_time / args.time_measurement_interval
print(f"Step {step}, loss: {loss}, avg step time: {avg_step_time:.2f}s")
if args.log and jax.process_index() == 0:
wandb.log({"avg_step_time_s": avg_step_time, "step": step})
start_time = time.time()
else:
print(f"Step {step}, loss: {loss}")
step += 1

# --- Logging ---
Expand Down
14 changes: 13 additions & 1 deletion train_tokenizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
from dataclasses import dataclass, field
import os
import time

import einops
from flax.training import orbax_utils
Expand Down Expand Up @@ -60,6 +61,7 @@ class Args:
log_checkpoint_interval: int = 10000
log_checkpoint_keep_period: int = 20000
log_gradients: bool = False
time_measurement_interval: int = 50


args = tyro.cli(Args)
Expand Down Expand Up @@ -249,6 +251,7 @@ def train_step(state, inputs):
print(f"Restored dataloader and model state from step {step}")

# --- TRAIN LOOP ---
start_time = time.time()
dataloader = (jax.make_array_from_process_local_data(videos_sharding, elem) for elem in grain_iterator) # type: ignore
print(f"Starting training from step {step}...")
while step < args.num_steps:
Expand All @@ -258,7 +261,16 @@ def train_step(state, inputs):

inputs = dict(videos=videos, rng=_rng, dropout_rng=_rng_dropout)
train_state, loss, recon, metrics = train_step(train_state, inputs)
print(f"Step {step}, loss: {loss}")
if step % args.time_measurement_interval == 0:
jax.block_until_ready(train_state)
elapsed_time = (time.time() - start_time)
avg_step_time = elapsed_time / args.time_measurement_interval
print(f"Step {step}, loss: {loss}, avg step time: {avg_step_time:.2f}s")
if args.log and jax.process_index() == 0:
wandb.log({"avg_step_time_s": avg_step_time, "step": step})
start_time = time.time()
else:
print(f"Step {step}, loss: {loss}")
step += 1

# --- Logging ---
Expand Down