Skip to content

Add: Multiple changes to original Axlearn repo to make it work. For DAG-01 #1

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

Closed
camiloCienet wants to merge 1 commit into from
Closed

Add: Multiple changes to original Axlearn repo to make it work. For DAG-01 #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 5 additions & 6 deletions axlearn/common/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1116,12 +1116,11 @@ def _run_step(
self._trainer_state, outputs = compiled_train_step_fn(self.trainer_state, input_batch)

n = self._config.log_every_n_steps or 100
if self.step % n == 0 or 0 <= self.step <= 5:
self._step_log(
"loss=%s aux=%s",
outputs["loss"],
jax.tree.map(lambda x: x.item() if x.ndim == 0 else f"T{x.shape}", outputs["aux"]),
)
self._step_log(
"loss=%s aux=%s",
outputs["loss"],
jax.tree.map(lambda x: x.item() if x.ndim == 0 else f"T{x.shape}", outputs["aux"]),
)

self.summary_writer(self.step, {"loss": outputs["loss"], **outputs["summaries"]})
# Aggregate summaries across evalers.
Expand Down
4 changes: 2 additions & 2 deletions axlearn/experiments/text/gpt/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -413,7 +413,7 @@ def adamw_decoupled_learner_config(
peak_lr: float,
max_step: int,
weight_decay: float,
lr_warmup_steps: int = 2000,
lr_warmup_steps: int = 50,
alpha: float = 0.1,
b1: float = 0.9,
b2: float = 0.95,
Expand Down Expand Up @@ -451,7 +451,7 @@ def adastar_learner_config(
*,
peak_lr: float,
max_step: int,
lr_warmup_steps: int = 2000,
lr_warmup_steps: int = 50,
alpha: float = 0.005,
weight_decay: float = 3.16e-4,
b1: float = 0.95,
Expand Down
7 changes: 5 additions & 2 deletions axlearn/experiments/text/gpt/fuji.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,10 +291,12 @@ def get_trainer_kwargs(
tokens_per_batch = TOKENS_PER_BATCH[version]
max_step = TOTAL_TOKENS[version][model_size] // tokens_per_batch
max_sequence_length = MAX_SEQUENCE_LENGTH[version]
train_batch_size = tokens_per_batch // max_sequence_length
# train_batch_size = tokens_per_batch // max_sequence_length
train_batch_size = 128

# Whether to use grouped query attention.
num_kv_heads = None
max_step = 300
if version in (Version.V3, Version.V3_TIKTOKEN):
num_kv_heads = 8

Expand Down Expand Up @@ -412,6 +414,7 @@ def get_trainer_kwargs(
max_sequence_length=max_sequence_length,
train_batch_size=train_batch_size,
max_step=max_step,
save_every_n_steps=100,
mesh_shape=mesh_shape_from_axes(data=-1, fsdp=8),
mesh_rules=(
# Step time:
Expand Down Expand Up @@ -504,7 +507,7 @@ def get_trainer_kwargs(
config_modifiers=[
MeshShapeModifier.default_config().set(
# fsdp=8 is also ok, only 2% slower step time.
mesh_shape=mesh_shape_from_axes(data=-1, fsdp=64)
mesh_shape=mesh_shape_from_axes(data=1, fsdp=128)
),
RematSpecModifier.default_config().set(
remat_policies={
Expand Down
Loading