Skip to content

Use Tokamax's representative group sizes. #3434

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

Open
copybara-service wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Use Tokamax's representative group sizes. #3434

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
5 changes: 1 addition & 4 deletions src/maxtext/layers/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -900,10 +900,7 @@ def gmm(
if self.config.using_pipeline_parallelism and self.config.pipeline_fsdp_ag_per_repeat:
tokamax_group_sizes = group_sizes
else:
tokamax_group_sizes = tokamax.RaggedDotGroupSizes(
group_sizes,
max_utils.generate_representative_group_sizes(inputs.shape[0], kernel.shape[0]),
)
tokamax_group_sizes = tokamax.RaggedDotGroupSizes(group_sizes, self.config.wi_tile_fwd_batch_seq)
pad_length = self.config.wi_tile_fwd_batch_seq
hs_shape = inputs.shape
# pad length is the 1st dimension of tiling size in gmm call
Expand Down
6 changes: 1 addition & 5 deletions src/maxtext/models/deepseek_batchsplit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,6 @@
from maxtext.layers import attention_op
from maxtext.layers import moe as moe_lib
from maxtext.layers import quantizations
from maxtext.utils import max_utils
import qwix.pallas as qpl
import tokamax

Expand Down Expand Up @@ -970,10 +969,7 @@ def gmm(
output = tokamax.ragged_dot(
lhs=inputs,
rhs=kernel,
group_sizes=tokamax.RaggedDotGroupSizes(
group_sizes,
max_utils.generate_representative_group_sizes(inputs.shape[0], kernel.shape[0]),
),
group_sizes=tokamax.RaggedDotGroupSizes(group_sizes, config.wi_tile_fwd_batch_seq),
precision=jax.lax.Precision.DEFAULT,
preferred_element_type=preferred_element_type,
implementation="mosaic",
Expand Down
10 changes: 0 additions & 10 deletions src/maxtext/utils/max_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1078,13 +1078,3 @@ def transformer_engine_context():
yield
except (ImportError, AttributeError):
yield


def generate_representative_group_sizes(target_m: int, g: int) -> tuple[int, ...]:
"""Generate group sizes for a given target m."""
np.random.seed(0)
repr_val = np.random.uniform(size=(g,))
repr_val = np.random.binomial(1, 0.9, (g,)) * repr_val
repr_val = np.int32((repr_val / np.sum(repr_val)) * target_m)
repr_val[0] += target_m - np.sum(repr_val)
return tuple(map(int, repr_val))
Loading