Skip to content

Fix megatron grad norm explosion #648

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
Kovbo wants to merge 5 commits into
base: main
Choose a base branch
from
+4 −2
Open

Fix megatron grad norm explosion #648

Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
3b382dd
fix drad norm explosion
Kovbo Apr 9, 2026
df48179
fix types
Kovbo Apr 10, 2026
003d4da
Use aot_eager for shared-prefix block masks
FurtherAI Apr 13, 2026
103a6d7
Merge branch 'main' of github.com:OpenPipe/ART into fix/megatron-grad…
Kovbo Apr 14, 2026
2e4bb3d
add coordinate_descent_tuning back
Kovbo Apr 14, 2026
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
12 changes: 7 additions & 5 deletions src/art/megatron/flex_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
from torch import Tensor
from torch.nn.attention.flex_attention import (
BlockMask,
FlexKernelOptions,
create_block_mask,
flex_attention,
)
Expand All @@ -29,17 +30,16 @@ class FlexAttentionWrapper(torch.nn.Module):
"""Compiled `flex_attention` wrapper with Torchtitan-style inductor options."""

# Torchtitan inductor options for compiling flex attention.
_compile_options = {
_compile_options: dict[str, Any] = {
"max_autotune": True,
"coordinate_descent_tuning": True,
"triton.cudagraphs": False,
}
# Skip Inductor's flex_decoding specialization: it has triggered both
# shared-memory OOMs (triton_flex_decoding) and symbolic-shape assertion
# failures (create_flex_decoding_kernel). The regular flex_attention
# kernel autotunes against the actual hardware smem budget, so this
# stays GPU-agnostic.
_kernel_options = {
_kernel_options: FlexKernelOptions = {
"FORCE_USE_FLEX_ATTENTION": True,
}
_compiled_flex_attention: ClassVar = torch.compile(
Expand Down Expand Up @@ -72,14 +72,16 @@ def forward(
)


_compiled_create_block_mask = torch.compile(create_block_mask)
# Sequence-length churn can break the Inductor backend here. Keep this
# on aot_eager instead.
_compiled_create_block_mask = torch.compile(create_block_mask, backend="aot_eager")


def create_shared_prefix_attention_state(
group_ids: Tensor,
parent_ids: Tensor,
) -> SharedPrefixAttentionState:
"""Build a compiled block mask for ART shared-prefix packing.
"""Build a block mask for ART shared-prefix packing.

Initialized on the device of the group_ids tensor.

Expand Down