Skip to content

[NPU] FIX fused_linear_jsd ub overflow and OOM on NPU #1043

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
MAYUNHUI666 wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[NPU] FIX fused_linear_jsd ub overflow and OOM on NPU #1043

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
21a6747
【NPU】fixed oom error for benchmark_fused_linear_jsd.py
MAYUNHUI666 Jan 27, 2026
1e8e5f8
[NPU] FIX fused_linear_jsd ub overflow on NPU
MAYUNHUI666 Jan 27, 2026
74b77d1
Update benchmark_fused_linear_jsd.py
MAYUNHUI666 Jan 27, 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
10 changes: 8 additions & 2 deletions benchmark/scripts/benchmark_fused_linear_jsd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@

from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD
from liger_kernel.utils import infer_device
from liger_kernel.utils import get_total_gpu_memory

device = infer_device()

Expand Down Expand Up @@ -233,14 +234,19 @@ def full():

if __name__ == "__main__":
args = parse_benchmark_script_args()

gpu_memory_gbs = get_total_gpu_memory()
if gpu_memory_gbs >= 69:
vocab_size = 128256
else:
vocab_size = 65536

common_configs = {
"kernel_name": "fused_linear_jsd",
"x_name": "BT",
"x_label": "B x T",
"x_values": [2**i for i in range(10, 14)],
"kernel_providers": ["liger", "torch"],
"extra_benchmark_configs": [{"H": 4096, "V": 128256, "mode": "forward", "dtype": torch.bfloat16}],
"extra_benchmark_configs": [{"H": 4096, "V": vocab_size, "mode": "forward", "dtype": torch.bfloat16}],
Comment on lines +237 to +249
Copy link
Collaborator

@Tcc0403 Tcc0403 Jan 29, 2026

Choose a reason for hiding this comment

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

Let's lower the upper bound of x_values instead of vocab_size for now.

We can discuss what configs should be scalable if there's memory constraint, see #1051

"overwrite": args.overwrite,
}

Expand Down
2 changes: 1 addition & 1 deletion src/liger_kernel/ops/fused_linear_jsd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
# The optimal maximum block size depends on your hardware, your kernel, and your dtype
MAX_FUSED_SIZE = 4096 if infer_device() == "xpu" else 65536 // 2
MAX_FUSED_SIZE = 4096 if infer_device() == "npu" else 65536 // 2
Copy link
Collaborator

@Tcc0403 Tcc0403 Jan 29, 2026

Choose a reason for hiding this comment

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

append instead of replace

Suggested change
MAX_FUSED_SIZE = 4096 if infer_device() == "npu" else 65536 // 2
MAX_FUSED_SIZE = 4096 if infer_device() in ["npu", "xpu"] else 65536 // 2



def fused_linear_jsd_forward(
Expand Down