Skip to content

perf(rocm): enable Liger fused RMSNorm/SwiGLU for AMD DFlash configs #779

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
zhentaocc wants to merge 1 commit into
base: main
Choose a base branch
from
Open

perf(rocm): enable Liger fused RMSNorm/SwiGLU for AMD DFlash configs #779

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
3 changes: 3 additions & 0 deletions examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,9 @@ model:
trust_remote_code: true
embedding_key: "model.language_model.embed_tokens.weight"
torch_dtype: "bfloat16"
# Liger fused RMSNorm/SwiGLU: validated on MI355X (+10% tokens/s, -15% peak
# mem, loss parity). Requires the flex-backend fix (torch<2.11) to run.
use_liger_kernel: true
data:
hidden_states_path: "./cache/hidden_states/qwen3.5-4b-dflash-sharegpt"
max_length: 2048
Expand Down
3 changes: 3 additions & 0 deletions examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,9 @@ model:
trust_remote_code: true
embedding_key: "model.language_model.embed_tokens.weight"
torch_dtype: "bfloat16"
# Liger fused RMSNorm/SwiGLU: validated on MI355X (+10% tokens/s, -15% peak
# mem, loss parity). Requires the flex-backend fix (torch<2.11) to run.
use_liger_kernel: true
data:
train_data_path: "./cache/dataset/sharegpt_train.jsonl"
max_length: 2048
Expand Down
Loading