Skip to content

Add option to construct Llama model with Transformer Engine op fuser #13776

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

Merged
chtruong814 merged 3 commits into from
Jun 3, 2025
Merged

Add option to construct Llama model with Transformer Engine op fuser #13776

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
1f9856b
Add option to construct Llama model with TE op fuser
timmoon10 May 30, 2025
2b2c1bf
Merge branch 'main' into mcore-with-te-ops
chtruong814 May 31, 2025
3e48270
Merge branch 'main' into mcore-with-te-ops
gautham-kollu Jun 2, 2025
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
15 changes: 9 additions & 6 deletions nemo/collections/llm/gpt/model/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,12 +153,15 @@ def transformer_engine_layer_spec(config: "GPTConfig") -> ModuleSpec:
"""
from megatron.core.models.gpt import gpt_layer_specs

return gpt_layer_specs.get_gpt_layer_with_transformer_engine_spec(
num_experts=config.num_moe_experts,
moe_grouped_gemm=config.moe_grouped_gemm,
qk_layernorm=config.qk_layernorm,
fp8=bool(config.num_moe_experts and (config.fp8 is not None)),
)
kwargs = {
"num_experts": config.num_moe_experts,
"moe_grouped_gemm": config.moe_grouped_gemm,
"qk_layernorm": config.qk_layernorm,
"fp8": bool(config.num_moe_experts and (config.fp8 is not None)),
}
if getattr(config, "use_transformer_engine_op_fuser", None) is not None:

@gautham-kollu gautham-kollu May 30, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

@hemildesai Is backward compatibility necessary here ? Can we just do config.use_transformer_engine_op_fuser

@timmoon10 timmoon10 Jun 2, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

Alternative implementation that maintains backward compatibility and avoids the None option: https://github.com/timmoon10/NeMo/blob/5094732a45deeb5872ffd72c69ac9ccdb710460e/nemo/collections/llm/gpt/model/base.py#L162-L164

kwargs["use_te_op_fuser"] = config.use_transformer_engine_op_fuser
return gpt_layer_specs.get_gpt_layer_with_transformer_engine_spec(**kwargs)


def transformer_engine_full_layer_spec(config: "GPTConfig") -> ModuleSpec:
Expand Down
1 change: 1 addition & 0 deletions nemo/collections/llm/gpt/model/llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,7 @@ class LlamaConfig(GPTConfig):
persist_layer_norm: bool = True
bias_dropout_fusion: bool = True
apply_rope_fusion: bool = True
use_transformer_engine_op_fuser: Optional[bool] = None

@gautham-kollu gautham-kollu May 30, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

False seems more simpler ?



@dataclass
Expand Down