Skip to content

feat: Add TP to embed_tokens and lm_head for Gemma models #879

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
terrykong merged 1 commit into from
Aug 14, 2025
Merged

feat: Add TP to embed_tokens and lm_head for Gemma models #879

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
23 changes: 7 additions & 16 deletions nemo_rl/models/dtensor/parallelize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,8 +31,6 @@
from torch.distributed.tensor.parallel import (
ColwiseParallel,
ParallelStyle,
PrepareModuleInput,
PrepareModuleOutput,
RowwiseParallel,
SequenceParallel,
parallelize_module,
Expand Down Expand Up @@ -93,32 +91,27 @@ def _parallelize_gemma3(
model: Union[Gemma3ForCausalLM, Gemma3ForConditionalGeneration],
sequence_parallel: bool = False,
) -> dict[str, ParallelStyle]:
"""Parallelizes a Gemma3ForCausalLM model across data parallel dimensions.

Tensor parallelism is not supported for Gemma3 models because of tied word embeddings.
"""
"""Parallelizes a Gemma3ForCausalLM model across data and tensor parallel dimensions."""
if isinstance(model, Gemma3ForConditionalGeneration):
model_prefix = "model.language_model"
else:
model_prefix = "model"

# For gemma3 models, we don't include the model.embed_tokens and lm_head in the
# parallelization plans because they have tied weights.
base_model_tp_plan: dict[str, ParallelStyle] = {
f"{model_prefix}.embed_tokens": RowwiseParallel(input_layouts=Replicate()),
f"{model_prefix}.layers.*.self_attn.q_proj": ColwiseParallel(),
f"{model_prefix}.layers.*.self_attn.k_proj": ColwiseParallel(),
f"{model_prefix}.layers.*.self_attn.v_proj": ColwiseParallel(),
f"{model_prefix}.layers.*.self_attn.o_proj": RowwiseParallel(),
f"{model_prefix}.layers.*.mlp.up_proj": ColwiseParallel(),
f"{model_prefix}.layers.*.mlp.gate_proj": ColwiseParallel(),
f"{model_prefix}.layers.*.mlp.down_proj": RowwiseParallel(),
"lm_head": ColwiseParallel(output_layouts=Shard(-1), use_local_output=False),
}

base_model_sp_plan = {
f"{model_prefix}.embed_tokens": PrepareModuleOutput(
output_layouts=Replicate(),
desired_output_layouts=Shard(1),
use_local_output=False,
f"{model_prefix}.embed_tokens": RowwiseParallel(
input_layouts=Replicate(), output_layouts=Shard(1)
),
f"{model_prefix}.rotary_emb": RotaryEmbedParallel(use_local_output=True),
f"{model_prefix}.rotary_emb_local": RotaryEmbedParallel(use_local_output=True),
Expand All @@ -133,10 +126,8 @@ def _parallelize_gemma3(
),
f"{model_prefix}.layers.*.post_feedforward_layernorm": SequenceParallel(),
f"{model_prefix}.norm": SequenceParallel(),
"lm_head": PrepareModuleInput(
input_layouts=(Shard(1),),
desired_input_layouts=(Replicate(),),
use_local_output=True,
"lm_head": ColwiseParallel(
input_layouts=Shard(1), output_layouts=Shard(-1), use_local_output=False
),
}

Expand Down
Loading