Skip to content

fix: [nvbugs/5310520] disable embed_tokens's TP when DP enabled for llama model. #4758

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
merged 1 commit into from
May 30, 2025
Merged

fix: [nvbugs/5310520] disable embed_tokens's TP when DP enabled for llama model. #4758

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
31 changes: 20 additions & 11 deletions tensorrt_llm/_torch/models/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -678,9 +678,11 @@ def __init__(self, model_config: ModelConfig[LlamaConfig]):
self.aux_stream = torch.cuda.Stream()

if self.model_config.mapping.enable_attention_dp:
self.embed_tokens = Embedding(config.vocab_size,
config.hidden_size,
dtype=config.torch_dtype)
self.embed_tokens = Embedding(
config.vocab_size,
config.hidden_size,
dtype=config.torch_dtype,
)
else:
self.embed_tokens = Embedding(
config.vocab_size,
Expand Down Expand Up @@ -756,14 +758,21 @@ def __init__(self, model_config: ModelConfig[LlamaConfig]):
weight = lora_loader.embed_tokens
self.has_custom_embed_tokens = True

self.embed_tokens = Embedding(
vocab_size,
config.hidden_size,
dtype=config.torch_dtype,
mapping=model_config.mapping,
tensor_parallel_mode=TensorParallelMode.COLUMN,
gather_output=True,
)
if self.model_config.mapping.enable_attention_dp:
self.embed_tokens = Embedding(
vocab_size,
config.hidden_size,
dtype=config.torch_dtype,
)
else:
self.embed_tokens = Embedding(
vocab_size,
config.hidden_size,
dtype=config.torch_dtype,
mapping=model_config.mapping,
tensor_parallel_mode=TensorParallelMode.COLUMN,
gather_output=True,
)

if self.has_custom_embed_tokens:
with torch.no_grad():
Expand Down
9 changes: 0 additions & 9 deletions tensorrt_llm/_torch/models/modeling_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,11 +11,9 @@
from tqdm import tqdm

from tensorrt_llm.lora_manager import HfLoraLoader
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.models.convert_utils import split_matrix_tp

from ...logger import logger
from ...mapping import Mapping
from ...models.modeling_utils import QuantConfig
from ..attention_backend import AttentionMetadata
from ..distributed.communicator import pp_recv, pp_send
Expand Down Expand Up @@ -356,13 +354,6 @@ def __init__(self, model: TModel, *, config: ModelConfig[TConfig],
vocab_size,
hidden_size,
dtype=config.pretrained_config.torch_dtype,
mapping=Mapping(
world_size=1,
tp_size=1,
rank=0,
),
tensor_parallel_mode=None,
gather_output=False,
)
else:
# TODO(zhenhuanc): Currently lm_head Linear will not accept QuantConfig
Expand Down