Skip to content

[https://nvbugs/5517023][fix] Pass allreduce strategy and force NCCL on pre-Blackwell arch #7768

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
Superjomn merged 1 commit into from
Sep 17, 2025
Merged

[https://nvbugs/5517023][fix] Pass allreduce strategy and force NCCL on pre-Blackwell arch #7768

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
10 changes: 8 additions & 2 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 @@ -11,7 +11,8 @@
from transformers.models.llama4.modeling_llama4 import Llama4MultiModalProjector

from tensorrt_llm._torch.distributed import (AllReduce, AllReduceFusionOp,
AllReduceParams, MoEAllReduce)
AllReduceParams, AllReduceStrategy,
MoEAllReduce)
from tensorrt_llm._torch.models.checkpoints.base_weight_mapper import \
BaseWeightMapper
from tensorrt_llm._utils import get_sm_version
Expand Down Expand Up @@ -646,7 +647,12 @@ def __init__(
eps=config.rms_norm_eps,
dtype=config.torch_dtype)

self.all_reduce = AllReduce(mapping=model_config.mapping)
# TODO: This is a temporary fix to disable oneshot kernel for pre-Blackwell arch to avoid perf regressions
self.all_reduce = AllReduce(
strategy=model_config.allreduce_strategy
if get_sm_version() >= 100 else AllReduceStrategy.NCCL,
mapping=model_config.mapping,
)

self.next_layer_layernorm: RMSNorm = None
self.next_attn: LlamaAttention = None
Expand Down