Skip to content

[https://nvbugs/6450338][fix] Default to FP32 accumulation in allreduce fusion kernels #16753

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
dongfengy wants to merge 1 commit into
base: main
Choose a base branch
from
+4 −2
Open

[https://nvbugs/6450338][fix] Default to FP32 accumulation in allreduce fusion kernels #16753

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
6 changes: 4 additions & 2 deletions cpp/tensorrt_llm/kernels/communicationKernels/allReduceFusionKernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -719,9 +719,11 @@ void allreduce_fusion_kernel_launcher(AllReduceFusionParams const& params)

bool use_fp32_acc()
{
// we use fp16 acc type by default due to keep align with nccl
// FP32 accumulation is the default: native-dtype accumulation measurably hurts
// accuracy on models with large residual-stream outliers.
// Opt out with ALL_REDUCE_FUSION_KERNEL_ACC_FP32=0 to restore the legacy behavior.
static char* fp32_acc = std::getenv("ALL_REDUCE_FUSION_KERNEL_ACC_FP32");
return fp32_acc != nullptr;
return fp32_acc == nullptr || fp32_acc[0] != '0';
}

void allreduce_fusion_op(AllReduceFusionParams const& params)
Expand Down
Loading