Skip to content
Open
Changes from all commits
Commits
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
7 changes: 6 additions & 1 deletion megatron/core/transformer/moe/experts.py
Original file line number Diff line number Diff line change
Expand Up @@ -222,7 +222,12 @@ def forward(
permuted_probs: torch.Tensor,
):
"""Forward step of the GroupedMLP."""
assert self.config.bf16, "Currently GroupedMLP for MoE only supports bf16."
assert self.config.bf16, (
"The legacy GroupedMLP only supports bf16. "
"For FP16/FP8 support, please use TEGroupedMLP instead, which is adopted by default "
"when TransformerEngine >= 1.9 is installed and '--moe-use-legacy-grouped-gemm' is "
"*not* set."
)
if self.activation_recompute:
self.activation_checkpoint = tensor_parallel.CheckpointWithoutOutput()

Expand Down