Add support for LoRA adapters in Nemotron-H models (#30802)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
@@ -227,6 +227,11 @@ class FlashInferExperts(mk.FusedMoEPermuteExpertsUnpermute):
|
||||
use_deepseek_fp8_block_scale=self.use_deepseek_fp8_block_scale,
|
||||
)
|
||||
|
||||
def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None:
|
||||
# No support for LoRA in flashinfer_cutlass_fused_moe.
|
||||
# See TODOs in flashinfer functions runMoe and runMoeMinLantency.
|
||||
raise NotImplementedError("LoRA is not supported for flashinfer_cutlass_moe")
|
||||
|
||||
|
||||
def flashinfer_cutlass_moe_fp4(
|
||||
hidden_states: torch.Tensor,
|
||||
|
||||
Reference in New Issue
Block a user