[AMD] Use aiter_dsv3_router_gemm kernel if number of experts <= 256. (#18451)
This commit is contained in:
@@ -326,7 +326,11 @@ class MoEGate(nn.Module):
|
|||||||
logits = dsv3_router_gemm(
|
logits = dsv3_router_gemm(
|
||||||
hidden_states, self.weight, out_dtype=torch.float32
|
hidden_states, self.weight, out_dtype=torch.float32
|
||||||
)
|
)
|
||||||
elif _use_aiter_gfx95 and hidden_states.shape[0] <= 256:
|
elif (
|
||||||
|
_use_aiter_gfx95
|
||||||
|
and hidden_states.shape[0] <= 256
|
||||||
|
and self.weight.shape[0] <= 256
|
||||||
|
):
|
||||||
logits = aiter_dsv3_router_gemm(
|
logits = aiter_dsv3_router_gemm(
|
||||||
hidden_states, self.weight, gemm_output_zero_allocator
|
hidden_states, self.weight, gemm_output_zero_allocator
|
||||||
)
|
)
|
||||||
|
|||||||
Reference in New Issue
Block a user