Fix GLM-5.3 forget-gate shape for nvCUTEDSL verify (#40607)
This commit is contained in:
@@ -1583,6 +1583,7 @@ class KDAAttnBackend(MambaAttnBackendBase):
|
|||||||
onorm_eps = None
|
onorm_eps = None
|
||||||
onorm_gate = None
|
onorm_gate = None
|
||||||
|
|
||||||
|
a = a.reshape(1, seq_len, h, layer.head_k_dim)
|
||||||
out = fused_kda_decode_mtp_dspark(
|
out = fused_kda_decode_mtp_dspark(
|
||||||
x_q=x_q,
|
x_q=x_q,
|
||||||
x_k=x_k,
|
x_k=x_k,
|
||||||
|
|||||||
Reference in New Issue
Block a user