Disable DualGemm fusion for Inkling dense MLP (#43447)
This commit is contained in:
@@ -120,6 +120,8 @@ class InklingDenseMLP(LlamaMLP):
|
||||
fused = fused and not lora_compatible_layout_enabled()
|
||||
self.layer_id = layer_id
|
||||
self.act_fn = InklingSwiglu(interleaved=fused)
|
||||
# DualGemm assumes [gate||up] and SiluAndMul; it would bypass act_fn.
|
||||
self.dual_gemm.mode = None
|
||||
self.scattered_sconv = get_exec().comm.enable_scattered_sconv
|
||||
|
||||
def forward(
|
||||
|
||||
Reference in New Issue
Block a user