[Perf] zeros -> empty to remove additional fill (#42988)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -1094,7 +1094,6 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
|
||||
)
|
||||
|
||||
core_attn_out.zero_()
|
||||
z_out.zero_()
|
||||
num_tokens_all = qkvz.shape[0]
|
||||
mixed_qkv, z, b, a = self.prepare_gdn_attention_core_inputs(
|
||||
qkvz, ba, num_tokens_all
|
||||
|
||||
@@ -255,7 +255,7 @@ def silu_mul_quant_fp8_packed_triton(
|
||||
if output_q is None:
|
||||
output_q = torch.empty((M, N_2), dtype=fp8_dtype, device=input.device)
|
||||
|
||||
output_scale_packed = torch.zeros(
|
||||
output_scale_packed = torch.empty(
|
||||
(num_packed_groups, tma_aligned_M),
|
||||
dtype=torch.int32,
|
||||
device=input.device,
|
||||
|
||||
Reference in New Issue
Block a user