[Perf] zeros -> empty to remove additional fill (#42988)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-05-21 16:00:29 -04:00
committed by GitHub
parent 9b54e50e2c
commit b29cbf0652
4 changed files with 4 additions and 5 deletions
+2 -2
View File
@@ -57,11 +57,11 @@ def create_fp4_scale_tensor(
rounded_m = round_up(m, 128)
scale_n = n // block_size
rounded_n = round_up(scale_n, 4)
return torch.zeros(
return torch.empty(
(rounded_m, rounded_n // 4), device=device, dtype=torch.int32
)
else:
return torch.zeros((m, n // block_size), device=device, dtype=torch.uint8)
return torch.empty((m, n // block_size), device=device, dtype=torch.uint8)
def create_fp4_output_tensors(
@@ -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,
@@ -454,7 +454,7 @@ class TurboQuantAttentionImpl(AttentionImpl["TurboQuantMetadata"]):
)
else:
# Mixed batch: decodes first (guaranteed by reorder_batch).
attn_out = torch.zeros(
attn_out = torch.empty(
N, self.num_heads, self.head_size, device=device, dtype=q.dtype
)