A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 03:00 UTC
[Perf] `zeros` -> `empty` to remove additional fill (#42988)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
May 21, 2026 at 16:00 UTC
b29cbf06525254693f29d98686e038eaf225be8c
4 files changed
+4
-5
vllm/_custom_ops.py
+2
-2
@@ -57,11 +57,11 @@ def create_fp4_scale_tensor(
57
rounded_m = round_up(m, 128)
58
scale_n = n // block_size
59
rounded_n = round_up(scale_n, 4)
60
- return torch.zeros(
60
+ return torch.empty(
61
(rounded_m, rounded_n // 4), device=device, dtype=torch.int32
62
)
63
else:
64
- return torch.zeros((m, n // block_size), device=device, dtype=torch.uint8)
64
+ return torch.empty((m, n // block_size), device=device, dtype=torch.uint8)
65
66
67
def create_fp4_output_tensors(
vllm/model_executor/layers/mamba/gdn_linear_attn.py
-1
@@ -1094,7 +1094,6 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
1094
)
1095
1096
core_attn_out.zero_()
1097
- z_out.zero_()
1097
num_tokens_all = qkvz.shape[0]
1098
mixed_qkv, z, b, a = self.prepare_gdn_attention_core_inputs(
1099
qkvz, ba, num_tokens_all
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+1
-1
@@ -255,7 +255,7 @@ def silu_mul_quant_fp8_packed_triton(
255
if output_q is None:
256
output_q = torch.empty((M, N_2), dtype=fp8_dtype, device=input.device)
257
258
- output_scale_packed = torch.zeros(
258
+ output_scale_packed = torch.empty(
259
(num_packed_groups, tma_aligned_M),
260
dtype=torch.int32,
261
device=input.device,
vllm/v1/attention/backends/turboquant_attn.py
+1
-1
@@ -454,7 +454,7 @@ class TurboQuantAttentionImpl(AttentionImpl["TurboQuantMetadata"]):
454
)
455
else:
456
# Mixed batch: decodes first (guaranteed by reorder_batch).
457
- attn_out = torch.zeros(
457
+ attn_out = torch.empty(
458
N, self.num_heads, self.head_size, device=device, dtype=q.dtype
459
)
460