biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Fix] Use torch.empty for output in attention+quant fusion (#31785)

Browse Source 
Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
This commit is contained in:
elvischenv
2026-03-11 12:26:14 +08:00
committed by GitHub
parent a8ff2cca92
commit 7d6abdd022
1 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/compilation/passes/fusion/attn_quant_fusion.py
View File

@@ -170,9 +170,8 @@ class AttentionFp8StaticQuantPattern(AttentionQuantPattern):
kv_cache_dummy_dep: torch.Tensor,
) -> torch.Tensor:
# attn output in quant_dtype
output_attn = torch.ops.aten.full.default(
output_attn = torch.empty(
[q.shape[0], self.num_heads, self.head_size],
0.0,
dtype=self.quant_dtype,
device=q.device,
)
@@ -271,9 +270,8 @@ class AttentionNvfp4QuantPattern(AttentionQuantPattern):
kv_cache_dummy_dep: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
# attention output in quant_dtype
output_attn = torch.ops.aten.full.default(
output_attn = torch.empty(
[q.shape[0], self.num_heads, self.head_size // 2],
0.0,
dtype=self.quant_dtype,
device=q.device,
)