forked from Karylab-cklius/vllm
Update cudagraph_utils.py
Signed-off-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
@@ -522,8 +522,7 @@ class ModelCudaGraphManager(CudaGraphManager):
|
||||
batch_descriptor=batch_descriptor,
|
||||
is_padding=input_buffers.is_padding[:num_tokens],
|
||||
skip_attention=(
|
||||
cg_mode == CUDAGraphMode.PIECEWISE
|
||||
and not self.use_breakable_cg
|
||||
cg_mode == CUDAGraphMode.PIECEWISE and not self.use_breakable_cg
|
||||
),
|
||||
):
|
||||
if cg_mode == CUDAGraphMode.PIECEWISE:
|
||||
|
||||
Reference in New Issue
Block a user