DeepSeekV4-Pro enable cuda graph full and piecewise mode (#42604)
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
@@ -5,7 +5,6 @@ import torch
|
||||
# this import will also register the custom ops
|
||||
import vllm.model_executor.kernels.mhc as mhc_kernels
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
# --8<-- [start:mhc_pre]
|
||||
@@ -190,8 +189,6 @@ class HCHeadOp(CustomOp):
|
||||
)
|
||||
return out.view(*outer_shape, hidden_size)
|
||||
|
||||
# This @torch.compile is necessary for accuracy as well as performance.
|
||||
@torch.compile(backend=current_platform.simple_compile_backend)
|
||||
def forward_hip(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
|
||||
Reference in New Issue
Block a user