[Perf] Fuse MiniMax H3 SwiGLU activation (#6283)
Signed-off-by: Canlin Guo <canlinguosdu@gmail.com>
This commit is contained in:
1 parent
07ceb3324a
commit
9683b7f4b4
2 files changed
+40
-2
No files matched your search
@@ -0,0 +1,37 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
"""Custom activation functions for diffusion models."""
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
from vllm_omni.diffusion.layers.custom_op import CustomOp
|
||||
from vllm_omni.platforms import current_omni_platform
|
||||
|
||||
|
||||
class SiluAndMul(CustomOp):
|
||||
"""SwiGLU activation dispatched by the current vLLM-Omni platform."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
super().__init__()
|
||||
if current_omni_platform.is_cuda() or current_omni_platform.is_rocm() or current_omni_platform.is_musa():
|
||||
self.op = torch.ops._C.silu_and_mul
|
||||
|
||||
@staticmethod
|
||||
def forward_native(x: torch.Tensor) -> torch.Tensor:
|
||||
d = x.shape[-1] // 2
|
||||
return F.silu(x[..., :d]) * x[..., d:]
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
if x.device.type != "cuda":
|
||||
return self.forward_native(x)
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = x.shape[:-1] + (d,)
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
self.op(out, x)
|
||||
return out
|
||||
|
||||
def forward_npu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.forward_native(x)
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.forward_native(x)
|
||||
Reference in new issue
Block a user