f40256b697
### What this PR does / why we need it? This PR integrates the `npu_add_rms_norm` fused kernel for RMSNorm operations with residual connections on 310P devices. This change optimizes the computation by replacing a two-step process (manual residual addition followed by RMSNorm) with a single, more efficient fused operation. This is needed to improve the performance of models utilizing RMSNorm with residual connections on the 310P architecture. Fixes # ### Does this PR introduce _any_ user-facing change? No, this PR introduces an internal optimization and does not change any user-facing APIs or behaviors. ### How was this patch tested? This patch was tested with updated unit tests (`test_RMSNorm_forward_310p`) that mock the `npu_add_rms_norm` operation to verify the correctness of the fused kernel integration. --------- Signed-off-by: Tflowers-0129 <2906339855@qq.com>
40 lines
1.3 KiB
Python
40 lines
1.3 KiB
Python
import torch
|
|
import torch_npu
|
|
|
|
from vllm_ascend.ops.layernorm import AscendGemmaRMSNorm, AscendRMSNorm
|
|
|
|
|
|
class AscendRMSNorm310(AscendRMSNorm):
|
|
def forward_oot(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: torch.Tensor | None = None,
|
|
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
|
|
if residual is not None:
|
|
x, _, residual = torch_npu.npu_add_rms_norm(x, residual, self.weight, self.variance_epsilon)
|
|
if self.bias is not None:
|
|
x.add_(self.bias)
|
|
return x, residual
|
|
|
|
x, _ = torch_npu.npu_rms_norm(x, self.weight, self.variance_epsilon)
|
|
if self.bias is not None:
|
|
x.add_(self.bias)
|
|
return x
|
|
|
|
|
|
class AscendGemmaRMSNorm310(AscendGemmaRMSNorm):
|
|
def forward_oot(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: torch.Tensor | None = None,
|
|
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
|
|
if residual is not None:
|
|
orig_dtype = residual.dtype
|
|
x = x + residual.to(x.dtype)
|
|
residual = x.to(orig_dtype)
|
|
x, _ = torch_npu.npu_rms_norm(x, 1.0 + self.weight, self.variance_epsilon)
|
|
return x, residual
|
|
|
|
x, _ = torch_npu.npu_rms_norm(x, 1.0 + self.weight, self.variance_epsilon)
|
|
return x
|