Use aiter triton fused_add_rmsnorm_pad for gpt-oss (#30976)

Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
This commit is contained in:
Rohan Potdar
2026-01-28 20:47:47 +00:00
committed by GitHub
parent 3e440786af
commit 59bcc5b6f2
9 changed files with 327 additions and 11 deletions
+46
View File
@@ -790,6 +790,41 @@ def _rocm_aiter_act_mul_and_fp8_group_quant_fake(
return x_fp8, out_bs
def _rocm_aiter_triton_add_rmsnorm_pad_impl(
x: torch.Tensor,
weight: torch.Tensor,
variance_epsilon: float,
residual: torch.Tensor,
x_pad_to_multiple: int,
) -> tuple[torch.Tensor, torch.Tensor]:
from aiter.ops.triton.fused_add_rmsnorm_pad import fused_add_rmsnorm_pad
return fused_add_rmsnorm_pad(
x,
weight,
variance_epsilon,
residual,
x_pad_to_multiple=x_pad_to_multiple,
)
def _rocm_aiter_triton_add_rmsnorm_pad_fake(
x: torch.Tensor,
weight: torch.Tensor,
variance_epsilon: float,
residual: torch.Tensor,
x_pad_to_multiple: int,
) -> tuple[torch.Tensor, torch.Tensor]:
M, N = x.shape
if x_pad_to_multiple > 0:
N_out = (N + x_pad_to_multiple - 1) // x_pad_to_multiple * x_pad_to_multiple
else:
N_out = N
out = torch.empty((M, N_out), dtype=x.dtype, device=x.device)
residual_out = torch.empty_like(residual)
return out, residual_out
# Global flag to ensure ops are registered only once
_OPS_REGISTERED = False
@@ -1108,6 +1143,13 @@ class rocm_aiter_ops:
fake_impl=_rocm_aiter_act_mul_and_fp8_group_quant_fake,
)
direct_register_custom_op(
op_name="rocm_aiter_triton_add_rmsnorm_pad",
op_func=_rocm_aiter_triton_add_rmsnorm_pad_impl,
fake_impl=_rocm_aiter_triton_add_rmsnorm_pad_fake,
dispatch_key=current_platform.dispatch_key,
)
direct_register_custom_op(
op_name="rocm_aiter_group_fp8_quant",
op_func=_rocm_aiter_group_fp8_quant_impl,
@@ -1175,6 +1217,10 @@ class rocm_aiter_ops:
def get_act_mul_fused_fp8_group_quant_op() -> OpOverload:
return torch.ops.vllm.rocm_aiter_act_mul_and_fp8_group_quant.default
@staticmethod
def get_triton_add_rmsnorm_pad_op() -> OpOverload:
return torch.ops.vllm.rocm_aiter_triton_add_rmsnorm_pad.default
@staticmethod
def rms_norm(
x: torch.Tensor, weight: torch.Tensor, variance_epsilon: float