[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2026-04-11 00:20:20 +08:00
committed by GitHub
parent 8f121f7879
commit ecd1ea1363
14 changed files with 1861 additions and 4 deletions
+35
View File
@@ -3491,3 +3491,38 @@ if hasattr(torch.ops._C, "hadacore_transform"):
@register_fake("_C::hadacore_transform")
def _hadacore_transform_fake(x: torch.Tensor, inplace: bool) -> torch.Tensor:
return torch.empty_like(x) if not inplace else x
if hasattr(torch.ops._C, "minimax_allreduce_rms"):
@register_fake("_C::minimax_allreduce_rms")
def _minimax_allreduce_rms_fake(
input: torch.Tensor,
norm_weight: torch.Tensor,
workspace: torch.Tensor,
rank: int,
nranks: int,
eps: float,
) -> torch.Tensor:
return torch.empty_like(input)
if hasattr(torch.ops._C, "minimax_allreduce_rms_qk"):
@register_fake("_C::minimax_allreduce_rms_qk")
def _minimax_allreduce_rms_qk_fake(
qkv: torch.Tensor,
norm_weight_q: torch.Tensor,
norm_weight_k: torch.Tensor,
workspace: torch.Tensor,
q_size: int,
kv_size: int,
rank: int,
nranks: int,
eps: float,
) -> tuple[torch.Tensor, torch.Tensor]:
token_num = qkv.shape[0]
return (
torch.empty([token_num, q_size], dtype=qkv.dtype, device=qkv.device),
torch.empty([token_num, kv_size], dtype=qkv.dtype, device=qkv.device),
)